Adding upstream version 16.2.11+ds.upstream/16.2.11+dsupstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

36 files changed, 27221 insertions, 0 deletions

diff --git a/src/client/CMakeLists.txt b/src/client/CMakeLists.txt
new file mode 100644
index 000000000..8897ada7b
--- /dev/null
+++ b/src/client/CMakeLists.txt
@@ -0,0 +1,13 @@
+set(libclient_srcs
+  Client.cc
+  Dentry.cc
+  Fh.cc
+  Inode.cc
+  MetaRequest.cc
+  ClientSnapRealm.cc
+  MetaSession.cc
+  Trace.cc
+  posix_acl.cc
+  Delegation.cc)
+add_library(client STATIC ${libclient_srcs})
+target_link_libraries(client osdc)
diff --git a/src/client/Client.cc b/src/client/Client.cc
new file mode 100644
index 000000000..811cfa8ec
--- /dev/null
+++ b/src/client/Client.cc
@@ -0,0 +1,16029 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+
+// unix-ey fs stuff
+#include <unistd.h>
+#include <sys/types.h>
+#include <time.h>
+#include <utime.h>
+#include <string.h>
+#include <sys/stat.h>
+#include <sys/param.h>
+#include <fcntl.h>
+#include <sys/file.h>
+#ifndef _WIN32
+#include <sys/utsname.h>
+#endif
+#include <sys/uio.h>
+
+#include <boost/lexical_cast.hpp>
+#include <boost/fusion/include/std_pair.hpp>
+
+#include "common/async/waiter.h"
+
+#if defined(__FreeBSD__) || defined(_WIN32)
+#define XATTR_CREATE    0x1
+#define XATTR_REPLACE   0x2
+#else
+#include <sys/xattr.h>
+#endif
+
+#if defined(__linux__)
+#include <linux/falloc.h>
+#endif
+
+#include <sys/statvfs.h>
+
+#include "common/config.h"
+#include "common/version.h"
+#include "common/async/blocked_completion.h"
+
+#include "mon/MonClient.h"
+
+#include "messages/MClientCaps.h"
+#include "messages/MClientLease.h"
+#include "messages/MClientQuota.h"
+#include "messages/MClientReclaim.h"
+#include "messages/MClientReclaimReply.h"
+#include "messages/MClientReconnect.h"
+#include "messages/MClientReply.h"
+#include "messages/MClientRequest.h"
+#include "messages/MClientRequestForward.h"
+#include "messages/MClientSession.h"
+#include "messages/MClientSnap.h"
+#include "messages/MClientMetrics.h"
+#include "messages/MCommandReply.h"
+#include "messages/MFSMap.h"
+#include "messages/MFSMapUser.h"
+#include "messages/MMDSMap.h"
+#include "messages/MOSDMap.h"
+
+#include "mds/flock.h"
+#include "mds/cephfs_features.h"
+#include "osd/OSDMap.h"
+#include "osdc/Filer.h"
+
+#include "common/Cond.h"
+#include "common/perf_counters.h"
+#include "common/admin_socket.h"
+#include "common/errno.h"
+#include "include/str_list.h"
+
+#define dout_subsys ceph_subsys_client
+
+#include "include/lru.h"
+#include "include/compat.h"
+#include "include/stringify.h"
+#include "include/random.h"
+
+#include "Client.h"
+#include "Inode.h"
+#include "Dentry.h"
+#include "Delegation.h"
+#include "Dir.h"
+#include "ClientSnapRealm.h"
+#include "Fh.h"
+#include "MetaSession.h"
+#include "MetaRequest.h"
+#include "ObjecterWriteback.h"
+#include "posix_acl.h"
+
+#include "include/ceph_assert.h"
+#include "include/stat.h"
+
+#include "include/cephfs/ceph_ll_client.h"
+
+#if HAVE_GETGROUPLIST
+#include <grp.h>
+#include <pwd.h>
+#include <unistd.h>
+#endif
+
+#undef dout_prefix
+#define dout_prefix *_dout << "client." << whoami << " "
+
+#define  tout(cct)       if (!cct->_conf->client_trace.empty()) traceout
+
+// FreeBSD fails to define this
+#ifndef O_DSYNC
+#define O_DSYNC 0x0
+#endif
+// Darwin fails to define this
+#ifndef O_RSYNC
+#define O_RSYNC 0x0
+#endif
+
+#ifndef O_DIRECT
+#define O_DIRECT 0x0
+#endif
+
+// Windows doesn't define those values. While the Posix compatibilty layer
+// doesn't support those values, the Windows native functions do provide
+// similar flags. Special care should be taken if we're going to use those
+// flags in ceph-dokan. The current values are no-ops, while propagating
+// them to the rest of the code might cause the Windows functions to reject
+// them as invalid.
+#ifndef O_NOFOLLOW
+#define O_NOFOLLOW 0x0
+#endif
+
+#ifndef O_SYNC
+#define O_SYNC 0x0
+#endif
+
+#define DEBUG_GETATTR_CAPS (CEPH_CAP_XATTR_SHARED)
+
+#ifndef S_IXUGO
+#define S_IXUGO	(S_IXUSR|S_IXGRP|S_IXOTH)
+#endif
+
+using namespace TOPNSPC::common;
+
+namespace bs = boost::system;
+namespace ca = ceph::async;
+
+void client_flush_set_callback(void *p, ObjectCacher::ObjectSet *oset)
+{
+  Client *client = static_cast<Client*>(p);
+  client->flush_set_callback(oset);
+}
+
+bool Client::is_reserved_vino(vinodeno_t &vino) {
+  if (MDS_IS_PRIVATE_INO(vino.ino)) {
+    ldout(cct, -1) << __func__ << " attempt to access reserved inode number " << vino << dendl;
+    return true;
+  }
+  return false;
+}
+
+// running average and standard deviation -- presented in
+// Donald Knuth's TAoCP, Volume II.
+double calc_average(double old_avg, double value, uint64_t count) {
+  double new_avg;
+  if (count == 1) {
+    new_avg = value;
+  } else {
+    new_avg = old_avg + ((value - old_avg) / count);
+  }
+
+  return new_avg;
+}
+
+double calc_sq_sum(double old_sq_sum, double old_mean, double new_mean,
+                   double value, uint64_t count) {
+  double new_sq_sum;
+  if (count == 1) {
+    new_sq_sum = 0.0;
+  } else {
+    new_sq_sum = old_sq_sum + (value - old_mean)*(value - new_mean);
+  }
+
+  return new_sq_sum;
+}
+
+// -------------
+
+Client::CommandHook::CommandHook(Client *client) :
+  m_client(client)
+{
+}
+
+int Client::CommandHook::call(
+  std::string_view command,
+  const cmdmap_t& cmdmap,
+  Formatter *f,
+  std::ostream& errss,
+  bufferlist& out)
+{
+  f->open_object_section("result");
+  {
+    std::scoped_lock l{m_client->client_lock};
+    if (command == "mds_requests")
+      m_client->dump_mds_requests(f);
+    else if (command == "mds_sessions") {
+      bool cap_dump = false;
+      cmd_getval(cmdmap, "cap_dump", cap_dump);
+      m_client->dump_mds_sessions(f, cap_dump);
+    } else if (command == "dump_cache")
+      m_client->dump_cache(f);
+    else if (command == "kick_stale_sessions")
+      m_client->_kick_stale_sessions();
+    else if (command == "status")
+      m_client->dump_status(f);
+    else
+      ceph_abort_msg("bad command registered");
+  }
+  f->close_section();
+  return 0;
+}
+
+
+// -------------
+
+int Client::get_fd_inode(int fd, InodeRef *in) {
+  int r = 0;
+  if (fd == CEPHFS_AT_FDCWD) {
+    *in = cwd;
+  } else {
+    Fh *f = get_filehandle(fd);
+    if (!f) {
+      r = -CEPHFS_EBADF;
+    } else {
+      *in = f->inode;
+    }
+  }
+  return r;
+}
+
+dir_result_t::dir_result_t(Inode *in, const UserPerm& perms)
+  : inode(in), offset(0), next_offset(2),
+    release_count(0), ordered_count(0), cache_index(0), start_shared_gen(0),
+    perms(perms)
+  { }
+
+void Client::_reset_faked_inos()
+{
+  ino_t start = 1024;
+  free_faked_inos.clear();
+  free_faked_inos.insert(start, (uint32_t)-1 - start + 1);
+  last_used_faked_ino = 0;
+  last_used_faked_root = 0;
+  #ifdef _WIN32
+  // On Windows, sizeof(ino_t) is just 2. Despite that, most "native"
+  // Windows structures, including Dokan ones, are using 64B identifiers.
+  _use_faked_inos = false;
+  #else
+  _use_faked_inos = sizeof(ino_t) < 8 || cct->_conf->client_use_faked_inos;
+  #endif
+}
+
+void Client::_assign_faked_ino(Inode *in)
+{
+  if (0 == last_used_faked_ino)
+    last_used_faked_ino = last_used_faked_ino + 2048; // start(1024)~2048 reserved for _assign_faked_root
+  interval_set<ino_t>::const_iterator it = free_faked_inos.lower_bound(last_used_faked_ino + 1);
+  if (it == free_faked_inos.end() && last_used_faked_ino > 0) {
+    last_used_faked_ino = 2048;
+    it = free_faked_inos.lower_bound(last_used_faked_ino + 1);
+  }
+  ceph_assert(it != free_faked_inos.end());
+  if (last_used_faked_ino < it.get_start()) {
+    ceph_assert(it.get_len() > 0);
+    last_used_faked_ino = it.get_start();
+  } else {
+    ++last_used_faked_ino;
+    ceph_assert(it.get_start() + it.get_len() > last_used_faked_ino);
+  }
+  in->faked_ino = last_used_faked_ino;
+  free_faked_inos.erase(in->faked_ino);
+  faked_ino_map[in->faked_ino] = in->vino();
+}
+
+/*
+ * In the faked mode, if you export multiple subdirectories,
+ * you will see that the inode numbers of the exported subdirectories
+ * are the same. so we distinguish the mount point by reserving 
+ * the "fake ids" between "1024~2048" and combining the last 
+ * 10bits(0x3ff) of the "root inodes".
+*/
+void Client::_assign_faked_root(Inode *in)
+{
+  interval_set<ino_t>::const_iterator it = free_faked_inos.lower_bound(last_used_faked_root + 1);
+  if (it == free_faked_inos.end() && last_used_faked_root > 0) {
+    last_used_faked_root = 0;
+    it = free_faked_inos.lower_bound(last_used_faked_root + 1);
+  }
+  assert(it != free_faked_inos.end());
+  vinodeno_t inode_info = in->vino();
+  uint64_t inode_num = (uint64_t)inode_info.ino;
+  ldout(cct, 10) << "inode_num " << inode_num << "inode_num & 0x3ff=" << (inode_num & 0x3ff)<< dendl;
+  last_used_faked_root = it.get_start()  + (inode_num & 0x3ff); // 0x3ff mask and get_start will not exceed 2048
+  assert(it.get_start() + it.get_len() > last_used_faked_root);
+
+  in->faked_ino = last_used_faked_root;
+  free_faked_inos.erase(in->faked_ino);
+  faked_ino_map[in->faked_ino] = in->vino();
+}
+
+void Client::_release_faked_ino(Inode *in)
+{
+  free_faked_inos.insert(in->faked_ino);
+  faked_ino_map.erase(in->faked_ino);
+}
+
+vinodeno_t Client::_map_faked_ino(ino_t ino)
+{
+  vinodeno_t vino;
+  if (ino == 1)
+    vino = root->vino();
+  else if (faked_ino_map.count(ino))
+    vino = faked_ino_map[ino];
+  else
+    vino = vinodeno_t(0, CEPH_NOSNAP);
+  ldout(cct, 10) << __func__ << " " << ino << " -> " << vino << dendl;
+  return vino;
+}
+
+vinodeno_t Client::map_faked_ino(ino_t ino)
+{
+  std::scoped_lock lock(client_lock);
+  return _map_faked_ino(ino);
+}
+
+// cons/des
+
+Client::Client(Messenger *m, MonClient *mc, Objecter *objecter_)
+  : Dispatcher(m->cct->get()),
+    timer(m->cct, timer_lock, false),
+    messenger(m),
+    monclient(mc),
+    objecter(objecter_),
+    whoami(mc->get_global_id()),
+    mount_state(CLIENT_UNMOUNTED, "Client::mountstate_lock"),
+    initialize_state(CLIENT_NEW, "Client::initstate_lock"),
+    cct_deleter{m->cct, [](CephContext *p) {p->put();}},
+    async_ino_invalidator(m->cct),
+    async_dentry_invalidator(m->cct),
+    interrupt_finisher(m->cct),
+    remount_finisher(m->cct),
+    async_ino_releasor(m->cct),
+    objecter_finisher(m->cct),
+    m_command_hook(this),
+    fscid(0)
+{
+  _reset_faked_inos();
+
+  user_id = cct->_conf->client_mount_uid;
+  group_id = cct->_conf->client_mount_gid;
+  fuse_default_permissions = cct->_conf.get_val<bool>(
+    "fuse_default_permissions");
+
+  _collect_and_send_global_metrics = cct->_conf.get_val<bool>(
+    "client_collect_and_send_global_metrics");
+
+  mount_timeout = cct->_conf.get_val<std::chrono::seconds>(
+    "client_mount_timeout");
+
+  caps_release_delay = cct->_conf.get_val<std::chrono::seconds>(
+    "client_caps_release_delay");
+
+  if (cct->_conf->client_acl_type == "posix_acl")
+    acl_type = POSIX_ACL;
+
+  lru.lru_set_midpoint(cct->_conf->client_cache_mid);
+
+  // file handles
+  free_fd_set.insert(10, 1<<30);
+
+  mdsmap.reset(new MDSMap);
+
+  // osd interfaces
+  writeback_handler.reset(new ObjecterWriteback(objecter, &objecter_finisher,
+					    &client_lock));
+  objectcacher.reset(new ObjectCacher(cct, "libcephfs", *writeback_handler, client_lock,
+				  client_flush_set_callback,    // all commit callback
+				  (void*)this,
+				  cct->_conf->client_oc_size,
+				  cct->_conf->client_oc_max_objects,
+				  cct->_conf->client_oc_max_dirty,
+				  cct->_conf->client_oc_target_dirty,
+				  cct->_conf->client_oc_max_dirty_age,
+				  true));
+}
+
+
+Client::~Client()
+{
+  ceph_assert(ceph_mutex_is_not_locked(client_lock));
+
+  // If the task is crashed or aborted and doesn't
+  // get any chance to run the umount and shutdow.
+  {
+    std::scoped_lock l{client_lock};
+    tick_thread_stopped = true;
+    upkeep_cond.notify_one();
+  }
+
+  if (upkeeper.joinable())
+    upkeeper.join();
+
+  // It is necessary to hold client_lock, because any inode destruction
+  // may call into ObjectCacher, which asserts that it's lock (which is
+  // client_lock) is held.
+  std::scoped_lock l{client_lock};
+  tear_down_cache();
+}
+
+void Client::tear_down_cache()
+{
+  // fd's
+  for (auto &[fd, fh] : fd_map) {
+    ldout(cct, 1) << __func__ << " forcing close of fh " << fd << " ino " << fh->inode->ino << dendl;
+    _release_fh(fh);
+  }
+  fd_map.clear();
+
+  while (!opened_dirs.empty()) {
+    dir_result_t *dirp = *opened_dirs.begin();
+    ldout(cct, 1) << __func__ << " forcing close of dir " << dirp << " ino " << dirp->inode->ino << dendl;
+    _closedir(dirp);
+  }
+
+  // caps!
+  // *** FIXME ***
+
+  // empty lru
+  trim_cache();
+  ceph_assert(lru.lru_get_size() == 0);
+
+  // close root ino
+  ceph_assert(inode_map.size() <= 1 + root_parents.size());
+  if (root && inode_map.size() == 1 + root_parents.size()) {
+    root.reset();
+  }
+
+  ceph_assert(inode_map.empty());
+}
+
+inodeno_t Client::get_root_ino()
+{
+  std::scoped_lock l(client_lock);
+  if (use_faked_inos())
+    return root->faked_ino;
+  else
+    return root->ino;
+}
+
+Inode *Client::get_root()
+{
+  std::scoped_lock l(client_lock);
+  root->ll_get();
+  return root.get();
+}
+
+
+// debug crapola
+
+void Client::dump_inode(Formatter *f, Inode *in, set<Inode*>& did, bool disconnected)
+{
+  filepath path;
+  in->make_long_path(path);
+  ldout(cct, 1) << "dump_inode: "
+		<< (disconnected ? "DISCONNECTED ":"")
+		<< "inode " << in->ino
+		<< " " << path
+		<< " ref " << in->get_nref()
+		<< " " << *in << dendl;
+
+  if (f) {
+    f->open_object_section("inode");
+    f->dump_stream("path") << path;
+    if (disconnected)
+      f->dump_int("disconnected", 1);
+    in->dump(f);
+    f->close_section();
+  }
+
+  did.insert(in);
+  if (in->dir) {
+    ldout(cct, 1) << "  dir " << in->dir << " size " << in->dir->dentries.size() << dendl;
+    for (ceph::unordered_map<string, Dentry*>::iterator it = in->dir->dentries.begin();
+         it != in->dir->dentries.end();
+         ++it) {
+      ldout(cct, 1) << "   " << in->ino << " dn " << it->first << " " << it->second << " ref " << it->second->ref << dendl;
+      if (f) {
+	f->open_object_section("dentry");
+	it->second->dump(f);
+	f->close_section();
+      }	
+      if (it->second->inode)
+	dump_inode(f, it->second->inode.get(), did, false);
+    }
+  }
+}
+
+void Client::dump_cache(Formatter *f)
+{
+  set<Inode*> did;
+
+  ldout(cct, 1) << __func__ << dendl;
+
+  if (f)
+    f->open_array_section("cache");
+
+  if (root)
+    dump_inode(f, root.get(), did, true);
+
+  // make a second pass to catch anything disconnected
+  for (ceph::unordered_map<vinodeno_t, Inode*>::iterator it = inode_map.begin();
+       it != inode_map.end();
+       ++it) {
+    if (did.count(it->second))
+      continue;
+    dump_inode(f, it->second, did, true);
+  }
+
+  if (f)
+    f->close_section();
+}
+
+void Client::dump_status(Formatter *f)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  ldout(cct, 1) << __func__ << dendl;
+
+  const epoch_t osd_epoch
+    = objecter->with_osdmap(std::mem_fn(&OSDMap::get_epoch));
+
+  if (f) {
+    f->open_object_section("metadata");
+    for (const auto& kv : metadata)
+      f->dump_string(kv.first.c_str(), kv.second);
+    f->close_section();
+
+    f->dump_int("dentry_count", lru.lru_get_size());
+    f->dump_int("dentry_pinned_count", lru.lru_get_num_pinned());
+    f->dump_int("id", get_nodeid().v);
+    entity_inst_t inst(messenger->get_myname(), messenger->get_myaddr_legacy());
+    f->dump_object("inst", inst);
+    f->dump_object("addr", inst.addr);
+    f->dump_stream("inst_str") << inst.name << " " << inst.addr.get_legacy_str();
+    f->dump_string("addr_str", inst.addr.get_legacy_str());
+    f->dump_int("inode_count", inode_map.size());
+    f->dump_int("mds_epoch", mdsmap->get_epoch());
+    f->dump_int("osd_epoch", osd_epoch);
+    f->dump_int("osd_epoch_barrier", cap_epoch_barrier);
+    f->dump_bool("blocklisted", blocklisted);
+    f->dump_string("fs_name", mdsmap->get_fs_name());
+  }
+}
+
+void Client::_pre_init()
+{
+  timer.init();
+
+  objecter_finisher.start();
+  filer.reset(new Filer(objecter, &objecter_finisher));
+
+  objectcacher->start();
+}
+
+int Client::init()
+{
+  RWRef_t iref_writer(initialize_state, CLIENT_INITIALIZING, false);
+  ceph_assert(iref_writer.is_first_writer());
+
+  _pre_init();
+  {
+    std::scoped_lock l{client_lock};
+    messenger->add_dispatcher_tail(this);
+  }
+  _finish_init();
+  iref_writer.update_state(CLIENT_INITIALIZED);
+  return 0;
+}
+
+void Client::_finish_init()
+{
+  {
+    std::scoped_lock l{client_lock};
+    // logger
+    PerfCountersBuilder plb(cct, "client", l_c_first, l_c_last);
+    plb.add_time_avg(l_c_reply, "reply", "Latency of receiving a reply on metadata request");
+    plb.add_time_avg(l_c_lat, "lat", "Latency of processing a metadata request");
+    plb.add_time_avg(l_c_wrlat, "wrlat", "Latency of a file data write operation");
+    plb.add_time_avg(l_c_read, "rdlat", "Latency of a file data read operation");
+    plb.add_time_avg(l_c_fsync, "fsync", "Latency of a file sync operation");
+    // average, standard deviation mds/r/w/ latencies
+    plb.add_time(l_c_md_avg, "mdavg", "Average latency for processing metadata requests");
+    plb.add_u64(l_c_md_sqsum, "mdsqsum", "Sum of squares (to calculate variability/stdev) for metadata requests");
+    plb.add_u64(l_c_md_ops, "mdops", "Total metadata IO operations");
+    plb.add_time(l_c_rd_avg, "readavg", "Average latency for processing read requests");
+    plb.add_u64(l_c_rd_sqsum, "readsqsum", "Sum of squares ((to calculate variability/stdev) for read requests");
+    plb.add_u64(l_c_rd_ops, "rdops", "Total read IO operations");
+    plb.add_time(l_c_wr_avg, "writeavg", "Average latency for processing write requests");
+    plb.add_u64(l_c_wr_sqsum, "writesqsum", "Sum of squares ((to calculate variability/stdev) for write requests");
+    plb.add_u64(l_c_wr_ops, "rdops", "Total write IO operations");
+    logger.reset(plb.create_perf_counters());
+    cct->get_perfcounters_collection()->add(logger.get());
+  }
+
+  cct->_conf.add_observer(this);
+
+  AdminSocket* admin_socket = cct->get_admin_socket();
+  int ret = admin_socket->register_command("mds_requests",
+					   &m_command_hook,
+					   "show in-progress mds requests");
+  if (ret < 0) {
+    lderr(cct) << "error registering admin socket command: "
+	       << cpp_strerror(-ret) << dendl;
+  }
+  ret = admin_socket->register_command("mds_sessions "
+		                       "name=cap_dump,type=CephBool,req=false",
+				       &m_command_hook,
+				       "show mds session state");
+  if (ret < 0) {
+    lderr(cct) << "error registering admin socket command: "
+	       << cpp_strerror(-ret) << dendl;
+  }
+  ret = admin_socket->register_command("dump_cache",
+				       &m_command_hook,
+				       "show in-memory metadata cache contents");
+  if (ret < 0) {
+    lderr(cct) << "error registering admin socket command: "
+	       << cpp_strerror(-ret) << dendl;
+  }
+  ret = admin_socket->register_command("kick_stale_sessions",
+				       &m_command_hook,
+				       "kick sessions that were remote reset");
+  if (ret < 0) {
+    lderr(cct) << "error registering admin socket command: "
+	       << cpp_strerror(-ret) << dendl;
+  }
+  ret = admin_socket->register_command("status",
+				       &m_command_hook,
+				       "show overall client status");
+  if (ret < 0) {
+    lderr(cct) << "error registering admin socket command: "
+	       << cpp_strerror(-ret) << dendl;
+  }
+}
+
+void Client::shutdown() 
+{
+  ldout(cct, 1) << __func__ << dendl;
+
+  // If we were not mounted, but were being used for sending
+  // MDS commands, we may have sessions that need closing.
+  {
+    std::scoped_lock l{client_lock};
+
+    // To make sure the tick thread will be stoppped before
+    // destructing the Client, just in case like the _mount()
+    // failed but didn't not get a chance to stop the tick
+    // thread
+    tick_thread_stopped = true;
+    upkeep_cond.notify_one();
+
+    _close_sessions();
+  }
+  cct->_conf.remove_observer(this);
+
+  cct->get_admin_socket()->unregister_commands(&m_command_hook);
+
+  if (ino_invalidate_cb) {
+    ldout(cct, 10) << "shutdown stopping cache invalidator finisher" << dendl;
+    async_ino_invalidator.wait_for_empty();
+    async_ino_invalidator.stop();
+  }
+
+  if (dentry_invalidate_cb) {
+    ldout(cct, 10) << "shutdown stopping dentry invalidator finisher" << dendl;
+    async_dentry_invalidator.wait_for_empty();
+    async_dentry_invalidator.stop();
+  }
+
+  if (switch_interrupt_cb) {
+    ldout(cct, 10) << "shutdown stopping interrupt finisher" << dendl;
+    interrupt_finisher.wait_for_empty();
+    interrupt_finisher.stop();
+  }
+
+  if (remount_cb) {
+    ldout(cct, 10) << "shutdown stopping remount finisher" << dendl;
+    remount_finisher.wait_for_empty();
+    remount_finisher.stop();
+  }
+
+  if (ino_release_cb) {
+    ldout(cct, 10) << "shutdown stopping inode release finisher" << dendl;
+    async_ino_releasor.wait_for_empty();
+    async_ino_releasor.stop();
+  }
+
+  objectcacher->stop();  // outside of client_lock! this does a join.
+
+  /*
+   * We are shuting down the client.
+   *
+   * Just declare the state to CLIENT_NEW to block and fail any
+   * new comming "reader" and then try to wait all the in-flight
+   * "readers" to finish.
+   */
+  RWRef_t iref_writer(initialize_state, CLIENT_NEW, false);
+  if (!iref_writer.is_first_writer())
+    return;
+  iref_writer.wait_readers_done();
+
+  {
+    std::scoped_lock l(timer_lock);
+    timer.shutdown();
+  }
+
+  objecter_finisher.wait_for_empty();
+  objecter_finisher.stop();
+
+  if (logger) {
+    cct->get_perfcounters_collection()->remove(logger.get());
+    logger.reset();
+  }
+}
+
+void Client::update_io_stat_metadata(utime_t latency) {
+  auto lat_nsec = latency.to_nsec();
+  // old values are used to compute new ones
+  auto o_avg = logger->tget(l_c_md_avg).to_nsec();
+  auto o_sqsum = logger->get(l_c_md_sqsum);
+
+  auto n_avg = calc_average(o_avg, lat_nsec, nr_metadata_request);
+  auto n_sqsum = calc_sq_sum(o_sqsum, o_avg, n_avg, lat_nsec,
+                              nr_metadata_request);
+
+  logger->tinc(l_c_lat, latency);
+  logger->tinc(l_c_reply, latency);
+
+  utime_t avg;
+  avg.set_from_double(n_avg / 1000000000);
+  logger->tset(l_c_md_avg, avg);
+  logger->set(l_c_md_sqsum, n_sqsum);
+  logger->set(l_c_md_ops, nr_metadata_request);
+}
+
+void Client::update_io_stat_read(utime_t latency) {
+  auto lat_nsec = latency.to_nsec();
+  // old values are used to compute new ones
+  auto o_avg = logger->tget(l_c_rd_avg).to_nsec();
+  auto o_sqsum = logger->get(l_c_rd_sqsum);
+
+  auto n_avg = calc_average(o_avg, lat_nsec, nr_read_request);
+  auto n_sqsum = calc_sq_sum(o_sqsum, o_avg, n_avg, lat_nsec,
+                              nr_read_request);
+
+  logger->tinc(l_c_read, latency);
+
+  utime_t avg;
+  avg.set_from_double(n_avg / 1000000000);
+  logger->tset(l_c_rd_avg, avg);
+  logger->set(l_c_rd_sqsum, n_sqsum);
+  logger->set(l_c_rd_ops, nr_read_request);
+}
+
+void Client::update_io_stat_write(utime_t latency) {
+  auto lat_nsec = latency.to_nsec();
+  // old values are used to compute new ones
+  auto o_avg = logger->tget(l_c_wr_avg).to_nsec();
+  auto o_sqsum = logger->get(l_c_wr_sqsum);
+
+  auto n_avg = calc_average(o_avg, lat_nsec, nr_write_request);
+  auto n_sqsum = calc_sq_sum(o_sqsum, o_avg, n_avg, lat_nsec,
+                              nr_write_request);
+
+  logger->tinc(l_c_wrlat, latency);
+
+  utime_t avg;
+  avg.set_from_double(n_avg / 1000000000);
+  logger->tset(l_c_wr_avg, avg);
+  logger->set(l_c_wr_sqsum, n_sqsum);
+  logger->set(l_c_wr_ops, nr_write_request);
+}
+
+// ===================
+// metadata cache stuff
+
+void Client::trim_cache(bool trim_kernel_dcache)
+{
+  uint64_t max = cct->_conf->client_cache_size;
+  ldout(cct, 20) << "trim_cache size " << lru.lru_get_size() << " max " << max << dendl;
+  unsigned last = 0;
+  while (lru.lru_get_size() != last) {
+    last = lru.lru_get_size();
+
+    if (!is_unmounting() && lru.lru_get_size() <= max)  break;
+
+    // trim!
+    Dentry *dn = static_cast<Dentry*>(lru.lru_get_next_expire());
+    if (!dn)
+      break;  // done
+
+    trim_dentry(dn);
+  }
+
+  if (trim_kernel_dcache && lru.lru_get_size() > max)
+    _invalidate_kernel_dcache();
+
+  // hose root?
+  if (lru.lru_get_size() == 0 && root && root->get_nref() == 1 && inode_map.size() == 1 + root_parents.size()) {
+    ldout(cct, 15) << "trim_cache trimmed root " << root << dendl;
+    root.reset();
+  }
+}
+
+void Client::trim_cache_for_reconnect(MetaSession *s)
+{
+  mds_rank_t mds = s->mds_num;
+  ldout(cct, 20) << __func__ << " mds." << mds << dendl;
+
+  int trimmed = 0;
+  list<Dentry*> skipped;
+  while (lru.lru_get_size() > 0) {
+    Dentry *dn = static_cast<Dentry*>(lru.lru_expire());
+    if (!dn)
+      break;
+
+    if ((dn->inode && dn->inode->caps.count(mds)) ||
+	dn->dir->parent_inode->caps.count(mds)) {
+      trim_dentry(dn);
+      trimmed++;
+    } else
+      skipped.push_back(dn);
+  }
+
+  for(list<Dentry*>::iterator p = skipped.begin(); p != skipped.end(); ++p)
+    lru.lru_insert_mid(*p);
+
+  ldout(cct, 20) << __func__ << " mds." << mds
+		 << " trimmed " << trimmed << " dentries" << dendl;
+
+  if (s->caps.size() > 0)
+    _invalidate_kernel_dcache();
+}
+
+void Client::trim_dentry(Dentry *dn)
+{
+  ldout(cct, 15) << "trim_dentry unlinking dn " << dn->name 
+		 << " in dir "
+		 << std::hex << dn->dir->parent_inode->ino << std::dec
+		 << dendl;
+  if (dn->inode) {
+    Inode *diri = dn->dir->parent_inode;
+    clear_dir_complete_and_ordered(diri, true);
+  }
+  unlink(dn, false, false);  // drop dir, drop dentry
+}
+
+
+void Client::update_inode_file_size(Inode *in, int issued, uint64_t size,
+				    uint64_t truncate_seq, uint64_t truncate_size)
+{
+  uint64_t prior_size = in->size;
+
+  if (truncate_seq > in->truncate_seq ||
+      (truncate_seq == in->truncate_seq && size > in->size)) {
+    ldout(cct, 10) << "size " << in->size << " -> " << size << dendl;
+    in->size = size;
+    in->reported_size = size;
+    if (truncate_seq != in->truncate_seq) {
+      ldout(cct, 10) << "truncate_seq " << in->truncate_seq << " -> "
+	       << truncate_seq << dendl;
+      in->truncate_seq = truncate_seq;
+      in->oset.truncate_seq = truncate_seq;
+
+      // truncate cached file data
+      if (prior_size > size) {
+	_invalidate_inode_cache(in, truncate_size, prior_size - truncate_size);
+      }
+    }
+
+    // truncate inline data
+    if (in->inline_version < CEPH_INLINE_NONE) {
+      uint32_t len = in->inline_data.length();
+      if (size < len)
+        in->inline_data.splice(size, len - size);
+    }
+  }
+  if (truncate_seq >= in->truncate_seq &&
+      in->truncate_size != truncate_size) {
+    if (in->is_file()) {
+      ldout(cct, 10) << "truncate_size " << in->truncate_size << " -> "
+	       << truncate_size << dendl;
+      in->truncate_size = truncate_size;
+      in->oset.truncate_size = truncate_size;
+    } else {
+      ldout(cct, 0) << "Hmmm, truncate_seq && truncate_size changed on non-file inode!" << dendl;
+    }
+  }
+}
+
+void Client::update_inode_file_time(Inode *in, int issued, uint64_t time_warp_seq,
+				    utime_t ctime, utime_t mtime, utime_t atime)
+{
+  ldout(cct, 10) << __func__ << " " << *in << " " << ccap_string(issued)
+		 << " ctime " << ctime << " mtime " << mtime << dendl;
+
+  if (time_warp_seq > in->time_warp_seq)
+    ldout(cct, 10) << " mds time_warp_seq " << time_warp_seq
+		   << " is higher than local time_warp_seq "
+		   << in->time_warp_seq << dendl;
+
+  int warn = false;
+  // be careful with size, mtime, atime
+  if (issued & (CEPH_CAP_FILE_EXCL|
+		CEPH_CAP_FILE_WR|
+		CEPH_CAP_FILE_BUFFER|
+		CEPH_CAP_AUTH_EXCL|
+		CEPH_CAP_XATTR_EXCL)) {
+    ldout(cct, 30) << "Yay have enough caps to look at our times" << dendl;
+    if (ctime > in->ctime) 
+      in->ctime = ctime;
+    if (time_warp_seq > in->time_warp_seq) {
+      //the mds updated times, so take those!
+      in->mtime = mtime;
+      in->atime = atime;
+      in->time_warp_seq = time_warp_seq;
+    } else if (time_warp_seq == in->time_warp_seq) {
+      //take max times
+      if (mtime > in->mtime)
+	in->mtime = mtime;
+      if (atime > in->atime)
+	in->atime = atime;
+    } else if (issued & CEPH_CAP_FILE_EXCL) {
+      //ignore mds values as we have a higher seq
+    } else warn = true;
+  } else {
+    ldout(cct, 30) << "Don't have enough caps, just taking mds' time values" << dendl;
+    if (time_warp_seq >= in->time_warp_seq) {
+      in->ctime = ctime;
+      in->mtime = mtime;
+      in->atime = atime;
+      in->time_warp_seq = time_warp_seq;
+    } else warn = true;
+  }
+  if (warn) {
+    ldout(cct, 0) << "WARNING: " << *in << " mds time_warp_seq "
+	    << time_warp_seq << " is lower than local time_warp_seq "
+	    << in->time_warp_seq
+	    << dendl;
+  }
+}
+
+void Client::_fragmap_remove_non_leaves(Inode *in)
+{
+  for (map<frag_t,int>::iterator p = in->fragmap.begin(); p != in->fragmap.end(); )
+    if (!in->dirfragtree.is_leaf(p->first))
+      in->fragmap.erase(p++);
+    else
+      ++p;
+}
+
+void Client::_fragmap_remove_stopped_mds(Inode *in, mds_rank_t mds)
+{
+  for (auto p = in->fragmap.begin(); p != in->fragmap.end(); )
+    if (p->second == mds)
+      in->fragmap.erase(p++);
+    else
+      ++p;
+}
+
+Inode * Client::add_update_inode(InodeStat *st, utime_t from,
+				 MetaSession *session,
+				 const UserPerm& request_perms)
+{
+  Inode *in;
+  bool was_new = false;
+  if (inode_map.count(st->vino)) {
+    in = inode_map[st->vino];
+    ldout(cct, 12) << __func__ << " had " << *in << " caps " << ccap_string(st->cap.caps) << dendl;
+  } else {
+    in = new Inode(this, st->vino, &st->layout);
+    inode_map[st->vino] = in;
+
+    if (use_faked_inos())
+      _assign_faked_ino(in);
+
+    if (!root) {
+      root = in;
+      if (use_faked_inos())
+        _assign_faked_root(root.get());
+      root_ancestor = in;
+      cwd = root;
+    } else if (is_mounting()) {
+      root_parents[root_ancestor] = in;
+      root_ancestor = in;
+    }
+
+    // immutable bits
+    in->ino = st->vino.ino;
+    in->snapid = st->vino.snapid;
+    in->mode = st->mode & S_IFMT;
+    was_new = true;
+  }
+
+  in->rdev = st->rdev;
+  if (in->is_symlink())
+    in->symlink = st->symlink;
+
+  // only update inode if mds info is strictly newer, or it is the same and projected (odd).
+  bool new_version = false;
+  if (in->version == 0 ||
+      ((st->cap.flags & CEPH_CAP_FLAG_AUTH) &&
+       (in->version & ~1) < st->version))
+    new_version = true;
+
+  int issued;
+  in->caps_issued(&issued);
+  issued |= in->caps_dirty();
+  int new_issued = ~issued & (int)st->cap.caps;
+
+  bool need_snapdir_attr_refresh = false;
+  if ((new_version || (new_issued & CEPH_CAP_AUTH_SHARED)) &&
+      !(issued & CEPH_CAP_AUTH_EXCL)) {
+    in->mode = st->mode;
+    in->uid = st->uid;
+    in->gid = st->gid;
+    in->btime = st->btime;
+    in->snap_btime = st->snap_btime;
+    in->snap_metadata = st->snap_metadata;
+    need_snapdir_attr_refresh = true;
+  }
+
+  if ((new_version || (new_issued & CEPH_CAP_LINK_SHARED)) &&
+      !(issued & CEPH_CAP_LINK_EXCL)) {
+    in->nlink = st->nlink;
+  }
+
+  if (new_version || (new_issued & CEPH_CAP_ANY_RD)) {
+    need_snapdir_attr_refresh = true;
+    update_inode_file_time(in, issued, st->time_warp_seq,
+			   st->ctime, st->mtime, st->atime);
+  }
+
+  if (new_version ||
+      (new_issued & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR))) {
+    in->layout = st->layout;
+    update_inode_file_size(in, issued, st->size, st->truncate_seq, st->truncate_size);
+  }
+
+  if (in->is_dir()) {
+    if (new_version || (new_issued & CEPH_CAP_FILE_SHARED)) {
+      in->dirstat = st->dirstat;
+    }
+    // dir_layout/rstat/quota are not tracked by capability, update them only if
+    // the inode stat is from auth mds
+    if (new_version || (st->cap.flags & CEPH_CAP_FLAG_AUTH)) {
+      in->dir_layout = st->dir_layout;
+      ldout(cct, 20) << " dir hash is " << (int)in->dir_layout.dl_dir_hash << dendl;
+      in->rstat = st->rstat;
+      in->quota = st->quota;
+      in->dir_pin = st->dir_pin;
+    }
+    // move me if/when version reflects fragtree changes.
+    if (in->dirfragtree != st->dirfragtree) {
+      in->dirfragtree = st->dirfragtree;
+      _fragmap_remove_non_leaves(in);
+    }
+  }
+
+  if ((in->xattr_version  == 0 || !(issued & CEPH_CAP_XATTR_EXCL)) &&
+      st->xattrbl.length() &&
+      st->xattr_version > in->xattr_version) {
+    auto p = st->xattrbl.cbegin();
+    decode(in->xattrs, p);
+    in->xattr_version = st->xattr_version;
+    need_snapdir_attr_refresh = true;
+  }
+
+  if (st->inline_version > in->inline_version) {
+    in->inline_data = st->inline_data;
+    in->inline_version = st->inline_version;
+  }
+
+  /* always take a newer change attr */
+  ldout(cct, 12) << __func__ << " client inode change_attr: " << in->change_attr << " , mds inodestat change_attr:  " << st->change_attr << dendl;
+  if (st->change_attr > in->change_attr)
+    in->change_attr = st->change_attr;
+
+  if (st->version > in->version)
+    in->version = st->version;
+
+  if (was_new)
+    ldout(cct, 12) << __func__ << " adding " << *in << " caps " << ccap_string(st->cap.caps) << dendl;
+
+  if (!st->cap.caps)
+    return in;   // as with readdir returning indoes in different snaprealms (no caps!)
+
+  if (in->snapid == CEPH_NOSNAP) {
+    add_update_cap(in, session, st->cap.cap_id, st->cap.caps, st->cap.wanted,
+		   st->cap.seq, st->cap.mseq, inodeno_t(st->cap.realm),
+		   st->cap.flags, request_perms);
+    if (in->auth_cap && in->auth_cap->session == session) {
+      in->max_size = st->max_size;
+      in->rstat = st->rstat;
+    }
+
+    // setting I_COMPLETE needs to happen after adding the cap
+    if (in->is_dir() &&
+	(st->cap.caps & CEPH_CAP_FILE_SHARED) &&
+	(issued & CEPH_CAP_FILE_EXCL) == 0 &&
+	in->dirstat.nfiles == 0 &&
+	in->dirstat.nsubdirs == 0) {
+      ldout(cct, 10) << " marking (I_COMPLETE|I_DIR_ORDERED) on empty dir " << *in << dendl;
+      in->flags |= I_COMPLETE | I_DIR_ORDERED;
+      if (in->dir) {
+	ldout(cct, 10) << " dir is open on empty dir " << in->ino << " with "
+		       << in->dir->dentries.size() << " entries, marking all dentries null" << dendl;
+	in->dir->readdir_cache.clear();
+	for (const auto& p : in->dir->dentries) {
+	  unlink(p.second, true, true);  // keep dir, keep dentry
+	}
+	if (in->dir->dentries.empty())
+	  close_dir(in->dir);
+      }
+    }
+  } else {
+    in->snap_caps |= st->cap.caps;
+  }
+
+  in->fscrypt = st->fscrypt;
+  if (need_snapdir_attr_refresh && in->is_dir() && in->snapid == CEPH_NOSNAP) {
+    vinodeno_t vino(in->ino, CEPH_SNAPDIR);
+    if (inode_map.count(vino)) {
+      refresh_snapdir_attrs(inode_map[vino], in);
+    }
+  }
+
+  return in;
+}
+
+
+/*
+ * insert_dentry_inode - insert + link a single dentry + inode into the metadata cache.
+ */
+Dentry *Client::insert_dentry_inode(Dir *dir, const string& dname, LeaseStat *dlease, 
+				    Inode *in, utime_t from, MetaSession *session,
+				    Dentry *old_dentry)
+{
+  Dentry *dn = NULL;
+  if (dir->dentries.count(dname))
+    dn = dir->dentries[dname];
+
+  ldout(cct, 12) << __func__ << " '" << dname << "' vino " << in->vino()
+		 << " in dir " << dir->parent_inode->vino() << " dn " << dn
+		 << dendl;
+
+  if (dn && dn->inode) {
+    if (dn->inode->vino() == in->vino()) {
+      touch_dn(dn);
+      ldout(cct, 12) << " had dentry " << dname
+	       << " with correct vino " << dn->inode->vino()
+	       << dendl;
+    } else {
+      ldout(cct, 12) << " had dentry " << dname
+	       << " with WRONG vino " << dn->inode->vino()
+	       << dendl;
+      unlink(dn, true, true);  // keep dir, keep dentry
+    }
+  }
+  
+  if (!dn || !dn->inode) {
+    InodeRef tmp_ref(in);
+    if (old_dentry) {
+      if (old_dentry->dir != dir) {
+	Inode *old_diri = old_dentry->dir->parent_inode;
+	clear_dir_complete_and_ordered(old_diri, false);
+      }
+      unlink(old_dentry, dir == old_dentry->dir, false);  // drop dentry, keep dir open if its the same dir
+    }
+    Inode *diri = dir->parent_inode;
+    clear_dir_complete_and_ordered(diri, false);
+    dn = link(dir, dname, in, dn);
+  }
+
+  update_dentry_lease(dn, dlease, from, session);
+  return dn;
+}
+
+void Client::update_dentry_lease(Dentry *dn, LeaseStat *dlease, utime_t from, MetaSession *session)
+{
+  utime_t dttl = from;
+  dttl += (float)dlease->duration_ms / 1000.0;
+
+  ldout(cct, 15) << __func__ << " " << *dn << " " << *dlease << " from " << from << dendl;
+  
+  ceph_assert(dn);
+
+  if (dlease->mask & CEPH_LEASE_VALID) {
+    if (dttl > dn->lease_ttl) {
+      ldout(cct, 10) << "got dentry lease on " << dn->name
+	       << " dur " << dlease->duration_ms << "ms ttl " << dttl << dendl;
+      dn->lease_ttl = dttl;
+      dn->lease_mds = session->mds_num;
+      dn->lease_seq = dlease->seq;
+      dn->lease_gen = session->cap_gen;
+    }
+  }
+  dn->cap_shared_gen = dn->dir->parent_inode->shared_gen;
+  if (dlease->mask & CEPH_LEASE_PRIMARY_LINK)
+    dn->mark_primary();
+  dn->alternate_name = std::move(dlease->alternate_name);
+}
+
+
+/*
+ * update MDS location cache for a single inode
+ */
+void Client::update_dir_dist(Inode *in, DirStat *dst, mds_rank_t from)
+{
+  // auth
+  ldout(cct, 20) << "got dirfrag map for " << in->ino << " frag " << dst->frag << " to mds " << dst->auth << dendl;
+  if (dst->auth >= 0) {
+    in->fragmap[dst->frag] = dst->auth;
+  } else {
+    in->fragmap.erase(dst->frag);
+  }
+  if (!in->dirfragtree.is_leaf(dst->frag)) {
+    in->dirfragtree.force_to_leaf(cct, dst->frag);
+    _fragmap_remove_non_leaves(in);
+  }
+
+  // replicated, only update from auth mds reply
+  if (from == dst->auth) {
+    in->dir_replicated = !dst->dist.empty();
+    if (!dst->dist.empty())
+      in->frag_repmap[dst->frag].assign(dst->dist.begin(), dst->dist.end()) ;
+    else
+      in->frag_repmap.erase(dst->frag);
+  }
+}
+
+void Client::clear_dir_complete_and_ordered(Inode *diri, bool complete)
+{
+  if (complete)
+    diri->dir_release_count++;
+  else
+    diri->dir_ordered_count++;
+  if (diri->flags & I_COMPLETE) {
+    if (complete) {
+      ldout(cct, 10) << " clearing (I_COMPLETE|I_DIR_ORDERED) on " << *diri << dendl;
+      diri->flags &= ~(I_COMPLETE | I_DIR_ORDERED);
+    } else {
+      if (diri->flags & I_DIR_ORDERED) {
+	ldout(cct, 10) << " clearing I_DIR_ORDERED on " << *diri << dendl;
+	diri->flags &= ~I_DIR_ORDERED;
+      }
+    }
+    if (diri->dir)
+      diri->dir->readdir_cache.clear();
+  }
+}
+
+/*
+ * insert results from readdir or lssnap into the metadata cache.
+ */
+void Client::insert_readdir_results(MetaRequest *request, MetaSession *session, Inode *diri) {
+
+  auto& reply = request->reply;
+  ConnectionRef con = request->reply->get_connection();
+  uint64_t features;
+  if(session->mds_features.test(CEPHFS_FEATURE_REPLY_ENCODING)) {
+    features = (uint64_t)-1;
+  }
+  else {
+    features = con->get_features();
+  }
+
+  dir_result_t *dirp = request->dirp;
+  ceph_assert(dirp);
+
+  // the extra buffer list is only set for readdir and lssnap replies
+  auto p = reply->get_extra_bl().cbegin();
+  if (!p.end()) {
+    // snapdir?
+    if (request->head.op == CEPH_MDS_OP_LSSNAP) {
+      ceph_assert(diri);
+      diri = open_snapdir(diri);
+    }
+
+    // only open dir if we're actually adding stuff to it!
+    Dir *dir = diri->open_dir();
+    ceph_assert(dir);
+
+    // dirstat
+    DirStat dst(p, features);
+    __u32 numdn;
+    __u16 flags;
+    decode(numdn, p);
+    decode(flags, p);
+
+    bool end = ((unsigned)flags & CEPH_READDIR_FRAG_END);
+    bool hash_order = ((unsigned)flags & CEPH_READDIR_HASH_ORDER);
+
+    frag_t fg = (unsigned)request->head.args.readdir.frag;
+    unsigned readdir_offset = dirp->next_offset;
+    string readdir_start = dirp->last_name;
+    ceph_assert(!readdir_start.empty() || readdir_offset == 2);
+
+    unsigned last_hash = 0;
+    if (hash_order) {
+      if (!readdir_start.empty()) {
+	last_hash = ceph_frag_value(diri->hash_dentry_name(readdir_start));
+      } else if (flags & CEPH_READDIR_OFFSET_HASH) {
+	/* mds understands offset_hash */
+	last_hash = (unsigned)request->head.args.readdir.offset_hash;
+      }
+    }
+
+    if (fg != dst.frag) {
+      ldout(cct, 10) << "insert_trace got new frag " << fg << " -> " << dst.frag << dendl;
+      fg = dst.frag;
+      if (!hash_order) {
+	readdir_offset = 2;
+	readdir_start.clear();
+	dirp->offset = dir_result_t::make_fpos(fg, readdir_offset, false);
+      }
+    }
+
+    ldout(cct, 10) << __func__ << " " << numdn << " readdir items, end=" << end
+		   << ", hash_order=" << hash_order
+		   << ", readdir_start " << readdir_start
+		   << ", last_hash " << last_hash
+		   << ", next_offset " << readdir_offset << dendl;
+
+    if (diri->snapid != CEPH_SNAPDIR &&
+	fg.is_leftmost() && readdir_offset == 2 &&
+	!(hash_order && last_hash)) {
+      dirp->release_count = diri->dir_release_count;
+      dirp->ordered_count = diri->dir_ordered_count;
+      dirp->start_shared_gen = diri->shared_gen;
+      dirp->cache_index = 0;
+    }
+
+    dirp->buffer_frag = fg;
+
+    _readdir_drop_dirp_buffer(dirp);
+    dirp->buffer.reserve(numdn);
+
+    string dname;
+    LeaseStat dlease;
+    for (unsigned i=0; i<numdn; i++) {
+      decode(dname, p);
+      dlease.decode(p, features);
+      InodeStat ist(p, features);
+
+      ldout(cct, 15) << "" << i << ": '" << dname << "'" << dendl;
+
+      Inode *in = add_update_inode(&ist, request->sent_stamp, session,
+				   request->perms);
+      Dentry *dn;
+      if (diri->dir->dentries.count(dname)) {
+	Dentry *olddn = diri->dir->dentries[dname];
+	if (olddn->inode != in) {
+	  // replace incorrect dentry
+	  unlink(olddn, true, true);  // keep dir, dentry
+	  dn = link(dir, dname, in, olddn);
+	  ceph_assert(dn == olddn);
+	} else {
+	  // keep existing dn
+	  dn = olddn;
+	  touch_dn(dn);
+	}
+      } else {
+	// new dn
+	dn = link(dir, dname, in, NULL);
+      }
+      dn->alternate_name = std::move(dlease.alternate_name);
+
+      update_dentry_lease(dn, &dlease, request->sent_stamp, session);
+      if (hash_order) {
+	unsigned hash = ceph_frag_value(diri->hash_dentry_name(dname));
+	if (hash != last_hash)
+	  readdir_offset = 2;
+	last_hash = hash;
+	dn->offset = dir_result_t::make_fpos(hash, readdir_offset++, true);
+      } else {
+	dn->offset = dir_result_t::make_fpos(fg, readdir_offset++, false);
+      }
+      // add to readdir cache
+      if (dirp->release_count == diri->dir_release_count &&
+	  dirp->ordered_count == diri->dir_ordered_count &&
+	  dirp->start_shared_gen == diri->shared_gen) {
+	if (dirp->cache_index == dir->readdir_cache.size()) {
+	  if (i == 0) {
+	    ceph_assert(!dirp->inode->is_complete_and_ordered());
+	    dir->readdir_cache.reserve(dirp->cache_index + numdn);
+	  }
+	  dir->readdir_cache.push_back(dn);
+	} else if (dirp->cache_index < dir->readdir_cache.size()) {
+	  if (dirp->inode->is_complete_and_ordered())
+	    ceph_assert(dir->readdir_cache[dirp->cache_index] == dn);
+	  else
+	    dir->readdir_cache[dirp->cache_index] = dn;
+	} else {
+	  ceph_abort_msg("unexpected readdir buffer idx");
+	}
+	dirp->cache_index++;
+      }
+      // add to cached result list
+      dirp->buffer.push_back(dir_result_t::dentry(dn->offset, dname, dn->alternate_name, in));
+      ldout(cct, 15) << __func__ << "  " << hex << dn->offset << dec << ": '" << dname << "' -> " << in->ino << dendl;
+    }
+
+    if (numdn > 0)
+      dirp->last_name = dname;
+    if (end)
+      dirp->next_offset = 2;
+    else
+      dirp->next_offset = readdir_offset;
+
+    if (dir->is_empty())
+      close_dir(dir);
+  }
+}
+
+/** insert_trace
+ *
+ * insert a trace from a MDS reply into the cache.
+ */
+Inode* Client::insert_trace(MetaRequest *request, MetaSession *session)
+{
+  auto& reply = request->reply;
+  int op = request->get_op();
+
+  ldout(cct, 10) << "insert_trace from " << request->sent_stamp << " mds." << session->mds_num
+	   << " is_target=" << (int)reply->head.is_target
+	   << " is_dentry=" << (int)reply->head.is_dentry
+	   << dendl;
+
+  auto p = reply->get_trace_bl().cbegin();
+  if (request->got_unsafe) {
+    ldout(cct, 10) << "insert_trace -- already got unsafe; ignoring" << dendl;
+    ceph_assert(p.end());
+    return NULL;
+  }
+
+  if (p.end()) {
+    ldout(cct, 10) << "insert_trace -- no trace" << dendl;
+
+    Dentry *d = request->dentry();
+    if (d) {
+      Inode *diri = d->dir->parent_inode;
+      clear_dir_complete_and_ordered(diri, true);
+    }
+
+    if (d && reply->get_result() == 0) {
+      if (op == CEPH_MDS_OP_RENAME) {
+	// rename
+	Dentry *od = request->old_dentry();
+	ldout(cct, 10) << " unlinking rename src dn " << od << " for traceless reply" << dendl;
+	ceph_assert(od);
+	unlink(od, true, true);  // keep dir, dentry
+      } else if (op == CEPH_MDS_OP_RMDIR ||
+		 op == CEPH_MDS_OP_UNLINK) {
+	// unlink, rmdir
+	ldout(cct, 10) << " unlinking unlink/rmdir dn " << d << " for traceless reply" << dendl;
+	unlink(d, true, true);  // keep dir, dentry
+      }
+    }
+    return NULL;
+  }
+
+  ConnectionRef con = request->reply->get_connection();
+  uint64_t features;
+  if (session->mds_features.test(CEPHFS_FEATURE_REPLY_ENCODING)) {
+    features = (uint64_t)-1;
+  }
+  else {
+    features = con->get_features();
+  }
+  ldout(cct, 10) << " features 0x" << hex << features << dec << dendl;
+
+  // snap trace
+  SnapRealm *realm = NULL;
+  if (reply->snapbl.length())
+    update_snap_trace(reply->snapbl, &realm);
+
+  ldout(cct, 10) << " hrm " 
+	   << " is_target=" << (int)reply->head.is_target
+	   << " is_dentry=" << (int)reply->head.is_dentry
+	   << dendl;
+
+  InodeStat dirst;
+  DirStat dst;
+  string dname;
+  LeaseStat dlease;
+  InodeStat ist;
+
+  if (reply->head.is_dentry) {
+    dirst.decode(p, features);
+    dst.decode(p, features);
+    decode(dname, p);
+    dlease.decode(p, features);
+  }
+
+  Inode *in = 0;
+  if (reply->head.is_target) {
+    ist.decode(p, features);
+    if (cct->_conf->client_debug_getattr_caps) {
+      unsigned wanted = 0;
+      if (op == CEPH_MDS_OP_GETATTR || op == CEPH_MDS_OP_LOOKUP)
+	wanted = request->head.args.getattr.mask;
+      else if (op == CEPH_MDS_OP_OPEN || op == CEPH_MDS_OP_CREATE)
+	wanted = request->head.args.open.mask;
+
+      if ((wanted & CEPH_CAP_XATTR_SHARED) &&
+	  !(ist.xattr_version > 0 && ist.xattrbl.length() > 0))
+	ceph_abort_msg("MDS reply does not contain xattrs");
+    }
+
+    in = add_update_inode(&ist, request->sent_stamp, session,
+			  request->perms);
+  }
+
+  Inode *diri = NULL;
+  if (reply->head.is_dentry) {
+    diri = add_update_inode(&dirst, request->sent_stamp, session,
+			    request->perms);
+    mds_rank_t from_mds = mds_rank_t(reply->get_source().num());
+    update_dir_dist(diri, &dst, from_mds);  // dir stat info is attached to ..
+
+    if (in) {
+      Dir *dir = diri->open_dir();
+      insert_dentry_inode(dir, dname, &dlease, in, request->sent_stamp, session,
+                          (op == CEPH_MDS_OP_RENAME) ? request->old_dentry() : NULL);
+    } else {
+      Dentry *dn = NULL;
+      if (diri->dir && diri->dir->dentries.count(dname)) {
+	dn = diri->dir->dentries[dname];
+	if (dn->inode) {
+	  clear_dir_complete_and_ordered(diri, false);
+	  unlink(dn, true, true);  // keep dir, dentry
+	}
+      }
+      if (dlease.duration_ms > 0) {
+	if (!dn) {
+	  Dir *dir = diri->open_dir();
+	  dn = link(dir, dname, NULL, NULL);
+	}
+	update_dentry_lease(dn, &dlease, request->sent_stamp, session);
+      }
+    }
+  } else if (op == CEPH_MDS_OP_LOOKUPSNAP ||
+	     op == CEPH_MDS_OP_MKSNAP) {
+    ldout(cct, 10) << " faking snap lookup weirdness" << dendl;
+    // fake it for snap lookup
+    vinodeno_t vino = ist.vino;
+    vino.snapid = CEPH_SNAPDIR;
+    ceph_assert(inode_map.count(vino));
+    diri = inode_map[vino];
+    
+    string dname = request->path.last_dentry();
+    
+    LeaseStat dlease;
+    dlease.duration_ms = 0;
+
+    if (in) {
+      Dir *dir = diri->open_dir();
+      insert_dentry_inode(dir, dname, &dlease, in, request->sent_stamp, session);
+    } else {
+      if (diri->dir && diri->dir->dentries.count(dname)) {
+	Dentry *dn = diri->dir->dentries[dname];
+	if (dn->inode)
+	  unlink(dn, true, true);  // keep dir, dentry
+      }
+    }
+  }
+
+  if (in) {
+    if (op == CEPH_MDS_OP_READDIR ||
+	op == CEPH_MDS_OP_LSSNAP) {
+      insert_readdir_results(request, session, in);
+    } else if (op == CEPH_MDS_OP_LOOKUPNAME) {
+      // hack: return parent inode instead
+      in = diri;
+    }
+
+    if (request->dentry() == NULL && in != request->inode()) {
+      // pin the target inode if its parent dentry is not pinned
+      request->set_other_inode(in);
+    }
+  }
+
+  if (realm)
+    put_snap_realm(realm);
+
+  request->target = in;
+  return in;
+}
+
+// -------
+
+mds_rank_t Client::choose_target_mds(MetaRequest *req, Inode** phash_diri)
+{
+  mds_rank_t mds = MDS_RANK_NONE;
+  __u32 hash = 0;
+  bool is_hash = false;
+  int issued = 0;
+
+  Inode *in = NULL;
+  Dentry *de = NULL;
+
+  if (req->resend_mds >= 0) {
+    mds = req->resend_mds;
+    req->resend_mds = -1;
+    ldout(cct, 10) << __func__ << " resend_mds specified as mds." << mds << dendl;
+    goto out;
+  }
+
+  if (cct->_conf->client_use_random_mds)
+    goto random_mds;
+
+  in = req->inode();
+  de = req->dentry();
+  if (in) {
+    ldout(cct, 20) << __func__ << " starting with req->inode " << *in << dendl;
+    if (req->path.depth()) {
+      hash = in->hash_dentry_name(req->path[0]);
+      ldout(cct, 20) << __func__ << " inode dir hash is " << (int)in->dir_layout.dl_dir_hash
+	       << " on " << req->path[0]
+	       << " => " << hash << dendl;
+      is_hash = true;
+    }
+  } else if (de) {
+    if (de->inode) {
+      in = de->inode.get();
+      ldout(cct, 20) << __func__ << " starting with req->dentry inode " << *in << dendl;
+    } else {
+      in = de->dir->parent_inode;
+      hash = in->hash_dentry_name(de->name);
+      ldout(cct, 20) << __func__ << " dentry dir hash is " << (int)in->dir_layout.dl_dir_hash
+	       << " on " << de->name
+	       << " => " << hash << dendl;
+      is_hash = true;
+    }
+  }
+  if (in) {
+    if (in->snapid != CEPH_NOSNAP) {
+      ldout(cct, 10) << __func__ << " " << *in << " is snapped, using nonsnap parent" << dendl;
+      while (in->snapid != CEPH_NOSNAP) {
+        if (in->snapid == CEPH_SNAPDIR)
+	  in = in->snapdir_parent.get();
+        else if (!in->dentries.empty())
+          /* In most cases there will only be one dentry, so getting it
+           * will be the correct action. If there are multiple hard links,
+           * I think the MDS should be able to redirect as needed*/
+	  in = in->get_first_parent()->dir->parent_inode;
+        else {
+          ldout(cct, 10) << "got unlinked inode, can't look at parent" << dendl;
+          break;
+        }
+      }
+      is_hash = false;
+    }
+  
+    ldout(cct, 20) << __func__ << " " << *in << " is_hash=" << is_hash
+             << " hash=" << hash << dendl;
+  
+    if (req->get_op() == CEPH_MDS_OP_GETATTR)
+      issued = req->inode()->caps_issued();
+
+    if (is_hash && S_ISDIR(in->mode) && (!in->fragmap.empty() || !in->frag_repmap.empty())) {
+      frag_t fg = in->dirfragtree[hash];
+      if (!req->auth_is_best(issued)) {
+        auto repmapit = in->frag_repmap.find(fg);
+        if (repmapit != in->frag_repmap.end()) {
+          auto& repmap = repmapit->second;
+          auto r = ceph::util::generate_random_number<uint64_t>(0, repmap.size()-1);
+          mds = repmap.at(r);
+        }
+      } else if (in->fragmap.count(fg)) {
+	mds = in->fragmap[fg];
+	if (phash_diri)
+	  *phash_diri = in;
+      } else if (in->auth_cap) {
+	req->send_to_auth = true;
+	mds = in->auth_cap->session->mds_num;
+      }
+      if (mds >= 0) {
+	ldout(cct, 10) << __func__ << " from dirfragtree hash" << dendl;
+	goto out;
+      }
+    }
+  
+    if (in->auth_cap && req->auth_is_best(issued)) {
+      mds = in->auth_cap->session->mds_num;
+    } else if (!in->caps.empty()) {
+      mds = in->caps.begin()->second.session->mds_num;
+    } else {
+      goto random_mds;
+    }
+    ldout(cct, 10) << __func__ << " from caps on inode " << *in << dendl;
+  
+    goto out;
+  }
+
+random_mds:
+  if (mds < 0) {
+    mds = _get_random_up_mds();
+    ldout(cct, 10) << "did not get mds through better means, so chose random mds " << mds << dendl;
+  }
+
+out:
+  ldout(cct, 20) << "mds is " << mds << dendl;
+  return mds;
+}
+
+void Client::connect_mds_targets(mds_rank_t mds)
+{
+  ldout(cct, 10) << __func__ << " for mds." << mds << dendl;
+  ceph_assert(mds_sessions.count(mds));
+  const MDSMap::mds_info_t& info = mdsmap->get_mds_info(mds);
+  for (const auto &rank : info.export_targets) {
+    if (mds_sessions.count(rank) == 0 &&
+	mdsmap->is_clientreplay_or_active_or_stopping(rank)) {
+      ldout(cct, 10) << "check_mds_sessions opening mds." << mds
+		     << " export target mds." << rank << dendl;
+      _open_mds_session(rank);
+    }
+  }
+}
+
+void Client::dump_mds_sessions(Formatter *f, bool cap_dump)
+{
+  f->dump_int("id", get_nodeid().v);
+  entity_inst_t inst(messenger->get_myname(), messenger->get_myaddr_legacy());
+  f->dump_object("inst", inst);
+  f->dump_stream("inst_str") << inst;
+  f->dump_stream("addr_str") << inst.addr;
+  f->open_array_section("sessions");
+  for (const auto &p : mds_sessions) {
+    f->open_object_section("session");
+    p.second.dump(f, cap_dump);
+    f->close_section();
+  }
+  f->close_section();
+  f->dump_int("mdsmap_epoch", mdsmap->get_epoch());
+}
+
+void Client::dump_mds_requests(Formatter *f)
+{
+  for (map<ceph_tid_t, MetaRequest*>::iterator p = mds_requests.begin();
+       p != mds_requests.end();
+       ++p) {
+    f->open_object_section("request");
+    p->second->dump(f);
+    f->close_section();
+  }
+}
+
+int Client::verify_reply_trace(int r, MetaSession *session,
+			       MetaRequest *request, const MConstRef<MClientReply>& reply,
+			       InodeRef *ptarget, bool *pcreated,
+			       const UserPerm& perms)
+{
+  // check whether this request actually did the create, and set created flag
+  bufferlist extra_bl;
+  inodeno_t created_ino;
+  bool got_created_ino = false;
+  ceph::unordered_map<vinodeno_t, Inode*>::iterator p;
+
+  extra_bl = reply->get_extra_bl();
+  if (extra_bl.length() >= 8) {
+    if (session->mds_features.test(CEPHFS_FEATURE_DELEG_INO)) {
+     struct openc_response_t	ocres;
+
+     decode(ocres, extra_bl);
+     created_ino = ocres.created_ino;
+     /*
+      * The userland cephfs client doesn't have a way to do an async create
+      * (yet), so just discard delegated_inos for now. Eventually we should
+      * store them and use them in create calls, even if they are synchronous,
+      * if only for testing purposes.
+      */
+     ldout(cct, 10) << "delegated_inos: " << ocres.delegated_inos << dendl;
+    } else {
+     // u64 containing number of created ino
+     decode(created_ino, extra_bl);
+    }
+    ldout(cct, 10) << "make_request created ino " << created_ino << dendl;
+    got_created_ino = true;
+  }
+
+  if (pcreated)
+    *pcreated = got_created_ino;
+
+  if (request->target) {
+    *ptarget = request->target;
+    ldout(cct, 20) << "make_request target is " << *ptarget->get() << dendl;
+  } else {
+    if (got_created_ino && (p = inode_map.find(vinodeno_t(created_ino, CEPH_NOSNAP))) != inode_map.end()) {
+      (*ptarget) = p->second;
+      ldout(cct, 20) << "make_request created, target is " << *ptarget->get() << dendl;
+    } else {
+      // we got a traceless reply, and need to look up what we just
+      // created.  for now, do this by name.  someday, do this by the
+      // ino... which we know!  FIXME.
+      InodeRef target;
+      Dentry *d = request->dentry();
+      if (d) {
+	if (d->dir) {
+	  ldout(cct, 10) << "make_request got traceless reply, looking up #"
+			 << d->dir->parent_inode->ino << "/" << d->name
+			 << " got_ino " << got_created_ino
+			 << " ino " << created_ino
+			 << dendl;
+	  r = _do_lookup(d->dir->parent_inode, d->name, request->regetattr_mask,
+			 &target, perms);
+	} else {
+	  // if the dentry is not linked, just do our best. see #5021.
+	  ceph_abort_msg("how did this happen?  i want logs!");
+	}
+      } else {
+	Inode *in = request->inode();
+	ldout(cct, 10) << "make_request got traceless reply, forcing getattr on #"
+		       << in->ino << dendl;
+	r = _getattr(in, request->regetattr_mask, perms, true);
+	target = in;
+      }
+      if (r >= 0) {
+	// verify ino returned in reply and trace_dist are the same
+	if (got_created_ino &&
+	    created_ino.val != target->ino.val) {
+	  ldout(cct, 5) << "create got ino " << created_ino << " but then failed on lookup; EINTR?" << dendl;
+	  r = -CEPHFS_EINTR;
+	}
+	if (ptarget)
+	  ptarget->swap(target);
+      }
+    }
+  }
+
+  return r;
+}
+
+
+/**
+ * make a request
+ *
+ * Blocking helper to make an MDS request.
+ *
+ * If the ptarget flag is set, behavior changes slightly: the caller
+ * expects to get a pointer to the inode we are creating or operating
+ * on.  As a result, we will follow up any traceless mutation reply
+ * with a getattr or lookup to transparently handle a traceless reply
+ * from the MDS (as when the MDS restarts and the client has to replay
+ * a request).
+ *
+ * @param request the MetaRequest to execute
+ * @param perms The user uid/gid to execute as (eventually, full group lists?)
+ * @param ptarget [optional] address to store a pointer to the target inode we want to create or operate on
+ * @param pcreated [optional; required if ptarget] where to store a bool of whether our create atomically created a file
+ * @param use_mds [optional] prefer a specific mds (-1 for default)
+ * @param pdirbl [optional; disallowed if ptarget] where to pass extra reply payload to the caller
+ */
+int Client::make_request(MetaRequest *request,
+			 const UserPerm& perms,
+			 InodeRef *ptarget, bool *pcreated,
+			 mds_rank_t use_mds,
+			 bufferlist *pdirbl,
+			 size_t feature_needed)
+{
+  int r = 0;
+
+  // assign a unique tid
+  ceph_tid_t tid = ++last_tid;
+  request->set_tid(tid);
+
+  // and timestamp
+  request->op_stamp = ceph_clock_now();
+  request->created = ceph::coarse_mono_clock::now();
+
+  // make note
+  mds_requests[tid] = request->get();
+  if (oldest_tid == 0 && request->get_op() != CEPH_MDS_OP_SETFILELOCK)
+    oldest_tid = tid;
+
+  request->set_caller_perms(perms);
+
+  if (cct->_conf->client_inject_fixed_oldest_tid) {
+    ldout(cct, 20) << __func__ << " injecting fixed oldest_client_tid(1)" << dendl;
+    request->set_oldest_client_tid(1);
+  } else {
+    request->set_oldest_client_tid(oldest_tid);
+  }
+
+  // hack target mds?
+  if (use_mds >= 0)
+    request->resend_mds = use_mds;
+
+  MetaSession *session = NULL;
+  while (1) {
+    if (request->aborted())
+      break;
+
+    if (blocklisted) {
+      request->abort(-CEPHFS_EBLOCKLISTED);
+      break;
+    }
+
+    // set up wait cond
+    ceph::condition_variable caller_cond;
+    request->caller_cond = &caller_cond;
+
+    // choose mds
+    Inode *hash_diri = NULL;
+    mds_rank_t mds = choose_target_mds(request, &hash_diri);
+    int mds_state = (mds == MDS_RANK_NONE) ? MDSMap::STATE_NULL : mdsmap->get_state(mds);
+    if (mds_state != MDSMap::STATE_ACTIVE && mds_state != MDSMap::STATE_STOPPING) {
+      if (mds_state == MDSMap::STATE_NULL && mds >= mdsmap->get_max_mds()) {
+	if (hash_diri) {
+	  ldout(cct, 10) << " target mds." << mds << " has stopped, remove it from fragmap" << dendl;
+	  _fragmap_remove_stopped_mds(hash_diri, mds);
+	} else {
+	  ldout(cct, 10) << " target mds." << mds << " has stopped, trying a random mds" << dendl;
+	  request->resend_mds = _get_random_up_mds();
+	}
+      } else {
+	ldout(cct, 10) << " target mds." << mds << " not active, waiting for new mdsmap" << dendl;
+	wait_on_list(waiting_for_mdsmap);
+      }
+      continue;
+    }
+
+    // open a session?
+    if (!have_open_session(mds)) {
+      session = _get_or_open_mds_session(mds);
+      if (session->state == MetaSession::STATE_REJECTED) {
+	request->abort(-CEPHFS_EPERM);
+	break;
+      }
+      // wait
+      if (session->state == MetaSession::STATE_OPENING) {
+	ldout(cct, 10) << "waiting for session to mds." << mds << " to open" << dendl;
+	wait_on_context_list(session->waiting_for_open);
+	continue;
+      }
+
+      if (!have_open_session(mds))
+	continue;
+    } else {
+      session = &mds_sessions.at(mds);
+    }
+
+    if (feature_needed != ULONG_MAX && !session->mds_features.test(feature_needed)) {
+      request->abort(-CEPHFS_EOPNOTSUPP);
+      break;
+    }
+
+    // send request.
+    send_request(request, session);
+
+    // wait for signal
+    ldout(cct, 20) << "awaiting reply|forward|kick on " << &caller_cond << dendl;
+    request->kick = false;
+    std::unique_lock l{client_lock, std::adopt_lock};
+    caller_cond.wait(l, [request] {
+      return (request->reply ||	          // reply
+	      request->resend_mds >= 0 || // forward
+	      request->kick);
+    });
+    l.release();
+    request->caller_cond = nullptr;
+
+    // did we get a reply?
+    if (request->reply)
+      break;
+  }
+
+  if (!request->reply) {
+    ceph_assert(request->aborted());
+    ceph_assert(!request->got_unsafe);
+    r = request->get_abort_code();
+    request->item.remove_myself();
+    unregister_request(request);
+    put_request(request);
+    return r;
+  }
+
+  // got it!
+  auto reply = std::move(request->reply);
+  r = reply->get_result();
+  if (r >= 0)
+    request->success = true;
+
+  // kick dispatcher (we've got it!)
+  ceph_assert(request->dispatch_cond);
+  request->dispatch_cond->notify_all();
+  ldout(cct, 20) << "sendrecv kickback on tid " << tid << " " << request->dispatch_cond << dendl;
+  request->dispatch_cond = 0;
+  
+  if (r >= 0 && ptarget)
+    r = verify_reply_trace(r, session, request, reply, ptarget, pcreated, perms);
+
+  if (pdirbl)
+    *pdirbl = reply->get_extra_bl();
+
+  // -- log times --
+  utime_t lat = ceph_clock_now();
+  lat -= request->sent_stamp;
+  ldout(cct, 20) << "lat " << lat << dendl;
+
+  ++nr_metadata_request;
+  update_io_stat_metadata(lat);
+
+  put_request(request);
+  return r;
+}
+
+void Client::unregister_request(MetaRequest *req)
+{
+  mds_requests.erase(req->tid);
+  if (req->tid == oldest_tid) {
+    map<ceph_tid_t, MetaRequest*>::iterator p = mds_requests.upper_bound(oldest_tid);
+    while (true) {
+      if (p == mds_requests.end()) {
+	oldest_tid = 0;
+	break;
+      }
+      if (p->second->get_op() != CEPH_MDS_OP_SETFILELOCK) {
+	oldest_tid = p->first;
+	break;
+      }
+      ++p;
+    }
+  }
+  put_request(req);
+}
+
+void Client::put_request(MetaRequest *request)
+{
+  if (request->_put()) {
+    int op = -1;
+    if (request->success)
+      op = request->get_op();
+    InodeRef other_in;
+    request->take_other_inode(&other_in);
+    delete request;
+
+    if (other_in &&
+	(op == CEPH_MDS_OP_RMDIR ||
+	 op == CEPH_MDS_OP_RENAME ||
+	 op == CEPH_MDS_OP_RMSNAP)) {
+      _try_to_trim_inode(other_in.get(), false);
+    }
+  }
+}
+
+int Client::encode_inode_release(Inode *in, MetaRequest *req,
+			 mds_rank_t mds, int drop,
+			 int unless, int force)
+{
+  ldout(cct, 20) << __func__ << " enter(in:" << *in << ", req:" << req
+	   << " mds:" << mds << ", drop:" << ccap_string(drop) << ", unless:" << ccap_string(unless)
+	   << ", force:" << force << ")" << dendl;
+  int released = 0;
+  auto it = in->caps.find(mds);
+  if (it != in->caps.end()) {
+    Cap &cap = it->second;
+    drop &= ~(in->dirty_caps | get_caps_used(in));
+    if ((drop & cap.issued) &&
+	!(unless & cap.issued)) {
+      ldout(cct, 25) << "dropping caps " << ccap_string(drop) << dendl;
+      cap.issued &= ~drop;
+      cap.implemented &= ~drop;
+      released = 1;
+    } else {
+      released = force;
+    }
+    if (released) {
+      cap.wanted = in->caps_wanted();
+      if (&cap == in->auth_cap &&
+	  !(cap.wanted & CEPH_CAP_ANY_FILE_WR)) {
+	in->requested_max_size = 0;
+	ldout(cct, 25) << "reset requested_max_size due to not wanting any file write cap" << dendl;
+      }
+      ceph_mds_request_release rel;
+      rel.ino = in->ino;
+      rel.cap_id = cap.cap_id;
+      rel.seq = cap.seq;
+      rel.issue_seq = cap.issue_seq;
+      rel.mseq = cap.mseq;
+      rel.caps = cap.implemented;
+      rel.wanted = cap.wanted;
+      rel.dname_len = 0;
+      rel.dname_seq = 0;
+      req->cap_releases.push_back(MClientRequest::Release(rel,""));
+    }
+  }
+  ldout(cct, 25) << __func__ << " exit(in:" << *in << ") released:"
+	   << released << dendl;
+  return released;
+}
+
+void Client::encode_dentry_release(Dentry *dn, MetaRequest *req,
+			   mds_rank_t mds, int drop, int unless)
+{
+  ldout(cct, 20) << __func__ << " enter(dn:"
+	   << dn << ")" << dendl;
+  int released = 0;
+  if (dn->dir)
+    released = encode_inode_release(dn->dir->parent_inode, req,
+				    mds, drop, unless, 1);
+  if (released && dn->lease_mds == mds) {
+    ldout(cct, 25) << "preemptively releasing dn to mds" << dendl;
+    auto& rel = req->cap_releases.back();
+    rel.item.dname_len = dn->name.length();
+    rel.item.dname_seq = dn->lease_seq;
+    rel.dname = dn->name;
+    dn->lease_mds = -1;
+  }
+  ldout(cct, 25) << __func__ << " exit(dn:"
+	   << dn << ")" << dendl;
+}
+
+
+/*
+ * This requires the MClientRequest *request member to be set.
+ * It will error out horribly without one.
+ * Additionally, if you set any *drop member, you'd better have
+ * set the corresponding dentry!
+ */
+void Client::encode_cap_releases(MetaRequest *req, mds_rank_t mds)
+{
+  ldout(cct, 20) << __func__ << " enter (req: "
+		 << req << ", mds: " << mds << ")" << dendl;
+  if (req->inode_drop && req->inode())
+    encode_inode_release(req->inode(), req,
+			 mds, req->inode_drop,
+			 req->inode_unless);
+  
+  if (req->old_inode_drop && req->old_inode())
+    encode_inode_release(req->old_inode(), req,
+			 mds, req->old_inode_drop,
+			 req->old_inode_unless);
+  if (req->other_inode_drop && req->other_inode())
+    encode_inode_release(req->other_inode(), req,
+			 mds, req->other_inode_drop,
+			 req->other_inode_unless);
+  
+  if (req->dentry_drop && req->dentry())
+    encode_dentry_release(req->dentry(), req,
+			  mds, req->dentry_drop,
+			  req->dentry_unless);
+  
+  if (req->old_dentry_drop && req->old_dentry())
+    encode_dentry_release(req->old_dentry(), req,
+			  mds, req->old_dentry_drop,
+			  req->old_dentry_unless);
+  ldout(cct, 25) << __func__ << " exit (req: "
+	   << req << ", mds " << mds <<dendl;
+}
+
+bool Client::have_open_session(mds_rank_t mds)
+{
+  const auto &it = mds_sessions.find(mds);
+  return it != mds_sessions.end() &&
+    (it->second.state == MetaSession::STATE_OPEN ||
+     it->second.state == MetaSession::STATE_STALE);
+}
+
+MetaSession *Client::_get_mds_session(mds_rank_t mds, Connection *con)
+{
+  const auto &it = mds_sessions.find(mds);
+  if (it == mds_sessions.end() || it->second.con != con) {
+    return NULL;
+  } else {
+    return &it->second;
+  }
+}
+
+MetaSession *Client::_get_or_open_mds_session(mds_rank_t mds)
+{
+  auto it = mds_sessions.find(mds);
+  return it == mds_sessions.end() ? _open_mds_session(mds) : &it->second;
+}
+
+/**
+ * Populate a map of strings with client-identifying metadata,
+ * such as the hostname.  Call this once at initialization.
+ */
+void Client::populate_metadata(const std::string &mount_root)
+{
+  // Hostname
+#ifdef _WIN32
+  // TODO: move this to compat.h
+  char hostname[64];
+  DWORD hostname_sz = 64;
+  GetComputerNameA(hostname, &hostname_sz);
+  metadata["hostname"] = hostname;
+#else
+  struct utsname u;
+  int r = uname(&u);
+  if (r >= 0) {
+    metadata["hostname"] = u.nodename;
+    ldout(cct, 20) << __func__ << " read hostname '" << u.nodename << "'" << dendl;
+  } else {
+    ldout(cct, 1) << __func__ << " failed to read hostname (" << cpp_strerror(r) << ")" << dendl;
+  }
+#endif
+
+  metadata["pid"] = stringify(getpid());
+
+  // Ceph entity id (the '0' in "client.0")
+  metadata["entity_id"] = cct->_conf->name.get_id();
+
+  // Our mount position
+  if (!mount_root.empty()) {
+    metadata["root"] = mount_root;
+  }
+
+  // Ceph version
+  metadata["ceph_version"] = pretty_version_to_str();
+  metadata["ceph_sha1"] = git_version_to_str();
+
+  // Apply any metadata from the user's configured overrides
+  std::vector<std::string> tokens;
+  get_str_vec(cct->_conf->client_metadata, ",", tokens);
+  for (const auto &i : tokens) {
+    auto eqpos = i.find("=");
+    // Throw out anything that isn't of the form "<str>=<str>"
+    if (eqpos == 0 || eqpos == std::string::npos || eqpos == i.size()) {
+      lderr(cct) << "Invalid metadata keyval pair: '" << i << "'" << dendl;
+      continue;
+    }
+    metadata[i.substr(0, eqpos)] = i.substr(eqpos + 1);
+  }
+}
+
+/**
+ * Optionally add or override client metadata fields.
+ */
+void Client::update_metadata(std::string const &k, std::string const &v)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  ceph_assert(iref_reader.is_state_satisfied());
+
+  std::scoped_lock l(client_lock);
+
+  auto it = metadata.find(k);
+  if (it != metadata.end()) {
+    ldout(cct, 1) << __func__ << " warning, overriding metadata field '" << k
+		  << "' from '" << it->second << "' to '" << v << "'" << dendl;
+  }
+
+  metadata[k] = v;
+}
+
+MetaSession *Client::_open_mds_session(mds_rank_t mds)
+{
+  ldout(cct, 10) << __func__ << " mds." << mds << dendl;
+  auto addrs = mdsmap->get_addrs(mds);
+  auto em = mds_sessions.emplace(std::piecewise_construct,
+      std::forward_as_tuple(mds),
+      std::forward_as_tuple(mds, messenger->connect_to_mds(addrs), addrs));
+  ceph_assert(em.second); /* not already present */
+  MetaSession *session = &em.first->second;
+
+  auto m = make_message<MClientSession>(CEPH_SESSION_REQUEST_OPEN);
+  m->metadata = metadata;
+  m->supported_features = feature_bitset_t(CEPHFS_FEATURES_CLIENT_SUPPORTED);
+  m->metric_spec = feature_bitset_t(CEPHFS_METRIC_FEATURES_ALL);
+  session->con->send_message2(std::move(m));
+  return session;
+}
+
+void Client::_close_mds_session(MetaSession *s)
+{
+  ldout(cct, 2) << __func__ << " mds." << s->mds_num << " seq " << s->seq << dendl;
+  s->state = MetaSession::STATE_CLOSING;
+  s->con->send_message2(make_message<MClientSession>(CEPH_SESSION_REQUEST_CLOSE, s->seq));
+}
+
+void Client::_closed_mds_session(MetaSession *s, int err, bool rejected)
+{
+  ldout(cct, 5) << __func__ << " mds." << s->mds_num << " seq " << s->seq << dendl;
+  if (rejected && s->state != MetaSession::STATE_CLOSING)
+    s->state = MetaSession::STATE_REJECTED;
+  else
+    s->state = MetaSession::STATE_CLOSED;
+  s->con->mark_down();
+  signal_context_list(s->waiting_for_open);
+  mount_cond.notify_all();
+  remove_session_caps(s, err);
+  kick_requests_closed(s);
+  mds_ranks_closing.erase(s->mds_num);
+  if (s->state == MetaSession::STATE_CLOSED)
+    mds_sessions.erase(s->mds_num);
+}
+
+void Client::handle_client_session(const MConstRef<MClientSession>& m)
+{
+  mds_rank_t from = mds_rank_t(m->get_source().num());
+  ldout(cct, 10) << __func__ << " " << *m << " from mds." << from << dendl;
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(from, m->get_connection().get());
+  if (!session) {
+    ldout(cct, 10) << " discarding session message from sessionless mds " << m->get_source_inst() << dendl;
+    return;
+  }
+
+  switch (m->get_op()) {
+  case CEPH_SESSION_OPEN:
+    {
+      if (session->state == MetaSession::STATE_OPEN) {
+        ldout(cct, 10) << "mds." << from << " already opened, ignore it"
+                       << dendl;
+        return;
+      }
+      /*
+       * The connection maybe broken and the session in client side
+       * has been reinitialized, need to update the seq anyway.
+       */
+      if (!session->seq && m->get_seq())
+        session->seq = m->get_seq();
+
+      feature_bitset_t missing_features(CEPHFS_FEATURES_CLIENT_REQUIRED);
+      missing_features -= m->supported_features;
+      if (!missing_features.empty()) {
+	lderr(cct) << "mds." << from << " lacks required features '"
+		   << missing_features << "', closing session " << dendl;
+	_close_mds_session(session);
+	_closed_mds_session(session, -CEPHFS_EPERM, true);
+	break;
+      }
+      session->mds_features = std::move(m->supported_features);
+      session->mds_metric_flags = std::move(m->metric_spec.metric_flags);
+
+      renew_caps(session);
+      session->state = MetaSession::STATE_OPEN;
+      if (is_unmounting())
+	mount_cond.notify_all();
+      else
+	connect_mds_targets(from);
+      signal_context_list(session->waiting_for_open);
+      break;
+    }
+
+  case CEPH_SESSION_CLOSE:
+    _closed_mds_session(session);
+    break;
+
+  case CEPH_SESSION_RENEWCAPS:
+    if (session->cap_renew_seq == m->get_seq()) {
+      bool was_stale = ceph_clock_now() >= session->cap_ttl;
+      session->cap_ttl =
+	session->last_cap_renew_request + mdsmap->get_session_timeout();
+      if (was_stale)
+	wake_up_session_caps(session, false);
+    }
+    break;
+
+  case CEPH_SESSION_STALE:
+    // invalidate session caps/leases
+    session->cap_gen++;
+    session->cap_ttl = ceph_clock_now();
+    session->cap_ttl -= 1;
+    renew_caps(session);
+    break;
+
+  case CEPH_SESSION_RECALL_STATE:
+    /*
+     * Call the renew caps and flush cap releases just before
+     * triming the caps in case the tick() won't get a chance
+     * to run them, which could cause the client to be blocklisted
+     * and MDS daemons trying to recall the caps again and
+     * again.
+     *
+     * In most cases it will do nothing, and the new cap releases
+     * added by trim_caps() followed will be deferred flushing
+     * by tick().
+     */
+    renew_and_flush_cap_releases();
+    trim_caps(session, m->get_max_caps());
+    break;
+
+  case CEPH_SESSION_FLUSHMSG:
+    /* flush cap release */
+    if (auto& m = session->release; m) {
+      session->con->send_message2(std::move(m));
+    }
+    session->con->send_message2(make_message<MClientSession>(CEPH_SESSION_FLUSHMSG_ACK, m->get_seq()));
+    break;
+
+  case CEPH_SESSION_FORCE_RO:
+    force_session_readonly(session);
+    break;
+
+  case CEPH_SESSION_REJECT:
+    {
+      std::string_view error_str;
+      auto it = m->metadata.find("error_string");
+      if (it != m->metadata.end())
+	error_str = it->second;
+      else
+	error_str = "unknown error";
+      lderr(cct) << "mds." << from << " rejected us (" << error_str << ")" << dendl;
+
+      _closed_mds_session(session, -CEPHFS_EPERM, true);
+    }
+    break;
+
+  default:
+    ceph_abort();
+  }
+}
+
+bool Client::_any_stale_sessions() const
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  for (const auto &p : mds_sessions) {
+    if (p.second.state == MetaSession::STATE_STALE) {
+      return true;
+    }
+  }
+
+  return false;
+}
+
+void Client::_kick_stale_sessions()
+{
+  ldout(cct, 1) << __func__ << dendl;
+
+  for (auto it = mds_sessions.begin(); it != mds_sessions.end(); ) {
+    MetaSession &s = it->second;
+    if (s.state == MetaSession::STATE_REJECTED) {
+      mds_sessions.erase(it++);
+      continue;
+    }
+    ++it;
+    if (s.state == MetaSession::STATE_STALE)
+      _closed_mds_session(&s);
+  }
+}
+
+void Client::send_request(MetaRequest *request, MetaSession *session,
+			  bool drop_cap_releases)
+{
+  // make the request
+  mds_rank_t mds = session->mds_num;
+  ldout(cct, 10) << __func__ << " rebuilding request " << request->get_tid()
+		 << " for mds." << mds << dendl;
+  auto r = build_client_request(request);
+  if (request->dentry()) {
+    r->set_dentry_wanted();
+  }
+  if (request->got_unsafe) {
+    r->set_replayed_op();
+    if (request->target)
+      r->head.ino = request->target->ino;
+  } else {
+    encode_cap_releases(request, mds);
+    if (drop_cap_releases) // we haven't send cap reconnect yet, drop cap releases
+      request->cap_releases.clear();
+    else
+      r->releases.swap(request->cap_releases);
+  }
+  r->set_mdsmap_epoch(mdsmap->get_epoch());
+  if (r->head.op == CEPH_MDS_OP_SETXATTR) {
+    objecter->with_osdmap([r](const OSDMap& o) {
+	r->set_osdmap_epoch(o.get_epoch());
+      });
+  }
+
+  if (request->mds == -1) {
+    request->sent_stamp = ceph_clock_now();
+    ldout(cct, 20) << __func__ << " set sent_stamp to " << request->sent_stamp << dendl;
+  }
+  request->mds = mds;
+
+  Inode *in = request->inode();
+  if (in) {
+    auto it = in->caps.find(mds);
+    if (it != in->caps.end()) {
+      request->sent_on_mseq = it->second.mseq;
+    }
+  }
+
+  session->requests.push_back(&request->item);
+
+  ldout(cct, 10) << __func__ << " " << *r << " to mds." << mds << dendl;
+  session->con->send_message2(std::move(r));
+}
+
+ref_t<MClientRequest> Client::build_client_request(MetaRequest *request)
+{
+  auto req = make_message<MClientRequest>(request->get_op());
+  req->set_tid(request->tid);
+  req->set_stamp(request->op_stamp);
+  memcpy(&req->head, &request->head, sizeof(ceph_mds_request_head));
+
+  // if the filepath's haven't been set, set them!
+  if (request->path.empty()) {
+    Inode *in = request->inode();
+    Dentry *de = request->dentry();
+    if (in)
+      in->make_nosnap_relative_path(request->path);
+    else if (de) {
+      if (de->inode)
+	de->inode->make_nosnap_relative_path(request->path);
+      else if (de->dir) {
+	de->dir->parent_inode->make_nosnap_relative_path(request->path);
+	request->path.push_dentry(de->name);
+      }
+      else ldout(cct, 1) << "Warning -- unable to construct a filepath!"
+		   << " No path, inode, or appropriately-endowed dentry given!"
+		   << dendl;
+    } else ldout(cct, 1) << "Warning -- unable to construct a filepath!"
+		   << " No path, inode, or dentry given!"
+		   << dendl;
+  }
+  req->set_filepath(request->get_filepath());
+  req->set_filepath2(request->get_filepath2());
+  req->set_alternate_name(request->alternate_name);
+  req->set_data(request->data);
+  req->set_retry_attempt(request->retry_attempt++);
+  req->head.num_fwd = request->num_fwd;
+  const gid_t *_gids;
+  int gid_count = request->perms.get_gids(&_gids);
+  req->set_gid_list(gid_count, _gids);
+  return req;
+}
+
+
+
+void Client::handle_client_request_forward(const MConstRef<MClientRequestForward>& fwd)
+{
+  mds_rank_t mds = mds_rank_t(fwd->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds, fwd->get_connection().get());
+  if (!session) {
+    return;
+  }
+  ceph_tid_t tid = fwd->get_tid();
+
+  if (mds_requests.count(tid) == 0) {
+    ldout(cct, 10) << __func__ << " no pending request on tid " << tid << dendl;
+    return;
+  }
+
+  MetaRequest *request = mds_requests[tid];
+  ceph_assert(request);
+
+  /*
+   * The type of 'num_fwd' in ceph 'MClientRequestForward'
+   * is 'int32_t', while in 'ceph_mds_request_head' the
+   * type is '__u8'. So in case the request bounces between
+   * MDSes exceeding 256 times, the client will get stuck.
+   *
+   * In this case it's ususally a bug in MDS and continue
+   * bouncing the request makes no sense.
+   *
+   * In future this could be fixed in ceph code, so avoid
+   * using the hardcode here.
+   */
+  int max_fwd = sizeof(((struct ceph_mds_request_head*)0)->num_fwd);
+  max_fwd = 1 << (max_fwd * CHAR_BIT) - 1;
+  auto num_fwd = fwd->get_num_fwd();
+  if (num_fwd <= request->num_fwd || num_fwd >= max_fwd) {
+    if (request->num_fwd >= max_fwd || num_fwd >= max_fwd) {
+      request->abort(-EMULTIHOP);
+      request->caller_cond->notify_all();
+      ldout(cct, 1) << __func__ << " tid " << tid << " seq overflow"
+                    << ", abort it" << dendl;
+    } else {
+      ldout(cct, 10) << __func__ << " tid " << tid
+                     << " old fwd seq " << fwd->get_num_fwd()
+                     << " <= req fwd " << request->num_fwd
+                     << ", ignore it" << dendl;
+    }
+    return;
+  }
+
+  // reset retry counter
+  request->retry_attempt = 0;
+
+  // request not forwarded, or dest mds has no session.
+  // resend.
+  ldout(cct, 10) << __func__ << " tid " << tid
+	   << " fwd " << fwd->get_num_fwd() 
+	   << " to mds." << fwd->get_dest_mds() 
+	   << ", resending to " << fwd->get_dest_mds()
+	   << dendl;
+  
+  request->mds = -1;
+  request->item.remove_myself();
+  request->num_fwd = num_fwd;
+  request->resend_mds = fwd->get_dest_mds();
+  request->caller_cond->notify_all();
+}
+
+bool Client::is_dir_operation(MetaRequest *req)
+{
+  int op = req->get_op();
+  if (op == CEPH_MDS_OP_MKNOD || op == CEPH_MDS_OP_LINK ||
+      op == CEPH_MDS_OP_UNLINK || op == CEPH_MDS_OP_RENAME ||
+      op == CEPH_MDS_OP_MKDIR || op == CEPH_MDS_OP_RMDIR ||
+      op == CEPH_MDS_OP_SYMLINK || op == CEPH_MDS_OP_CREATE)
+    return true;
+  return false;
+}
+
+void Client::handle_client_reply(const MConstRef<MClientReply>& reply)
+{
+  mds_rank_t mds_num = mds_rank_t(reply->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds_num, reply->get_connection().get());
+  if (!session) {
+    return;
+  }
+
+  ceph_tid_t tid = reply->get_tid();
+  bool is_safe = reply->is_safe();
+
+  if (mds_requests.count(tid) == 0) {
+    lderr(cct) << __func__ << " no pending request on tid " << tid
+	       << " safe is:" << is_safe << dendl;
+    return;
+  }
+  MetaRequest *request = mds_requests.at(tid);
+
+  ldout(cct, 20) << __func__ << " got a reply. Safe:" << is_safe
+		 << " tid " << tid << dendl;
+
+  if (request->got_unsafe && !is_safe) {
+    //duplicate response
+    ldout(cct, 0) << "got a duplicate reply on tid " << tid << " from mds "
+	    << mds_num << " safe:" << is_safe << dendl;
+    return;
+  }
+
+  ceph_assert(!request->reply);
+  request->reply = reply;
+  insert_trace(request, session);
+
+  // Handle unsafe reply
+  if (!is_safe) {
+    request->got_unsafe = true;
+    session->unsafe_requests.push_back(&request->unsafe_item);
+    if (is_dir_operation(request)) {
+      Inode *dir = request->inode();
+      ceph_assert(dir);
+      dir->unsafe_ops.push_back(&request->unsafe_dir_item);
+    }
+    if (request->target) {
+      InodeRef &in = request->target;
+      in->unsafe_ops.push_back(&request->unsafe_target_item);
+    }
+  }
+
+  // Only signal the caller once (on the first reply):
+  // Either its an unsafe reply, or its a safe reply and no unsafe reply was sent.
+  if (!is_safe || !request->got_unsafe) {
+    ceph::condition_variable cond;
+    request->dispatch_cond = &cond;
+
+    // wake up waiter
+    ldout(cct, 20) << __func__ << " signalling caller " << (void*)request->caller_cond << dendl;
+    request->caller_cond->notify_all();
+
+    // wake for kick back
+    std::unique_lock l{client_lock, std::adopt_lock};
+    cond.wait(l, [tid, request, &cond, this] {
+      if (request->dispatch_cond) {
+        ldout(cct, 20) << "handle_client_reply awaiting kickback on tid "
+		       << tid << " " << &cond << dendl;
+      }
+      return !request->dispatch_cond;
+    });
+    l.release();
+  }
+
+  if (is_safe) {
+    // the filesystem change is committed to disk
+    // we're done, clean up
+    if (request->got_unsafe) {
+      request->unsafe_item.remove_myself();
+      request->unsafe_dir_item.remove_myself();
+      request->unsafe_target_item.remove_myself();
+      signal_cond_list(request->waitfor_safe);
+    }
+    request->item.remove_myself();
+    unregister_request(request);
+  }
+  if (is_unmounting())
+    mount_cond.notify_all();
+}
+
+void Client::_handle_full_flag(int64_t pool)
+{
+  ldout(cct, 1) << __func__ << ": FULL: cancelling outstanding operations "
+    << "on " << pool << dendl;
+  // Cancel all outstanding ops in this pool with -CEPHFS_ENOSPC: it is necessary
+  // to do this rather than blocking, because otherwise when we fill up we
+  // potentially lock caps forever on files with dirty pages, and we need
+  // to be able to release those caps to the MDS so that it can delete files
+  // and free up space.
+  epoch_t cancelled_epoch = objecter->op_cancel_writes(-CEPHFS_ENOSPC, pool);
+
+  // For all inodes with layouts in this pool and a pending flush write op
+  // (i.e. one of the ones we will cancel), we've got to purge_set their data
+  // from ObjectCacher so that it doesn't re-issue the write in response to
+  // the ENOSPC error.
+  // Fortunately since we're cancelling everything in a given pool, we don't
+  // need to know which ops belong to which ObjectSet, we can just blow all
+  // the un-flushed cached data away and mark any dirty inodes' async_err
+  // field with -CEPHFS_ENOSPC as long as we're sure all the ops we cancelled were
+  // affecting this pool, and all the objectsets we're purging were also
+  // in this pool.
+  for (unordered_map<vinodeno_t,Inode*>::iterator i = inode_map.begin();
+       i != inode_map.end(); ++i)
+  {
+    Inode *inode = i->second;
+    if (inode->oset.dirty_or_tx
+        && (pool == -1 || inode->layout.pool_id == pool)) {
+      ldout(cct, 4) << __func__ << ": FULL: inode 0x" << std::hex << i->first << std::dec
+        << " has dirty objects, purging and setting ENOSPC" << dendl;
+      objectcacher->purge_set(&inode->oset);
+      inode->set_async_err(-CEPHFS_ENOSPC);
+    }
+  }
+
+  if (cancelled_epoch != (epoch_t)-1) {
+    set_cap_epoch_barrier(cancelled_epoch);
+  }
+}
+
+void Client::handle_osd_map(const MConstRef<MOSDMap>& m)
+{
+  std::scoped_lock cl(client_lock);
+
+  const auto myaddrs = messenger->get_myaddrs();
+  bool new_blocklist = objecter->with_osdmap(
+    [&](const OSDMap& o) {
+      return o.is_blocklisted(myaddrs);
+    });
+  
+  if (new_blocklist && !blocklisted) {
+    auto epoch = objecter->with_osdmap([](const OSDMap &o){
+        return o.get_epoch();
+        });
+    lderr(cct) << "I was blocklisted at osd epoch " << epoch << dendl;
+    blocklisted = true;
+
+    _abort_mds_sessions(-CEPHFS_EBLOCKLISTED);
+
+    // Since we know all our OSD ops will fail, cancel them all preemtively,
+    // so that on an unhealthy cluster we can umount promptly even if e.g.
+    // some PGs were inaccessible.
+    objecter->op_cancel_writes(-CEPHFS_EBLOCKLISTED);
+
+  } 
+
+  if (blocklisted) {
+    // Handle case where we were blocklisted but no longer are
+    blocklisted = objecter->with_osdmap([myaddrs](const OSDMap &o){
+        return o.is_blocklisted(myaddrs);});
+  }
+
+  // Always subscribe to next osdmap for blocklisted client
+  // until this client is not blocklisted.
+  if (blocklisted) {
+    objecter->maybe_request_map();
+  }
+
+  if (objecter->osdmap_full_flag()) {
+    _handle_full_flag(-1);
+  } else {
+    // Accumulate local list of full pools so that I can drop
+    // the objecter lock before re-entering objecter in
+    // cancel_writes
+    std::vector<int64_t> full_pools;
+
+    objecter->with_osdmap([&full_pools](const OSDMap &o) {
+	for (const auto& kv : o.get_pools()) {
+	  if (kv.second.has_flag(pg_pool_t::FLAG_FULL)) {
+	    full_pools.push_back(kv.first);
+	  }
+	}
+      });
+
+    for (auto p : full_pools)
+      _handle_full_flag(p);
+
+    // Subscribe to subsequent maps to watch for the full flag going
+    // away.  For the global full flag objecter does this for us, but
+    // it pays no attention to the per-pool full flag so in this branch
+    // we do it ourselves.
+    if (!full_pools.empty()) {
+      objecter->maybe_request_map();
+    }
+  }
+}
+
+
+// ------------------------
+// incoming messages
+
+
+bool Client::ms_dispatch2(const MessageRef &m)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  if (!iref_reader.is_state_satisfied()) {
+    ldout(cct, 10) << "inactive, discarding " << *m << dendl;
+    return true;
+  }
+
+  switch (m->get_type()) {
+    // mounting and mds sessions
+  case CEPH_MSG_MDS_MAP:
+    handle_mds_map(ref_cast<MMDSMap>(m));
+    break;
+  case CEPH_MSG_FS_MAP:
+    handle_fs_map(ref_cast<MFSMap>(m));
+    break;
+  case CEPH_MSG_FS_MAP_USER:
+    handle_fs_map_user(ref_cast<MFSMapUser>(m));
+    break;
+  case CEPH_MSG_CLIENT_SESSION:
+    handle_client_session(ref_cast<MClientSession>(m));
+    break;
+
+  case CEPH_MSG_OSD_MAP:
+    handle_osd_map(ref_cast<MOSDMap>(m));
+    break;
+
+    // requests
+  case CEPH_MSG_CLIENT_REQUEST_FORWARD:
+    handle_client_request_forward(ref_cast<MClientRequestForward>(m));
+    break;
+  case CEPH_MSG_CLIENT_REPLY:
+    handle_client_reply(ref_cast<MClientReply>(m));
+    break;
+
+  // reclaim reply
+  case CEPH_MSG_CLIENT_RECLAIM_REPLY:
+    handle_client_reclaim_reply(ref_cast<MClientReclaimReply>(m));
+    break;
+
+  case CEPH_MSG_CLIENT_SNAP:
+    handle_snap(ref_cast<MClientSnap>(m));
+    break;
+  case CEPH_MSG_CLIENT_CAPS:
+    handle_caps(ref_cast<MClientCaps>(m));
+    break;
+  case CEPH_MSG_CLIENT_LEASE:
+    handle_lease(ref_cast<MClientLease>(m));
+    break;
+  case MSG_COMMAND_REPLY:
+    if (m->get_source().type() == CEPH_ENTITY_TYPE_MDS) {
+      handle_command_reply(ref_cast<MCommandReply>(m));
+    } else {
+      return false;
+    }
+    break;
+  case CEPH_MSG_CLIENT_QUOTA:
+    handle_quota(ref_cast<MClientQuota>(m));
+    break;
+
+  default:
+    return false;
+  }
+
+  // unmounting?
+  std::scoped_lock cl(client_lock);
+  if (is_unmounting()) {
+    ldout(cct, 10) << "unmounting: trim pass, size was " << lru.lru_get_size() 
+             << "+" << inode_map.size() << dendl;
+    uint64_t size = lru.lru_get_size() + inode_map.size();
+    trim_cache();
+    if (size > lru.lru_get_size() + inode_map.size()) {
+      ldout(cct, 10) << "unmounting: trim pass, cache shrank, poking unmount()" << dendl;
+      mount_cond.notify_all();
+    } else {
+      ldout(cct, 10) << "unmounting: trim pass, size still " << lru.lru_get_size() 
+               << "+" << inode_map.size() << dendl;
+    }
+  }
+
+  return true;
+}
+
+void Client::handle_fs_map(const MConstRef<MFSMap>& m)
+{
+  std::scoped_lock cl(client_lock);
+  fsmap.reset(new FSMap(m->get_fsmap()));
+
+  signal_cond_list(waiting_for_fsmap);
+
+  monclient->sub_got("fsmap", fsmap->get_epoch());
+}
+
+void Client::handle_fs_map_user(const MConstRef<MFSMapUser>& m)
+{
+  std::scoped_lock cl(client_lock);
+  fsmap_user.reset(new FSMapUser);
+  *fsmap_user = m->get_fsmap();
+
+  monclient->sub_got("fsmap.user", fsmap_user->get_epoch());
+  signal_cond_list(waiting_for_fsmap);
+}
+
+// Cancel all the commands for missing or laggy GIDs
+void Client::cancel_commands(const MDSMap& newmap)
+{
+  std::vector<ceph_tid_t> cancel_ops;
+
+  std::scoped_lock cmd_lock(command_lock);
+  auto &commands = command_table.get_commands();
+  for (const auto &[tid, op] : commands) {
+    const mds_gid_t op_mds_gid = op.mds_gid;
+    if (newmap.is_dne_gid(op_mds_gid) || newmap.is_laggy_gid(op_mds_gid)) {
+      ldout(cct, 1) << __func__ << ": cancelling command op " << tid << dendl;
+      cancel_ops.push_back(tid);
+      if (op.outs) {
+        std::ostringstream ss;
+        ss << "MDS " << op_mds_gid << " went away";
+        *(op.outs) = ss.str();
+      }
+      /*
+       * No need to make the con->mark_down under
+       * client_lock here, because the con will
+       * has its own lock.
+       */
+      op.con->mark_down();
+      if (op.on_finish)
+        op.on_finish->complete(-CEPHFS_ETIMEDOUT);
+    }
+  }
+
+  for (const auto &tid : cancel_ops)
+    command_table.erase(tid);
+}
+
+void Client::handle_mds_map(const MConstRef<MMDSMap>& m)
+{
+  std::unique_lock cl(client_lock);
+  if (m->get_epoch() <= mdsmap->get_epoch()) {
+    ldout(cct, 1) << __func__ << " epoch " << m->get_epoch()
+                  << " is identical to or older than our "
+                  << mdsmap->get_epoch() << dendl;
+    return;
+  }
+
+  cl.unlock();
+  ldout(cct, 1) << __func__ << " epoch " << m->get_epoch() << dendl;
+  std::unique_ptr<MDSMap> _mdsmap(new MDSMap);
+  _mdsmap->decode(m->get_encoded());
+  cancel_commands(*_mdsmap.get());
+  cl.lock();
+
+  _mdsmap.swap(mdsmap);
+
+  // reset session
+  for (auto p = mds_sessions.begin(); p != mds_sessions.end(); ) {
+    mds_rank_t mds = p->first;
+    MetaSession *session = &p->second;
+    ++p;
+
+    int oldstate = _mdsmap->get_state(mds);
+    int newstate = mdsmap->get_state(mds);
+    if (!mdsmap->is_up(mds)) {
+      session->con->mark_down();
+    } else if (mdsmap->get_addrs(mds) != session->addrs) {
+      auto old_inc = _mdsmap->get_incarnation(mds);
+      auto new_inc = mdsmap->get_incarnation(mds);
+      if (old_inc != new_inc) {
+        ldout(cct, 1) << "mds incarnation changed from "
+		      << old_inc << " to " << new_inc << dendl;
+        oldstate = MDSMap::STATE_NULL;
+      }
+      session->con->mark_down();
+      session->addrs = mdsmap->get_addrs(mds);
+      // When new MDS starts to take over, notify kernel to trim unused entries
+      // in its dcache/icache. Hopefully, the kernel will release some unused
+      // inodes before the new MDS enters reconnect state.
+      trim_cache_for_reconnect(session);
+    } else if (oldstate == newstate)
+      continue;  // no change
+
+    session->mds_state = newstate;
+    if (newstate == MDSMap::STATE_RECONNECT) {
+      session->con = messenger->connect_to_mds(session->addrs);
+      send_reconnect(session);
+    } else if (newstate > MDSMap::STATE_RECONNECT) {
+      if (oldstate < MDSMap::STATE_RECONNECT) {
+	ldout(cct, 1) << "we may miss the MDSMap::RECONNECT, close mds session ... " << dendl;
+	_closed_mds_session(session);
+	continue;
+      }
+      if (newstate >= MDSMap::STATE_ACTIVE) {
+	if (oldstate < MDSMap::STATE_ACTIVE) {
+	  // kick new requests
+	  kick_requests(session);
+	  kick_flushing_caps(session);
+	  signal_context_list(session->waiting_for_open);
+	  wake_up_session_caps(session, true);
+	}
+	connect_mds_targets(mds);
+      }
+    } else if (newstate == MDSMap::STATE_NULL &&
+	       mds >= mdsmap->get_max_mds()) {
+      _closed_mds_session(session);
+    }
+  }
+
+  // kick any waiting threads
+  signal_cond_list(waiting_for_mdsmap);
+
+  monclient->sub_got("mdsmap", mdsmap->get_epoch());
+}
+
+void Client::send_reconnect(MetaSession *session)
+{
+  mds_rank_t mds = session->mds_num;
+  ldout(cct, 10) << __func__ << " to mds." << mds << dendl;
+
+  // trim unused caps to reduce MDS's cache rejoin time
+  trim_cache_for_reconnect(session);
+
+  session->readonly = false;
+
+  session->release.reset();
+
+  // reset my cap seq number
+  session->seq = 0;
+  //connect to the mds' offload targets
+  connect_mds_targets(mds);
+  //make sure unsafe requests get saved
+  resend_unsafe_requests(session);
+
+  early_kick_flushing_caps(session);
+
+  auto m = make_message<MClientReconnect>();
+  bool allow_multi = session->mds_features.test(CEPHFS_FEATURE_MULTI_RECONNECT);
+
+  // i have an open session.
+  ceph::unordered_set<inodeno_t> did_snaprealm;
+  for (ceph::unordered_map<vinodeno_t, Inode*>::iterator p = inode_map.begin();
+       p != inode_map.end();
+       ++p) {
+    Inode *in = p->second;
+    auto it = in->caps.find(mds);
+    if (it != in->caps.end()) {
+      if (allow_multi &&
+	  m->get_approx_size() >=
+	  static_cast<size_t>((std::numeric_limits<int>::max() >> 1))) {
+	m->mark_more();
+	session->con->send_message2(std::move(m));
+
+	m = make_message<MClientReconnect>();
+      }
+
+      Cap &cap = it->second;
+      ldout(cct, 10) << " caps on " << p->first
+	       << " " << ccap_string(cap.issued)
+	       << " wants " << ccap_string(in->caps_wanted())
+	       << dendl;
+      filepath path;
+      in->make_short_path(path);
+      ldout(cct, 10) << "    path " << path << dendl;
+
+      bufferlist flockbl;
+      _encode_filelocks(in, flockbl);
+
+      cap.seq = 0;  // reset seq.
+      cap.issue_seq = 0;  // reset seq.
+      cap.mseq = 0;  // reset seq.
+      // cap gen should catch up with session cap_gen
+      if (cap.gen < session->cap_gen) {
+	cap.gen = session->cap_gen;
+	cap.issued = cap.implemented = CEPH_CAP_PIN;
+      } else {
+	cap.issued = cap.implemented;
+      }
+      snapid_t snap_follows = 0;
+      if (!in->cap_snaps.empty())
+	snap_follows = in->cap_snaps.begin()->first;
+
+      m->add_cap(p->first.ino, 
+		 cap.cap_id,
+		 path.get_ino(), path.get_path(),   // ino
+		 in->caps_wanted(), // wanted
+		 cap.issued,     // issued
+		 in->snaprealm->ino,
+		 snap_follows,
+		 flockbl);
+
+      if (did_snaprealm.count(in->snaprealm->ino) == 0) {
+	ldout(cct, 10) << " snaprealm " << *in->snaprealm << dendl;
+	m->add_snaprealm(in->snaprealm->ino, in->snaprealm->seq, in->snaprealm->parent);
+	did_snaprealm.insert(in->snaprealm->ino);
+      }
+    }
+  }
+
+  if (!allow_multi)
+    m->set_encoding_version(0); // use connection features to choose encoding
+  session->con->send_message2(std::move(m));
+
+  mount_cond.notify_all();
+
+  if (session->reclaim_state == MetaSession::RECLAIMING)
+    signal_cond_list(waiting_for_reclaim);
+}
+
+
+void Client::kick_requests(MetaSession *session)
+{
+  ldout(cct, 10) << __func__ << " for mds." << session->mds_num << dendl;
+  for (map<ceph_tid_t, MetaRequest*>::iterator p = mds_requests.begin();
+       p != mds_requests.end();
+       ++p) {
+    MetaRequest *req = p->second;
+    if (req->got_unsafe)
+      continue;
+    if (req->aborted()) {
+      if (req->caller_cond) {
+	req->kick = true;
+	req->caller_cond->notify_all();
+      }
+      continue;
+    }
+    if (req->retry_attempt > 0)
+      continue; // new requests only
+    if (req->mds == session->mds_num) {
+      send_request(p->second, session);
+    }
+  }
+}
+
+void Client::resend_unsafe_requests(MetaSession *session)
+{
+  for (xlist<MetaRequest*>::iterator iter = session->unsafe_requests.begin();
+       !iter.end();
+       ++iter)
+    send_request(*iter, session);
+
+  // also re-send old requests when MDS enters reconnect stage. So that MDS can
+  // process completed requests in clientreplay stage.
+  for (map<ceph_tid_t, MetaRequest*>::iterator p = mds_requests.begin();
+       p != mds_requests.end();
+       ++p) {
+    MetaRequest *req = p->second;
+    if (req->got_unsafe)
+      continue;
+    if (req->aborted())
+      continue;
+    if (req->retry_attempt == 0)
+      continue; // old requests only
+    if (req->mds == session->mds_num)
+      send_request(req, session, true);
+  }
+}
+
+void Client::wait_unsafe_requests()
+{
+  list<MetaRequest*> last_unsafe_reqs;
+  for (const auto &p : mds_sessions) {
+    const MetaSession &s = p.second;
+    if (!s.unsafe_requests.empty()) {
+      MetaRequest *req = s.unsafe_requests.back();
+      req->get();
+      last_unsafe_reqs.push_back(req);
+    }
+  }
+
+  for (list<MetaRequest*>::iterator p = last_unsafe_reqs.begin();
+       p != last_unsafe_reqs.end();
+       ++p) {
+    MetaRequest *req = *p;
+    if (req->unsafe_item.is_on_list())
+      wait_on_list(req->waitfor_safe);
+    put_request(req);
+  }
+}
+
+void Client::kick_requests_closed(MetaSession *session)
+{
+  ldout(cct, 10) << __func__ << " for mds." << session->mds_num << dendl;
+  for (map<ceph_tid_t, MetaRequest*>::iterator p = mds_requests.begin();
+       p != mds_requests.end(); ) {
+    MetaRequest *req = p->second;
+    ++p;
+    if (req->mds == session->mds_num) {
+      if (req->caller_cond) {
+	req->kick = true;
+	req->caller_cond->notify_all();
+      }
+      req->item.remove_myself();
+      if (req->got_unsafe) {
+	lderr(cct) << __func__ << " removing unsafe request " << req->get_tid() << dendl;
+	req->unsafe_item.remove_myself();
+	if (is_dir_operation(req)) {
+	  Inode *dir = req->inode();
+	  assert(dir);
+	  dir->set_async_err(-CEPHFS_EIO);
+	  lderr(cct) << "kick_requests_closed drop req of inode(dir) : "
+		     <<  dir->ino  << " " << req->get_tid() << dendl;
+	  req->unsafe_dir_item.remove_myself();
+	}
+	if (req->target) {
+	  InodeRef &in = req->target;
+	  in->set_async_err(-CEPHFS_EIO);
+	  lderr(cct) << "kick_requests_closed drop req of inode : "
+		     <<  in->ino  << " " << req->get_tid() << dendl;
+	  req->unsafe_target_item.remove_myself();
+	}
+	signal_cond_list(req->waitfor_safe);
+	unregister_request(req);
+      }
+    }
+  }
+  ceph_assert(session->requests.empty());
+  ceph_assert(session->unsafe_requests.empty());
+}
+
+
+
+
+/************
+ * leases
+ */
+
+void Client::got_mds_push(MetaSession *s)
+{
+  s->seq++;
+  ldout(cct, 10) << " mds." << s->mds_num << " seq now " << s->seq << dendl;
+  if (s->state == MetaSession::STATE_CLOSING) {
+    s->con->send_message2(make_message<MClientSession>(CEPH_SESSION_REQUEST_CLOSE, s->seq));
+  }
+}
+
+void Client::handle_lease(const MConstRef<MClientLease>& m)
+{
+  ldout(cct, 10) << __func__ << " " << *m << dendl;
+
+  ceph_assert(m->get_action() == CEPH_MDS_LEASE_REVOKE);
+  mds_rank_t mds = mds_rank_t(m->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds, m->get_connection().get());
+  if (!session) {
+    return;
+  }
+
+  got_mds_push(session);
+
+  ceph_seq_t seq = m->get_seq();
+
+  Inode *in;
+  vinodeno_t vino(m->get_ino(), CEPH_NOSNAP);
+  if (inode_map.count(vino) == 0) {
+    ldout(cct, 10) << " don't have vino " << vino << dendl;
+    goto revoke;
+  }
+  in = inode_map[vino];
+
+  if (m->get_mask() & CEPH_LEASE_VALID) {
+    if (!in->dir || in->dir->dentries.count(m->dname) == 0) {
+      ldout(cct, 10) << " don't have dir|dentry " << m->get_ino() << "/" << m->dname <<dendl;
+      goto revoke;
+    }
+    Dentry *dn = in->dir->dentries[m->dname];
+    ldout(cct, 10) << " revoked DN lease on " << dn << dendl;
+    dn->lease_mds = -1;
+  }
+
+ revoke:
+  {
+    auto reply = make_message<MClientLease>(CEPH_MDS_LEASE_RELEASE, seq,
+					    m->get_mask(), m->get_ino(),
+					    m->get_first(), m->get_last(), m->dname);
+    m->get_connection()->send_message2(std::move(reply));
+  }
+}
+
+void Client::_put_inode(Inode *in, int n)
+{
+  ldout(cct, 10) << __func__ << " on " << *in << " n = " << n << dendl;
+
+  int left = in->get_nref();
+  ceph_assert(left >= n + 1);
+  in->iput(n);
+  left -= n;
+  if (left == 1) { // the last one will be held by the inode_map
+    // release any caps
+    remove_all_caps(in);
+
+    ldout(cct, 10) << __func__ << " deleting " << *in << dendl;
+    bool unclean = objectcacher->release_set(&in->oset);
+    ceph_assert(!unclean);
+    inode_map.erase(in->vino());
+    if (use_faked_inos())
+      _release_faked_ino(in);
+
+    if (root == nullptr) {
+      root_ancestor = 0;
+      while (!root_parents.empty())
+        root_parents.erase(root_parents.begin());
+    }
+
+    in->iput();
+  }
+}
+
+void Client::delay_put_inodes(bool wakeup)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  std::map<Inode*,int> release;
+  {
+    std::scoped_lock dl(delay_i_lock);
+    release.swap(delay_i_release);
+  }
+
+  if (release.empty())
+    return;
+
+  for (auto &[in, cnt] : release)
+    _put_inode(in, cnt);
+
+  if (wakeup)
+    mount_cond.notify_all();
+}
+
+void Client::put_inode(Inode *in, int n)
+{
+  ldout(cct, 20) << __func__ << " on " << *in << " n = " << n << dendl;
+
+  std::scoped_lock dl(delay_i_lock);
+  delay_i_release[in] += n;
+}
+
+void Client::close_dir(Dir *dir)
+{
+  Inode *in = dir->parent_inode;
+  ldout(cct, 15) << __func__ << " dir " << dir << " on " << in << dendl;
+  ceph_assert(dir->is_empty());
+  ceph_assert(in->dir == dir);
+  ceph_assert(in->dentries.size() < 2);     // dirs can't be hard-linked
+  if (!in->dentries.empty())
+    in->get_first_parent()->put();   // unpin dentry
+  
+  delete in->dir;
+  in->dir = 0;
+  put_inode(in);               // unpin inode
+}
+
+  /**
+   * Don't call this with in==NULL, use get_or_create for that
+   * leave dn set to default NULL unless you're trying to add
+   * a new inode to a pre-created Dentry
+   */
+Dentry* Client::link(Dir *dir, const string& name, Inode *in, Dentry *dn)
+{
+  if (!dn) {
+    // create a new Dentry
+    dn = new Dentry(dir, name);
+
+    lru.lru_insert_mid(dn);    // mid or top?
+
+    ldout(cct, 15) << "link dir " << dir->parent_inode << " '" << name << "' to inode " << in
+		   << " dn " << dn << " (new dn)" << dendl;
+  } else {
+    ceph_assert(!dn->inode);
+    ldout(cct, 15) << "link dir " << dir->parent_inode << " '" << name << "' to inode " << in
+		   << " dn " << dn << " (old dn)" << dendl;
+  }
+
+  if (in) {    // link to inode
+    InodeRef tmp_ref;
+    // only one parent for directories!
+    if (in->is_dir() && !in->dentries.empty()) {
+      tmp_ref = in; // prevent unlink below from freeing the inode.
+      Dentry *olddn = in->get_first_parent();
+      ceph_assert(olddn->dir != dir || olddn->name != name);
+      Inode *old_diri = olddn->dir->parent_inode;
+      clear_dir_complete_and_ordered(old_diri, true);
+      unlink(olddn, true, true);  // keep dir, dentry
+    }
+
+    dn->link(in);
+    inc_dentry_nr();
+    ldout(cct, 20) << "link  inode " << in << " parents now " << in->dentries << dendl;
+  }
+  
+  return dn;
+}
+
+void Client::unlink(Dentry *dn, bool keepdir, bool keepdentry)
+{
+  InodeRef in(dn->inode);
+  ldout(cct, 15) << "unlink dir " << dn->dir->parent_inode << " '" << dn->name << "' dn " << dn
+		 << " inode " << dn->inode << dendl;
+
+  // unlink from inode
+  if (dn->inode) {
+    dn->unlink();
+    dec_dentry_nr();
+    ldout(cct, 20) << "unlink  inode " << in << " parents now " << in->dentries << dendl;
+  }
+
+  if (keepdentry) {
+    dn->lease_mds = -1;
+  } else {
+    ldout(cct, 15) << "unlink  removing '" << dn->name << "' dn " << dn << dendl;
+
+    // unlink from dir
+    Dir *dir = dn->dir;
+    dn->detach();
+
+    // delete den
+    lru.lru_remove(dn);
+    dn->put();
+
+    if (dir->is_empty() && !keepdir)
+      close_dir(dir);
+  }
+}
+
+/**
+ * For asynchronous flushes, check for errors from the IO and
+ * update the inode if necessary
+ */
+class C_Client_FlushComplete : public Context {
+private:
+  Client *client;
+  InodeRef inode;
+public:
+  C_Client_FlushComplete(Client *c, Inode *in) : client(c), inode(in) { }
+  void finish(int r) override {
+    ceph_assert(ceph_mutex_is_locked_by_me(client->client_lock));
+    if (r != 0) {
+      client_t const whoami = client->whoami;  // For the benefit of ldout prefix
+      ldout(client->cct, 1) << "I/O error from flush on inode " << inode
+        << " 0x" << std::hex << inode->ino << std::dec
+        << ": " << r << "(" << cpp_strerror(r) << ")" << dendl;
+      inode->set_async_err(r);
+    }
+  }
+};
+
+
+/****
+ * caps
+ */
+
+void Client::get_cap_ref(Inode *in, int cap)
+{
+  if ((cap & CEPH_CAP_FILE_BUFFER) &&
+      in->cap_refs[CEPH_CAP_FILE_BUFFER] == 0) {
+    ldout(cct, 5) << __func__ << " got first FILE_BUFFER ref on " << *in << dendl;
+    in->iget();
+  }
+  if ((cap & CEPH_CAP_FILE_CACHE) &&
+      in->cap_refs[CEPH_CAP_FILE_CACHE] == 0) {
+    ldout(cct, 5) << __func__ << " got first FILE_CACHE ref on " << *in << dendl;
+    in->iget();
+  }
+  in->get_cap_ref(cap);
+}
+
+void Client::put_cap_ref(Inode *in, int cap)
+{
+  int last = in->put_cap_ref(cap);
+  if (last) {
+    int put_nref = 0;
+    int drop = last & ~in->caps_issued();
+    if (in->snapid == CEPH_NOSNAP) {
+      if ((last & (CEPH_CAP_FILE_WR | CEPH_CAP_FILE_BUFFER)) &&
+	  !in->cap_snaps.empty() &&
+	  in->cap_snaps.rbegin()->second.writing) {
+	ldout(cct, 10) << __func__ << " finishing pending cap_snap on " << *in << dendl;
+	in->cap_snaps.rbegin()->second.writing = 0;
+	finish_cap_snap(in, in->cap_snaps.rbegin()->second, get_caps_used(in));
+	signal_cond_list(in->waitfor_caps);  // wake up blocked sync writers
+      }
+      if (last & CEPH_CAP_FILE_BUFFER) {
+	for (auto &p : in->cap_snaps)
+	  p.second.dirty_data = 0;
+	signal_cond_list(in->waitfor_commit);
+	ldout(cct, 5) << __func__ << " dropped last FILE_BUFFER ref on " << *in << dendl;
+	++put_nref;
+      }
+    }
+    if (last & CEPH_CAP_FILE_CACHE) {
+      ldout(cct, 5) << __func__ << " dropped last FILE_CACHE ref on " << *in << dendl;
+      ++put_nref;
+    }
+    if (drop)
+      check_caps(in, 0);
+    if (put_nref)
+      put_inode(in, put_nref);
+  }
+}
+
+// get caps for a given file handle -- the inode should have @need caps
+// issued by the mds and @want caps not revoked (or not under revocation).
+// this routine blocks till the cap requirement is satisfied. also account
+// (track) for capability hit when required (when cap requirement succeedes).
+int Client::get_caps(Fh *fh, int need, int want, int *phave, loff_t endoff)
+{
+  Inode *in = fh->inode.get();
+
+  int r = check_pool_perm(in, need);
+  if (r < 0)
+    return r;
+
+  while (1) {
+    int file_wanted = in->caps_file_wanted();
+    if ((file_wanted & need) != need) {
+      ldout(cct, 10) << "get_caps " << *in << " need " << ccap_string(need)
+		     << " file_wanted " << ccap_string(file_wanted) << ", EBADF "
+		     << dendl;
+      return -CEPHFS_EBADF;
+    }
+
+    if ((fh->mode & CEPH_FILE_MODE_WR) && fh->gen != fd_gen)
+      return -CEPHFS_EBADF;
+
+    if ((in->flags & I_ERROR_FILELOCK) && fh->has_any_filelocks())
+      return -CEPHFS_EIO;
+
+    int implemented;
+    int have = in->caps_issued(&implemented);
+
+    bool waitfor_caps = false;
+    bool waitfor_commit = false;
+
+    if (have & need & CEPH_CAP_FILE_WR) {
+      if (endoff > 0) {
+	 if ((endoff >= (loff_t)in->max_size ||
+	      endoff > (loff_t)(in->size << 1)) &&
+	     endoff > (loff_t)in->wanted_max_size) {
+	   ldout(cct, 10) << "wanted_max_size " << in->wanted_max_size << " -> " << endoff << dendl;
+	   in->wanted_max_size = endoff;
+	 }
+	 if (in->wanted_max_size > in->max_size &&
+	     in->wanted_max_size > in->requested_max_size)
+	   check_caps(in, 0);
+      }
+
+      if (endoff >= 0 && endoff > (loff_t)in->max_size) {
+	ldout(cct, 10) << "waiting on max_size, endoff " << endoff << " max_size " << in->max_size << " on " << *in << dendl;
+	waitfor_caps = true;
+      }
+      if (!in->cap_snaps.empty()) {
+	if (in->cap_snaps.rbegin()->second.writing) {
+	  ldout(cct, 10) << "waiting on cap_snap write to complete" << dendl;
+	  waitfor_caps = true;
+	}
+	for (auto &p : in->cap_snaps) {
+	  if (p.second.dirty_data) {
+	    waitfor_commit = true;
+	    break;
+	  }
+        }
+	if (waitfor_commit) {
+	  _flush(in, new C_Client_FlushComplete(this, in));
+	  ldout(cct, 10) << "waiting for WRBUFFER to get dropped" << dendl;
+	}
+      }
+    }
+
+    if (!waitfor_caps && !waitfor_commit) {
+      if ((have & need) == need) {
+	int revoking = implemented & ~have;
+	ldout(cct, 10) << "get_caps " << *in << " have " << ccap_string(have)
+		 << " need " << ccap_string(need) << " want " << ccap_string(want)
+		 << " revoking " << ccap_string(revoking)
+		 << dendl;
+	if ((revoking & want) == 0) {
+	  *phave = need | (have & want);
+	  in->get_cap_ref(need);
+	  cap_hit();
+	  return 0;
+	}
+      }
+      ldout(cct, 10) << "waiting for caps " << *in << " need " << ccap_string(need) << " want " << ccap_string(want) << dendl;
+      waitfor_caps = true;
+    }
+
+    if ((need & CEPH_CAP_FILE_WR) && in->auth_cap &&
+	in->auth_cap->session->readonly)
+      return -CEPHFS_EROFS;
+
+    if (in->flags & I_CAP_DROPPED) {
+      int mds_wanted = in->caps_mds_wanted();
+      if ((mds_wanted & need) != need) {
+	int ret = _renew_caps(in);
+	if (ret < 0)
+	  return ret;
+	continue;
+      }
+      if (!(file_wanted & ~mds_wanted))
+	in->flags &= ~I_CAP_DROPPED;
+    }
+
+    if (waitfor_caps)
+      wait_on_list(in->waitfor_caps);
+    else if (waitfor_commit)
+      wait_on_list(in->waitfor_commit);
+  }
+}
+
+int Client::get_caps_used(Inode *in)
+{
+  unsigned used = in->caps_used();
+  if (!(used & CEPH_CAP_FILE_CACHE) &&
+      !objectcacher->set_is_empty(&in->oset))
+    used |= CEPH_CAP_FILE_CACHE;
+  return used;
+}
+
+void Client::cap_delay_requeue(Inode *in)
+{
+  ldout(cct, 10) << __func__ << " on " << *in << dendl;
+
+  in->hold_caps_until = ceph::coarse_mono_clock::now() + caps_release_delay;
+  delayed_list.push_back(&in->delay_cap_item);
+}
+
+void Client::send_cap(Inode *in, MetaSession *session, Cap *cap,
+		      int flags, int used, int want, int retain,
+		      int flush, ceph_tid_t flush_tid)
+{
+  int held = cap->issued | cap->implemented;
+  int revoking = cap->implemented & ~cap->issued;
+  retain &= ~revoking;
+  int dropping = cap->issued & ~retain;
+  int op = CEPH_CAP_OP_UPDATE;
+
+  ldout(cct, 10) << __func__ << " " << *in
+	   << " mds." << session->mds_num << " seq " << cap->seq
+	   << " used " << ccap_string(used)
+	   << " want " << ccap_string(want)
+	   << " flush " << ccap_string(flush)
+	   << " retain " << ccap_string(retain)
+	   << " held "<< ccap_string(held)
+	   << " revoking " << ccap_string(revoking)
+	   << " dropping " << ccap_string(dropping)
+	   << dendl;
+
+  if (cct->_conf->client_inject_release_failure && revoking) {
+    const int would_have_issued = cap->issued & retain;
+    const int would_have_implemented = cap->implemented & (cap->issued | used);
+    // Simulated bug:
+    //  - tell the server we think issued is whatever they issued plus whatever we implemented
+    //  - leave what we have implemented in place
+    ldout(cct, 20) << __func__ << " injecting failure to release caps" << dendl;
+    cap->issued = cap->issued | cap->implemented;
+
+    // Make an exception for revoking xattr caps: we are injecting
+    // failure to release other caps, but allow xattr because client
+    // will block on xattr ops if it can't release these to MDS (#9800)
+    const int xattr_mask = CEPH_CAP_XATTR_SHARED | CEPH_CAP_XATTR_EXCL;
+    cap->issued ^= xattr_mask & revoking;
+    cap->implemented ^= xattr_mask & revoking;
+
+    ldout(cct, 20) << __func__ << " issued " << ccap_string(cap->issued) << " vs " << ccap_string(would_have_issued) << dendl;
+    ldout(cct, 20) << __func__ << " implemented " << ccap_string(cap->implemented) << " vs " << ccap_string(would_have_implemented) << dendl;
+  } else {
+    // Normal behaviour
+    cap->issued &= retain;
+    cap->implemented &= cap->issued | used;
+  }
+
+  snapid_t follows = 0;
+
+  if (flush)
+    follows = in->snaprealm->get_snap_context().seq;
+  
+  auto m = make_message<MClientCaps>(op,
+				   in->ino,
+				   0,
+				   cap->cap_id, cap->seq,
+				   cap->implemented,
+				   want,
+				   flush,
+				   cap->mseq,
+                                   cap_epoch_barrier);
+  m->caller_uid = in->cap_dirtier_uid;
+  m->caller_gid = in->cap_dirtier_gid;
+
+  m->head.issue_seq = cap->issue_seq;
+  m->set_tid(flush_tid);
+
+  m->head.uid = in->uid;
+  m->head.gid = in->gid;
+  m->head.mode = in->mode;
+  
+  m->head.nlink = in->nlink;
+  
+  if (flush & CEPH_CAP_XATTR_EXCL) {
+    encode(in->xattrs, m->xattrbl);
+    m->head.xattr_version = in->xattr_version;
+  }
+  
+  m->size = in->size;
+  m->max_size = in->max_size;
+  m->truncate_seq = in->truncate_seq;
+  m->truncate_size = in->truncate_size;
+  m->mtime = in->mtime;
+  m->atime = in->atime;
+  m->ctime = in->ctime;
+  m->btime = in->btime;
+  m->time_warp_seq = in->time_warp_seq;
+  m->change_attr = in->change_attr;
+
+  if (!(flags & MClientCaps::FLAG_PENDING_CAPSNAP) &&
+      !in->cap_snaps.empty() &&
+      in->cap_snaps.rbegin()->second.flush_tid == 0)
+    flags |= MClientCaps::FLAG_PENDING_CAPSNAP;
+  m->flags = flags;
+
+  if (flush & CEPH_CAP_FILE_WR) {
+    m->inline_version = in->inline_version;
+    m->inline_data = in->inline_data;
+  }
+
+  in->reported_size = in->size;
+  m->set_snap_follows(follows);
+  cap->wanted = want;
+  if (cap == in->auth_cap) {
+    if (want & CEPH_CAP_ANY_FILE_WR) {
+      m->set_max_size(in->wanted_max_size);
+      in->requested_max_size = in->wanted_max_size;
+      ldout(cct, 15) << "auth cap, requesting max_size " << in->requested_max_size << dendl;
+    } else {
+      in->requested_max_size = 0;
+      ldout(cct, 15) << "auth cap, reset requested_max_size due to not wanting any file write cap" << dendl;
+    }
+  }
+
+  if (!session->flushing_caps_tids.empty())
+    m->set_oldest_flush_tid(*session->flushing_caps_tids.begin());
+
+  session->con->send_message2(std::move(m));
+}
+
+static bool is_max_size_approaching(Inode *in)
+{
+  /* mds will adjust max size according to the reported size */
+  if (in->flushing_caps & CEPH_CAP_FILE_WR)
+    return false;
+  if (in->size >= in->max_size)
+    return true;
+  /* half of previous max_size increment has been used */
+  if (in->max_size > in->reported_size &&
+      (in->size << 1) >= in->max_size + in->reported_size)
+    return true;
+  return false;
+}
+
+static int adjust_caps_used_for_lazyio(int used, int issued, int implemented)
+{
+  if (!(used & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_BUFFER)))
+    return used;
+  if (!(implemented & CEPH_CAP_FILE_LAZYIO))
+    return used;
+
+  if (issued & CEPH_CAP_FILE_LAZYIO) {
+    if (!(issued & CEPH_CAP_FILE_CACHE)) {
+      used &= ~CEPH_CAP_FILE_CACHE;
+      used |= CEPH_CAP_FILE_LAZYIO;
+    }
+    if (!(issued & CEPH_CAP_FILE_BUFFER)) {
+      used &= ~CEPH_CAP_FILE_BUFFER;
+      used |= CEPH_CAP_FILE_LAZYIO;
+    }
+  } else {
+    if (!(implemented & CEPH_CAP_FILE_CACHE)) {
+      used &= ~CEPH_CAP_FILE_CACHE;
+      used |= CEPH_CAP_FILE_LAZYIO;
+    }
+    if (!(implemented & CEPH_CAP_FILE_BUFFER)) {
+      used &= ~CEPH_CAP_FILE_BUFFER;
+      used |= CEPH_CAP_FILE_LAZYIO;
+    }
+  }
+  return used;
+}
+
+/**
+ * check_caps
+ *
+ * Examine currently used and wanted versus held caps. Release, flush or ack
+ * revoked caps to the MDS as appropriate.
+ *
+ * @param in the inode to check
+ * @param flags flags to apply to cap check
+ */
+void Client::check_caps(Inode *in, unsigned flags)
+{
+  unsigned wanted = in->caps_wanted();
+  unsigned used = get_caps_used(in);
+  unsigned cap_used;
+
+  int implemented;
+  int issued = in->caps_issued(&implemented);
+  int revoking = implemented & ~issued;
+
+  int orig_used = used;
+  used = adjust_caps_used_for_lazyio(used, issued, implemented);
+
+  int retain = wanted | used | CEPH_CAP_PIN;
+  if (!is_unmounting() && in->nlink > 0) {
+    if (wanted) {
+      retain |= CEPH_CAP_ANY;
+    } else if (in->is_dir() &&
+	       (issued & CEPH_CAP_FILE_SHARED) &&
+	       (in->flags & I_COMPLETE)) {
+      // we do this here because we don't want to drop to Fs (and then
+      // drop the Fs if we do a create!) if that alone makes us send lookups
+      // to the MDS. Doing it in in->caps_wanted() has knock-on effects elsewhere
+      wanted = CEPH_CAP_ANY_SHARED | CEPH_CAP_FILE_EXCL;
+      retain |= wanted;
+    } else {
+      retain |= CEPH_CAP_ANY_SHARED;
+      // keep RD only if we didn't have the file open RW,
+      // because then the mds would revoke it anyway to
+      // journal max_size=0.
+      if (in->max_size == 0)
+	retain |= CEPH_CAP_ANY_RD;
+    }
+  }
+
+  ldout(cct, 10) << __func__ << " on " << *in
+	   << " wanted " << ccap_string(wanted)
+	   << " used " << ccap_string(used)
+	   << " issued " << ccap_string(issued)
+	   << " revoking " << ccap_string(revoking)
+	   << " flags=" << flags
+	   << dendl;
+
+  if (in->snapid != CEPH_NOSNAP)
+    return; //snap caps last forever, can't write
+
+  if (in->caps.empty())
+    return;   // guard if at end of func
+
+  if (!(orig_used & CEPH_CAP_FILE_BUFFER) &&
+      (revoking & used & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO))) {
+    if (_release(in))
+      used &= ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO);
+  }
+
+
+  for (auto &p : in->caps) {
+    mds_rank_t mds = p.first;
+    Cap &cap = p.second;
+
+    MetaSession *session = &mds_sessions.at(mds);
+
+    cap_used = used;
+    if (in->auth_cap && &cap != in->auth_cap)
+      cap_used &= ~in->auth_cap->issued;
+
+    revoking = cap.implemented & ~cap.issued;
+    
+    ldout(cct, 10) << " cap mds." << mds
+	     << " issued " << ccap_string(cap.issued)
+	     << " implemented " << ccap_string(cap.implemented)
+	     << " revoking " << ccap_string(revoking) << dendl;
+
+    if (in->wanted_max_size > in->max_size &&
+	in->wanted_max_size > in->requested_max_size &&
+	&cap == in->auth_cap)
+      goto ack;
+
+    /* approaching file_max? */
+    if ((cap.issued & CEPH_CAP_FILE_WR) &&
+	&cap == in->auth_cap &&
+	is_max_size_approaching(in)) {
+      ldout(cct, 10) << "size " << in->size << " approaching max_size " << in->max_size
+		     << ", reported " << in->reported_size << dendl;
+      goto ack;
+    }
+
+    /* completed revocation? */
+    if (revoking && (revoking & cap_used) == 0) {
+      ldout(cct, 10) << "completed revocation of " << ccap_string(cap.implemented & ~cap.issued) << dendl;
+      goto ack;
+    }
+
+    /* want more caps from mds? */
+    if (wanted & ~(cap.wanted | cap.issued))
+      goto ack;
+
+    if (!revoking && is_unmounting() && (cap_used == 0))
+      goto ack;
+
+    if ((cap.issued & ~retain) == 0 && // and we don't have anything we wouldn't like
+	!in->dirty_caps)               // and we have no dirty caps
+      continue;
+
+    if (!(flags & CHECK_CAPS_NODELAY)) {
+      ldout(cct, 10) << "delaying cap release" << dendl;
+      cap_delay_requeue(in);
+      continue;
+    }
+
+  ack:
+    if (&cap == in->auth_cap) {
+      if (in->flags & I_KICK_FLUSH) {
+	ldout(cct, 20) << " reflushing caps (check_caps) on " << *in
+		       << " to mds." << mds << dendl;
+	kick_flushing_caps(in, session);
+      }
+      if (!in->cap_snaps.empty() &&
+	  in->cap_snaps.rbegin()->second.flush_tid == 0)
+	flush_snaps(in);
+    }
+
+    int flushing;
+    int msg_flags = 0;
+    ceph_tid_t flush_tid;
+    if (in->auth_cap == &cap && in->dirty_caps) {
+      flushing = mark_caps_flushing(in, &flush_tid);
+      if (flags & CHECK_CAPS_SYNCHRONOUS)
+	msg_flags |= MClientCaps::FLAG_SYNC;
+    } else {
+      flushing = 0;
+      flush_tid = 0;
+    }
+
+    send_cap(in, session, &cap, msg_flags, cap_used, wanted, retain,
+	     flushing, flush_tid);
+  }
+}
+
+
+void Client::queue_cap_snap(Inode *in, SnapContext& old_snapc)
+{
+  int used = get_caps_used(in);
+  int dirty = in->caps_dirty();
+  ldout(cct, 10) << __func__ << " " << *in << " snapc " << old_snapc << " used " << ccap_string(used) << dendl;
+
+  if (in->cap_snaps.size() &&
+      in->cap_snaps.rbegin()->second.writing) {
+    ldout(cct, 10) << __func__ << " already have pending cap_snap on " << *in << dendl;
+    return;
+  } else if (in->caps_dirty() ||
+            (used & CEPH_CAP_FILE_WR) ||
+	     (dirty & CEPH_CAP_ANY_WR)) {
+    const auto &capsnapem = in->cap_snaps.emplace(std::piecewise_construct, std::make_tuple(old_snapc.seq), std::make_tuple(in));
+    ceph_assert(capsnapem.second); /* element inserted */
+    CapSnap &capsnap = capsnapem.first->second;
+    capsnap.context = old_snapc;
+    capsnap.issued = in->caps_issued();
+    capsnap.dirty = in->caps_dirty();
+
+    capsnap.dirty_data = (used & CEPH_CAP_FILE_BUFFER);
+
+    capsnap.uid = in->uid;
+    capsnap.gid = in->gid;
+    capsnap.mode = in->mode;
+    capsnap.btime = in->btime;
+    capsnap.xattrs = in->xattrs;
+    capsnap.xattr_version = in->xattr_version;
+    capsnap.cap_dirtier_uid = in->cap_dirtier_uid;
+    capsnap.cap_dirtier_gid = in->cap_dirtier_gid;
+
+    if (used & CEPH_CAP_FILE_WR) {
+      ldout(cct, 10) << __func__ << " WR used on " << *in << dendl;
+      capsnap.writing = 1;
+    } else {
+      finish_cap_snap(in, capsnap, used);
+    }
+  } else {
+    ldout(cct, 10) << __func__ << " not dirty|writing on " << *in << dendl;
+  }
+}
+
+void Client::finish_cap_snap(Inode *in, CapSnap &capsnap, int used)
+{
+  ldout(cct, 10) << __func__ << " " << *in << " capsnap " << (void *)&capsnap << " used " << ccap_string(used) << dendl;
+  capsnap.size = in->size;
+  capsnap.mtime = in->mtime;
+  capsnap.atime = in->atime;
+  capsnap.ctime = in->ctime;
+  capsnap.time_warp_seq = in->time_warp_seq;
+  capsnap.change_attr = in->change_attr;
+  capsnap.dirty |= in->caps_dirty();
+
+  /* Only reset it if it wasn't set before */
+  if (capsnap.cap_dirtier_uid == -1) {
+    capsnap.cap_dirtier_uid = in->cap_dirtier_uid;
+    capsnap.cap_dirtier_gid = in->cap_dirtier_gid;
+  }
+
+  if (capsnap.dirty & CEPH_CAP_FILE_WR) {
+    capsnap.inline_data = in->inline_data;
+    capsnap.inline_version = in->inline_version;
+  }
+
+  if (used & CEPH_CAP_FILE_BUFFER) {
+    capsnap.writing = 1;
+    ldout(cct, 10) << __func__ << " " << *in << " cap_snap " << &capsnap << " used " << used
+	     << " WRBUFFER, delaying" << dendl;
+  } else {
+    capsnap.dirty_data = 0;
+    flush_snaps(in);
+  }
+}
+
+void Client::send_flush_snap(Inode *in, MetaSession *session,
+			     snapid_t follows, CapSnap& capsnap)
+{
+  auto m = make_message<MClientCaps>(CEPH_CAP_OP_FLUSHSNAP,
+				     in->ino, in->snaprealm->ino, 0,
+				     in->auth_cap->mseq, cap_epoch_barrier);
+  m->caller_uid = capsnap.cap_dirtier_uid;
+  m->caller_gid = capsnap.cap_dirtier_gid;
+
+  m->set_client_tid(capsnap.flush_tid);
+  m->head.snap_follows = follows;
+
+  m->head.caps = capsnap.issued;
+  m->head.dirty = capsnap.dirty;
+
+  m->head.uid = capsnap.uid;
+  m->head.gid = capsnap.gid;
+  m->head.mode = capsnap.mode;
+  m->btime = capsnap.btime;
+
+  m->size = capsnap.size;
+
+  m->head.xattr_version = capsnap.xattr_version;
+  encode(capsnap.xattrs, m->xattrbl);
+
+  m->ctime = capsnap.ctime;
+  m->btime = capsnap.btime;
+  m->mtime = capsnap.mtime;
+  m->atime = capsnap.atime;
+  m->time_warp_seq = capsnap.time_warp_seq;
+  m->change_attr = capsnap.change_attr;
+
+  if (capsnap.dirty & CEPH_CAP_FILE_WR) {
+    m->inline_version = in->inline_version;
+    m->inline_data = in->inline_data;
+  }
+
+  ceph_assert(!session->flushing_caps_tids.empty());
+  m->set_oldest_flush_tid(*session->flushing_caps_tids.begin());
+
+  session->con->send_message2(std::move(m));
+}
+
+void Client::flush_snaps(Inode *in)
+{
+  ldout(cct, 10) << "flush_snaps on " << *in << dendl;
+  ceph_assert(in->cap_snaps.size());
+
+  // pick auth mds
+  ceph_assert(in->auth_cap);
+  MetaSession *session = in->auth_cap->session;
+
+  for (auto &p : in->cap_snaps) {
+    CapSnap &capsnap = p.second;
+    // only do new flush
+    if (capsnap.flush_tid > 0)
+      continue;
+
+    ldout(cct, 10) << "flush_snaps mds." << session->mds_num
+	     << " follows " << p.first
+	     << " size " << capsnap.size
+	     << " mtime " << capsnap.mtime
+	     << " dirty_data=" << capsnap.dirty_data
+	     << " writing=" << capsnap.writing
+	     << " on " << *in << dendl;
+    if (capsnap.dirty_data || capsnap.writing)
+      break;
+
+    capsnap.flush_tid = ++last_flush_tid;
+    session->flushing_caps_tids.insert(capsnap.flush_tid);
+    in->flushing_cap_tids[capsnap.flush_tid] = 0;
+    if (!in->flushing_cap_item.is_on_list())
+      session->flushing_caps.push_back(&in->flushing_cap_item);
+
+    send_flush_snap(in, session, p.first, capsnap);
+  }
+}
+
+void Client::wait_on_list(list<ceph::condition_variable*>& ls)
+{
+  ceph::condition_variable cond;
+  ls.push_back(&cond);
+  std::unique_lock l{client_lock, std::adopt_lock};
+  cond.wait(l);
+  l.release();
+  ls.remove(&cond);
+}
+
+void Client::signal_cond_list(list<ceph::condition_variable*>& ls)
+{
+  for (auto cond : ls) {
+    cond->notify_all();
+  }
+}
+
+void Client::wait_on_context_list(list<Context*>& ls)
+{
+  ceph::condition_variable cond;
+  bool done = false;
+  int r;
+  ls.push_back(new C_Cond(cond, &done, &r));
+  std::unique_lock l{client_lock, std::adopt_lock};
+  cond.wait(l, [&done] { return done;});
+  l.release();
+}
+
+void Client::signal_context_list(list<Context*>& ls)
+{
+  while (!ls.empty()) {
+    ls.front()->complete(0);
+    ls.pop_front();
+  }
+}
+
+void Client::wake_up_session_caps(MetaSession *s, bool reconnect)
+{
+  for (const auto &cap : s->caps) {
+    auto &in = cap->inode;
+    if (reconnect) {
+      in.requested_max_size = 0;
+      in.wanted_max_size = 0;
+    } else {
+      if (cap->gen < s->cap_gen) {
+	// mds did not re-issue stale cap.
+	cap->issued = cap->implemented = CEPH_CAP_PIN;
+	// make sure mds knows what we want.
+	if (in.caps_file_wanted() & ~cap->wanted)
+	  in.flags |= I_CAP_DROPPED;
+      }
+    }
+    signal_cond_list(in.waitfor_caps);
+  }
+}
+
+
+// flush dirty data (from objectcache)
+
+class C_Client_CacheInvalidate : public Context  {
+private:
+  Client *client;
+  vinodeno_t ino;
+  int64_t offset, length;
+public:
+  C_Client_CacheInvalidate(Client *c, Inode *in, int64_t off, int64_t len) :
+    client(c), offset(off), length(len) {
+    if (client->use_faked_inos())
+      ino = vinodeno_t(in->faked_ino, CEPH_NOSNAP);
+    else
+      ino = in->vino();
+  }
+  void finish(int r) override {
+    // _async_invalidate takes the lock when it needs to, call this back from outside of lock.
+    ceph_assert(ceph_mutex_is_not_locked_by_me(client->client_lock));
+    client->_async_invalidate(ino, offset, length);
+  }
+};
+
+void Client::_async_invalidate(vinodeno_t ino, int64_t off, int64_t len)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  ldout(cct, 10) << __func__ << " " << ino << " " << off << "~" << len << dendl;
+  ino_invalidate_cb(callback_handle, ino, off, len);
+}
+
+void Client::_schedule_invalidate_callback(Inode *in, int64_t off, int64_t len) {
+
+  if (ino_invalidate_cb)
+    // we queue the invalidate, which calls the callback and decrements the ref
+    async_ino_invalidator.queue(new C_Client_CacheInvalidate(this, in, off, len));
+}
+
+void Client::_invalidate_inode_cache(Inode *in)
+{
+  ldout(cct, 10) << __func__ << " " << *in << dendl;
+
+  // invalidate our userspace inode cache
+  if (cct->_conf->client_oc) {
+    objectcacher->release_set(&in->oset);
+    if (!objectcacher->set_is_empty(&in->oset))
+      lderr(cct) << "failed to invalidate cache for " << *in << dendl;
+  }
+
+  _schedule_invalidate_callback(in, 0, 0);
+}
+
+void Client::_invalidate_inode_cache(Inode *in, int64_t off, int64_t len)
+{
+  ldout(cct, 10) << __func__ << " " << *in << " " << off << "~" << len << dendl;
+
+  // invalidate our userspace inode cache
+  if (cct->_conf->client_oc) {
+    vector<ObjectExtent> ls;
+    Striper::file_to_extents(cct, in->ino, &in->layout, off, len, in->truncate_size, ls);
+    objectcacher->discard_writeback(&in->oset, ls, nullptr);
+  }
+
+  _schedule_invalidate_callback(in, off, len);
+}
+
+bool Client::_release(Inode *in)
+{
+  ldout(cct, 20) << "_release " << *in << dendl;
+  if (in->cap_refs[CEPH_CAP_FILE_CACHE] == 0) {
+    _invalidate_inode_cache(in);
+    return true;
+  }
+  return false;
+}
+
+bool Client::_flush(Inode *in, Context *onfinish)
+{
+  ldout(cct, 10) << "_flush " << *in << dendl;
+
+  if (!in->oset.dirty_or_tx) {
+    ldout(cct, 10) << " nothing to flush" << dendl;
+    onfinish->complete(0);
+    return true;
+  }
+
+  if (objecter->osdmap_pool_full(in->layout.pool_id)) {
+    ldout(cct, 8) << __func__ << ": FULL, purging for ENOSPC" << dendl;
+    objectcacher->purge_set(&in->oset);
+    if (onfinish) {
+      onfinish->complete(-CEPHFS_ENOSPC);
+    }
+    return true;
+  }
+
+  return objectcacher->flush_set(&in->oset, onfinish);
+}
+
+void Client::_flush_range(Inode *in, int64_t offset, uint64_t size)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+  if (!in->oset.dirty_or_tx) {
+    ldout(cct, 10) << " nothing to flush" << dendl;
+    return;
+  }
+
+  C_SaferCond onflush("Client::_flush_range flock");
+  bool ret = objectcacher->file_flush(&in->oset, &in->layout, in->snaprealm->get_snap_context(),
+				      offset, size, &onflush);
+  if (!ret) {
+    // wait for flush
+    client_lock.unlock();
+    onflush.wait();
+    client_lock.lock();
+  }
+}
+
+void Client::flush_set_callback(ObjectCacher::ObjectSet *oset)
+{
+  //  std::scoped_lock l(client_lock);
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));   // will be called via dispatch() -> objecter -> ...
+  Inode *in = static_cast<Inode *>(oset->parent);
+  ceph_assert(in);
+  _flushed(in);
+}
+
+void Client::_flushed(Inode *in)
+{
+  ldout(cct, 10) << "_flushed " << *in << dendl;
+
+  put_cap_ref(in, CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_BUFFER);
+}
+
+
+
+// checks common to add_update_cap, handle_cap_grant
+void Client::check_cap_issue(Inode *in, unsigned issued)
+{
+  unsigned had = in->caps_issued();
+
+  if ((issued & CEPH_CAP_FILE_CACHE) &&
+      !(had & CEPH_CAP_FILE_CACHE))
+    in->cache_gen++;
+
+  if ((issued & CEPH_CAP_FILE_SHARED) !=
+      (had & CEPH_CAP_FILE_SHARED)) {
+    if (issued & CEPH_CAP_FILE_SHARED)
+      in->shared_gen++;
+    if (in->is_dir())
+      clear_dir_complete_and_ordered(in, true);
+  }
+}
+
+void Client::add_update_cap(Inode *in, MetaSession *mds_session, uint64_t cap_id,
+			    unsigned issued, unsigned wanted, unsigned seq, unsigned mseq,
+			    inodeno_t realm, int flags, const UserPerm& cap_perms)
+{
+  if (!in->is_any_caps()) {
+    ceph_assert(in->snaprealm == 0);
+    in->snaprealm = get_snap_realm(realm);
+    in->snaprealm->inodes_with_caps.push_back(&in->snaprealm_item);
+    ldout(cct, 15) << __func__ << " first one, opened snaprealm " << in->snaprealm << dendl;
+  } else {
+    ceph_assert(in->snaprealm);
+    if ((flags & CEPH_CAP_FLAG_AUTH) &&
+	realm != inodeno_t(-1) && in->snaprealm->ino != realm) {
+      in->snaprealm_item.remove_myself();
+      auto oldrealm = in->snaprealm;
+      in->snaprealm = get_snap_realm(realm);
+      in->snaprealm->inodes_with_caps.push_back(&in->snaprealm_item);
+      put_snap_realm(oldrealm);
+    }
+  }
+
+  mds_rank_t mds = mds_session->mds_num;
+  const auto &capem = in->caps.emplace(std::piecewise_construct, std::forward_as_tuple(mds), std::forward_as_tuple(*in, mds_session));
+  Cap &cap = capem.first->second;
+  if (!capem.second) {
+    if (cap.gen < mds_session->cap_gen)
+      cap.issued = cap.implemented = CEPH_CAP_PIN;
+
+    /*
+     * auth mds of the inode changed. we received the cap export
+     * message, but still haven't received the cap import message.
+     * handle_cap_export() updated the new auth MDS' cap.
+     *
+     * "ceph_seq_cmp(seq, cap->seq) <= 0" means we are processing
+     * a message that was send before the cap import message. So
+     * don't remove caps.
+     */
+    if (ceph_seq_cmp(seq, cap.seq) <= 0) {
+      if (&cap != in->auth_cap)
+         ldout(cct, 0) << "WARNING: " <<  "inode " << *in << " caps on mds." << mds << " != auth_cap." << dendl;
+
+      ceph_assert(cap.cap_id == cap_id);
+      seq = cap.seq;
+      mseq = cap.mseq;
+      issued |= cap.issued;
+      flags |= CEPH_CAP_FLAG_AUTH;
+    }
+  } else {
+    inc_pinned_icaps();
+  }
+
+  check_cap_issue(in, issued);
+
+  if (flags & CEPH_CAP_FLAG_AUTH) {
+    if (in->auth_cap != &cap &&
+        (!in->auth_cap || ceph_seq_cmp(in->auth_cap->mseq, mseq) < 0)) {
+      if (in->auth_cap && in->flushing_cap_item.is_on_list()) {
+	ldout(cct, 10) << __func__ << " changing auth cap: "
+		       << "add myself to new auth MDS' flushing caps list" << dendl;
+	adjust_session_flushing_caps(in, in->auth_cap->session, mds_session);
+      }
+      in->auth_cap = &cap;
+    }
+  }
+
+  unsigned old_caps = cap.issued;
+  cap.cap_id = cap_id;
+  cap.issued = issued;
+  cap.implemented |= issued;
+  if (ceph_seq_cmp(mseq, cap.mseq) > 0)
+    cap.wanted = wanted;
+  else
+    cap.wanted |= wanted;
+  cap.seq = seq;
+  cap.issue_seq = seq;
+  cap.mseq = mseq;
+  cap.gen = mds_session->cap_gen;
+  cap.latest_perms = cap_perms;
+  ldout(cct, 10) << __func__ << " issued " << ccap_string(old_caps) << " -> " << ccap_string(cap.issued)
+	   << " from mds." << mds
+	   << " on " << *in
+	   << dendl;
+
+  if ((issued & ~old_caps) && in->auth_cap == &cap) {
+    // non-auth MDS is revoking the newly grant caps ?
+    for (auto &p : in->caps) {
+      if (&p.second == &cap)
+	continue;
+      if (p.second.implemented & ~p.second.issued & issued) {
+	check_caps(in, CHECK_CAPS_NODELAY);
+	break;
+      }
+    }
+  }
+
+  if (issued & ~old_caps)
+    signal_cond_list(in->waitfor_caps);
+}
+
+void Client::remove_cap(Cap *cap, bool queue_release)
+{
+  auto &in = cap->inode;
+  MetaSession *session = cap->session;
+  mds_rank_t mds = cap->session->mds_num;
+
+  ldout(cct, 10) << __func__ << " mds." << mds << " on " << in << dendl;
+  
+  if (queue_release) {
+    session->enqueue_cap_release(
+      in.ino,
+      cap->cap_id,
+      cap->issue_seq,
+      cap->mseq,
+      cap_epoch_barrier);
+  } else {
+    dec_pinned_icaps();
+  }
+
+
+  if (in.auth_cap == cap) {
+    if (in.flushing_cap_item.is_on_list()) {
+      ldout(cct, 10) << " removing myself from flushing_cap list" << dendl;
+      in.flushing_cap_item.remove_myself();
+    }
+    in.auth_cap = NULL;
+  }
+  size_t n = in.caps.erase(mds);
+  ceph_assert(n == 1);
+  cap = nullptr;
+
+  if (!in.is_any_caps()) {
+    ldout(cct, 15) << __func__ << " last one, closing snaprealm " << in.snaprealm << dendl;
+    in.snaprealm_item.remove_myself();
+    put_snap_realm(in.snaprealm);
+    in.snaprealm = 0;
+  }
+}
+
+void Client::remove_all_caps(Inode *in)
+{
+  while (!in->caps.empty())
+    remove_cap(&in->caps.begin()->second, true);
+}
+
+void Client::remove_session_caps(MetaSession *s, int err)
+{
+  ldout(cct, 10) << __func__ << " mds." << s->mds_num << dendl;
+
+  while (s->caps.size()) {
+    Cap *cap = *s->caps.begin();
+    InodeRef in(&cap->inode);
+    bool dirty_caps = false;
+    if (in->auth_cap == cap) {
+      dirty_caps = in->dirty_caps | in->flushing_caps;
+      in->wanted_max_size = 0;
+      in->requested_max_size = 0;
+      if (in->has_any_filelocks())
+	in->flags |= I_ERROR_FILELOCK;
+    }
+    auto caps = cap->implemented;
+    if (cap->wanted | cap->issued)
+      in->flags |= I_CAP_DROPPED;
+    remove_cap(cap, false);
+    in->cap_snaps.clear();
+    if (dirty_caps) {
+      lderr(cct) << __func__ << " still has dirty|flushing caps on " << *in << dendl;
+      if (in->flushing_caps) {
+	num_flushing_caps--;
+	in->flushing_cap_tids.clear();
+      }
+      in->flushing_caps = 0;
+      in->mark_caps_clean();
+      put_inode(in.get());
+    }
+    caps &= CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_BUFFER;
+    if (caps && !in->caps_issued_mask(caps, true)) {
+      if (err == -CEPHFS_EBLOCKLISTED) {
+	if (in->oset.dirty_or_tx) {
+	  lderr(cct) << __func__ << " still has dirty data on " << *in << dendl;
+	  in->set_async_err(err);
+	}
+	objectcacher->purge_set(&in->oset);
+      } else {
+	objectcacher->release_set(&in->oset);
+      }
+      _schedule_invalidate_callback(in.get(), 0, 0);
+    }
+
+    signal_cond_list(in->waitfor_caps);
+  }
+  s->flushing_caps_tids.clear();
+  sync_cond.notify_all();
+}
+
+std::pair<int, bool> Client::_do_remount(bool retry_on_error)
+{
+  uint64_t max_retries = cct->_conf.get_val<uint64_t>("mds_max_retries_on_remount_failure");
+  bool abort_on_failure = false;
+
+  errno = 0;
+  int r = remount_cb(callback_handle);
+  if (r == 0) {
+    retries_on_invalidate = 0;
+  } else {
+    int e = errno;
+    client_t whoami = get_nodeid();
+    if (r == -1) {
+      lderr(cct) <<
+          "failed to remount (to trim kernel dentries): "
+          "errno = " << e << " (" << strerror(e) << ")" << dendl;
+    } else {
+      lderr(cct) <<
+          "failed to remount (to trim kernel dentries): "
+          "return code = " << r << dendl;
+    }
+    bool should_abort =
+      (cct->_conf.get_val<bool>("client_die_on_failed_remount") ||
+       cct->_conf.get_val<bool>("client_die_on_failed_dentry_invalidate")) &&
+      !(retry_on_error && (++retries_on_invalidate < max_retries));
+    if (should_abort && !is_unmounting()) {
+      lderr(cct) << "failed to remount for kernel dentry trimming; quitting!" << dendl;
+      abort_on_failure = true;
+    }
+  }
+  return std::make_pair(r, abort_on_failure);
+}
+
+class C_Client_Remount : public Context  {
+private:
+  Client *client;
+public:
+  explicit C_Client_Remount(Client *c) : client(c) {}
+  void finish(int r) override {
+    ceph_assert(r == 0);
+    auto result = client->_do_remount(true);
+    if (result.second) {
+      ceph_abort();
+    }
+  }
+};
+
+void Client::_invalidate_kernel_dcache()
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  if (can_invalidate_dentries) {
+    if (dentry_invalidate_cb && root->dir) {
+      for (ceph::unordered_map<string, Dentry*>::iterator p = root->dir->dentries.begin();
+         p != root->dir->dentries.end();
+         ++p) {
+       if (p->second->inode)
+        _schedule_invalidate_dentry_callback(p->second, false);
+      }
+    }
+  } else if (remount_cb) {
+    // Hacky:
+    // when remounting a file system, linux kernel trims all unused dentries in the fs
+    remount_finisher.queue(new C_Client_Remount(this));
+  }
+}
+
+void Client::_trim_negative_child_dentries(InodeRef& in)
+{
+  if (!in->is_dir())
+    return;
+
+  Dir* dir = in->dir;
+  if (dir && dir->dentries.size() == dir->num_null_dentries) {
+    for (auto p = dir->dentries.begin(); p != dir->dentries.end(); ) {
+      Dentry *dn = p->second;
+      ++p;
+      ceph_assert(!dn->inode);
+      if (dn->lru_is_expireable())
+	unlink(dn, true, false);  // keep dir, drop dentry
+    }
+    if (dir->dentries.empty()) {
+      close_dir(dir);
+    }
+  }
+
+  if (in->flags & I_SNAPDIR_OPEN) {
+    InodeRef snapdir = open_snapdir(in.get());
+    _trim_negative_child_dentries(snapdir);
+  }
+}
+
+class C_Client_CacheRelease : public Context  {
+private:
+  Client *client;
+  vinodeno_t ino;
+public:
+  C_Client_CacheRelease(Client *c, Inode *in) :
+    client(c) {
+    if (client->use_faked_inos())
+      ino = vinodeno_t(in->faked_ino, CEPH_NOSNAP);
+    else
+      ino = in->vino();
+  }
+  void finish(int r) override {
+    ceph_assert(ceph_mutex_is_not_locked_by_me(client->client_lock));
+    client->_async_inode_release(ino);
+  }
+};
+
+void Client::_async_inode_release(vinodeno_t ino)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  ldout(cct, 10) << __func__ << " " << ino << dendl;
+  ino_release_cb(callback_handle, ino);
+}
+
+void Client::_schedule_ino_release_callback(Inode *in) {
+
+  if (ino_release_cb)
+    // we queue the invalidate, which calls the callback and decrements the ref
+    async_ino_releasor.queue(new C_Client_CacheRelease(this, in));
+}
+
+void Client::trim_caps(MetaSession *s, uint64_t max)
+{
+  mds_rank_t mds = s->mds_num;
+  size_t caps_size = s->caps.size();
+  ldout(cct, 10) << __func__ << " mds." << mds << " max " << max 
+    << " caps " << caps_size << dendl;
+
+  uint64_t trimmed = 0;
+  auto p = s->caps.begin();
+  std::set<Dentry *> to_trim; /* this avoids caps other than the one we're
+                               * looking at from getting deleted during traversal. */
+  while ((caps_size - trimmed) > max && !p.end()) {
+    Cap *cap = *p;
+    InodeRef in(&cap->inode);
+
+    // Increment p early because it will be invalidated if cap
+    // is deleted inside remove_cap
+    ++p;
+
+    if (in->caps.size() > 1 && cap != in->auth_cap) {
+      int mine = cap->issued | cap->implemented;
+      int oissued = in->auth_cap ? in->auth_cap->issued : 0;
+      // disposable non-auth cap
+      if (!(get_caps_used(in.get()) & ~oissued & mine)) {
+	ldout(cct, 20) << " removing unused, unneeded non-auth cap on " << *in << dendl;
+	cap = (remove_cap(cap, true), nullptr);
+	trimmed++;
+      }
+    } else {
+      ldout(cct, 20) << " trying to trim dentries for " << *in << dendl;
+      _trim_negative_child_dentries(in);
+      bool all = true;
+      auto q = in->dentries.begin();
+      while (q != in->dentries.end()) {
+        Dentry *dn = *q;
+        ++q;
+	if (dn->lru_is_expireable()) {
+	  if (can_invalidate_dentries &&
+	      dn->dir->parent_inode->ino == CEPH_INO_ROOT) {
+	    // Only issue one of these per DN for inodes in root: handle
+	    // others more efficiently by calling for root-child DNs at
+	    // the end of this function.
+	    _schedule_invalidate_dentry_callback(dn, true);
+	  }
+          ldout(cct, 20) << " queueing dentry for trimming: " << dn->name << dendl;
+          to_trim.insert(dn);
+        } else {
+          ldout(cct, 20) << "  not expirable: " << dn->name << dendl;
+	  all = false;
+        }
+      }
+      if (in->ll_ref == 1 && in->ino != CEPH_INO_ROOT) {
+         _schedule_ino_release_callback(in.get());
+      }
+      if (all && in->ino != CEPH_INO_ROOT) {
+        ldout(cct, 20) << __func__ << " counting as trimmed: " << *in << dendl;
+	trimmed++;
+      }
+    }
+  }
+  ldout(cct, 20) << " trimming queued dentries: " << dendl;
+  for (const auto &dn : to_trim) {
+    trim_dentry(dn);
+  }
+  to_trim.clear();
+
+  caps_size = s->caps.size();
+  if (caps_size > (size_t)max)
+    _invalidate_kernel_dcache();
+}
+
+void Client::force_session_readonly(MetaSession *s)
+{
+  s->readonly = true;
+  for (xlist<Cap*>::iterator p = s->caps.begin(); !p.end(); ++p) {
+    auto &in = (*p)->inode;
+    if (in.caps_wanted() & CEPH_CAP_FILE_WR)
+      signal_cond_list(in.waitfor_caps);
+  }
+}
+
+int Client::mark_caps_flushing(Inode *in, ceph_tid_t* ptid)
+{
+  MetaSession *session = in->auth_cap->session;
+
+  int flushing = in->dirty_caps;
+  ceph_assert(flushing);
+
+  ceph_tid_t flush_tid = ++last_flush_tid;
+  in->flushing_cap_tids[flush_tid] = flushing;
+
+  if (!in->flushing_caps) {
+    ldout(cct, 10) << __func__ << " " << ccap_string(flushing) << " " << *in << dendl;
+    num_flushing_caps++;
+  } else {
+    ldout(cct, 10) << __func__ << " (more) " << ccap_string(flushing) << " " << *in << dendl;
+  }
+
+  in->flushing_caps |= flushing;
+  in->mark_caps_clean();
+ 
+  if (!in->flushing_cap_item.is_on_list())
+    session->flushing_caps.push_back(&in->flushing_cap_item);
+  session->flushing_caps_tids.insert(flush_tid);
+
+  *ptid = flush_tid;
+  return flushing;
+}
+
+void Client::adjust_session_flushing_caps(Inode *in, MetaSession *old_s,  MetaSession *new_s)
+{
+  for (auto &p : in->cap_snaps) {
+    CapSnap &capsnap = p.second;
+    if (capsnap.flush_tid > 0) {
+      old_s->flushing_caps_tids.erase(capsnap.flush_tid);
+      new_s->flushing_caps_tids.insert(capsnap.flush_tid);
+    }
+  }
+  for (map<ceph_tid_t, int>::iterator it = in->flushing_cap_tids.begin();
+       it != in->flushing_cap_tids.end();
+       ++it) {
+    old_s->flushing_caps_tids.erase(it->first);
+    new_s->flushing_caps_tids.insert(it->first);
+  }
+  new_s->flushing_caps.push_back(&in->flushing_cap_item);
+}
+
+/*
+ * Flush all caps back to the MDS. Because the callers generally wait on the
+ * result of this function (syncfs and umount cases), we set
+ * CHECK_CAPS_SYNCHRONOUS on the last check_caps call.
+ */
+void Client::flush_caps_sync()
+{
+  ldout(cct, 10) << __func__ << dendl;
+  xlist<Inode*>::iterator p = delayed_list.begin();
+  while (!p.end()) {
+    unsigned flags = CHECK_CAPS_NODELAY;
+    Inode *in = *p;
+
+    ++p;
+    delayed_list.pop_front();
+    if (p.end() && dirty_list.empty())
+      flags |= CHECK_CAPS_SYNCHRONOUS;
+    check_caps(in, flags);
+  }
+
+  // other caps, too
+  p = dirty_list.begin();
+  while (!p.end()) {
+    unsigned flags = CHECK_CAPS_NODELAY;
+    Inode *in = *p;
+
+    ++p;
+    if (p.end())
+      flags |= CHECK_CAPS_SYNCHRONOUS;
+    check_caps(in, flags);
+  }
+}
+
+void Client::wait_sync_caps(Inode *in, ceph_tid_t want)
+{
+  while (in->flushing_caps) {
+    map<ceph_tid_t, int>::iterator it = in->flushing_cap_tids.begin();
+    ceph_assert(it != in->flushing_cap_tids.end());
+    if (it->first > want)
+      break;
+    ldout(cct, 10) << __func__ << " on " << *in << " flushing "
+		   << ccap_string(it->second) << " want " << want
+		   << " last " << it->first << dendl;
+    wait_on_list(in->waitfor_caps);
+  }
+}
+
+void Client::wait_sync_caps(ceph_tid_t want)
+{
+ retry:
+  ldout(cct, 10) << __func__ << " want " << want  << " (last is " << last_flush_tid << ", "
+	   << num_flushing_caps << " total flushing)" << dendl;
+  for (auto &p : mds_sessions) {
+    MetaSession *s = &p.second;
+    if (s->flushing_caps_tids.empty())
+	continue;
+    ceph_tid_t oldest_tid = *s->flushing_caps_tids.begin();
+    if (oldest_tid <= want) {
+      ldout(cct, 10) << " waiting on mds." << p.first << " tid " << oldest_tid
+		     << " (want " << want << ")" << dendl;
+      std::unique_lock l{client_lock, std::adopt_lock};
+      sync_cond.wait(l);
+      l.release();
+      goto retry;
+    }
+  }
+}
+
+void Client::kick_flushing_caps(Inode *in, MetaSession *session)
+{
+  in->flags &= ~I_KICK_FLUSH;
+
+  Cap *cap = in->auth_cap;
+  ceph_assert(cap->session == session);
+
+  ceph_tid_t last_snap_flush = 0;
+  for (auto p = in->flushing_cap_tids.rbegin();
+       p != in->flushing_cap_tids.rend();
+       ++p) {
+    if (!p->second) {
+      last_snap_flush = p->first;
+      break;
+    }
+  }
+
+  int wanted = in->caps_wanted();
+  int used = get_caps_used(in) | in->caps_dirty();
+  auto it = in->cap_snaps.begin();
+  for (auto& p : in->flushing_cap_tids) {
+    if (p.second) {
+      int msg_flags = p.first < last_snap_flush ? MClientCaps::FLAG_PENDING_CAPSNAP : 0;
+      send_cap(in, session, cap, msg_flags, used, wanted, (cap->issued | cap->implemented),
+	       p.second, p.first);
+    } else {
+      ceph_assert(it != in->cap_snaps.end());
+      ceph_assert(it->second.flush_tid == p.first);
+      send_flush_snap(in, session, it->first, it->second);
+      ++it;
+    }
+  }
+}
+
+void Client::kick_flushing_caps(MetaSession *session)
+{
+  mds_rank_t mds = session->mds_num;
+  ldout(cct, 10) << __func__ << " mds." << mds << dendl;
+
+  for (xlist<Inode*>::iterator p = session->flushing_caps.begin(); !p.end(); ++p) {
+    Inode *in = *p;
+    if (in->flags & I_KICK_FLUSH) {
+      ldout(cct, 20) << " reflushing caps on " << *in << " to mds." << mds << dendl;
+      kick_flushing_caps(in, session);
+    }
+  }
+}
+
+void Client::early_kick_flushing_caps(MetaSession *session)
+{
+  for (xlist<Inode*>::iterator p = session->flushing_caps.begin(); !p.end(); ++p) {
+    Inode *in = *p;
+    Cap *cap = in->auth_cap;
+    ceph_assert(cap);
+
+    // if flushing caps were revoked, we re-send the cap flush in client reconnect
+    // stage. This guarantees that MDS processes the cap flush message before issuing
+    // the flushing caps to other client.
+    if ((in->flushing_caps & in->auth_cap->issued) == in->flushing_caps) {
+      in->flags |= I_KICK_FLUSH;
+      continue;
+    }
+
+    ldout(cct, 20) << " reflushing caps (early_kick) on " << *in
+		   << " to mds." << session->mds_num << dendl;
+    // send_reconnect() also will reset these sequence numbers. make sure
+    // sequence numbers in cap flush message match later reconnect message.
+    cap->seq = 0;
+    cap->issue_seq = 0;
+    cap->mseq = 0;
+    cap->issued = cap->implemented;
+
+    kick_flushing_caps(in, session);
+  }
+}
+
+void SnapRealm::build_snap_context()
+{
+  set<snapid_t> snaps;
+  snapid_t max_seq = seq;
+  
+  // start with prior_parents?
+  for (unsigned i=0; i<prior_parent_snaps.size(); i++)
+    snaps.insert(prior_parent_snaps[i]);
+
+  // current parent's snaps
+  if (pparent) {
+    const SnapContext& psnapc = pparent->get_snap_context();
+    for (unsigned i=0; i<psnapc.snaps.size(); i++)
+      if (psnapc.snaps[i] >= parent_since)
+	snaps.insert(psnapc.snaps[i]);
+    if (psnapc.seq > max_seq)
+      max_seq = psnapc.seq;
+  }
+
+  // my snaps
+  for (unsigned i=0; i<my_snaps.size(); i++)
+    snaps.insert(my_snaps[i]);
+
+  // ok!
+  cached_snap_context.seq = max_seq;
+  cached_snap_context.snaps.resize(0);
+  cached_snap_context.snaps.reserve(snaps.size());
+  for (set<snapid_t>::reverse_iterator p = snaps.rbegin(); p != snaps.rend(); ++p)
+    cached_snap_context.snaps.push_back(*p);
+}
+
+void Client::invalidate_snaprealm_and_children(SnapRealm *realm)
+{
+  list<SnapRealm*> q;
+  q.push_back(realm);
+
+  while (!q.empty()) {
+    realm = q.front();
+    q.pop_front();
+
+    ldout(cct, 10) << __func__ << " " << *realm << dendl;
+    realm->invalidate_cache();
+
+    for (set<SnapRealm*>::iterator p = realm->pchildren.begin();
+	 p != realm->pchildren.end(); 
+	 ++p)
+      q.push_back(*p);
+  }
+}
+
+SnapRealm *Client::get_snap_realm(inodeno_t r)
+{
+  SnapRealm *realm = snap_realms[r];
+  if (!realm)
+    snap_realms[r] = realm = new SnapRealm(r);
+  ldout(cct, 20) << __func__ << " " << r << " " << realm << " " << realm->nref << " -> " << (realm->nref + 1) << dendl;
+  realm->nref++;
+  return realm;
+}
+
+SnapRealm *Client::get_snap_realm_maybe(inodeno_t r)
+{
+  if (snap_realms.count(r) == 0) {
+    ldout(cct, 20) << __func__ << " " << r << " fail" << dendl;
+    return NULL;
+  }
+  SnapRealm *realm = snap_realms[r];
+  ldout(cct, 20) << __func__ << " " << r << " " << realm << " " << realm->nref << " -> " << (realm->nref + 1) << dendl;
+  realm->nref++;
+  return realm;
+}
+
+void Client::put_snap_realm(SnapRealm *realm)
+{
+  ldout(cct, 20) << __func__ << " " << realm->ino << " " << realm
+		 << " " << realm->nref << " -> " << (realm->nref - 1) << dendl;
+  if (--realm->nref == 0) {
+    snap_realms.erase(realm->ino);
+    if (realm->pparent) {
+      realm->pparent->pchildren.erase(realm);
+      put_snap_realm(realm->pparent);
+    }
+    delete realm;
+  }
+}
+
+bool Client::adjust_realm_parent(SnapRealm *realm, inodeno_t parent)
+{
+  if (realm->parent != parent) {
+    ldout(cct, 10) << __func__ << " " << *realm
+	     << " " << realm->parent << " -> " << parent << dendl;
+    realm->parent = parent;
+    if (realm->pparent) {
+      realm->pparent->pchildren.erase(realm);
+      put_snap_realm(realm->pparent);
+    }
+    realm->pparent = get_snap_realm(parent);
+    realm->pparent->pchildren.insert(realm);
+    return true;
+  }
+  return false;
+}
+
+static bool has_new_snaps(const SnapContext& old_snapc,
+			  const SnapContext& new_snapc)
+{
+  return !new_snapc.snaps.empty() && new_snapc.snaps[0] > old_snapc.seq;
+}
+
+
+void Client::update_snap_trace(const bufferlist& bl, SnapRealm **realm_ret, bool flush)
+{
+  SnapRealm *first_realm = NULL;
+  ldout(cct, 10) << __func__ << " len " << bl.length() << dendl;
+
+  map<SnapRealm*, SnapContext> dirty_realms;
+
+  auto p = bl.cbegin();
+  while (!p.end()) {
+    SnapRealmInfo info;
+    decode(info, p);
+    SnapRealm *realm = get_snap_realm(info.ino());
+
+    bool invalidate = false;
+
+    if (info.seq() > realm->seq) {
+      ldout(cct, 10) << __func__ << " " << *realm << " seq " << info.seq() << " > " << realm->seq
+	       << dendl;
+
+      if (flush) {
+	// writeback any dirty caps _before_ updating snap list (i.e. with old snap info)
+	//  flush me + children
+	list<SnapRealm*> q;
+	q.push_back(realm);
+	while (!q.empty()) {
+	  SnapRealm *realm = q.front();
+	  q.pop_front();
+
+	  for (set<SnapRealm*>::iterator p = realm->pchildren.begin(); 
+	       p != realm->pchildren.end();
+	       ++p)
+	    q.push_back(*p);
+
+	  if (dirty_realms.count(realm) == 0) {
+	    realm->nref++;
+	    dirty_realms[realm] = realm->get_snap_context();
+	  }
+	}
+      }
+
+      // update
+      realm->seq = info.seq();
+      realm->created = info.created();
+      realm->parent_since = info.parent_since();
+      realm->prior_parent_snaps = info.prior_parent_snaps;
+      realm->my_snaps = info.my_snaps;
+      invalidate = true;
+    }
+
+    // _always_ verify parent
+    if (adjust_realm_parent(realm, info.parent()))
+      invalidate = true;
+
+    if (invalidate) {
+      invalidate_snaprealm_and_children(realm);
+      ldout(cct, 15) << __func__ << " " << *realm << " self|parent updated" << dendl;
+      ldout(cct, 15) << "  snapc " << realm->get_snap_context() << dendl;
+    } else {
+      ldout(cct, 10) << __func__ << " " << *realm << " seq " << info.seq()
+	       << " <= " << realm->seq << " and same parent, SKIPPING" << dendl;
+    }
+
+    if (!first_realm)
+      first_realm = realm;
+    else
+      put_snap_realm(realm);
+  }
+
+  for (auto &[realm, snapc] : dirty_realms) {
+    // if there are new snaps ?
+    if (has_new_snaps(snapc, realm->get_snap_context())) {
+      ldout(cct, 10) << " flushing caps on " << *realm << dendl;
+      for (auto&& in : realm->inodes_with_caps) {
+	queue_cap_snap(in, snapc);
+      }
+    } else {
+      ldout(cct, 10) << " no new snap on " << *realm << dendl;
+    }
+    put_snap_realm(realm);
+  }
+
+  if (realm_ret)
+    *realm_ret = first_realm;
+  else
+    put_snap_realm(first_realm);
+}
+
+void Client::handle_snap(const MConstRef<MClientSnap>& m)
+{
+  ldout(cct, 10) << __func__ << " " << *m << dendl;
+  mds_rank_t mds = mds_rank_t(m->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds, m->get_connection().get());
+  if (!session) {
+    return;
+  }
+
+  got_mds_push(session);
+
+  map<Inode*, SnapContext> to_move;
+  SnapRealm *realm = 0;
+
+  if (m->head.op == CEPH_SNAP_OP_SPLIT) {
+    ceph_assert(m->head.split);
+    SnapRealmInfo info;
+    auto p = m->bl.cbegin();
+    decode(info, p);
+    ceph_assert(info.ino() == m->head.split);
+    
+    // flush, then move, ino's.
+    realm = get_snap_realm(info.ino());
+    ldout(cct, 10) << " splitting off " << *realm << dendl;
+    for (auto& ino : m->split_inos) {
+      vinodeno_t vino(ino, CEPH_NOSNAP);
+      if (inode_map.count(vino)) {
+	Inode *in = inode_map[vino];
+	if (!in->snaprealm || in->snaprealm == realm)
+	  continue;
+	if (in->snaprealm->created > info.created()) {
+	  ldout(cct, 10) << " NOT moving " << *in << " from _newer_ realm " 
+		   << *in->snaprealm << dendl;
+	  continue;
+	}
+	ldout(cct, 10) << " moving " << *in << " from " << *in->snaprealm << dendl;
+
+
+	in->snaprealm_item.remove_myself();
+	to_move[in] = in->snaprealm->get_snap_context();
+	put_snap_realm(in->snaprealm);
+      }
+    }
+
+    // move child snaprealms, too
+    for (auto& child_realm : m->split_realms) {
+      ldout(cct, 10) << "adjusting snaprealm " << child_realm << " parent" << dendl;
+      SnapRealm *child = get_snap_realm_maybe(child_realm);
+      if (!child)
+	continue;
+      adjust_realm_parent(child, realm->ino);
+      put_snap_realm(child);
+    }
+  }
+
+  update_snap_trace(m->bl, NULL, m->head.op != CEPH_SNAP_OP_DESTROY);
+
+  if (realm) {
+    for (auto p = to_move.begin(); p != to_move.end(); ++p) {
+      Inode *in = p->first;
+      in->snaprealm = realm;
+      realm->inodes_with_caps.push_back(&in->snaprealm_item);
+      realm->nref++;
+      // queue for snap writeback
+      if (has_new_snaps(p->second, realm->get_snap_context()))
+	queue_cap_snap(in, p->second);
+    }
+    put_snap_realm(realm);
+  }
+}
+
+void Client::handle_quota(const MConstRef<MClientQuota>& m)
+{
+  mds_rank_t mds = mds_rank_t(m->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds, m->get_connection().get());
+  if (!session) {
+    return;
+  }
+
+  got_mds_push(session);
+
+  ldout(cct, 10) << __func__ << " " << *m << " from mds." << mds << dendl;
+
+  vinodeno_t vino(m->ino, CEPH_NOSNAP);
+  if (inode_map.count(vino)) {
+    Inode *in = NULL;
+    in = inode_map[vino];
+
+    if (in) {
+      in->quota = m->quota;
+      in->rstat = m->rstat;
+    }
+  }
+}
+
+void Client::handle_caps(const MConstRef<MClientCaps>& m)
+{
+  mds_rank_t mds = mds_rank_t(m->get_source().num());
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(mds, m->get_connection().get());
+  if (!session) {
+    return;
+  }
+
+  if (m->osd_epoch_barrier && !objecter->have_map(m->osd_epoch_barrier)) {
+    // Pause RADOS operations until we see the required epoch
+    objecter->set_epoch_barrier(m->osd_epoch_barrier);
+  }
+
+  if (m->osd_epoch_barrier > cap_epoch_barrier) {
+    // Record the barrier so that we will transmit it to MDS when releasing
+    set_cap_epoch_barrier(m->osd_epoch_barrier);
+  }
+
+  got_mds_push(session);
+
+  Inode *in;
+  vinodeno_t vino(m->get_ino(), CEPH_NOSNAP);
+  if (auto it = inode_map.find(vino); it != inode_map.end()) {
+    in = it->second;
+  } else {
+    if (m->get_op() == CEPH_CAP_OP_IMPORT) {
+      ldout(cct, 5) << __func__ << " don't have vino " << vino << " on IMPORT, immediately releasing" << dendl;
+      session->enqueue_cap_release(
+        m->get_ino(),
+        m->get_cap_id(),
+        m->get_seq(),
+        m->get_mseq(),
+        cap_epoch_barrier);
+    } else {
+      ldout(cct, 5) << __func__ << " don't have vino " << vino << ", dropping" << dendl;
+    }
+
+    // in case the mds is waiting on e.g. a revocation
+    flush_cap_releases();
+    return;
+  }
+
+  switch (m->get_op()) {
+    case CEPH_CAP_OP_EXPORT: return handle_cap_export(session, in, m);
+    case CEPH_CAP_OP_FLUSHSNAP_ACK: return handle_cap_flushsnap_ack(session, in, m);
+    case CEPH_CAP_OP_IMPORT: /* no return */ handle_cap_import(session, in, m);
+  }
+
+  if (auto it = in->caps.find(mds); it != in->caps.end()) {
+    Cap &cap = in->caps.at(mds);
+
+    switch (m->get_op()) {
+      case CEPH_CAP_OP_TRUNC: return handle_cap_trunc(session, in, m);
+      case CEPH_CAP_OP_IMPORT:
+      case CEPH_CAP_OP_REVOKE:
+      case CEPH_CAP_OP_GRANT: return handle_cap_grant(session, in, &cap, m);
+      case CEPH_CAP_OP_FLUSH_ACK: return handle_cap_flush_ack(session, in, &cap, m);
+    }
+  } else {
+    ldout(cct, 5) << __func__ << " don't have " << *in << " cap on mds." << mds << dendl;
+    return;
+  }
+}
+
+void Client::handle_cap_import(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m)
+{
+  mds_rank_t mds = session->mds_num;
+
+  ldout(cct, 5) << __func__ << " ino " << m->get_ino() << " mseq " << m->get_mseq()
+		<< " IMPORT from mds." << mds << dendl;
+
+  const mds_rank_t peer_mds = mds_rank_t(m->peer.mds);
+  Cap *cap = NULL;
+  UserPerm cap_perms;
+  if (auto it = in->caps.find(peer_mds); m->peer.cap_id && it != in->caps.end()) {
+    cap = &it->second;
+    cap_perms = cap->latest_perms;
+  }
+
+  // add/update it
+  SnapRealm *realm = NULL;
+  update_snap_trace(m->snapbl, &realm);
+
+  int issued = m->get_caps();
+  int wanted = m->get_wanted();
+  add_update_cap(in, session, m->get_cap_id(),
+		 issued, wanted, m->get_seq(), m->get_mseq(),
+		 m->get_realm(), CEPH_CAP_FLAG_AUTH, cap_perms);
+  
+  if (cap && cap->cap_id == m->peer.cap_id) {
+      remove_cap(cap, (m->peer.flags & CEPH_CAP_FLAG_RELEASE));
+  }
+
+  if (realm)
+    put_snap_realm(realm);
+  
+  if (in->auth_cap && in->auth_cap->session == session) {
+    if (!(wanted & CEPH_CAP_ANY_FILE_WR) ||
+	in->requested_max_size > m->get_max_size()) {
+      in->requested_max_size = 0;
+      ldout(cct, 15) << "reset requested_max_size after cap import" << dendl;
+    }
+    // reflush any/all caps (if we are now the auth_cap)
+    kick_flushing_caps(in, session);
+  }
+}
+
+void Client::handle_cap_export(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m)
+{
+  mds_rank_t mds = session->mds_num;
+
+  ldout(cct, 5) << __func__ << " ino " << m->get_ino() << " mseq " << m->get_mseq()
+		<< " EXPORT from mds." << mds << dendl;
+
+  auto it = in->caps.find(mds);
+  if (it != in->caps.end()) {
+    Cap &cap = it->second;
+    if (cap.cap_id == m->get_cap_id()) {
+      if (m->peer.cap_id) {
+	const auto peer_mds = mds_rank_t(m->peer.mds);
+        MetaSession *tsession = _get_or_open_mds_session(peer_mds);
+        auto it = in->caps.find(peer_mds);
+        if (it != in->caps.end()) {
+	  Cap &tcap = it->second;
+	  if (tcap.cap_id == m->peer.cap_id &&
+	      ceph_seq_cmp(tcap.seq, m->peer.seq) < 0) {
+	    tcap.cap_id = m->peer.cap_id;
+	    tcap.seq = m->peer.seq - 1;
+	    tcap.issue_seq = tcap.seq;
+	    tcap.issued |= cap.issued;
+	    tcap.implemented |= cap.issued;
+	    if (&cap == in->auth_cap)
+	      in->auth_cap = &tcap;
+	    if (in->auth_cap == &tcap && in->flushing_cap_item.is_on_list())
+	      adjust_session_flushing_caps(in, session, tsession);
+	  }
+        } else {
+	  add_update_cap(in, tsession, m->peer.cap_id, cap.issued, 0,
+		         m->peer.seq - 1, m->peer.mseq, (uint64_t)-1,
+		         &cap == in->auth_cap ? CEPH_CAP_FLAG_AUTH : 0,
+		         cap.latest_perms);
+        }
+      } else {
+	if (cap.wanted | cap.issued)
+	  in->flags |= I_CAP_DROPPED;
+      }
+
+      remove_cap(&cap, false);
+    }
+  }
+}
+
+void Client::handle_cap_trunc(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m)
+{
+  mds_rank_t mds = session->mds_num;
+  ceph_assert(in->caps.count(mds));
+
+  ldout(cct, 10) << __func__ << " on ino " << *in
+	   << " size " << in->size << " -> " << m->get_size()
+	   << dendl;
+  
+  int issued;
+  in->caps_issued(&issued);
+  issued |= in->caps_dirty();
+  update_inode_file_size(in, issued, m->get_size(),
+			 m->get_truncate_seq(), m->get_truncate_size());
+}
+
+void Client::handle_cap_flush_ack(MetaSession *session, Inode *in, Cap *cap, const MConstRef<MClientCaps>& m)
+{
+  ceph_tid_t flush_ack_tid = m->get_client_tid();
+  int dirty = m->get_dirty();
+  int cleaned = 0;
+  int flushed = 0;
+
+  auto it = in->flushing_cap_tids.begin();
+  if (it->first < flush_ack_tid) {
+       ldout(cct, 0) << __func__ << " mds." << session->mds_num
+                   << " got unexpected flush ack tid " << flush_ack_tid
+                   << " expected is " << it->first << dendl;
+  }
+  for (; it != in->flushing_cap_tids.end(); ) {
+    if (!it->second) {
+      // cap snap
+      ++it;
+      continue;
+    }
+    if (it->first == flush_ack_tid)
+      cleaned = it->second;
+    if (it->first <= flush_ack_tid) {
+      session->flushing_caps_tids.erase(it->first);
+      in->flushing_cap_tids.erase(it++);
+      ++flushed;
+      continue;
+    }
+    cleaned &= ~it->second;
+    if (!cleaned)
+      break;
+    ++it;
+  }
+
+  ldout(cct, 5) << __func__ << " mds." << session->mds_num
+	  << " cleaned " << ccap_string(cleaned) << " on " << *in
+	  << " with " << ccap_string(dirty) << dendl;
+
+  if (flushed) {
+    signal_cond_list(in->waitfor_caps);
+    if (session->flushing_caps_tids.empty() ||
+	*session->flushing_caps_tids.begin() > flush_ack_tid)
+      sync_cond.notify_all();
+  }
+
+  if (!dirty) {
+    in->cap_dirtier_uid = -1;
+    in->cap_dirtier_gid = -1;
+  }
+
+  if (!cleaned) {
+    ldout(cct, 10) << " tid " << m->get_client_tid() << " != any cap bit tids" << dendl;
+  } else {
+    if (in->flushing_caps) {
+      ldout(cct, 5) << "  flushing_caps " << ccap_string(in->flushing_caps)
+	      << " -> " << ccap_string(in->flushing_caps & ~cleaned) << dendl;
+      in->flushing_caps &= ~cleaned;
+      if (in->flushing_caps == 0) {
+	ldout(cct, 10) << " " << *in << " !flushing" << dendl;
+	num_flushing_caps--;
+       if (in->flushing_cap_tids.empty())
+	  in->flushing_cap_item.remove_myself();
+      }
+      if (!in->caps_dirty())
+	put_inode(in);
+    }
+  }
+}
+
+
+void Client::handle_cap_flushsnap_ack(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m)
+{
+  ceph_tid_t flush_ack_tid = m->get_client_tid();
+  mds_rank_t mds = session->mds_num;
+  ceph_assert(in->caps.count(mds));
+  snapid_t follows = m->get_snap_follows();
+
+  if (auto it = in->cap_snaps.find(follows); it != in->cap_snaps.end()) {
+    auto& capsnap = it->second;
+    if (flush_ack_tid != capsnap.flush_tid) {
+      ldout(cct, 10) << " tid " << flush_ack_tid << " != " << capsnap.flush_tid << dendl;
+    } else {
+      InodeRef tmp_ref(in);
+      ldout(cct, 5) << __func__ << " mds." << mds << " flushed snap follows " << follows
+	      << " on " << *in << dendl;
+      session->flushing_caps_tids.erase(capsnap.flush_tid);
+      in->flushing_cap_tids.erase(capsnap.flush_tid);
+      if (in->flushing_caps == 0 && in->flushing_cap_tids.empty())
+	in->flushing_cap_item.remove_myself();
+      in->cap_snaps.erase(it);
+
+      signal_cond_list(in->waitfor_caps);
+      if (session->flushing_caps_tids.empty() ||
+	  *session->flushing_caps_tids.begin() > flush_ack_tid)
+	sync_cond.notify_all();
+    }
+  } else {
+    ldout(cct, 5) << __func__ << " DUP(?) mds." << mds << " flushed snap follows " << follows
+	    << " on " << *in << dendl;
+    // we may not have it if we send multiple FLUSHSNAP requests and (got multiple FLUSHEDSNAPs back)
+  }
+}
+
+class C_Client_DentryInvalidate : public Context  {
+private:
+  Client *client;
+  vinodeno_t dirino;
+  vinodeno_t ino;
+  string name;
+public:
+  C_Client_DentryInvalidate(Client *c, Dentry *dn, bool del) :
+    client(c), name(dn->name) {
+      if (client->use_faked_inos()) {
+	dirino.ino = dn->dir->parent_inode->faked_ino;
+	if (del)
+	  ino.ino = dn->inode->faked_ino;
+      } else {
+	dirino = dn->dir->parent_inode->vino();
+	if (del)
+	  ino = dn->inode->vino();
+      }
+      if (!del)
+	ino.ino = inodeno_t();
+  }
+  void finish(int r) override {
+    // _async_dentry_invalidate is responsible for its own locking
+    ceph_assert(ceph_mutex_is_not_locked_by_me(client->client_lock));
+    client->_async_dentry_invalidate(dirino, ino, name);
+  }
+};
+
+void Client::_async_dentry_invalidate(vinodeno_t dirino, vinodeno_t ino, string& name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  ldout(cct, 10) << __func__ << " '" << name << "' ino " << ino
+		 << " in dir " << dirino << dendl;
+  dentry_invalidate_cb(callback_handle, dirino, ino, name.c_str(), name.length());
+}
+
+void Client::_schedule_invalidate_dentry_callback(Dentry *dn, bool del)
+{
+  if (dentry_invalidate_cb && dn->inode->ll_ref > 0)
+    async_dentry_invalidator.queue(new C_Client_DentryInvalidate(this, dn, del));
+}
+
+void Client::_try_to_trim_inode(Inode *in, bool sched_inval)
+{
+  int ref = in->get_nref();
+  ldout(cct, 5) << __func__ << " in " << *in <<dendl;
+
+  if (in->dir && !in->dir->dentries.empty()) {
+    for (auto p = in->dir->dentries.begin();
+	 p != in->dir->dentries.end(); ) {
+      Dentry *dn = p->second;
+      ++p;
+      /* rmsnap removes whole subtree, need trim inodes recursively.
+       * we don't need to invalidate dentries recursively. because
+       * invalidating a directory dentry effectively invalidate
+       * whole subtree */
+      if (in->snapid != CEPH_NOSNAP && dn->inode && dn->inode->is_dir())
+	_try_to_trim_inode(dn->inode.get(), false);
+
+      if (dn->lru_is_expireable())
+	unlink(dn, true, false);  // keep dir, drop dentry
+    }
+    if (in->dir->dentries.empty()) {
+      close_dir(in->dir);
+      --ref;
+    }
+  }
+
+  if (ref > 1 && (in->flags & I_SNAPDIR_OPEN)) {
+    InodeRef snapdir = open_snapdir(in);
+    _try_to_trim_inode(snapdir.get(), false);
+    --ref;
+  }
+
+  if (ref > 1) {
+    auto q = in->dentries.begin();
+    while (q != in->dentries.end()) {
+      Dentry *dn = *q;
+      ++q;
+      if( in->ll_ref > 0 && sched_inval) {
+        // FIXME: we play lots of unlink/link tricks when handling MDS replies,
+        //        so in->dentries doesn't always reflect the state of kernel's dcache.
+        _schedule_invalidate_dentry_callback(dn, true);
+      }
+      unlink(dn, true, true);
+    }
+  }
+}
+
+void Client::handle_cap_grant(MetaSession *session, Inode *in, Cap *cap, const MConstRef<MClientCaps>& m)
+{
+  mds_rank_t mds = session->mds_num;
+  int used = get_caps_used(in);
+  int wanted = in->caps_wanted();
+  int flags = 0;
+
+  const unsigned new_caps = m->get_caps();
+  const bool was_stale = session->cap_gen > cap->gen;
+  ldout(cct, 5) << __func__ << " on in " << m->get_ino() 
+		<< " mds." << mds << " seq " << m->get_seq()
+		<< " caps now " << ccap_string(new_caps)
+		<< " was " << ccap_string(cap->issued)
+		<< (was_stale ? " (stale)" : "") << dendl;
+
+  if (was_stale)
+      cap->issued = cap->implemented = CEPH_CAP_PIN;
+  cap->seq = m->get_seq();
+  cap->gen = session->cap_gen;
+
+  check_cap_issue(in, new_caps);
+
+  // update inode
+  int issued;
+  in->caps_issued(&issued);
+  issued |= in->caps_dirty();
+
+  if ((new_caps & CEPH_CAP_AUTH_SHARED) &&
+      !(issued & CEPH_CAP_AUTH_EXCL)) {
+    in->mode = m->head.mode;
+    in->uid = m->head.uid;
+    in->gid = m->head.gid;
+    in->btime = m->btime;
+  }
+  bool deleted_inode = false;
+  if ((new_caps & CEPH_CAP_LINK_SHARED) &&
+      !(issued & CEPH_CAP_LINK_EXCL)) {
+    in->nlink = m->head.nlink;
+    if (in->nlink == 0 &&
+	(new_caps & (CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL)))
+      deleted_inode = true;
+  }
+  if (!(issued & CEPH_CAP_XATTR_EXCL) &&
+      m->xattrbl.length() &&
+      m->head.xattr_version > in->xattr_version) {
+    auto p = m->xattrbl.cbegin();
+    decode(in->xattrs, p);
+    in->xattr_version = m->head.xattr_version;
+  }
+
+  if ((new_caps & CEPH_CAP_FILE_SHARED) && m->dirstat_is_valid()) {
+    in->dirstat.nfiles = m->get_nfiles();
+    in->dirstat.nsubdirs = m->get_nsubdirs();
+  }
+
+  if (new_caps & CEPH_CAP_ANY_RD) {
+    update_inode_file_time(in, issued, m->get_time_warp_seq(),
+			   m->get_ctime(), m->get_mtime(), m->get_atime());
+  }
+
+  if (new_caps & (CEPH_CAP_ANY_FILE_RD | CEPH_CAP_ANY_FILE_WR)) {
+    in->layout = m->get_layout();
+    update_inode_file_size(in, issued, m->get_size(),
+			   m->get_truncate_seq(), m->get_truncate_size());
+  }
+
+  if (m->inline_version > in->inline_version) {
+    in->inline_data = m->inline_data;
+    in->inline_version = m->inline_version;
+  }
+
+  /* always take a newer change attr */
+  if (m->get_change_attr() > in->change_attr)
+    in->change_attr = m->get_change_attr();
+
+  // max_size
+  if (cap == in->auth_cap &&
+      (new_caps & CEPH_CAP_ANY_FILE_WR) &&
+      (m->get_max_size() != in->max_size)) {
+    ldout(cct, 10) << "max_size " << in->max_size << " -> " << m->get_max_size() << dendl;
+    in->max_size = m->get_max_size();
+    if (in->max_size > in->wanted_max_size) {
+      in->wanted_max_size = 0;
+      in->requested_max_size = 0;
+    }
+  }
+
+  bool check = false;
+  if ((was_stale || m->get_op() == CEPH_CAP_OP_IMPORT) &&
+      (wanted & ~(cap->wanted | new_caps))) {
+    // If mds is importing cap, prior cap messages that update 'wanted'
+    // may get dropped by mds (migrate seq mismatch).
+    //
+    // We don't send cap message to update 'wanted' if what we want are
+    // already issued. If mds revokes caps, cap message that releases caps
+    // also tells mds what we want. But if caps got revoked by mds forcedly
+    // (session stale). We may haven't told mds what we want.
+    check = true;
+  }
+
+
+  // update caps
+  auto revoked = cap->issued & ~new_caps;
+  if (revoked) {
+    ldout(cct, 10) << "  revocation of " << ccap_string(revoked) << dendl;
+    cap->issued = new_caps;
+    cap->implemented |= new_caps;
+
+    // recall delegations if we're losing caps necessary for them
+    if (revoked & ceph_deleg_caps_for_type(CEPH_DELEGATION_RD))
+      in->recall_deleg(false);
+    else if (revoked & ceph_deleg_caps_for_type(CEPH_DELEGATION_WR))
+      in->recall_deleg(true);
+
+    used = adjust_caps_used_for_lazyio(used, cap->issued, cap->implemented);
+    if ((used & revoked & (CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO)) &&
+	!_flush(in, new C_Client_FlushComplete(this, in))) {
+      // waitin' for flush
+    } else if (used & revoked & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO)) {
+      if (_release(in)) {
+        check = true;
+        flags = CHECK_CAPS_NODELAY;
+      }
+    } else {
+      cap->wanted = 0; // don't let check_caps skip sending a response to MDS
+      check = true;
+      flags = CHECK_CAPS_NODELAY;
+    }
+  } else if (cap->issued == new_caps) {
+    ldout(cct, 10) << "  caps unchanged at " << ccap_string(cap->issued) << dendl;
+  } else {
+    ldout(cct, 10) << "  grant, new caps are " << ccap_string(new_caps & ~cap->issued) << dendl;
+    cap->issued = new_caps;
+    cap->implemented |= new_caps;
+
+    if (cap == in->auth_cap) {
+      // non-auth MDS is revoking the newly grant caps ?
+      for (const auto &p : in->caps) {
+	if (&p.second == cap)
+	  continue;
+	if (p.second.implemented & ~p.second.issued & new_caps) {
+	  check = true;
+	  break;
+	}
+      }
+    }
+  }
+
+  if (check)
+    check_caps(in, flags);
+
+  // wake up waiters
+  if (new_caps)
+    signal_cond_list(in->waitfor_caps);
+
+  // may drop inode's last ref
+  if (deleted_inode)
+    _try_to_trim_inode(in, true);
+}
+
+int Client::inode_permission(Inode *in, const UserPerm& perms, unsigned want)
+{
+  if (perms.uid() == 0) {
+    // For directories, DACs are overridable.
+    // For files, Read/write DACs are always overridable but executable DACs are
+    // overridable when there is at least one exec bit set
+    if(!S_ISDIR(in->mode) && (want & MAY_EXEC) && !(in->mode & S_IXUGO))
+      return -CEPHFS_EACCES;
+    return 0;
+  }
+  
+  if (perms.uid() != in->uid && (in->mode & S_IRWXG)) {
+    int ret = _posix_acl_permission(in, perms, want);
+    if (ret != -CEPHFS_EAGAIN)
+      return ret;
+  }
+
+  // check permissions before doing anything else
+  if (!in->check_mode(perms, want))
+    return -CEPHFS_EACCES;
+  return 0;
+}
+
+int Client::xattr_permission(Inode *in, const char *name, unsigned want,
+			     const UserPerm& perms)
+{
+  int r = _getattr_for_perm(in, perms);
+  if (r < 0)
+    goto out;
+
+  r = 0;
+  if (strncmp(name, "system.", 7) == 0) {
+    if ((want & MAY_WRITE) && (perms.uid() != 0 && perms.uid() != in->uid))
+      r = -CEPHFS_EPERM;
+  } else {
+    r = inode_permission(in, perms, want);
+  }
+out:
+  ldout(cct, 5) << __func__ << " " << in << " = " << r <<  dendl;
+  return r;
+}
+
+ostream& operator<<(ostream &out, const UserPerm& perm) {
+  out << "UserPerm(uid: " << perm.uid() << ", gid: " << perm.gid() << ")";
+  return out;
+}
+
+int Client::may_setattr(Inode *in, struct ceph_statx *stx, int mask,
+			const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *in << "; " << perms << dendl;
+  int r = _getattr_for_perm(in, perms);
+  if (r < 0)
+    goto out;
+
+  if (mask & CEPH_SETATTR_SIZE) {
+    r = inode_permission(in, perms, MAY_WRITE);
+    if (r < 0)
+      goto out;
+  }
+
+  r = -CEPHFS_EPERM;
+  if (mask & CEPH_SETATTR_UID) {
+    if (perms.uid() != 0 && (perms.uid() != in->uid || stx->stx_uid != in->uid))
+      goto out;
+  }
+  if (mask & CEPH_SETATTR_GID) {
+    if (perms.uid() != 0 && (perms.uid() != in->uid ||
+      	       (!perms.gid_in_groups(stx->stx_gid) && stx->stx_gid != in->gid)))
+      goto out;
+  }
+
+  if (mask & CEPH_SETATTR_MODE) {
+    if (perms.uid() != 0 && perms.uid() != in->uid)
+      goto out;
+
+    gid_t i_gid = (mask & CEPH_SETATTR_GID) ? stx->stx_gid : in->gid;
+    if (perms.uid() != 0 && !perms.gid_in_groups(i_gid))
+      stx->stx_mode &= ~S_ISGID;
+  }
+
+  if (mask & (CEPH_SETATTR_CTIME | CEPH_SETATTR_BTIME |
+	      CEPH_SETATTR_MTIME | CEPH_SETATTR_ATIME)) {
+    if (perms.uid() != 0 && perms.uid() != in->uid) {
+      int check_mask = CEPH_SETATTR_CTIME | CEPH_SETATTR_BTIME;
+      if (!(mask & CEPH_SETATTR_MTIME_NOW))
+	check_mask |= CEPH_SETATTR_MTIME;
+      if (!(mask & CEPH_SETATTR_ATIME_NOW))
+	check_mask |= CEPH_SETATTR_ATIME;
+      if (check_mask & mask) {
+	goto out;
+      } else {
+	r = inode_permission(in, perms, MAY_WRITE);
+	if (r < 0)
+	  goto out;
+      }
+    }
+  }
+  r = 0;
+out:
+  ldout(cct, 3) << __func__ << " " << in << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::may_open(Inode *in, int flags, const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *in << "; " << perms << dendl;
+  unsigned want = 0;
+
+  if ((flags & O_ACCMODE) == O_WRONLY)
+    want = MAY_WRITE;
+  else if ((flags & O_ACCMODE) == O_RDWR)
+    want = MAY_READ | MAY_WRITE;
+  else if ((flags & O_ACCMODE) == O_RDONLY)
+    want = MAY_READ;
+  if (flags & O_TRUNC)
+    want |= MAY_WRITE;
+
+  int r = 0;
+  switch (in->mode & S_IFMT) {
+    case S_IFLNK:
+      r = -CEPHFS_ELOOP;
+      goto out;
+    case S_IFDIR:
+      if (want & MAY_WRITE) {
+	r = -CEPHFS_EISDIR;
+	goto out;
+      }
+      break;
+  }
+
+  r = _getattr_for_perm(in, perms);
+  if (r < 0)
+    goto out;
+
+  r = inode_permission(in, perms, want);
+out:
+  ldout(cct, 3) << __func__ << " " << in << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::may_lookup(Inode *dir, const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *dir << "; " << perms << dendl;
+  int r = _getattr_for_perm(dir, perms);
+  if (r < 0)
+    goto out;
+
+  r = inode_permission(dir, perms, MAY_EXEC);
+out:
+  ldout(cct, 3) << __func__ << " " << dir << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::may_create(Inode *dir, const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *dir << "; " << perms << dendl;
+  int r = _getattr_for_perm(dir, perms);
+  if (r < 0)
+    goto out;
+
+  r = inode_permission(dir, perms, MAY_EXEC | MAY_WRITE);
+out:
+  ldout(cct, 3) << __func__ << " " << dir << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::may_delete(Inode *dir, const char *name, const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *dir << "; " << "; name " << name << "; " << perms << dendl;
+  int r = _getattr_for_perm(dir, perms);
+  if (r < 0)
+    goto out;
+
+  r = inode_permission(dir, perms, MAY_EXEC | MAY_WRITE);
+  if (r < 0)
+    goto out;
+
+  /* 'name == NULL' means rmsnap w/o permission checks */
+  if (perms.uid() != 0 && name && (dir->mode & S_ISVTX)) {
+    InodeRef otherin;
+    r = _lookup(dir, name, CEPH_CAP_AUTH_SHARED, &otherin, perms);
+    if (r < 0)
+      goto out;
+    if (dir->uid != perms.uid() && otherin->uid != perms.uid())
+      r = -CEPHFS_EPERM;
+  }
+out:
+  ldout(cct, 3) << __func__ << " " << dir << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::may_delete(const char *relpath, const UserPerm& perms) {
+  ldout(cct, 20) << __func__ << " " << relpath << "; " << perms << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -ENOTCONN;
+
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+  InodeRef dir;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &dir, perms);
+  if (r < 0)
+    return r;
+  if (cct->_conf->client_permissions) {
+    int r = may_delete(dir.get(), name.c_str(), perms);
+    if (r < 0)
+      return r;
+  }
+
+  return 0;
+}
+
+int Client::may_hardlink(Inode *in, const UserPerm& perms)
+{
+  ldout(cct, 20) << __func__ << " " << *in << "; " << perms << dendl;
+  int r = _getattr_for_perm(in, perms);
+  if (r < 0)
+    goto out;
+
+  if (perms.uid() == 0 || perms.uid() == in->uid) {
+    r = 0;
+    goto out;
+  }
+
+  r = -CEPHFS_EPERM;
+  if (!S_ISREG(in->mode))
+    goto out;
+
+  if (in->mode & S_ISUID)
+    goto out;
+
+  if ((in->mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP))
+    goto out;
+
+  r = inode_permission(in, perms, MAY_READ | MAY_WRITE);
+out:
+  ldout(cct, 3) << __func__ << " " << in << " = " << r <<  dendl;
+  return r;
+}
+
+int Client::_getattr_for_perm(Inode *in, const UserPerm& perms)
+{
+  int mask = CEPH_STAT_CAP_MODE;
+  bool force = false;
+  if (acl_type != NO_ACL) {
+    mask |= CEPH_STAT_CAP_XATTR;
+    force = in->xattr_version == 0;
+  }
+  return _getattr(in, mask, perms, force);
+}
+
+vinodeno_t Client::_get_vino(Inode *in)
+{
+  /* The caller must hold the client lock */
+  return vinodeno_t(in->ino, in->snapid);
+}
+
+/**
+ * Resolve an MDS spec to a list of MDS daemon GIDs.
+ *
+ * The spec is a string representing a GID, rank, filesystem:rank, or name/id.
+ * It may be '*' in which case it matches all GIDs.
+ *
+ * If no error is returned, the `targets` vector will be populated with at least
+ * one MDS.
+ */
+int Client::resolve_mds(
+    const std::string &mds_spec,
+    std::vector<mds_gid_t> *targets)
+{
+  ceph_assert(fsmap);
+  ceph_assert(targets != nullptr);
+
+  mds_role_t role;
+  CachedStackStringStream css;
+  int role_r = fsmap->parse_role(mds_spec, &role, *css);
+  if (role_r == 0) {
+    // We got a role, resolve it to a GID
+    auto& info = fsmap->get_filesystem(role.fscid)->mds_map.get_info(role.rank);
+    ldout(cct, 10) << __func__ << ": resolved " << mds_spec << " to role '"
+      << role << "' aka " << info.human_name() << dendl;
+    targets->push_back(info.global_id);
+    return 0;
+  }
+
+  std::string strtol_err;
+  long long rank_or_gid = strict_strtoll(mds_spec.c_str(), 10, &strtol_err);
+  if (strtol_err.empty()) {
+    // It is a possible GID
+    const mds_gid_t mds_gid = mds_gid_t(rank_or_gid);
+    if (fsmap->gid_exists(mds_gid)) {
+      auto& info = fsmap->get_info_gid(mds_gid);
+      ldout(cct, 10) << __func__ << ": validated gid " << mds_gid << " aka "
+                     << info.human_name() << dendl;
+      targets->push_back(mds_gid);
+      return 0;
+    } else {
+      lderr(cct) << __func__ << ": gid " << mds_gid << " not in MDS map"
+                 << dendl;
+      lderr(cct) << "FSMap: " << *fsmap << dendl;
+      return -CEPHFS_ENOENT;
+    }
+  } else if (mds_spec == "*") {
+    // It is a wildcard: use all MDSs
+    const auto& mds_info = fsmap->get_mds_info();
+
+    ldout(cct, 10) << __func__ << ": resolving `*' to all MDS daemons" << dendl;
+    if (mds_info.empty()) {
+      lderr(cct) << __func__ << ": no MDS daemons found" << dendl;
+      lderr(cct) << "FSMap: " << *fsmap << dendl;
+      return -CEPHFS_ENOENT;
+    }
+
+    for (const auto& [gid, info] : mds_info) {
+      ldout(cct, 10) << __func__ << ": appending " << info.human_name() << " to targets" << dendl;
+      targets->push_back(gid);
+    }
+    return 0;
+  } else {
+    // It did not parse as an integer, it is not a wildcard, it must be a name
+    const mds_gid_t mds_gid = fsmap->find_mds_gid_by_name(mds_spec);
+    if (mds_gid == 0) {
+      lderr(cct) << __func__ << ": no MDS daemons found by name `" << mds_spec << "'" << dendl;
+      lderr(cct) << "FSMap: " << *fsmap << dendl;
+      return -CEPHFS_ENOENT;
+    } else {
+      auto& info = fsmap->get_info_gid(mds_gid);
+      ldout(cct, 10) << __func__ << ": resolved name '" << mds_spec
+                     << "' to " << info.human_name() << dendl;
+      targets->push_back(mds_gid);
+    }
+    return 0;
+  }
+}
+
+
+/**
+ * Authenticate with mon and establish global ID
+ */
+int Client::authenticate()
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  if (monclient->is_authenticated()) {
+    return 0;
+  }
+
+  client_lock.unlock();
+  int r = monclient->authenticate(std::chrono::duration<double>(mount_timeout).count());
+  client_lock.lock();
+  if (r < 0) {
+    return r;
+  }
+
+  whoami = monclient->get_global_id();
+  messenger->set_myname(entity_name_t::CLIENT(whoami.v));
+
+  return 0;
+}
+
+int Client::fetch_fsmap(bool user)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  // Retrieve FSMap to enable looking up daemon addresses.  We need FSMap
+  // rather than MDSMap because no one MDSMap contains all the daemons, and
+  // a `tell` can address any daemon.
+  version_t fsmap_latest;
+  bs::error_code ec;
+  do {
+    client_lock.unlock();
+    std::tie(fsmap_latest, std::ignore) =
+      monclient->get_version("fsmap", ca::use_blocked[ec]);
+    client_lock.lock();
+  } while (ec == bs::errc::resource_unavailable_try_again);
+
+  if (ec) {
+    lderr(cct) << "Failed to learn FSMap version: " << ec << dendl;
+    return ceph::from_error_code(ec);
+  }
+
+  ldout(cct, 10) << __func__ << " learned FSMap version " << fsmap_latest << dendl;
+
+  if (user) {
+    if (!fsmap_user || fsmap_user->get_epoch() < fsmap_latest) {
+      monclient->sub_want("fsmap.user", fsmap_latest, CEPH_SUBSCRIBE_ONETIME);
+      monclient->renew_subs();
+      wait_on_list(waiting_for_fsmap);
+    }
+    ceph_assert(fsmap_user);
+    ceph_assert(fsmap_user->get_epoch() >= fsmap_latest);
+  } else {
+    if (!fsmap || fsmap->get_epoch() < fsmap_latest) {
+      monclient->sub_want("fsmap", fsmap_latest, CEPH_SUBSCRIBE_ONETIME);
+      monclient->renew_subs();
+      wait_on_list(waiting_for_fsmap);
+    }
+    ceph_assert(fsmap);
+    ceph_assert(fsmap->get_epoch() >= fsmap_latest);
+  }
+  ldout(cct, 10) << __func__ << " finished waiting for FSMap version "
+		 << fsmap_latest << dendl;
+  return 0;
+}
+
+/**
+ *
+ * @mds_spec one of ID, rank, GID, "*"
+ *
+ */
+int Client::mds_command(
+    const std::string &mds_spec,
+    const vector<string>& cmd,
+    const bufferlist& inbl,
+    bufferlist *outbl,
+    string *outs,
+    Context *onfinish)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  if (!iref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::unique_lock cl(client_lock);
+
+  int r;
+  r = authenticate();
+  if (r < 0) {
+    return r;
+  }
+
+  r = fetch_fsmap(false);
+  if (r < 0) {
+    return r;
+  }
+
+  // Look up MDS target(s) of the command
+  std::vector<mds_gid_t> targets;
+  r = resolve_mds(mds_spec, &targets);
+  if (r < 0) {
+    return r;
+  }
+
+  // If daemons are laggy, we won't send them commands.  If all
+  // are laggy then we fail.
+  std::vector<mds_gid_t> non_laggy;
+  for (const auto& gid : targets) {
+    const auto info = fsmap->get_info_gid(gid);
+    if (!info.laggy()) {
+      non_laggy.push_back(gid);
+    }
+  }
+  if (non_laggy.size() == 0) {
+    *outs = "All targeted MDS daemons are laggy";
+    return -CEPHFS_ENOENT;
+  }
+
+  if (metadata.empty()) {
+    // We are called on an unmounted client, so metadata
+    // won't be initialized yet.
+    populate_metadata("");
+  }
+
+  // Send commands to targets
+  C_GatherBuilder gather(cct, onfinish);
+  for (const auto& target_gid : non_laggy) {
+    const auto info = fsmap->get_info_gid(target_gid);
+
+    // Open a connection to the target MDS
+    ConnectionRef conn = messenger->connect_to_mds(info.get_addrs());
+
+    cl.unlock();
+    {
+      std::scoped_lock cmd_lock(command_lock);
+      // Generate MDSCommandOp state
+      auto &op = command_table.start_command();
+
+      op.on_finish = gather.new_sub();
+      op.cmd = cmd;
+      op.outbl = outbl;
+      op.outs = outs;
+      op.inbl = inbl;
+      op.mds_gid = target_gid;
+      op.con = conn;
+
+      ldout(cct, 4) << __func__ << ": new command op to " << target_gid
+        << " tid=" << op.tid << cmd << dendl;
+
+      // Construct and send MCommand
+      MessageRef m = op.get_message(monclient->get_fsid());
+      conn->send_message2(std::move(m));
+    }
+    cl.lock();
+  }
+  gather.activate();
+
+  return 0;
+}
+
+void Client::handle_command_reply(const MConstRef<MCommandReply>& m)
+{
+  ceph_tid_t const tid = m->get_tid();
+
+  ldout(cct, 10) << __func__ << ": tid=" << m->get_tid() << dendl;
+
+  std::scoped_lock cmd_lock(command_lock);
+  if (!command_table.exists(tid)) {
+    ldout(cct, 1) << __func__ << ": unknown tid " << tid << ", dropping" << dendl;
+    return;
+  }
+
+  auto &op = command_table.get_command(tid);
+  if (op.outbl) {
+    *op.outbl = m->get_data();
+  }
+  if (op.outs) {
+    *op.outs = m->rs;
+  }
+
+  if (op.on_finish) {
+    op.on_finish->complete(m->r);
+  }
+
+  command_table.erase(tid);
+}
+
+// -------------------
+// MOUNT
+
+int Client::subscribe_mdsmap(const std::string &fs_name)
+{
+  int r = authenticate();
+  if (r < 0) {
+    lderr(cct) << "authentication failed: " << cpp_strerror(r) << dendl;
+    return r;
+  }
+
+  std::string resolved_fs_name;
+  if (fs_name.empty()) {
+    resolved_fs_name = cct->_conf.get_val<std::string>("client_fs");
+    if (resolved_fs_name.empty())
+	    // Try the backwards compatibility fs name option
+	    resolved_fs_name = cct->_conf.get_val<std::string>("client_mds_namespace");
+  } else {
+    resolved_fs_name = fs_name;
+  }
+
+  std::string want = "mdsmap";
+  if (!resolved_fs_name.empty()) {
+    r = fetch_fsmap(true);
+    if (r < 0)
+      return r;
+    fscid = fsmap_user->get_fs_cid(resolved_fs_name);
+    if (fscid == FS_CLUSTER_ID_NONE) {
+      return -CEPHFS_ENOENT;
+    }
+
+    std::ostringstream oss;
+    oss << want << "." << fscid;
+    want = oss.str();
+  }
+  ldout(cct, 10) << "Subscribing to map '" << want << "'" << dendl;
+
+  monclient->sub_want(want, 0, 0);
+  monclient->renew_subs();
+
+  return 0;
+}
+
+int Client::mount(const std::string &mount_root, const UserPerm& perms,
+		  bool require_mds, const std::string &fs_name)
+{
+  ceph_assert(is_initialized());
+
+  /*
+   * To make sure that the _unmount() must wait until the mount()
+   * is done.
+   */
+  RWRef_t mref_writer(mount_state, CLIENT_MOUNTING, false);
+  if (!mref_writer.is_first_writer()) // already mounting or mounted
+    return 0;
+
+  std::unique_lock cl(client_lock);
+
+  int r = subscribe_mdsmap(fs_name);
+  if (r < 0) {
+    lderr(cct) << "mdsmap subscription failed: " << cpp_strerror(r) << dendl;
+    return r;
+  }
+
+  start_tick_thread(); // start tick thread
+
+  if (require_mds) {
+    while (1) {
+      auto availability = mdsmap->is_cluster_available();
+      if (availability == MDSMap::STUCK_UNAVAILABLE) {
+        // Error out
+        ldout(cct, 10) << "mds cluster unavailable: epoch=" << mdsmap->get_epoch() << dendl;
+        return CEPH_FUSE_NO_MDS_UP;
+      } else if (availability == MDSMap::AVAILABLE) {
+        // Continue to mount
+        break;
+      } else if (availability == MDSMap::TRANSIENT_UNAVAILABLE) {
+        // Else, wait.  MDSMonitor will update the map to bring
+        // us to a conclusion eventually.
+        wait_on_list(waiting_for_mdsmap);
+      } else {
+        // Unexpected value!
+        ceph_abort();
+      }
+    }
+  }
+
+  populate_metadata(mount_root.empty() ? "/" : mount_root);
+
+  filepath fp(CEPH_INO_ROOT);
+  if (!mount_root.empty()) {
+    fp = filepath(mount_root.c_str());
+  }
+  while (true) {
+    MetaRequest *req = new MetaRequest(CEPH_MDS_OP_GETATTR);
+    req->set_filepath(fp);
+    req->head.args.getattr.mask = CEPH_STAT_CAP_INODE_ALL;
+    int res = make_request(req, perms);
+    if (res < 0) {
+      if (res == -CEPHFS_EACCES && root) {
+	ldout(cct, 1) << __func__ << " EACCES on parent of mount point; quotas may not work" << dendl;
+	break;
+      }
+      return res;
+    }
+
+    if (fp.depth())
+      fp.pop_dentry();
+    else
+      break;
+  }
+
+  ceph_assert(root);
+  _ll_get(root.get());
+
+  // trace?
+  if (!cct->_conf->client_trace.empty()) {
+    traceout.open(cct->_conf->client_trace.c_str());
+    if (traceout.is_open()) {
+      ldout(cct, 1) << "opened trace file '" << cct->_conf->client_trace << "'" << dendl;
+    } else {
+      ldout(cct, 1) << "FAILED to open trace file '" << cct->_conf->client_trace << "'" << dendl;
+    }
+  }
+
+  /*
+  ldout(cct, 3) << "op: // client trace data structs" << dendl;
+  ldout(cct, 3) << "op: struct stat st;" << dendl;
+  ldout(cct, 3) << "op: struct utimbuf utim;" << dendl;
+  ldout(cct, 3) << "op: int readlinkbuf_len = 1000;" << dendl;
+  ldout(cct, 3) << "op: char readlinkbuf[readlinkbuf_len];" << dendl;
+  ldout(cct, 3) << "op: map<string, inode_t*> dir_contents;" << dendl;
+  ldout(cct, 3) << "op: map<int, int> open_files;" << dendl;
+  ldout(cct, 3) << "op: int fd;" << dendl;
+  */
+
+  mref_writer.update_state(CLIENT_MOUNTED);
+  return 0;
+}
+
+// UNMOUNT
+
+void Client::_close_sessions()
+{
+  for (auto it = mds_sessions.begin(); it != mds_sessions.end(); ) {
+    if (it->second.state == MetaSession::STATE_REJECTED)
+      mds_sessions.erase(it++);
+    else
+      ++it;
+  }
+
+  while (!mds_sessions.empty()) {
+    // send session closes!
+    for (auto &p : mds_sessions) {
+      if (p.second.state != MetaSession::STATE_CLOSING) {
+	_close_mds_session(&p.second);
+	mds_ranks_closing.insert(p.first);
+      }
+    }
+
+    // wait for sessions to close
+    double timo = cct->_conf.get_val<std::chrono::seconds>("client_shutdown_timeout").count();
+    ldout(cct, 2) << "waiting for " << mds_ranks_closing.size() << " mds session(s) to close (timeout: "
+                  << timo << "s)" << dendl;
+    std::unique_lock l{client_lock, std::adopt_lock};
+    if (!timo) {
+      mount_cond.wait(l);
+    } else if (!mount_cond.wait_for(l, ceph::make_timespan(timo), [this] { return mds_ranks_closing.empty(); })) {
+      ldout(cct, 1) << mds_ranks_closing.size() << " mds(s) did not respond to session close -- timing out." << dendl;
+      while (!mds_ranks_closing.empty()) {
+        auto session = mds_sessions.at(*mds_ranks_closing.begin());
+        // this prunes entry from mds_sessions and mds_ranks_closing
+        _closed_mds_session(&session, -CEPHFS_ETIMEDOUT);
+      }
+    }
+
+    mds_ranks_closing.clear();
+    l.release();
+  }
+}
+
+void Client::flush_mdlog_sync(Inode *in)
+{
+  if (in->unsafe_ops.empty()) {
+    return;
+  }
+
+  std::set<mds_rank_t> anchor;
+  for (auto &&p : in->unsafe_ops) {
+    anchor.emplace(p->mds);
+  }
+  if (in->auth_cap) {
+    anchor.emplace(in->auth_cap->session->mds_num);
+  }
+
+  for (auto &rank : anchor) {
+    auto session = &mds_sessions.at(rank);
+    flush_mdlog(session);
+  }
+}
+
+void Client::flush_mdlog_sync()
+{
+  if (mds_requests.empty())
+    return;
+  for (auto &p : mds_sessions) {
+    flush_mdlog(&p.second);
+  }
+}
+
+void Client::flush_mdlog(MetaSession *session)
+{
+  // Only send this to Luminous or newer MDS daemons, older daemons
+  // will crash if they see an unknown CEPH_SESSION_* value in this msg.
+  const uint64_t features = session->con->get_features();
+  if (HAVE_FEATURE(features, SERVER_LUMINOUS)) {
+    auto m = make_message<MClientSession>(CEPH_SESSION_REQUEST_FLUSH_MDLOG);
+    session->con->send_message2(std::move(m));
+  }
+}
+
+
+void Client::_abort_mds_sessions(int err)
+{
+  for (auto p = mds_requests.begin(); p != mds_requests.end(); ) {
+    auto req = p->second;
+    ++p;
+    // unsafe requests will be removed during close session below.
+    if (req->got_unsafe)
+      continue;
+
+    req->abort(err);
+    if (req->caller_cond) {
+      req->kick = true;
+      req->caller_cond->notify_all();
+    }
+  }
+
+  // Process aborts on any requests that were on this waitlist.
+  // Any requests that were on a waiting_for_open session waitlist
+  // will get kicked during close session below.
+  signal_cond_list(waiting_for_mdsmap);
+
+  // Force-close all sessions
+  while(!mds_sessions.empty()) {
+    auto& session = mds_sessions.begin()->second;
+    _closed_mds_session(&session, err);
+  }
+}
+
+void Client::_unmount(bool abort)
+{
+  /*
+   * We are unmounting the client.
+   *
+   * Just declare the state to STATE_UNMOUNTING to block and fail
+   * any new comming "reader" and then try to wait all the in-flight
+   * "readers" to finish.
+   */
+  RWRef_t mref_writer(mount_state, CLIENT_UNMOUNTING, false);
+  if (!mref_writer.is_first_writer())
+    return;
+  mref_writer.wait_readers_done();
+
+  std::unique_lock lock{client_lock};
+
+  if (abort || blocklisted) {
+    ldout(cct, 2) << "unmounting (" << (abort ? "abort)" : "blocklisted)") << dendl;
+  } else {
+    ldout(cct, 2) << "unmounting" << dendl;
+  }
+
+  deleg_timeout = 0;
+
+  if (abort) {
+    mount_aborted = true;
+    // Abort all mds sessions
+    _abort_mds_sessions(-CEPHFS_ENOTCONN);
+
+    objecter->op_cancel_writes(-CEPHFS_ENOTCONN);
+  } else {
+    // flush the mdlog for pending requests, if any
+    flush_mdlog_sync();
+  }
+
+  mount_cond.wait(lock, [this] {
+    if (!mds_requests.empty()) {
+      ldout(cct, 10) << "waiting on " << mds_requests.size() << " requests"
+		     << dendl;
+    }
+    return mds_requests.empty();
+  });
+
+  cwd.reset();
+  root.reset();
+
+  // clean up any unclosed files
+  while (!fd_map.empty()) {
+    Fh *fh = fd_map.begin()->second;
+    fd_map.erase(fd_map.begin());
+    ldout(cct, 0) << " destroyed lost open file " << fh << " on " << *fh->inode << dendl;
+    _release_fh(fh);
+  }
+  
+  while (!ll_unclosed_fh_set.empty()) {
+    set<Fh*>::iterator it = ll_unclosed_fh_set.begin();
+    Fh *fh = *it;
+    ll_unclosed_fh_set.erase(fh);
+    ldout(cct, 0) << " destroyed lost open file " << fh << " on " << *(fh->inode) << dendl;
+    _release_fh(fh);
+  }
+
+  while (!opened_dirs.empty()) {
+    dir_result_t *dirp = *opened_dirs.begin();
+    ldout(cct, 0) << " destroyed lost open dir " << dirp << " on " << *dirp->inode << dendl;
+    _closedir(dirp);
+  }
+
+  _ll_drop_pins();
+
+  if (cct->_conf->client_oc) {
+    // flush/release all buffered data
+    std::list<InodeRef> anchor;
+    for (auto& p : inode_map) {
+      Inode *in = p.second;
+      if (!in) {
+	ldout(cct, 0) << "null inode_map entry ino " << p.first << dendl;
+	ceph_assert(in);
+      }
+
+      // prevent inode from getting freed
+      anchor.emplace_back(in);
+
+      if (abort || blocklisted) {
+        objectcacher->purge_set(&in->oset);
+      } else if (!in->caps.empty()) {
+	_release(in);
+	_flush(in, new C_Client_FlushComplete(this, in));
+      }
+    }
+  }
+
+  if (abort || blocklisted) {
+    for (auto p = dirty_list.begin(); !p.end(); ) {
+      Inode *in = *p;
+      ++p;
+      if (in->dirty_caps) {
+	ldout(cct, 0) << " drop dirty caps on " << *in << dendl;
+	in->mark_caps_clean();
+	put_inode(in);
+      }
+    }
+  } else {
+    flush_caps_sync();
+    wait_sync_caps(last_flush_tid);
+  }
+
+  // empty lru cache
+  trim_cache();
+
+  delay_put_inodes();
+
+  while (lru.lru_get_size() > 0 ||
+         !inode_map.empty()) {
+    ldout(cct, 2) << "cache still has " << lru.lru_get_size()
+            << "+" << inode_map.size() << " items"
+	    << ", waiting (for caps to release?)"
+            << dendl;
+
+    if (auto r = mount_cond.wait_for(lock, ceph::make_timespan(5));
+	r == std::cv_status::timeout) {
+      dump_cache(NULL);
+    }
+  }
+  ceph_assert(lru.lru_get_size() == 0);
+  ceph_assert(inode_map.empty());
+
+  // stop tracing
+  if (!cct->_conf->client_trace.empty()) {
+    ldout(cct, 1) << "closing trace file '" << cct->_conf->client_trace << "'" << dendl;
+    traceout.close();
+  }
+
+  // stop the tick thread
+  tick_thread_stopped = true;
+  upkeep_cond.notify_one();
+
+  _close_sessions();
+
+  mref_writer.update_state(CLIENT_UNMOUNTED);
+
+  /*
+   * Stop the remount_queue before clearing the mountpoint memory
+   * to avoid possible use-after-free bug.
+   */
+  if (remount_cb) {
+    ldout(cct, 10) << "unmount stopping remount finisher" << dendl;
+    remount_finisher.wait_for_empty();
+    remount_finisher.stop();
+    remount_cb = nullptr;
+  }
+
+  ldout(cct, 2) << "unmounted." << dendl;
+}
+
+void Client::unmount()
+{
+  _unmount(false);
+}
+
+void Client::abort_conn()
+{
+  _unmount(true);
+}
+
+void Client::flush_cap_releases()
+{
+  uint64_t nr_caps = 0;
+
+  // send any cap releases
+  for (auto &p : mds_sessions) {
+    auto &session = p.second;
+    if (session.release && mdsmap->is_clientreplay_or_active_or_stopping(
+          p.first)) {
+      nr_caps += session.release->caps.size();
+      if (cct->_conf->client_inject_release_failure) {
+        ldout(cct, 20) << __func__ << " injecting failure to send cap release message" << dendl;
+      } else {
+        session.con->send_message2(std::move(session.release));
+      }
+      session.release.reset();
+    }
+  }
+
+  if (nr_caps > 0) {
+    dec_pinned_icaps(nr_caps);
+  }
+}
+
+void Client::renew_and_flush_cap_releases()
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  if (!mount_aborted && mdsmap->get_epoch()) {
+    // renew caps?
+    auto el = ceph::coarse_mono_clock::now() - last_cap_renew;
+    if (unlikely(utime_t(el) > mdsmap->get_session_timeout() / 3.0))
+      renew_caps();
+
+    flush_cap_releases();
+  }
+}
+
+void Client::tick()
+{
+  ldout(cct, 20) << "tick" << dendl;
+
+  auto now = ceph::coarse_mono_clock::now();
+
+  /*
+   * If the mount() is not finished
+   */
+  if (is_mounting() && !mds_requests.empty()) {
+    MetaRequest *req = mds_requests.begin()->second;
+
+    if (req->created + mount_timeout < now) {
+      req->abort(-CEPHFS_ETIMEDOUT);
+      if (req->caller_cond) {
+        req->kick = true;
+        req->caller_cond->notify_all();
+      }
+      signal_cond_list(waiting_for_mdsmap);
+      for (auto &p : mds_sessions) {
+        signal_context_list(p.second.waiting_for_open);
+      }
+    }
+  }
+
+  renew_and_flush_cap_releases();
+
+  // delayed caps
+  xlist<Inode*>::iterator p = delayed_list.begin();
+  while (!p.end()) {
+    Inode *in = *p;
+    ++p;
+    if (!mount_aborted && in->hold_caps_until > now)
+      break;
+    delayed_list.pop_front();
+    if (!mount_aborted)
+      check_caps(in, CHECK_CAPS_NODELAY);
+  }
+
+  if (!mount_aborted)
+    collect_and_send_metrics();
+
+  delay_put_inodes(is_unmounting());
+  trim_cache(true);
+
+  if (blocklisted && (is_mounted() || is_unmounting()) &&
+      last_auto_reconnect + std::chrono::seconds(30 * 60) < now &&
+      cct->_conf.get_val<bool>("client_reconnect_stale")) {
+    messenger->client_reset();
+    fd_gen++; // invalidate open files
+    blocklisted = false;
+    _kick_stale_sessions();
+    last_auto_reconnect = now;
+  }
+}
+
+void Client::start_tick_thread()
+{
+  upkeeper = std::thread([this]() {
+    using time = ceph::coarse_mono_time;
+    using sec = std::chrono::seconds;
+
+    auto last_tick = time::min();
+
+    std::unique_lock cl(client_lock);
+    while (!tick_thread_stopped) {
+      auto now = clock::now();
+      auto since = now - last_tick;
+
+      auto t_interval = clock::duration(cct->_conf.get_val<sec>("client_tick_interval"));
+      auto d_interval = clock::duration(cct->_conf.get_val<sec>("client_debug_inject_tick_delay"));
+
+      auto interval = std::max(t_interval, d_interval);
+      if (likely(since >= interval*.90)) {
+        tick();
+        last_tick = clock::now();
+      } else {
+        interval -= since;
+      }
+
+      ldout(cct, 20) << "upkeep thread waiting interval " << interval << dendl;
+      if (!tick_thread_stopped)
+        upkeep_cond.wait_for(cl, interval);
+    }
+  });
+}
+
+void Client::collect_and_send_metrics() {
+  ldout(cct, 20) << __func__ << dendl;
+
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  // right now, we only track and send global metrics. its sufficient
+  // to send these metrics to MDS rank0.
+  collect_and_send_global_metrics();
+}
+
+void Client::collect_and_send_global_metrics() {
+  ldout(cct, 20) << __func__ << dendl;
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  if (!have_open_session((mds_rank_t)0)) {
+    ldout(cct, 5) << __func__ << ": no session with rank=0 -- not sending metric"
+                  << dendl;
+    return;
+  }
+  auto session = _get_or_open_mds_session((mds_rank_t)0);
+  if (!session->mds_features.test(CEPHFS_FEATURE_METRIC_COLLECT)) {
+    ldout(cct, 5) << __func__ << ": rank=0 does not support metrics" << dendl;
+    return;
+  }
+
+  ClientMetricMessage metric;
+  std::vector<ClientMetricMessage> message;
+
+  // read latency
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_READ_LATENCY)) {
+    metric = ClientMetricMessage(ReadLatencyPayload(logger->tget(l_c_read),
+                                                    logger->tget(l_c_rd_avg),
+                                                    logger->get(l_c_rd_sqsum),
+                                                    nr_read_request));
+    message.push_back(metric);
+  }
+
+  // write latency
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_WRITE_LATENCY)) {
+    metric = ClientMetricMessage(WriteLatencyPayload(logger->tget(l_c_wrlat),
+                                                    logger->tget(l_c_wr_avg),
+                                                    logger->get(l_c_wr_sqsum),
+                                                    nr_write_request));
+    message.push_back(metric);
+  }
+
+  // metadata latency
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_METADATA_LATENCY)) {
+    metric = ClientMetricMessage(MetadataLatencyPayload(logger->tget(l_c_lat),
+                                                        logger->tget(l_c_md_avg),
+                                                        logger->get(l_c_md_sqsum),
+                                                        nr_metadata_request));
+    message.push_back(metric);
+  }
+
+  // cap hit ratio -- nr_caps is unused right now
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_CAP_INFO)) {
+    auto [cap_hits, cap_misses] = get_cap_hit_rates();
+    metric = ClientMetricMessage(CapInfoPayload(cap_hits, cap_misses, 0));
+    message.push_back(metric);
+  }
+
+  // dentry lease hit ratio
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_DENTRY_LEASE)) {
+    auto [dlease_hits, dlease_misses, nr] = get_dlease_hit_rates();
+    metric = ClientMetricMessage(DentryLeasePayload(dlease_hits, dlease_misses, nr));
+    message.push_back(metric);
+  }
+
+  // opened files
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_OPENED_FILES)) {
+    auto [opened_files, total_inodes] = get_opened_files_rates();
+    metric = ClientMetricMessage(OpenedFilesPayload(opened_files, total_inodes));
+    message.push_back(metric);
+  }
+
+  // pinned i_caps
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_PINNED_ICAPS)) {
+    auto [pinned_icaps, total_inodes] = get_pinned_icaps_rates();
+    metric = ClientMetricMessage(PinnedIcapsPayload(pinned_icaps, total_inodes));
+    message.push_back(metric);
+  }
+
+  // opened inodes
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_OPENED_INODES)) {
+    auto [opened_inodes, total_inodes] = get_opened_inodes_rates();
+    metric = ClientMetricMessage(OpenedInodesPayload(opened_inodes, total_inodes));
+    message.push_back(metric);
+  }
+
+  // read io sizes
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_READ_IO_SIZES)) {
+    metric = ClientMetricMessage(ReadIoSizesPayload(total_read_ops,
+                                                    total_read_size));
+    message.push_back(metric);
+  }
+
+  // write io sizes
+  if (_collect_and_send_global_metrics ||
+      session->mds_metric_flags.test(CLIENT_METRIC_TYPE_WRITE_IO_SIZES)) {
+    metric = ClientMetricMessage(WriteIoSizesPayload(total_write_ops,
+                                                     total_write_size));
+    message.push_back(metric);
+  }
+
+  session->con->send_message2(make_message<MClientMetrics>(std::move(message)));
+}
+
+void Client::renew_caps()
+{
+  ldout(cct, 10) << "renew_caps()" << dendl;
+  last_cap_renew = ceph::coarse_mono_clock::now();
+
+  for (auto &p : mds_sessions) {
+    ldout(cct, 15) << "renew_caps requesting from mds." << p.first << dendl;
+    if (mdsmap->get_state(p.first) >= MDSMap::STATE_REJOIN)
+      renew_caps(&p.second);
+  }
+}
+
+void Client::renew_caps(MetaSession *session)
+{
+  ldout(cct, 10) << "renew_caps mds." << session->mds_num << dendl;
+  session->last_cap_renew_request = ceph_clock_now();
+  uint64_t seq = ++session->cap_renew_seq;
+  session->con->send_message2(make_message<MClientSession>(CEPH_SESSION_REQUEST_RENEWCAPS, seq));
+}
+
+
+// ===============================================================
+// high level (POSIXy) interface
+
+int Client::_do_lookup(Inode *dir, const string& name, int mask,
+		       InodeRef *target, const UserPerm& perms)
+{
+  int op = dir->snapid == CEPH_SNAPDIR ? CEPH_MDS_OP_LOOKUPSNAP : CEPH_MDS_OP_LOOKUP;
+  MetaRequest *req = new MetaRequest(op);
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+  req->set_inode(dir);
+  if (cct->_conf->client_debug_getattr_caps && op == CEPH_MDS_OP_LOOKUP)
+      mask |= DEBUG_GETATTR_CAPS;
+  req->head.args.getattr.mask = mask;
+
+  ldout(cct, 10) << __func__ << " on " << path << dendl;
+
+  int r = make_request(req, perms, target);
+  ldout(cct, 10) << __func__ << " res is " << r << dendl;
+  return r;
+}
+
+bool Client::_dentry_valid(const Dentry *dn)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  // is dn lease valid?
+  utime_t now = ceph_clock_now();
+  if (dn->lease_mds >= 0 && dn->lease_ttl > now &&
+      mds_sessions.count(dn->lease_mds)) {
+    MetaSession &s = mds_sessions.at(dn->lease_mds);
+    if (s.cap_ttl > now && s.cap_gen == dn->lease_gen) {
+      dlease_hit();
+      return true;
+    }
+
+    ldout(cct, 20) << " bad lease, cap_ttl " << s.cap_ttl << ", cap_gen " << s.cap_gen
+                   << " vs lease_gen " << dn->lease_gen << dendl;
+  }
+
+  dlease_miss();
+  return false;
+}
+
+int Client::_lookup(Inode *dir, const string& dname, int mask, InodeRef *target,
+		    const UserPerm& perms, std::string* alternate_name)
+{
+  int r = 0;
+  Dentry *dn = NULL;
+  bool did_lookup_request = false;
+  // can only request shared caps
+  mask &= CEPH_CAP_ANY_SHARED | CEPH_STAT_RSTAT;
+
+  if (dname == "..") {
+    if (dir->dentries.empty()) {
+      MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPPARENT);
+      filepath path(dir->ino);
+      req->set_filepath(path);
+
+      InodeRef tmptarget;
+      int r = make_request(req, perms, &tmptarget, NULL, rand() % mdsmap->get_num_in_mds());
+
+      if (r == 0) {
+	*target = std::move(tmptarget);
+	ldout(cct, 8) << __func__ << " found target " << (*target)->ino << dendl;
+      } else {
+	*target = dir;
+      }
+    }
+    else
+      *target = dir->get_first_parent()->dir->parent_inode; //dirs can't be hard-linked
+    goto done;
+  }
+
+  if (dname == ".") {
+    *target = dir;
+    goto done;
+  }
+
+  if (!dir->is_dir()) {
+    r = -CEPHFS_ENOTDIR;
+    goto done;
+  }
+
+  if (dname.length() > NAME_MAX) {
+    r = -CEPHFS_ENAMETOOLONG;
+    goto done;
+  }
+
+  if (dname == cct->_conf->client_snapdir &&
+      dir->snapid == CEPH_NOSNAP) {
+    *target = open_snapdir(dir);
+    goto done;
+  }
+
+relookup:
+  if (dir->dir &&
+      dir->dir->dentries.count(dname)) {
+    dn = dir->dir->dentries[dname];
+
+    ldout(cct, 20) << __func__ << " have " << *dn << " from mds." << dn->lease_mds
+        << " ttl " << dn->lease_ttl << " seq " << dn->lease_seq << dendl;
+
+    if (!dn->inode || dn->inode->caps_issued_mask(mask, true)) {
+      if (_dentry_valid(dn)) {
+        // touch this mds's dir cap too, even though we don't _explicitly_ use it here, to
+        // make trim_caps() behave.
+        dir->try_touch_cap(dn->lease_mds);
+          goto hit_dn;
+      }
+      // dir shared caps?
+      if (dir->caps_issued_mask(CEPH_CAP_FILE_SHARED, true)) {
+	if (dn->cap_shared_gen == dir->shared_gen &&
+	    (!dn->inode || dn->inode->caps_issued_mask(mask, true)))
+	      goto hit_dn;
+	if (!dn->inode && (dir->flags & I_COMPLETE)) {
+	  ldout(cct, 10) << __func__ << " concluded ENOENT locally for "
+			 << *dir << " dn '" << dname << "'" << dendl;
+	  return -CEPHFS_ENOENT;
+	}
+      }
+    } else {
+      ldout(cct, 20) << " no cap on " << dn->inode->vino() << dendl;
+    }
+  } else {
+    // can we conclude ENOENT locally?
+    if (dir->caps_issued_mask(CEPH_CAP_FILE_SHARED, true) &&
+	(dir->flags & I_COMPLETE)) {
+      ldout(cct, 10) << __func__ << " concluded ENOENT locally for " << *dir << " dn '" << dname << "'" << dendl;
+      return -CEPHFS_ENOENT;
+    }
+  }
+
+  if (did_lookup_request) {
+    r = 0;
+    goto done;
+  }
+  r = _do_lookup(dir, dname, mask, target, perms);
+  did_lookup_request = true;
+  if (r == 0) {
+    /* complete lookup to get dentry for alternate_name */
+    goto relookup;
+  } else {
+    goto done;
+  }
+
+ hit_dn:
+  if (dn->inode) {
+    *target = dn->inode;
+    if (alternate_name)
+      *alternate_name = dn->alternate_name;
+  } else {
+    r = -CEPHFS_ENOENT;
+  }
+  touch_dn(dn);
+  goto done;
+
+ done:
+  if (r < 0)
+    ldout(cct, 10) << __func__ << " " << *dir << " " << dname << " = " << r << dendl;
+  else
+    ldout(cct, 10) << __func__ << " " << *dir << " " << dname << " = " << **target << dendl;
+  return r;
+}
+
+int Client::get_or_create(Inode *dir, const char* name,
+			  Dentry **pdn, bool expect_null)
+{
+  // lookup
+  ldout(cct, 20) << __func__ << " " << *dir << " name " << name << dendl;
+  dir->open_dir();
+  if (dir->dir->dentries.count(name)) {
+    Dentry *dn = dir->dir->dentries[name];
+    if (_dentry_valid(dn)) {
+      if (expect_null)
+        return -CEPHFS_EEXIST;
+    }
+    *pdn = dn;
+  } else {
+    // otherwise link up a new one
+    *pdn = link(dir->dir, name, NULL, NULL);
+  }
+
+  // success
+  return 0;
+}
+
+int Client::walk(std::string_view path, walk_dentry_result* wdr, const UserPerm& perms, bool followsym)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 10) << __func__ << ": " << path << dendl;
+
+  std::scoped_lock lock(client_lock);
+
+  return path_walk(path, wdr, perms, followsym);
+}
+
+int Client::path_walk(const filepath& origpath, InodeRef *end,
+		      const UserPerm& perms, bool followsym, int mask, InodeRef dirinode)
+{
+  walk_dentry_result wdr;
+  int rc = path_walk(origpath, &wdr, perms, followsym, mask, dirinode);
+  *end = std::move(wdr.in);
+  return rc;
+}
+
+int Client::path_walk(const filepath& origpath, walk_dentry_result* result, const UserPerm& perms,
+                      bool followsym, int mask, InodeRef dirinode)
+{
+  filepath path = origpath;
+  InodeRef cur;
+  std::string alternate_name;
+  if (origpath.absolute())
+    cur = root;
+  else if (!dirinode)
+    cur = cwd;
+  else {
+    cur = dirinode;
+  }
+  ceph_assert(cur);
+
+  ldout(cct, 20) << __func__ << " cur=" << *cur << dendl;
+  ldout(cct, 10) << __func__ << " " << path << dendl;
+
+  int symlinks = 0;
+
+  unsigned i=0;
+  while (i < path.depth() && cur) {
+    int caps = 0;
+    const string &dname = path[i];
+    ldout(cct, 10) << " " << i << " " << *cur << " " << dname << dendl;
+    ldout(cct, 20) << "  (path is " << path << ")" << dendl;
+    InodeRef next;
+    if (cct->_conf->client_permissions) {
+      int r = may_lookup(cur.get(), perms);
+      if (r < 0)
+	return r;
+      caps = CEPH_CAP_AUTH_SHARED;
+    }
+
+    /* Get extra requested caps on the last component */
+    if (i == (path.depth() - 1))
+      caps |= mask;
+    int r = _lookup(cur.get(), dname, caps, &next, perms, &alternate_name);
+    if (r < 0)
+      return r;
+    // only follow trailing symlink if followsym.  always follow
+    // 'directory' symlinks.
+    if (next && next->is_symlink()) {
+      symlinks++;
+      ldout(cct, 20) << " symlink count " << symlinks << ", value is '" << next->symlink << "'" << dendl;
+      if (symlinks > MAXSYMLINKS) {
+	return -CEPHFS_ELOOP;
+      }
+
+      if (i < path.depth() - 1) {
+	// dir symlink
+	// replace consumed components of path with symlink dir target
+	filepath resolved(next->symlink.c_str());
+	resolved.append(path.postfixpath(i + 1));
+	path = resolved;
+	i = 0;
+	if (next->symlink[0] == '/') {
+	  cur = root;
+	}
+	continue;
+      } else if (followsym) {
+	if (next->symlink[0] == '/') {
+	  path = next->symlink.c_str();
+	  i = 0;
+	  // reset position
+	  cur = root;
+	} else {
+	  filepath more(next->symlink.c_str());
+	  // we need to remove the symlink component from off of the path
+	  // before adding the target that the symlink points to.  remain
+	  // at the same position in the path.
+	  path.pop_dentry();
+	  path.append(more);
+	}
+	continue;
+      }
+    }
+    cur.swap(next);
+    i++;
+  }
+  if (!cur)
+    return -CEPHFS_ENOENT;
+  if (result) {
+    result->in = std::move(cur);
+    result->alternate_name = std::move(alternate_name);
+  }
+  return 0;
+}
+
+
+// namespace ops
+
+int Client::link(const char *relexisting, const char *relpath, const UserPerm& perm, std::string alternate_name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "link" << std::endl;
+  tout(cct) << relexisting << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  filepath existing(relexisting);
+
+  InodeRef in, dir;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(existing, &in, perm, true);
+  if (r < 0)
+    return r;
+  if (std::string(relpath) == "/") {
+    r = -CEPHFS_EEXIST;
+    return r;
+  }
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+
+  r = path_walk(path, &dir, perm, true);
+  if (r < 0)
+    return r;
+  if (cct->_conf->client_permissions) {
+    if (S_ISDIR(in->mode)) {
+      r = -CEPHFS_EPERM;
+      return r;
+    }
+    r = may_hardlink(in.get(), perm);
+    if (r < 0)
+      return r;
+    r = may_create(dir.get(), perm);
+    if (r < 0)
+      return r;
+  }
+  r = _link(in.get(), dir.get(), name.c_str(), perm, std::move(alternate_name));
+  return r;
+}
+
+int Client::unlink(const char *relpath, const UserPerm& perm)
+{
+  return unlinkat(CEPHFS_AT_FDCWD, relpath, 0, perm);
+}
+
+int Client::unlinkat(int dirfd, const char *relpath, int flags, const UserPerm& perm)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << flags << std::endl;
+
+  if (std::string(relpath) == "/") {
+    return flags & AT_REMOVEDIR ? -CEPHFS_EBUSY : -CEPHFS_EISDIR;
+  }
+
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+  InodeRef dir;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef dirinode;
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  r = path_walk(path, &dir, perm, true, 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  if (cct->_conf->client_permissions) {
+    r = may_delete(dir.get(), name.c_str(), perm);
+    if (r < 0) {
+      return r;
+    }
+  }
+  if (flags & AT_REMOVEDIR) {
+    r = _rmdir(dir.get(), name.c_str(), perm);
+  } else {
+    r = _unlink(dir.get(), name.c_str(), perm);
+  }
+  return r;
+}
+
+int Client::rename(const char *relfrom, const char *relto, const UserPerm& perm, std::string alternate_name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relfrom << std::endl;
+  tout(cct) << relto << std::endl;
+
+  if (std::string(relfrom) == "/" || std::string(relto) == "/")
+    return -CEPHFS_EBUSY;
+
+  filepath from(relfrom);
+  filepath to(relto);
+  string fromname = from.last_dentry();
+  from.pop_dentry();
+  string toname = to.last_dentry();
+  to.pop_dentry();
+
+  InodeRef fromdir, todir;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(from, &fromdir, perm);
+  if (r < 0)
+    goto out;
+  r = path_walk(to, &todir, perm);
+  if (r < 0)
+    goto out;
+
+  if (cct->_conf->client_permissions) {
+    int r = may_delete(fromdir.get(), fromname.c_str(), perm);
+    if (r < 0)
+      return r;
+    r = may_delete(todir.get(), toname.c_str(), perm);
+    if (r < 0 && r != -CEPHFS_ENOENT)
+      return r;
+  }
+  r = _rename(fromdir.get(), fromname.c_str(), todir.get(), toname.c_str(), perm, std::move(alternate_name));
+out:
+  return r;
+}
+
+// dirs
+
+int Client::mkdir(const char *relpath, mode_t mode, const UserPerm& perm, std::string alternate_name)
+{
+  return mkdirat(CEPHFS_AT_FDCWD, relpath, mode, perm, alternate_name);
+}
+
+int Client::mkdirat(int dirfd, const char *relpath, mode_t mode, const UserPerm& perm,
+                    std::string alternate_name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mode << std::endl;
+  ldout(cct, 10) << __func__ << ": " << relpath << dendl;
+
+  if (std::string(relpath) == "/") {
+    return -CEPHFS_EEXIST;
+  }
+
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+  InodeRef dir;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef dirinode;
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  r = path_walk(path, &dir, perm, true, 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  if (cct->_conf->client_permissions) {
+    r = may_create(dir.get(), perm);
+    if (r < 0) {
+      return r;
+    }
+  }
+  return _mkdir(dir.get(), name.c_str(), mode, perm, 0, {}, std::move(alternate_name));
+}
+
+int Client::mkdirs(const char *relpath, mode_t mode, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 10) << "Client::mkdirs " << relpath << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mode << std::endl;
+
+  //get through existing parts of path
+  filepath path(relpath);
+  unsigned int i;
+  int r = 0, caps = 0;
+  InodeRef cur, next;
+
+  std::scoped_lock lock(client_lock);
+  cur = cwd;
+  for (i=0; i<path.depth(); ++i) {
+    if (cct->_conf->client_permissions) {
+      r = may_lookup(cur.get(), perms);
+      if (r < 0)
+	break;
+      caps = CEPH_CAP_AUTH_SHARED;
+    }
+    r = _lookup(cur.get(), path[i].c_str(), caps, &next, perms);
+    if (r < 0)
+      break;
+    cur.swap(next);
+  }
+  if (r!=-CEPHFS_ENOENT) return r;
+  ldout(cct, 20) << __func__ << " got through " << i << " directories on path " << relpath << dendl;
+  //make new directory at each level
+  for (; i<path.depth(); ++i) {
+    if (cct->_conf->client_permissions) {
+      r = may_create(cur.get(), perms);
+      if (r < 0)
+	return r;
+    }
+    //make new dir
+    r = _mkdir(cur.get(), path[i].c_str(), mode, perms, &next);
+    
+    //check proper creation/existence
+    if(-CEPHFS_EEXIST == r && i < path.depth() - 1) {
+      r = _lookup(cur.get(), path[i].c_str(), CEPH_CAP_AUTH_SHARED, &next, perms);
+    }	
+    if (r < 0) 
+      return r;
+    //move to new dir and continue
+    cur.swap(next);
+    ldout(cct, 20) << __func__ << ": successfully created directory "
+		   << filepath(cur->ino).get_path() << dendl;
+  }
+  return 0;
+}
+
+int Client::rmdir(const char *relpath, const UserPerm& perms)
+{
+  return unlinkat(CEPHFS_AT_FDCWD, relpath, AT_REMOVEDIR, perms);
+}
+
+int Client::mknod(const char *relpath, mode_t mode, const UserPerm& perms, dev_t rdev) 
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mode << std::endl;
+  tout(cct) << rdev << std::endl;
+
+  if (std::string(relpath) == "/")
+    return -CEPHFS_EEXIST;
+
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+  InodeRef dir;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &dir, perms);
+  if (r < 0)
+    return r;
+  if (cct->_conf->client_permissions) {
+    int r = may_create(dir.get(), perms);
+    if (r < 0)
+      return r;
+  }
+  return _mknod(dir.get(), name.c_str(), mode, rdev, perms);
+}
+
+// symlinks
+  
+int Client::symlink(const char *target, const char *relpath, const UserPerm& perms, std::string alternate_name)
+{
+  return symlinkat(target, CEPHFS_AT_FDCWD, relpath, perms, alternate_name);
+}
+
+int Client::symlinkat(const char *target, int dirfd, const char *relpath, const UserPerm& perms,
+                      std::string alternate_name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << target << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  if (std::string(relpath) == "/") {
+    return -CEPHFS_EEXIST;
+  }
+
+  filepath path(relpath);
+  string name = path.last_dentry();
+  path.pop_dentry();
+  InodeRef dir;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef dirinode;
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+  r = path_walk(path, &dir, perms, true, 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  if (cct->_conf->client_permissions) {
+    int r = may_create(dir.get(), perms);
+    if (r < 0) {
+      return r;
+    }
+  }
+  return _symlink(dir.get(), name.c_str(), target, perms, std::move(alternate_name));
+}
+
+int Client::readlink(const char *relpath, char *buf, loff_t size, const UserPerm& perms)
+{
+  return readlinkat(CEPHFS_AT_FDCWD, relpath, buf, size, perms);
+}
+
+int Client::readlinkat(int dirfd, const char *relpath, char *buf, loff_t size, const UserPerm& perms) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  InodeRef dirinode;
+  std::scoped_lock lock(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  InodeRef in;
+  filepath path(relpath);
+  r = path_walk(path, &in, perms, false, 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  return _readlink(in.get(), buf, size);
+}
+
+int Client::_readlink(Inode *in, char *buf, size_t size)
+{
+  if (!in->is_symlink())
+    return -CEPHFS_EINVAL;
+
+  // copy into buf (at most size bytes)
+  int r = in->symlink.length();
+  if (r > (int)size)
+    r = size;
+  memcpy(buf, in->symlink.c_str(), r);
+  return r;
+}
+
+
+// inode stuff
+
+int Client::_getattr(Inode *in, int mask, const UserPerm& perms, bool force)
+{
+  bool yes = in->caps_issued_mask(mask, true);
+
+  ldout(cct, 10) << __func__ << " mask " << ccap_string(mask) << " issued=" << yes << dendl;
+  if (yes && !force)
+    return 0;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_GETATTR);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_inode(in);
+  req->head.args.getattr.mask = mask;
+  
+  int res = make_request(req, perms);
+  ldout(cct, 10) << __func__ << " result=" << res << dendl;
+  return res;
+}
+
+int Client::_getvxattr(
+  Inode *in,
+  const UserPerm& perms,
+  const char *xattr_name,
+  ssize_t size,
+  void *value,
+  mds_rank_t rank)
+{
+  if (!xattr_name || strlen(xattr_name) <= 0 || strlen(xattr_name) > 255) {
+    return -CEPHFS_ENODATA;
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_GETVXATTR);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_inode(in);
+  req->set_string2(xattr_name);
+
+  bufferlist bl;
+  int res = make_request(req, perms, nullptr, nullptr, rank, &bl,
+                         CEPHFS_FEATURE_OP_GETVXATTR);
+  ldout(cct, 10) << __func__ << " result=" << res << dendl;
+
+  if (res < 0) {
+    if (res == -CEPHFS_EOPNOTSUPP) {
+      return -CEPHFS_ENODATA;
+    }
+    return res;
+  }
+
+  std::string buf;
+  auto p = bl.cbegin();
+
+  DECODE_START(1, p);
+  decode(buf, p);
+  DECODE_FINISH(p);
+
+  ssize_t len = buf.length();
+
+  res = len; // refer to man getxattr(2) for output buffer size == 0
+
+  if (size > 0) {
+    if (len > size) {
+      res = -CEPHFS_ERANGE; // insufficient output buffer space
+    } else {
+      memcpy(value, buf.c_str(), len);
+    }
+  }
+  return res;
+}
+
+int Client::_do_setattr(Inode *in, struct ceph_statx *stx, int mask,
+			const UserPerm& perms, InodeRef *inp)
+{
+  int issued = in->caps_issued();
+  union ceph_mds_request_args args;
+  bool kill_sguid = false;
+  int inode_drop = 0;
+
+  ldout(cct, 10) << __func__ << " mask " << mask << " issued " <<
+    ccap_string(issued) << dendl;
+
+  if (in->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+  if ((mask & CEPH_SETATTR_SIZE) &&
+      (uint64_t)stx->stx_size > in->size &&
+      is_quota_bytes_exceeded(in, (uint64_t)stx->stx_size - in->size,
+			      perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  memset(&args, 0, sizeof(args));
+
+  // make the change locally?
+  if ((in->cap_dirtier_uid >= 0 && perms.uid() != in->cap_dirtier_uid) ||
+      (in->cap_dirtier_gid >= 0 && perms.gid() != in->cap_dirtier_gid)) {
+    ldout(cct, 10) << __func__ << " caller " << perms.uid() << ":" << perms.gid()
+		   << " != cap dirtier " << in->cap_dirtier_uid << ":"
+		   << in->cap_dirtier_gid << ", forcing sync setattr"
+		   << dendl;
+    /*
+     * This works because we implicitly flush the caps as part of the
+     * request, so the cap update check will happen with the writeback
+     * cap context, and then the setattr check will happen with the
+     * caller's context.
+     *
+     * In reality this pattern is likely pretty rare (different users
+     * setattr'ing the same file).  If that turns out not to be the
+     * case later, we can build a more complex pipelined cap writeback
+     * infrastructure...
+     */
+    mask |= CEPH_SETATTR_CTIME;
+  }
+
+  if (!mask) {
+    // caller just needs us to bump the ctime
+    in->ctime = ceph_clock_now();
+    in->cap_dirtier_uid = perms.uid();
+    in->cap_dirtier_gid = perms.gid();
+    if (issued & CEPH_CAP_AUTH_EXCL)
+      in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+    else if (issued & CEPH_CAP_FILE_EXCL)
+      in->mark_caps_dirty(CEPH_CAP_FILE_EXCL);
+    else if (issued & CEPH_CAP_XATTR_EXCL)
+      in->mark_caps_dirty(CEPH_CAP_XATTR_EXCL);
+    else
+      mask |= CEPH_SETATTR_CTIME;
+  }
+
+  if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL)) {
+    kill_sguid = mask & (CEPH_SETATTR_SIZE|CEPH_SETATTR_KILL_SGUID);
+
+    mask &= ~CEPH_SETATTR_KILL_SGUID;
+  } else if (mask & CEPH_SETATTR_SIZE) {
+    /* If we don't have Ax, then we must ask the server to clear them on truncate */
+    mask |= CEPH_SETATTR_KILL_SGUID;
+    inode_drop |= CEPH_CAP_AUTH_SHARED;
+  }
+
+  if (mask & CEPH_SETATTR_UID) {
+    ldout(cct,10) << "changing uid to " << stx->stx_uid << dendl;
+
+    if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL)) {
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->uid = stx->stx_uid;
+      in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+      mask &= ~CEPH_SETATTR_UID;
+      kill_sguid = true;
+    } else if (!in->caps_issued_mask(CEPH_CAP_AUTH_SHARED) ||
+               in->uid != stx->stx_uid) {
+      args.setattr.uid = stx->stx_uid;
+      inode_drop |= CEPH_CAP_AUTH_SHARED;
+    } else {
+      mask &= ~CEPH_SETATTR_UID;
+    }
+  }
+
+  if (mask & CEPH_SETATTR_GID) {
+    ldout(cct,10) << "changing gid to " << stx->stx_gid << dendl;
+
+    if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL)) {
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->gid = stx->stx_gid;
+      in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+      mask &= ~CEPH_SETATTR_GID;
+      kill_sguid = true;
+    } else if (!in->caps_issued_mask(CEPH_CAP_AUTH_SHARED) ||
+               in->gid != stx->stx_gid) {
+      args.setattr.gid = stx->stx_gid;
+      inode_drop |= CEPH_CAP_AUTH_SHARED;
+    } else {
+      mask &= ~CEPH_SETATTR_GID;
+    }
+  }
+
+  if (mask & CEPH_SETATTR_MODE) {
+    ldout(cct,10) << "changing mode to " << stx->stx_mode << dendl;
+
+    if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL)) {
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->mode = (in->mode & ~07777) | (stx->stx_mode & 07777);
+      in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+      mask &= ~CEPH_SETATTR_MODE;
+    } else if (!in->caps_issued_mask(CEPH_CAP_AUTH_SHARED) ||
+               in->mode != stx->stx_mode) {
+      args.setattr.mode = stx->stx_mode;
+      inode_drop |= CEPH_CAP_AUTH_SHARED;
+    } else {
+      mask &= ~CEPH_SETATTR_MODE;
+    }
+  } else if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL) &&
+             kill_sguid && S_ISREG(in->mode) &&
+	     (in->mode & (S_IXUSR|S_IXGRP|S_IXOTH))) {
+    /* Must squash the any setuid/setgid bits with an ownership change */
+    in->mode &= ~(S_ISUID|S_ISGID);
+    in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+  }
+
+  if (mask & CEPH_SETATTR_BTIME) {
+    ldout(cct,10) << "changing btime to " << in->btime << dendl;
+
+    if (in->caps_issued_mask(CEPH_CAP_AUTH_EXCL)) {
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->btime = utime_t(stx->stx_btime);
+      in->mark_caps_dirty(CEPH_CAP_AUTH_EXCL);
+      mask &= ~CEPH_SETATTR_BTIME;
+    } else if (!in->caps_issued_mask(CEPH_CAP_AUTH_SHARED) ||
+               in->btime != utime_t(stx->stx_btime)) {
+      args.setattr.btime = utime_t(stx->stx_btime);
+      inode_drop |= CEPH_CAP_AUTH_SHARED;
+    } else {
+      mask &= ~CEPH_SETATTR_BTIME;
+    }
+  }
+
+  if (mask & CEPH_SETATTR_SIZE) {
+    if ((uint64_t)stx->stx_size >= mdsmap->get_max_filesize()) {
+      //too big!
+      ldout(cct,10) << "unable to set size to " << stx->stx_size << ". Too large!" << dendl;
+      return -CEPHFS_EFBIG;
+    }
+
+    ldout(cct,10) << "changing size to " << stx->stx_size << dendl;
+    if (in->caps_issued_mask(CEPH_CAP_FILE_EXCL) &&
+        !(mask & CEPH_SETATTR_KILL_SGUID) &&
+        stx->stx_size >= in->size) {
+      if (stx->stx_size > in->size) {
+        in->size = in->reported_size = stx->stx_size;
+        in->cap_dirtier_uid = perms.uid();
+        in->cap_dirtier_gid = perms.gid();
+        in->mark_caps_dirty(CEPH_CAP_FILE_EXCL);
+        mask &= ~(CEPH_SETATTR_SIZE);
+        mask |= CEPH_SETATTR_MTIME;
+      } else {
+        // ignore it when size doesn't change
+        mask &= ~(CEPH_SETATTR_SIZE);
+      }
+    } else {
+      args.setattr.size = stx->stx_size;
+      inode_drop |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
+                    CEPH_CAP_FILE_WR;
+    }
+  }
+
+  if (mask & CEPH_SETATTR_MTIME) {
+    if (in->caps_issued_mask(CEPH_CAP_FILE_EXCL)) {
+      in->mtime = utime_t(stx->stx_mtime);
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->time_warp_seq++;
+      in->mark_caps_dirty(CEPH_CAP_FILE_EXCL);
+      mask &= ~CEPH_SETATTR_MTIME;
+    } else if (in->caps_issued_mask(CEPH_CAP_FILE_WR) &&
+               utime_t(stx->stx_mtime) > in->mtime) {
+      in->mtime = utime_t(stx->stx_mtime);
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+      mask &= ~CEPH_SETATTR_MTIME;
+    } else if (!in->caps_issued_mask(CEPH_CAP_FILE_SHARED) ||
+	       in->mtime != utime_t(stx->stx_mtime)) {
+      args.setattr.mtime = utime_t(stx->stx_mtime);
+      inode_drop |= CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_RD |
+                    CEPH_CAP_FILE_WR;
+    } else {
+      mask &= ~CEPH_SETATTR_MTIME;
+    }
+  }
+
+  if (mask & CEPH_SETATTR_ATIME) {
+    if (in->caps_issued_mask(CEPH_CAP_FILE_EXCL)) {
+      in->atime = utime_t(stx->stx_atime);
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->time_warp_seq++;
+      in->mark_caps_dirty(CEPH_CAP_FILE_EXCL);
+      mask &= ~CEPH_SETATTR_ATIME;
+    } else if (in->caps_issued_mask(CEPH_CAP_FILE_WR) &&
+               utime_t(stx->stx_atime) > in->atime) {
+      in->atime = utime_t(stx->stx_atime);
+      in->ctime = ceph_clock_now();
+      in->cap_dirtier_uid = perms.uid();
+      in->cap_dirtier_gid = perms.gid();
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+      mask &= ~CEPH_SETATTR_ATIME;
+    } else if (!in->caps_issued_mask(CEPH_CAP_FILE_SHARED) ||
+	       in->atime != utime_t(stx->stx_atime)) {
+      args.setattr.atime = utime_t(stx->stx_atime);
+      inode_drop |= CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_RD |
+                    CEPH_CAP_FILE_WR;
+    } else {
+      mask &= ~CEPH_SETATTR_ATIME;
+    }
+  }
+
+  if (!mask) {
+    in->change_attr++;
+    if (in->is_dir() && in->snapid == CEPH_NOSNAP) {
+      vinodeno_t vino(in->ino, CEPH_SNAPDIR);
+      if (inode_map.count(vino)) {
+        refresh_snapdir_attrs(inode_map[vino], in);
+      }
+    }
+    return 0;
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_SETATTR);
+
+  filepath path;
+
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_inode(in);
+
+  req->head.args = args;
+  req->inode_drop = inode_drop;
+  req->head.args.setattr.mask = mask;
+  req->regetattr_mask = mask;
+
+  int res = make_request(req, perms, inp);
+  ldout(cct, 10) << "_setattr result=" << res << dendl;
+  return res;
+}
+
+/* Note that we only care about attrs that setattr cares about */
+void Client::stat_to_statx(struct stat *st, struct ceph_statx *stx)
+{
+  stx->stx_size = st->st_size;
+  stx->stx_mode = st->st_mode;
+  stx->stx_uid = st->st_uid;
+  stx->stx_gid = st->st_gid;
+#ifdef __APPLE__
+  stx->stx_mtime = st->st_mtimespec;
+  stx->stx_atime = st->st_atimespec;
+#elif __WIN32
+  stx->stx_mtime.tv_sec = st->st_mtime;
+  stx->stx_atime.tv_sec = st->st_atime;
+#else
+  stx->stx_mtime = st->st_mtim;
+  stx->stx_atime = st->st_atim;
+#endif
+}
+
+int Client::__setattrx(Inode *in, struct ceph_statx *stx, int mask,
+		       const UserPerm& perms, InodeRef *inp)
+{
+  int ret = _do_setattr(in, stx, mask, perms, inp);
+  if (ret < 0)
+   return ret;
+  if (mask & CEPH_SETATTR_MODE)
+    ret = _posix_acl_chmod(in, stx->stx_mode, perms);
+  return ret;
+}
+
+int Client::_setattrx(InodeRef &in, struct ceph_statx *stx, int mask,
+		      const UserPerm& perms)
+{
+  mask &= (CEPH_SETATTR_MODE | CEPH_SETATTR_UID |
+	   CEPH_SETATTR_GID | CEPH_SETATTR_MTIME |
+	   CEPH_SETATTR_ATIME | CEPH_SETATTR_SIZE |
+	   CEPH_SETATTR_CTIME | CEPH_SETATTR_BTIME);
+  if (cct->_conf->client_permissions) {
+    int r = may_setattr(in.get(), stx, mask, perms);
+    if (r < 0)
+      return r;
+  }
+  return __setattrx(in.get(), stx, mask, perms);
+}
+
+int Client::_setattr(InodeRef &in, struct stat *attr, int mask,
+		     const UserPerm& perms)
+{
+  struct ceph_statx stx;
+
+  stat_to_statx(attr, &stx);
+  mask &= ~CEPH_SETATTR_BTIME;
+
+  if ((mask & CEPH_SETATTR_UID) && attr->st_uid == static_cast<uid_t>(-1)) {
+    mask &= ~CEPH_SETATTR_UID;
+  }
+  if ((mask & CEPH_SETATTR_GID) && attr->st_gid == static_cast<uid_t>(-1)) {
+    mask &= ~CEPH_SETATTR_GID;
+  }
+
+  return _setattrx(in, &stx, mask, perms);
+}
+
+int Client::setattr(const char *relpath, struct stat *attr, int mask,
+		    const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mask  << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms);
+  if (r < 0)
+    return r;
+  return _setattr(in, attr, mask, perms);
+}
+
+int Client::setattrx(const char *relpath, struct ceph_statx *stx, int mask,
+		     const UserPerm& perms, int flags)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mask  << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW));
+  if (r < 0)
+    return r;
+  return _setattrx(in, stx, mask, perms);
+}
+
+int Client::fsetattr(int fd, struct stat *attr, int mask, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << mask  << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  return _setattr(f->inode, attr, mask, perms);
+}
+
+int Client::fsetattrx(int fd, struct ceph_statx *stx, int mask, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << mask  << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  return _setattrx(f->inode, stx, mask, perms);
+}
+
+int Client::stat(const char *relpath, struct stat *stbuf, const UserPerm& perms,
+		 frag_info_t *dirstat, int mask)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << " enter (relpath " << relpath << " mask " << mask << ")" << dendl;
+  tout(cct) << "stat" << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms, true, mask);
+  if (r < 0)
+    return r;
+  r = _getattr(in, mask, perms);
+  if (r < 0) {
+    ldout(cct, 3) << __func__ << " exit on error!" << dendl;
+    return r;
+  }
+  fill_stat(in, stbuf, dirstat);
+  ldout(cct, 3) << __func__ << " exit (relpath " << relpath << " mask " << mask << ")" << dendl;
+  return r;
+}
+
+unsigned Client::statx_to_mask(unsigned int flags, unsigned int want)
+{
+  unsigned mask = 0;
+
+  /* The AT_STATX_FORCE_SYNC is always in higher priority than AT_STATX_DONT_SYNC. */
+  if ((flags & AT_STATX_SYNC_TYPE) == AT_STATX_DONT_SYNC)
+    goto out;
+
+  /* Always set PIN to distinguish from AT_STATX_DONT_SYNC case */
+  mask |= CEPH_CAP_PIN;
+  if (want & (CEPH_STATX_MODE|CEPH_STATX_UID|CEPH_STATX_GID|CEPH_STATX_BTIME|CEPH_STATX_CTIME|CEPH_STATX_VERSION))
+    mask |= CEPH_CAP_AUTH_SHARED;
+  if (want & (CEPH_STATX_NLINK|CEPH_STATX_CTIME|CEPH_STATX_VERSION))
+    mask |= CEPH_CAP_LINK_SHARED;
+  if (want & (CEPH_STATX_NLINK|CEPH_STATX_ATIME|CEPH_STATX_MTIME|CEPH_STATX_CTIME|CEPH_STATX_SIZE|CEPH_STATX_BLOCKS|CEPH_STATX_VERSION))
+    mask |= CEPH_CAP_FILE_SHARED;
+  if (want & (CEPH_STATX_VERSION|CEPH_STATX_CTIME))
+    mask |= CEPH_CAP_XATTR_SHARED;
+out:
+  return mask;
+}
+
+int Client::statx(const char *relpath, struct ceph_statx *stx,
+		  const UserPerm& perms,
+		  unsigned int want, unsigned int flags)
+{
+  return statxat(CEPHFS_AT_FDCWD, relpath, stx, perms, want, flags);
+}
+
+int Client::lstat(const char *relpath, struct stat *stbuf,
+		  const UserPerm& perms, frag_info_t *dirstat, int mask)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << " enter (relpath " << relpath << " mask " << mask << ")" << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  // don't follow symlinks
+  int r = path_walk(path, &in, perms, false, mask);
+  if (r < 0)
+    return r;
+  r = _getattr(in, mask, perms);
+  if (r < 0) {
+    ldout(cct, 3) << __func__ << " exit on error!" << dendl;
+    return r;
+  }
+  fill_stat(in, stbuf, dirstat);
+  ldout(cct, 3) << __func__ << " exit (relpath " << relpath << " mask " << mask << ")" << dendl;
+  return r;
+}
+
+int Client::fill_stat(Inode *in, struct stat *st, frag_info_t *dirstat, nest_info_t *rstat)
+{
+  ldout(cct, 10) << __func__ << " on " << in->ino << " snap/dev" << in->snapid
+	   << " mode 0" << oct << in->mode << dec
+	   << " mtime " << in->mtime << " ctime " << in->ctime << dendl;
+  memset(st, 0, sizeof(struct stat));
+  if (use_faked_inos())
+    st->st_ino = in->faked_ino;
+  else
+    st->st_ino = in->ino;
+  st->st_dev = in->snapid;
+  st->st_mode = in->mode;
+  st->st_rdev = in->rdev;
+  if (in->is_dir()) {
+    switch (in->nlink) {
+      case 0:
+        st->st_nlink = 0; /* dir is unlinked */
+        break;
+      case 1:
+        st->st_nlink = 1 /* parent dentry */
+                       + 1 /* <dir>/. */
+                       + in->dirstat.nsubdirs; /* include <dir>/. self-reference */
+        break;
+      default:
+        ceph_abort();
+    }
+  } else {
+    st->st_nlink = in->nlink;
+  }
+  st->st_uid = in->uid;
+  st->st_gid = in->gid;
+  if (in->ctime > in->mtime) {
+    stat_set_ctime_sec(st, in->ctime.sec());
+    stat_set_ctime_nsec(st, in->ctime.nsec());
+  } else {
+    stat_set_ctime_sec(st, in->mtime.sec());
+    stat_set_ctime_nsec(st, in->mtime.nsec());
+  }
+  stat_set_atime_sec(st, in->atime.sec());
+  stat_set_atime_nsec(st, in->atime.nsec());
+  stat_set_mtime_sec(st, in->mtime.sec());
+  stat_set_mtime_nsec(st, in->mtime.nsec());
+  if (in->is_dir()) {
+    if (cct->_conf->client_dirsize_rbytes) {
+      st->st_size = in->rstat.rbytes;
+    } else if (in->snapid == CEPH_SNAPDIR) {
+      SnapRealm *realm = get_snap_realm_maybe(in->vino().ino);
+      if (realm) {
+        st->st_size = realm->my_snaps.size();
+        put_snap_realm(realm);
+      }
+    } else {
+      st->st_size = in->dirstat.size();
+    }
+// The Windows "stat" structure provides just a subset of the fields that are
+// available on Linux.
+#ifndef _WIN32
+    st->st_blocks = 1;
+#endif
+  } else {
+    st->st_size = in->size;
+#ifndef _WIN32
+    st->st_blocks = (in->size + 511) >> 9;
+#endif
+  }
+#ifndef _WIN32
+  st->st_blksize = std::max<uint32_t>(in->layout.stripe_unit, 4096);
+#endif
+
+  if (dirstat)
+    *dirstat = in->dirstat;
+  if (rstat)
+    *rstat = in->rstat;
+
+  return in->caps_issued();
+}
+
+void Client::fill_statx(Inode *in, unsigned int mask, struct ceph_statx *stx)
+{
+  ldout(cct, 10) << __func__ << " on " << in->ino << " snap/dev" << in->snapid
+	   << " mode 0" << oct << in->mode << dec
+	   << " mtime " << in->mtime << " ctime " << in->ctime << " change_attr " << in->change_attr << dendl;
+  memset(stx, 0, sizeof(struct ceph_statx));
+
+  /*
+   * If mask is 0, then the caller set AT_STATX_DONT_SYNC. Reset the mask
+   * so that all bits are set.
+   */
+  if (!mask)
+    mask = ~0;
+
+  /* These are always considered to be available */
+  stx->stx_dev = in->snapid;
+  stx->stx_blksize = std::max<uint32_t>(in->layout.stripe_unit, 4096);
+
+  /* Type bits are always set, even when CEPH_STATX_MODE is not */
+  stx->stx_mode = S_IFMT & in->mode;
+  stx->stx_ino = use_faked_inos() ? in->faked_ino : (ino_t)in->ino;
+  stx->stx_rdev = in->rdev;
+  stx->stx_mask |= (CEPH_STATX_INO|CEPH_STATX_RDEV);
+
+  if (mask & CEPH_CAP_AUTH_SHARED) {
+    stx->stx_uid = in->uid;
+    stx->stx_gid = in->gid;
+    stx->stx_mode = in->mode;
+    in->btime.to_timespec(&stx->stx_btime);
+    stx->stx_mask |= (CEPH_STATX_MODE|CEPH_STATX_UID|CEPH_STATX_GID|CEPH_STATX_BTIME);
+  }
+
+  if (mask & CEPH_CAP_LINK_SHARED) {
+    if (in->is_dir()) {
+      switch (in->nlink) {
+        case 0:
+          stx->stx_nlink = 0; /* dir is unlinked */
+          break;
+        case 1:
+          stx->stx_nlink = 1 /* parent dentry */
+                           + 1 /* <dir>/. */
+                           + in->dirstat.nsubdirs; /* include <dir>/. self-reference */
+          break;
+        default:
+          ceph_abort();
+      }
+    } else {
+      stx->stx_nlink = in->nlink;
+    }
+    stx->stx_mask |= CEPH_STATX_NLINK;
+  }
+
+  if (mask & CEPH_CAP_FILE_SHARED) {
+
+    in->atime.to_timespec(&stx->stx_atime);
+    in->mtime.to_timespec(&stx->stx_mtime);
+
+    if (in->is_dir()) {
+      if (cct->_conf->client_dirsize_rbytes) {
+	stx->stx_size = in->rstat.rbytes;
+      } else if (in->snapid == CEPH_SNAPDIR) {
+        SnapRealm *realm = get_snap_realm_maybe(in->vino().ino);
+	if (realm) {
+          stx->stx_size = realm->my_snaps.size();
+          put_snap_realm(realm);
+	}
+      } else {
+	stx->stx_size = in->dirstat.size();
+      }
+      stx->stx_blocks = 1;
+    } else {
+      stx->stx_size = in->size;
+      stx->stx_blocks = (in->size + 511) >> 9;
+    }
+    stx->stx_mask |= (CEPH_STATX_ATIME|CEPH_STATX_MTIME|
+		      CEPH_STATX_SIZE|CEPH_STATX_BLOCKS);
+  }
+
+  /* Change time and change_attr both require all shared caps to view */
+  if ((mask & CEPH_STAT_CAP_INODE_ALL) == CEPH_STAT_CAP_INODE_ALL) {
+    stx->stx_version = in->change_attr;
+    if (in->ctime > in->mtime)
+      in->ctime.to_timespec(&stx->stx_ctime);
+    else
+      in->mtime.to_timespec(&stx->stx_ctime);
+    stx->stx_mask |= (CEPH_STATX_CTIME|CEPH_STATX_VERSION);
+  }
+
+}
+
+void Client::touch_dn(Dentry *dn)
+{
+  lru.lru_touch(dn);
+}
+
+int Client::chmod(const char *relpath, mode_t mode, const UserPerm& perms)
+{
+  return chmodat(CEPHFS_AT_FDCWD, relpath, mode, 0, perms);
+}
+
+int Client::fchmod(int fd, mode_t mode, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << mode << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  struct stat attr;
+  attr.st_mode = mode;
+  return _setattr(f->inode, &attr, CEPH_SETATTR_MODE, perms);
+}
+
+int Client::chmodat(int dirfd, const char *relpath, mode_t mode, int flags,
+                    const UserPerm& perms) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << mode << std::endl;
+  tout(cct) << flags << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+  InodeRef dirinode;
+
+  std::scoped_lock lock(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  r = path_walk(path, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW), 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  struct stat attr;
+  attr.st_mode = mode;
+  return _setattr(in, &attr, CEPH_SETATTR_MODE, perms);
+}
+
+int Client::lchmod(const char *relpath, mode_t mode, const UserPerm& perms)
+{
+  return chmodat(CEPHFS_AT_FDCWD, relpath, mode, AT_SYMLINK_NOFOLLOW, perms);
+}
+
+int Client::chown(const char *relpath, uid_t new_uid, gid_t new_gid,
+		  const UserPerm& perms)
+{
+  return chownat(CEPHFS_AT_FDCWD, relpath, new_uid, new_gid, 0, perms);
+}
+
+int Client::fchown(int fd, uid_t new_uid, gid_t new_gid, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << new_uid << std::endl;
+  tout(cct) << new_gid << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  struct stat attr;
+  attr.st_uid = new_uid;
+  attr.st_gid = new_gid;
+  int mask = 0;
+  if (new_uid != static_cast<uid_t>(-1)) mask |= CEPH_SETATTR_UID;
+  if (new_gid != static_cast<gid_t>(-1)) mask |= CEPH_SETATTR_GID;
+  return _setattr(f->inode, &attr, mask, perms);
+}
+
+int Client::lchown(const char *relpath, uid_t new_uid, gid_t new_gid,
+		   const UserPerm& perms)
+{
+  return chownat(CEPHFS_AT_FDCWD, relpath, new_uid, new_gid, AT_SYMLINK_NOFOLLOW, perms);
+}
+
+int Client::chownat(int dirfd, const char *relpath, uid_t new_uid, gid_t new_gid,
+                    int flags, const UserPerm& perms) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << new_uid << std::endl;
+  tout(cct) << new_gid << std::endl;
+  tout(cct) << flags << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+  InodeRef dirinode;
+
+  std::scoped_lock lock(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  r = path_walk(path, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW), 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  struct stat attr;
+  attr.st_uid = new_uid;
+  attr.st_gid = new_gid;
+  return _setattr(in, &attr, CEPH_SETATTR_UID|CEPH_SETATTR_GID, perms);
+}
+
+static void attr_set_atime_and_mtime(struct stat *attr,
+                                     const utime_t &atime,
+                                     const utime_t &mtime)
+{
+  stat_set_atime_sec(attr, atime.tv.tv_sec);
+  stat_set_atime_nsec(attr, atime.tv.tv_nsec);
+  stat_set_mtime_sec(attr, mtime.tv.tv_sec);
+  stat_set_mtime_nsec(attr, mtime.tv.tv_nsec);
+}
+
+// for [l]utime() invoke the timeval variant as the timespec
+// variant are not yet implemented. for futime[s](), invoke
+// the timespec variant.
+int Client::utime(const char *relpath, struct utimbuf *buf,
+		  const UserPerm& perms)
+{
+  struct timeval tv[2];
+  tv[0].tv_sec  = buf->actime;
+  tv[0].tv_usec = 0;
+  tv[1].tv_sec  = buf->modtime;
+  tv[1].tv_usec = 0;
+
+  return utimes(relpath, tv, perms);
+}
+
+int Client::lutime(const char *relpath, struct utimbuf *buf,
+		   const UserPerm& perms)
+{
+  struct timeval tv[2];
+  tv[0].tv_sec  = buf->actime;
+  tv[0].tv_usec = 0;
+  tv[1].tv_sec  = buf->modtime;
+  tv[1].tv_usec = 0;
+
+  return lutimes(relpath, tv, perms);
+}
+
+int Client::futime(int fd, struct utimbuf *buf, const UserPerm& perms)
+{
+  struct timespec ts[2];
+  ts[0].tv_sec  = buf->actime;
+  ts[0].tv_nsec = 0;
+  ts[1].tv_sec  = buf->modtime;
+  ts[1].tv_nsec = 0;
+
+  return futimens(fd, ts, perms);
+}
+
+int Client::utimes(const char *relpath, struct timeval times[2],
+                   const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << "atime: " << times[0].tv_sec << "." << times[0].tv_usec
+            << std::endl;
+  tout(cct) << "mtime: " << times[1].tv_sec << "." << times[1].tv_usec
+            << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms);
+  if (r < 0)
+    return r;
+  struct stat attr;
+  utime_t atime(times[0]);
+  utime_t mtime(times[1]);
+
+  attr_set_atime_and_mtime(&attr, atime, mtime);
+  return _setattr(in, &attr, CEPH_SETATTR_MTIME|CEPH_SETATTR_ATIME, perms);
+}
+
+int Client::lutimes(const char *relpath, struct timeval times[2],
+                    const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << "atime: " << times[0].tv_sec << "." << times[0].tv_usec
+            << std::endl;
+  tout(cct) << "mtime: " << times[1].tv_sec << "." << times[1].tv_usec
+            << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms, false);
+  if (r < 0)
+    return r;
+  struct stat attr;
+  utime_t atime(times[0]);
+  utime_t mtime(times[1]);
+
+  attr_set_atime_and_mtime(&attr, atime, mtime);
+  return _setattr(in, &attr, CEPH_SETATTR_MTIME|CEPH_SETATTR_ATIME, perms);
+}
+
+int Client::futimes(int fd, struct timeval times[2], const UserPerm& perms)
+{
+  struct timespec ts[2];
+  ts[0].tv_sec  = times[0].tv_sec;
+  ts[0].tv_nsec = times[0].tv_usec * 1000;
+  ts[1].tv_sec  = times[1].tv_sec;
+  ts[1].tv_nsec = times[1].tv_usec * 1000;
+
+  return futimens(fd, ts, perms);
+}
+
+int Client::futimens(int fd, struct timespec times[2], const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << "atime: " << times[0].tv_sec << "." << times[0].tv_nsec
+            << std::endl;
+  tout(cct) << "mtime: " << times[1].tv_sec << "." << times[1].tv_nsec
+            << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  struct stat attr;
+  utime_t atime(times[0]);
+  utime_t mtime(times[1]);
+
+  attr_set_atime_and_mtime(&attr, atime, mtime);
+  return _setattr(f->inode, &attr, CEPH_SETATTR_MTIME|CEPH_SETATTR_ATIME, perms);
+}
+
+int Client::utimensat(int dirfd, const char *relpath, struct timespec times[2], int flags,
+                      const UserPerm& perms) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << "atime: " << times[0].tv_sec << "." << times[0].tv_nsec
+            << std::endl;
+  tout(cct) << "mtime: " << times[1].tv_sec << "." << times[1].tv_nsec
+            << std::endl;
+  tout(cct) << flags << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+  InodeRef dirinode;
+
+  std::scoped_lock lock(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+#if defined(__linux__) && defined(O_PATH)
+  if (flags & O_PATH) {
+    return -CEPHFS_EBADF;
+  }
+#endif
+
+  r = path_walk(path, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW), 0, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  struct stat attr;
+  utime_t atime(times[0]);
+  utime_t mtime(times[1]);
+
+  attr_set_atime_and_mtime(&attr, atime, mtime);
+  return _setattr(in, &attr, CEPH_SETATTR_MTIME|CEPH_SETATTR_ATIME, perms);
+}
+
+int Client::flock(int fd, int operation, uint64_t owner)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << operation << std::endl;
+  tout(cct) << owner << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+
+  return _flock(f, operation, owner);
+}
+
+int Client::opendir(const char *relpath, dir_result_t **dirpp, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms, true);
+  if (r < 0)
+    return r;
+  if (cct->_conf->client_permissions) {
+    int r = may_open(in.get(), O_RDONLY, perms);
+    if (r < 0)
+      return r;
+  }
+  r = _opendir(in.get(), dirpp, perms);
+  /* if ENOTDIR, dirpp will be an uninitialized point and it's very dangerous to access its value */
+  if (r != -CEPHFS_ENOTDIR)
+      tout(cct) << (uintptr_t)*dirpp << std::endl;
+  return r;
+}
+
+int Client::fdopendir(int dirfd, dir_result_t **dirpp, const UserPerm &perms) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << dirfd << std::endl;
+
+  InodeRef dirinode;
+  std::scoped_lock locker(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  if (cct->_conf->client_permissions) {
+    r = may_open(dirinode.get(), O_RDONLY, perms);
+    if (r < 0) {
+      return r;
+    }
+  }
+  r = _opendir(dirinode.get(), dirpp, perms);
+  /* if ENOTDIR, dirpp will be an uninitialized point and it's very dangerous to access its value */
+  if (r != -CEPHFS_ENOTDIR) {
+      tout(cct) << (uintptr_t)*dirpp << std::endl;
+  }
+  return r;
+}
+
+int Client::_opendir(Inode *in, dir_result_t **dirpp, const UserPerm& perms)
+{
+  if (!in->is_dir())
+    return -CEPHFS_ENOTDIR;
+  *dirpp = new dir_result_t(in, perms);
+  opened_dirs.insert(*dirpp);
+  ldout(cct, 8) << __func__ << "(" << in->ino << ") = " << 0 << " (" << *dirpp << ")" << dendl;
+  return 0;
+}
+
+
+int Client::closedir(dir_result_t *dir) 
+{
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << (uintptr_t)dir << std::endl;
+
+  ldout(cct, 3) << __func__ << "(" << dir << ") = 0" << dendl;
+  std::scoped_lock lock(client_lock);
+  _closedir(dir);
+  return 0;
+}
+
+void Client::_closedir(dir_result_t *dirp)
+{
+  ldout(cct, 10) << __func__ << "(" << dirp << ")" << dendl;
+
+  if (dirp->inode) {
+    ldout(cct, 10) << __func__ << " detaching inode " << dirp->inode << dendl;
+    dirp->inode.reset();
+  }
+  _readdir_drop_dirp_buffer(dirp);
+  opened_dirs.erase(dirp);
+  delete dirp;
+}
+
+void Client::rewinddir(dir_result_t *dirp)
+{
+  ldout(cct, 3) << __func__ << "(" << dirp << ")" << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  std::scoped_lock lock(client_lock);
+  dir_result_t *d = static_cast<dir_result_t*>(dirp);
+  _readdir_drop_dirp_buffer(d);
+  d->reset();
+}
+ 
+loff_t Client::telldir(dir_result_t *dirp)
+{
+  dir_result_t *d = static_cast<dir_result_t*>(dirp);
+  ldout(cct, 3) << __func__ << "(" << dirp << ") = " << d->offset << dendl;
+  return d->offset;
+}
+
+void Client::seekdir(dir_result_t *dirp, loff_t offset)
+{
+  ldout(cct, 3) << __func__ << "(" << dirp << ", " << offset << ")" << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  std::scoped_lock lock(client_lock);
+
+  if (offset == dirp->offset)
+    return;
+
+  if (offset > dirp->offset)
+    dirp->release_count = 0;   // bump if we do a forward seek
+  else
+    dirp->ordered_count = 0;   // disable filling readdir cache
+
+  if (dirp->hash_order()) {
+    if (dirp->offset > offset) {
+      _readdir_drop_dirp_buffer(dirp);
+      dirp->reset();
+    }
+  } else {
+    if (offset == 0 ||
+	dirp->buffer_frag != frag_t(dir_result_t::fpos_high(offset)) ||
+	dirp->offset_low() > dir_result_t::fpos_low(offset))  {
+      _readdir_drop_dirp_buffer(dirp);
+      dirp->reset();
+    }
+  }
+
+  dirp->offset = offset;
+}
+
+
+//struct dirent {
+//  ino_t          d_ino;       /* inode number */
+//  off_t          d_off;       /* offset to the next dirent */
+//  unsigned short d_reclen;    /* length of this record */
+//  unsigned char  d_type;      /* type of file */
+//  char           d_name[256]; /* filename */
+//};
+void Client::fill_dirent(struct dirent *de, const char *name, int type, uint64_t ino, loff_t next_off)
+{
+  strncpy(de->d_name, name, 255);
+  de->d_name[255] = '\0';
+#if !defined(__CYGWIN__) && !(defined(_WIN32))
+  de->d_ino = ino;
+#if !defined(__APPLE__) && !defined(__FreeBSD__)
+  de->d_off = next_off;
+#endif
+  de->d_reclen = 1;
+  de->d_type = IFTODT(type);
+  ldout(cct, 10) << __func__ << " '" << de->d_name << "' -> " << inodeno_t(de->d_ino)
+	   << " type " << (int)de->d_type << " w/ next_off " << hex << next_off << dec << dendl;
+#endif
+}
+
+void Client::_readdir_next_frag(dir_result_t *dirp)
+{
+  frag_t fg = dirp->buffer_frag;
+
+  if (fg.is_rightmost()) {
+    ldout(cct, 10) << __func__ << " advance from " << fg << " to END" << dendl;
+    dirp->set_end();
+    return;
+  }
+
+  // advance
+  fg = fg.next();
+  ldout(cct, 10) << __func__ << " advance from " << dirp->buffer_frag << " to " << fg << dendl;
+
+  if (dirp->hash_order()) {
+    // keep last_name
+    int64_t new_offset = dir_result_t::make_fpos(fg.value(), 2, true);
+    if (dirp->offset < new_offset) // don't decrease offset
+      dirp->offset = new_offset;
+  } else {
+    dirp->last_name.clear();
+    dirp->offset = dir_result_t::make_fpos(fg, 2, false);
+    _readdir_rechoose_frag(dirp);
+  }
+}
+
+void Client::_readdir_rechoose_frag(dir_result_t *dirp)
+{
+  ceph_assert(dirp->inode);
+
+  if (dirp->hash_order())
+    return;
+
+  frag_t cur = frag_t(dirp->offset_high());
+  frag_t fg = dirp->inode->dirfragtree[cur.value()];
+  if (fg != cur) {
+    ldout(cct, 10) << __func__ << " frag " << cur << " maps to " << fg << dendl;
+    dirp->offset = dir_result_t::make_fpos(fg, 2, false);
+    dirp->last_name.clear();
+    dirp->next_offset = 2;
+  }
+}
+
+void Client::_readdir_drop_dirp_buffer(dir_result_t *dirp)
+{
+  ldout(cct, 10) << __func__ << " " << dirp << dendl;
+  dirp->buffer.clear();
+}
+
+int Client::_readdir_get_frag(dir_result_t *dirp)
+{
+  ceph_assert(dirp);
+  ceph_assert(dirp->inode);
+
+  // get the current frag.
+  frag_t fg;
+  if (dirp->hash_order())
+    fg = dirp->inode->dirfragtree[dirp->offset_high()];
+  else
+    fg = frag_t(dirp->offset_high());
+  
+  ldout(cct, 10) << __func__ << " " << dirp << " on " << dirp->inode->ino << " fg " << fg
+		 << " offset " << hex << dirp->offset << dec << dendl;
+
+  int op = CEPH_MDS_OP_READDIR;
+  if (dirp->inode && dirp->inode->snapid == CEPH_SNAPDIR)
+    op = CEPH_MDS_OP_LSSNAP;
+
+  InodeRef& diri = dirp->inode;
+
+  MetaRequest *req = new MetaRequest(op);
+  filepath path;
+  diri->make_nosnap_relative_path(path);
+  req->set_filepath(path); 
+  req->set_inode(diri.get());
+  req->head.args.readdir.frag = fg;
+  req->head.args.readdir.flags = CEPH_READDIR_REPLY_BITFLAGS;
+  if (dirp->last_name.length()) {
+    req->path2.set_path(dirp->last_name);
+  } else if (dirp->hash_order()) {
+    req->head.args.readdir.offset_hash = dirp->offset_high();
+  }
+  req->dirp = dirp;
+  
+  bufferlist dirbl;
+  int res = make_request(req, dirp->perms, NULL, NULL, -1, &dirbl);
+  
+  if (res == -CEPHFS_EAGAIN) {
+    ldout(cct, 10) << __func__ << " got EAGAIN, retrying" << dendl;
+    _readdir_rechoose_frag(dirp);
+    return _readdir_get_frag(dirp);
+  }
+
+  if (res == 0) {
+    ldout(cct, 10) << __func__ << " " << dirp << " got frag " << dirp->buffer_frag
+		   << " size " << dirp->buffer.size() << dendl;
+  } else {
+    ldout(cct, 10) << __func__ << " got error " << res << ", setting end flag" << dendl;
+    dirp->set_end();
+  }
+
+  return res;
+}
+
+struct dentry_off_lt {
+  bool operator()(const Dentry* dn, int64_t off) const {
+    return dir_result_t::fpos_cmp(dn->offset, off) < 0;
+  }
+};
+
+int Client::_readdir_cache_cb(dir_result_t *dirp, add_dirent_cb_t cb, void *p,
+			      int caps, bool getref)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+  ldout(cct, 10) << __func__ << " " << dirp << " on " << dirp->inode->ino
+	   << " last_name " << dirp->last_name << " offset " << hex << dirp->offset << dec
+	   << dendl;
+  Dir *dir = dirp->inode->dir;
+
+  if (!dir) {
+    ldout(cct, 10) << " dir is empty" << dendl;
+    dirp->set_end();
+    return 0;
+  }
+
+  vector<Dentry*>::iterator pd = std::lower_bound(dir->readdir_cache.begin(),
+						  dir->readdir_cache.end(),
+						  dirp->offset, dentry_off_lt());
+
+  string dn_name;
+  while (true) {
+    int mask = caps;
+    if (!dirp->inode->is_complete_and_ordered())
+      return -CEPHFS_EAGAIN;
+    if (pd == dir->readdir_cache.end())
+      break;
+    Dentry *dn = *pd;
+    if (dn->inode == NULL) {
+      ldout(cct, 15) << " skipping null '" << dn->name << "'" << dendl;
+      ++pd;
+      continue;
+    }
+    if (dn->cap_shared_gen != dir->parent_inode->shared_gen) {
+      ldout(cct, 15) << " skipping mismatch shared gen '" << dn->name << "'" << dendl;
+      ++pd;
+      continue;
+    }
+
+    int idx = pd - dir->readdir_cache.begin();
+    if (dn->inode->is_dir()) {
+      mask |= CEPH_STAT_RSTAT;
+    }
+    int r = _getattr(dn->inode, mask, dirp->perms);
+    if (r < 0)
+      return r;
+    
+    // the content of readdir_cache may change after _getattr(), so pd may be invalid iterator    
+    pd = dir->readdir_cache.begin() + idx;
+    if (pd >= dir->readdir_cache.end() || *pd != dn)
+      return -CEPHFS_EAGAIN;
+
+    struct ceph_statx stx;
+    struct dirent de;
+    fill_statx(dn->inode, caps, &stx);
+
+    uint64_t next_off = dn->offset + 1;
+    fill_dirent(&de, dn->name.c_str(), stx.stx_mode, stx.stx_ino, next_off);
+    ++pd;
+    if (pd == dir->readdir_cache.end())
+      next_off = dir_result_t::END;
+
+    Inode *in = NULL;
+    if (getref) {
+      in = dn->inode.get();
+      _ll_get(in);
+    }
+
+    dn_name = dn->name; // fill in name while we have lock
+
+    client_lock.unlock();
+    r = cb(p, &de, &stx, next_off, in);  // _next_ offset
+    client_lock.lock();
+    ldout(cct, 15) << " de " << de.d_name << " off " << hex << dn->offset << dec
+		   << " = " << r << dendl;
+    if (r < 0) {
+      return r;
+    }
+
+    dirp->offset = next_off;
+    if (dirp->at_end())
+      dirp->next_offset = 2;
+    else
+      dirp->next_offset = dirp->offset_low();
+    dirp->last_name = dn_name; // we successfully returned this one; update!
+    dirp->release_count = 0; // last_name no longer match cache index
+    if (r > 0)
+      return r;
+  }
+
+  ldout(cct, 10) << __func__ << " " << dirp << " on " << dirp->inode->ino << " at end" << dendl;
+  dirp->set_end();
+  return 0;
+}
+
+int Client::readdir_r_cb(dir_result_t *d, add_dirent_cb_t cb, void *p,
+			 unsigned want, unsigned flags, bool getref)
+{
+  int caps = statx_to_mask(flags, want);
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::unique_lock cl(client_lock);
+
+  dir_result_t *dirp = static_cast<dir_result_t*>(d);
+
+  ldout(cct, 10) << __func__ << " " << *dirp->inode << " offset " << hex << dirp->offset
+		 << dec << " at_end=" << dirp->at_end()
+		 << " hash_order=" << dirp->hash_order() << dendl;
+
+  struct dirent de;
+  struct ceph_statx stx;
+  memset(&de, 0, sizeof(de));
+  memset(&stx, 0, sizeof(stx));
+
+  InodeRef& diri = dirp->inode;
+
+  if (dirp->at_end())
+    return 0;
+
+  if (dirp->offset == 0) {
+    ldout(cct, 15) << " including ." << dendl;
+    ceph_assert(diri->dentries.size() < 2); // can't have multiple hard-links to a dir
+    uint64_t next_off = 1;
+
+    int r;
+    r = _getattr(diri, caps | CEPH_STAT_RSTAT, dirp->perms);
+    if (r < 0)
+      return r;
+
+    fill_statx(diri, caps, &stx);
+    fill_dirent(&de, ".", S_IFDIR, stx.stx_ino, next_off);
+
+    Inode *inode = NULL;
+    if (getref) {
+      inode = diri.get();
+      _ll_get(inode);
+    }
+
+    cl.unlock();
+    r = cb(p, &de, &stx, next_off, inode);
+    cl.lock();
+    if (r < 0)
+      return r;
+
+    dirp->offset = next_off;
+    if (r > 0)
+      return r;
+  }
+  if (dirp->offset == 1) {
+    ldout(cct, 15) << " including .." << dendl;
+    uint64_t next_off = 2;
+    InodeRef in;
+    if (diri->dentries.empty())
+      in = diri;
+    else
+      in = diri->get_first_parent()->dir->parent_inode;
+
+    int r;
+    r = _getattr(in, caps | CEPH_STAT_RSTAT, dirp->perms);
+    if (r < 0)
+      return r;
+
+    fill_statx(in, caps, &stx);
+    fill_dirent(&de, "..", S_IFDIR, stx.stx_ino, next_off);
+
+    Inode *inode = NULL;
+    if (getref) {
+      inode = in.get();
+      _ll_get(inode);
+    }
+
+    cl.unlock();
+    r = cb(p, &de, &stx, next_off, inode);
+    cl.lock();
+    if (r < 0)
+      return r;
+
+    dirp->offset = next_off;
+    if (r > 0)
+      return r;
+  }
+
+  // can we read from our cache?
+  ldout(cct, 10) << "offset " << hex << dirp->offset << dec
+	   << " snapid " << dirp->inode->snapid << " (complete && ordered) "
+	   << dirp->inode->is_complete_and_ordered()
+	   << " issued " << ccap_string(dirp->inode->caps_issued())
+	   << dendl;
+  if (dirp->inode->snapid != CEPH_SNAPDIR &&
+      dirp->inode->is_complete_and_ordered() &&
+      dirp->inode->caps_issued_mask(CEPH_CAP_FILE_SHARED, true)) {
+    int err = _readdir_cache_cb(dirp, cb, p, caps, getref);
+    if (err != -CEPHFS_EAGAIN)
+      return err;
+  }
+
+  while (1) {
+    if (dirp->at_end())
+      return 0;
+
+    bool check_caps = true;
+    if (!dirp->is_cached()) {
+      int r = _readdir_get_frag(dirp);
+      if (r)
+	return r;
+      // _readdir_get_frag () may updates dirp->offset if the replied dirfrag is
+      // different than the requested one. (our dirfragtree was outdated)
+      check_caps = false;
+    }
+    frag_t fg = dirp->buffer_frag;
+
+    ldout(cct, 10) << "frag " << fg << " buffer size " << dirp->buffer.size()
+		   << " offset " << hex << dirp->offset << dendl;
+
+    for (auto it = std::lower_bound(dirp->buffer.begin(), dirp->buffer.end(),
+				    dirp->offset, dir_result_t::dentry_off_lt());
+	 it != dirp->buffer.end();
+	 ++it) {
+      dir_result_t::dentry &entry = *it;
+
+      uint64_t next_off = entry.offset + 1;
+
+      int r;
+      if (check_caps) {
+	int mask = caps;
+	if(entry.inode->is_dir()){
+          mask |= CEPH_STAT_RSTAT;
+	}
+	r = _getattr(entry.inode, mask, dirp->perms);
+	if (r < 0)
+	  return r;
+      }
+
+      fill_statx(entry.inode, caps, &stx);
+      fill_dirent(&de, entry.name.c_str(), stx.stx_mode, stx.stx_ino, next_off);
+
+      Inode *inode = NULL;
+      if (getref) {
+	inode = entry.inode.get();
+	_ll_get(inode);
+      }
+
+      cl.unlock();
+      r = cb(p, &de, &stx, next_off, inode);  // _next_ offset
+      cl.lock();
+
+      ldout(cct, 15) << " de " << de.d_name << " off " << hex << next_off - 1 << dec
+		     << " = " << r << dendl;
+      if (r < 0)
+	return r;
+
+      dirp->offset = next_off;
+      if (r > 0)
+	return r;
+    }
+
+    if (dirp->next_offset > 2) {
+      ldout(cct, 10) << " fetching next chunk of this frag" << dendl;
+      _readdir_drop_dirp_buffer(dirp);
+      continue;  // more!
+    }
+
+    if (!fg.is_rightmost()) {
+      // next frag!
+      _readdir_next_frag(dirp);
+      continue;
+    }
+
+    if (diri->shared_gen == dirp->start_shared_gen &&
+	diri->dir_release_count == dirp->release_count) {
+      if (diri->dir_ordered_count == dirp->ordered_count) {
+	ldout(cct, 10) << " marking (I_COMPLETE|I_DIR_ORDERED) on " << *diri << dendl;
+	if (diri->dir) {
+	  ceph_assert(diri->dir->readdir_cache.size() >= dirp->cache_index);
+	  diri->dir->readdir_cache.resize(dirp->cache_index);
+	}
+	diri->flags |= I_COMPLETE | I_DIR_ORDERED;
+      } else {
+	ldout(cct, 10) << " marking I_COMPLETE on " << *diri << dendl;
+	diri->flags |= I_COMPLETE;
+      }
+    }
+
+    dirp->set_end();
+    return 0;
+  }
+  ceph_abort();
+  return 0;
+}
+
+
+int Client::readdir_r(dir_result_t *d, struct dirent *de)
+{  
+  return readdirplus_r(d, de, 0, 0, 0, NULL);
+}
+
+/*
+ * readdirplus_r
+ *
+ * returns
+ *  1 if we got a dirent
+ *  0 for end of directory
+ * <0 on error
+ */
+
+struct single_readdir {
+  struct dirent *de;
+  struct ceph_statx *stx;
+  Inode *inode;
+  bool full;
+};
+
+static int _readdir_single_dirent_cb(void *p, struct dirent *de,
+				     struct ceph_statx *stx, off_t off,
+				     Inode *in)
+{
+  single_readdir *c = static_cast<single_readdir *>(p);
+
+  if (c->full)
+    return -1;  // already filled this dirent
+
+  *c->de = *de;
+  if (c->stx)
+    *c->stx = *stx;
+  c->inode = in;
+  c->full = true;
+  return 1;
+}
+
+struct dirent *Client::readdir(dir_result_t *d)
+{
+  int ret;
+  auto& de = d->de;
+  single_readdir sr;
+  sr.de = &de;
+  sr.stx = NULL;
+  sr.inode = NULL;
+  sr.full = false;
+
+  // our callback fills the dirent and sets sr.full=true on first
+  // call, and returns -1 the second time around.
+  ret = readdir_r_cb(d, _readdir_single_dirent_cb, (void *)&sr);
+  if (ret < -1) {
+    errno = -ret;  // this sucks.
+    return (dirent *) NULL;
+  }
+  if (sr.full) {
+    return &de;
+  }
+  return (dirent *) NULL;
+}
+
+int Client::readdirplus_r(dir_result_t *d, struct dirent *de,
+			  struct ceph_statx *stx, unsigned want,
+			  unsigned flags, Inode **out)
+{  
+  single_readdir sr;
+  sr.de = de;
+  sr.stx = stx;
+  sr.inode = NULL;
+  sr.full = false;
+
+  // our callback fills the dirent and sets sr.full=true on first
+  // call, and returns -1 the second time around.
+  int r = readdir_r_cb(d, _readdir_single_dirent_cb, (void *)&sr, want, flags, out);
+  if (r < -1)
+    return r;
+  if (out)
+    *out = sr.inode;
+  if (sr.full)
+    return 1;
+  return 0;
+}
+
+
+/* getdents */
+struct getdents_result {
+  char *buf;
+  int buflen;
+  int pos;
+  bool fullent;
+};
+
+static int _readdir_getdent_cb(void *p, struct dirent *de,
+			       struct ceph_statx *stx, off_t off, Inode *in)
+{
+  struct getdents_result *c = static_cast<getdents_result *>(p);
+
+  int dlen;
+  if (c->fullent)
+    dlen = sizeof(*de);
+  else
+    dlen = strlen(de->d_name) + 1;
+
+  if (c->pos + dlen > c->buflen)
+    return -1;  // doesn't fit
+
+  if (c->fullent) {
+    memcpy(c->buf + c->pos, de, sizeof(*de));
+  } else {
+    memcpy(c->buf + c->pos, de->d_name, dlen);
+  }
+  c->pos += dlen;
+  return 0;
+}
+
+int Client::_getdents(dir_result_t *dir, char *buf, int buflen, bool fullent)
+{
+  getdents_result gr;
+  gr.buf = buf;
+  gr.buflen = buflen;
+  gr.fullent = fullent;
+  gr.pos = 0;
+
+  int r = readdir_r_cb(dir, _readdir_getdent_cb, (void *)&gr);
+
+  if (r < 0) { // some error
+    if (r == -1) { // buffer ran out of space
+      if (gr.pos) { // but we got some entries already!
+        return gr.pos;
+      } // or we need a larger buffer
+      return -CEPHFS_ERANGE;
+    } else { // actual error, return it
+      return r;
+    }
+  }
+  return gr.pos;
+}
+
+
+/* getdir */
+struct getdir_result {
+  list<string> *contents;
+  int num;
+};
+
+static int _getdir_cb(void *p, struct dirent *de, struct ceph_statx *stx, off_t off, Inode *in)
+{
+  getdir_result *r = static_cast<getdir_result *>(p);
+
+  r->contents->push_back(de->d_name);
+  r->num++;
+  return 0;
+}
+
+int Client::getdir(const char *relpath, list<string>& contents,
+		   const UserPerm& perms)
+{
+  ldout(cct, 3) << "getdir(" << relpath << ")" << dendl;
+  tout(cct) << "getdir" << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  dir_result_t *d;
+  int r = opendir(relpath, &d, perms);
+  if (r < 0)
+    return r;
+
+  getdir_result gr;
+  gr.contents = &contents;
+  gr.num = 0;
+  r = readdir_r_cb(d, _getdir_cb, (void *)&gr);
+
+  closedir(d);
+
+  if (r < 0)
+    return r;
+  return gr.num;
+}
+
+
+/****** file i/o **********/
+
+// common parts for open and openat. call with client_lock locked.
+int Client::create_and_open(std::optional<int> dirfd, const char *relpath, int flags,
+                            const UserPerm& perms, mode_t mode, int stripe_unit,
+                            int stripe_count, int object_size, const char *data_pool,
+                            std::string alternate_name) {
+  ceph_assert(ceph_mutex_is_locked(client_lock));
+  int cflags = ceph_flags_sys2wire(flags);
+  tout(cct) << cflags << std::endl;
+
+  Fh *fh = NULL;
+
+#if defined(__linux__) && defined(O_PATH)
+  /* When the O_PATH is being specified, others flags than O_DIRECTORY
+   * and O_NOFOLLOW are ignored. Please refer do_entry_open() function
+   * in kernel (fs/open.c). */
+  if (flags & O_PATH)
+    flags &= O_DIRECTORY | O_NOFOLLOW | O_PATH;
+#endif
+
+  filepath path(relpath);
+  InodeRef in;
+  bool created = false;
+  /* O_CREATE with O_EXCL enforces O_NOFOLLOW. */
+  bool followsym = !((flags & O_NOFOLLOW) || ((flags & O_CREAT) && (flags & O_EXCL)));
+  int mask = ceph_caps_for_mode(ceph_flags_to_mode(cflags));
+
+  InodeRef dirinode = nullptr;
+  if (dirfd) {
+    int r = get_fd_inode(*dirfd, &dirinode);
+    if (r < 0) {
+      return r;
+    }
+  }
+
+  int r = path_walk(path, &in, perms, followsym, mask, dirinode);
+  if (r == 0 && (flags & O_CREAT) && (flags & O_EXCL))
+    return -CEPHFS_EEXIST;
+
+#if defined(__linux__) && defined(O_PATH)
+  if (r == 0 && in->is_symlink() && (flags & O_NOFOLLOW) && !(flags & O_PATH))
+#else
+    if (r == 0 && in->is_symlink() && (flags & O_NOFOLLOW))
+#endif
+    return -CEPHFS_ELOOP;
+
+  if (r == -CEPHFS_ENOENT && (flags & O_CREAT)) {
+    filepath dirpath = path;
+    string dname = dirpath.last_dentry();
+    dirpath.pop_dentry();
+    InodeRef dir;
+    r = path_walk(dirpath, &dir, perms, true,
+                  cct->_conf->client_permissions ? CEPH_CAP_AUTH_SHARED : 0, dirinode);
+    if (r < 0) {
+      goto out;
+    }
+    if (cct->_conf->client_permissions) {
+      r = may_create(dir.get(), perms);
+      if (r < 0)
+        goto out;
+    }
+    r = _create(dir.get(), dname.c_str(), flags, mode, &in, &fh, stripe_unit,
+                stripe_count, object_size, data_pool, &created, perms,
+                std::move(alternate_name));
+  }
+  if (r < 0)
+    goto out;
+
+  if (!created) {
+    // posix says we can only check permissions of existing files
+    if (cct->_conf->client_permissions) {
+      r = may_open(in.get(), flags, perms);
+      if (r < 0)
+        goto out;
+    }
+  }
+
+  if (!fh)
+    r = _open(in.get(), flags, mode, &fh, perms);
+  if (r >= 0) {
+    // allocate a integer file descriptor
+    ceph_assert(fh);
+    r = get_fd();
+    ceph_assert(fd_map.count(r) == 0);
+    fd_map[r] = fh;
+  }
+  
+ out:
+  return r;
+}
+
+int Client::open(const char *relpath, int flags, const UserPerm& perms,
+		 mode_t mode, int stripe_unit, int stripe_count,
+		 int object_size, const char *data_pool, std::string alternate_name)
+{
+  return openat(CEPHFS_AT_FDCWD, relpath, flags, perms, mode, stripe_unit,
+                stripe_count, object_size, data_pool, alternate_name);
+}
+
+int Client::_openat(int dirfd, const char *relpath, int flags, const UserPerm& perms,
+                    mode_t mode, std::string alternate_name) {
+  return create_and_open(dirfd, relpath, flags, perms, mode, 0, 0, 0, NULL, alternate_name);
+}
+
+int Client::openat(int dirfd, const char *relpath, int flags, const UserPerm& perms,
+                   mode_t mode, int stripe_unit, int stripe_count, int object_size,
+                   const char *data_pool, std::string alternate_name) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  ldout(cct, 3) << "openat enter(" << relpath << ")" << dendl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+  tout(cct) << flags << std::endl;
+  tout(cct) << mode << std::endl;
+
+  std::scoped_lock locker(client_lock);
+  int r =  create_and_open(dirfd, relpath, flags, perms, mode, stripe_unit, stripe_count,
+                           object_size, data_pool, alternate_name);
+
+  tout(cct) << r << std::endl;
+  ldout(cct, 3) << "openat exit(" << relpath << ")" << dendl;
+  return r;
+}
+
+int Client::lookup_hash(inodeno_t ino, inodeno_t dirino, const char *name,
+			const UserPerm& perms)
+{
+  ldout(cct, 3) << __func__ << " enter(" << ino << ", #" << dirino << "/" << name << ")" << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPHASH);
+  filepath path(ino);
+  req->set_filepath(path);
+
+  uint32_t h = ceph_str_hash(CEPH_STR_HASH_RJENKINS, name, strlen(name));
+  char f[30];
+  sprintf(f, "%u", h);
+  filepath path2(dirino);
+  path2.push_dentry(string(f));
+  req->set_filepath2(path2);
+
+  int r = make_request(req, perms, NULL, NULL,
+		       rand() % mdsmap->get_num_in_mds());
+  ldout(cct, 3) << __func__ << " exit(" << ino << ", #" << dirino << "/" << name << ") = " << r << dendl;
+  return r;
+}
+
+
+/**
+ * Load inode into local cache.
+ *
+ * If inode pointer is non-NULL, and take a reference on
+ * the resulting Inode object in one operation, so that caller
+ * can safely assume inode will still be there after return.
+ */
+int Client::_lookup_vino(vinodeno_t vino, const UserPerm& perms, Inode **inode)
+{
+  ldout(cct, 8) << __func__ << " enter(" << vino << ")" << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  if (is_reserved_vino(vino))
+    return -CEPHFS_ESTALE;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPINO);
+  filepath path(vino.ino);
+  req->set_filepath(path);
+
+  /*
+   * The MDS expects either a "real" snapid here or 0. The special value
+   * carveouts for the snapid are all at the end of the range so we can
+   * just look for any snapid below this value.
+   */
+  if (vino.snapid < CEPH_NOSNAP)
+    req->head.args.lookupino.snapid = vino.snapid;
+
+  int r = make_request(req, perms, NULL, NULL, rand() % mdsmap->get_num_in_mds());
+  if (r == 0 && inode != NULL) {
+    unordered_map<vinodeno_t,Inode*>::iterator p = inode_map.find(vino);
+    ceph_assert(p != inode_map.end());
+    *inode = p->second;
+    _ll_get(*inode);
+  }
+  ldout(cct, 8) << __func__ << " exit(" << vino << ") = " << r << dendl;
+  return r;
+}
+
+int Client::lookup_ino(inodeno_t ino, const UserPerm& perms, Inode **inode)
+{
+  vinodeno_t vino(ino, CEPH_NOSNAP);
+  std::scoped_lock lock(client_lock);
+  return _lookup_vino(vino, perms, inode);
+}
+
+/**
+ * Find the parent inode of `ino` and insert it into
+ * our cache.  Conditionally also set `parent` to a referenced
+ * Inode* if caller provides non-NULL value.
+ */
+int Client::_lookup_parent(Inode *ino, const UserPerm& perms, Inode **parent)
+{
+  ldout(cct, 8) << __func__ << " enter(" << ino->ino << ")" << dendl;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPPARENT);
+  filepath path(ino->ino);
+  req->set_filepath(path);
+
+  InodeRef target;
+  int r = make_request(req, perms, &target, NULL, rand() % mdsmap->get_num_in_mds());
+  // Give caller a reference to the parent ino if they provided a pointer.
+  if (parent != NULL) {
+    if (r == 0) {
+      *parent = target.get();
+      _ll_get(*parent);
+      ldout(cct, 8) << __func__ << " found parent " << (*parent)->ino << dendl;
+    } else {
+      *parent = NULL;
+    }
+  }
+  ldout(cct, 8) << __func__ << " exit(" << ino->ino << ") = " << r << dendl;
+  return r;
+}
+
+/**
+ * Populate the parent dentry for `ino`, provided it is
+ * a child of `parent`.
+ */
+int Client::_lookup_name(Inode *ino, Inode *parent, const UserPerm& perms)
+{
+  ceph_assert(parent->is_dir());
+  ldout(cct, 3) << __func__ << " enter(" << ino->ino << ")" << dendl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPNAME);
+  req->set_filepath2(filepath(parent->ino));
+  req->set_filepath(filepath(ino->ino));
+  req->set_inode(ino);
+
+  int r = make_request(req, perms, NULL, NULL, rand() % mdsmap->get_num_in_mds());
+  ldout(cct, 3) << __func__ << " exit(" << ino->ino << ") = " << r << dendl;
+  return r;
+}
+
+int Client::lookup_name(Inode *ino, Inode *parent, const UserPerm& perms)
+{
+  std::scoped_lock lock(client_lock);
+  return _lookup_name(ino, parent, perms);
+}
+
+Fh *Client::_create_fh(Inode *in, int flags, int cmode, const UserPerm& perms)
+{
+  ceph_assert(in);
+  Fh *f = new Fh(in, flags, cmode, fd_gen, perms);
+
+  ldout(cct, 10) << __func__ << " " << in->ino << " mode " << cmode << dendl;
+
+  if (in->snapid != CEPH_NOSNAP) {
+    in->snap_cap_refs++;
+    ldout(cct, 5) << "open success, fh is " << f << " combined IMMUTABLE SNAP caps " 
+	    << ccap_string(in->caps_issued()) << dendl;
+  }
+
+  const auto& conf = cct->_conf;
+  f->readahead.set_trigger_requests(1);
+  f->readahead.set_min_readahead_size(conf->client_readahead_min);
+  uint64_t max_readahead = Readahead::NO_LIMIT;
+  if (conf->client_readahead_max_bytes) {
+    max_readahead = std::min(max_readahead, (uint64_t)conf->client_readahead_max_bytes);
+  }
+  if (conf->client_readahead_max_periods) {
+    max_readahead = std::min(max_readahead, in->layout.get_period()*(uint64_t)conf->client_readahead_max_periods);
+  }
+  f->readahead.set_max_readahead_size(max_readahead);
+  vector<uint64_t> alignments;
+  alignments.push_back(in->layout.get_period());
+  alignments.push_back(in->layout.stripe_unit);
+  f->readahead.set_alignments(alignments);
+
+  return f;
+}
+
+int Client::_release_fh(Fh *f)
+{
+  //ldout(cct, 3) << "op: client->close(open_files[ " << fh << " ]);" << dendl;
+  //ldout(cct, 3) << "op: open_files.erase( " << fh << " );" << dendl;
+  Inode *in = f->inode.get();
+  ldout(cct, 8) << __func__ << " " << f << " mode " << f->mode << " on " << *in << dendl;
+
+  in->unset_deleg(f);
+
+  if (in->snapid == CEPH_NOSNAP) {
+    if (in->put_open_ref(f->mode)) {
+      _flush(in, new C_Client_FlushComplete(this, in));
+      check_caps(in, 0);
+    }
+  } else {
+    ceph_assert(in->snap_cap_refs > 0);
+    in->snap_cap_refs--;
+  }
+
+  _release_filelocks(f);
+
+  // Finally, read any async err (i.e. from flushes)
+  int err = f->take_async_err();
+  if (err != 0) {
+    ldout(cct, 1) << __func__ << " " << f << " on inode " << *in << " caught async_err = "
+                  << cpp_strerror(err) << dendl;
+  } else {
+    ldout(cct, 10) << __func__ << " " << f << " on inode " << *in << " no async_err state" << dendl;
+  }
+
+  _put_fh(f);
+
+  return err;
+}
+
+void Client::_put_fh(Fh *f)
+{
+  int left = f->put();
+  if (!left) {
+    delete f;
+  }
+}
+
+int Client::_open(Inode *in, int flags, mode_t mode, Fh **fhp,
+		  const UserPerm& perms)
+{
+  if (in->snapid != CEPH_NOSNAP &&
+      (flags & (O_WRONLY | O_RDWR | O_CREAT | O_TRUNC | O_APPEND))) {
+    return -CEPHFS_EROFS;
+  }
+
+  // use normalized flags to generate cmode
+  int cflags = ceph_flags_sys2wire(flags);
+  if (cct->_conf.get_val<bool>("client_force_lazyio"))
+    cflags |= CEPH_O_LAZY;
+
+  int cmode = ceph_flags_to_mode(cflags);
+  int want = ceph_caps_for_mode(cmode);
+  int result = 0;
+
+  in->get_open_ref(cmode);  // make note of pending open, since it effects _wanted_ caps.
+
+  if ((flags & O_TRUNC) == 0 && in->caps_issued_mask(want)) {
+    // update wanted?
+    check_caps(in, CHECK_CAPS_NODELAY);
+  } else {
+
+    MetaRequest *req = new MetaRequest(CEPH_MDS_OP_OPEN);
+    filepath path;
+    in->make_nosnap_relative_path(path);
+    req->set_filepath(path);
+    req->head.args.open.flags = cflags & ~CEPH_O_CREAT;
+    req->head.args.open.mode = mode;
+    req->head.args.open.pool = -1;
+    if (cct->_conf->client_debug_getattr_caps)
+      req->head.args.open.mask = DEBUG_GETATTR_CAPS;
+    else
+      req->head.args.open.mask = 0;
+    req->head.args.open.old_size = in->size;   // for O_TRUNC
+    req->set_inode(in);
+    result = make_request(req, perms);
+
+    /*
+     * NFS expects that delegations will be broken on a conflicting open,
+     * not just when there is actual conflicting access to the file. SMB leases
+     * and oplocks also have similar semantics.
+     *
+     * Ensure that clients that have delegations enabled will wait on minimal
+     * caps during open, just to ensure that other clients holding delegations
+     * return theirs first.
+     */
+    if (deleg_timeout && result == 0) {
+      int need = 0, have;
+
+      if (cmode & CEPH_FILE_MODE_WR)
+        need |= CEPH_CAP_FILE_WR;
+      if (cmode & CEPH_FILE_MODE_RD)
+        need |= CEPH_CAP_FILE_RD;
+
+      Fh fh(in, flags, cmode, fd_gen, perms);
+      result = get_caps(&fh, need, want, &have, -1);
+      if (result < 0) {
+	ldout(cct, 8) << "Unable to get caps after open of inode " << *in <<
+			  " . Denying open: " <<
+			  cpp_strerror(result) << dendl;
+      } else {
+	put_cap_ref(in, need);
+      }
+    }
+  }
+
+  // success?
+  if (result >= 0) {
+    if (fhp)
+      *fhp = _create_fh(in, flags, cmode, perms);
+  } else {
+    in->put_open_ref(cmode);
+  }
+
+  trim_cache();
+
+  return result;
+}
+
+int Client::_renew_caps(Inode *in)
+{
+  int wanted = in->caps_file_wanted();
+  if (in->is_any_caps() &&
+      ((wanted & CEPH_CAP_ANY_WR) == 0 || in->auth_cap)) {
+    check_caps(in, CHECK_CAPS_NODELAY);
+    return 0;
+  }
+
+  int flags = 0;
+  if ((wanted & CEPH_CAP_FILE_RD) && (wanted & CEPH_CAP_FILE_WR))
+    flags = O_RDWR;
+  else if (wanted & CEPH_CAP_FILE_RD)
+    flags = O_RDONLY;
+  else if (wanted & CEPH_CAP_FILE_WR)
+    flags = O_WRONLY;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_OPEN);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->head.args.open.flags = flags;
+  req->head.args.open.pool = -1;
+  if (cct->_conf->client_debug_getattr_caps)
+    req->head.args.open.mask = DEBUG_GETATTR_CAPS;
+  else
+    req->head.args.open.mask = 0;
+  req->set_inode(in);
+
+  // duplicate in case Cap goes away; not sure if that race is a concern?
+  const UserPerm *pperm = in->get_best_perms();
+  UserPerm perms;
+  if (pperm != NULL)
+    perms = *pperm;
+  int ret = make_request(req, perms);
+  return ret;
+}
+
+int Client::_close(int fd)
+{
+  ldout(cct, 3) << "close enter(" << fd << ")" << dendl;
+  tout(cct) << "close" << std::endl;
+  tout(cct) << fd << std::endl;
+
+  Fh *fh = get_filehandle(fd);
+  if (!fh)
+    return -CEPHFS_EBADF;
+  int err = _release_fh(fh);
+  fd_map.erase(fd);
+  put_fd(fd);
+  ldout(cct, 3) << "close exit(" << fd << ")" << dendl;
+  return err;
+}
+
+int Client::close(int fd) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+  return _close(fd);
+}
+
+// ------------
+// read, write
+
+loff_t Client::lseek(int fd, loff_t offset, int whence)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "lseek" << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << offset << std::endl;
+  tout(cct) << whence << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  return _lseek(f, offset, whence);
+}
+
+loff_t Client::_lseek(Fh *f, loff_t offset, int whence)
+{
+  Inode *in = f->inode.get();
+  bool whence_check = false;
+  loff_t pos = -1;
+
+  switch (whence) {
+  case SEEK_END:
+    whence_check = true;
+  break;
+
+#ifdef SEEK_DATA
+  case SEEK_DATA:
+    whence_check = true;
+  break;
+#endif
+
+#ifdef SEEK_HOLE
+  case SEEK_HOLE:
+    whence_check = true;
+  break;
+#endif
+  }
+
+  if (whence_check) {
+    int r = _getattr(in, CEPH_STAT_CAP_SIZE, f->actor_perms);
+    if (r < 0)
+      return r;
+  }
+
+  switch (whence) {
+  case SEEK_SET:
+    pos = offset;
+    break;
+
+  case SEEK_CUR:
+    pos = f->pos + offset;
+    break;
+
+  case SEEK_END:
+    pos = in->size + offset;
+    break;
+
+#ifdef SEEK_DATA
+  case SEEK_DATA:
+    if (offset < 0 || static_cast<uint64_t>(offset) >= in->size)
+      return -CEPHFS_ENXIO;
+    pos = offset;
+    break;
+#endif
+
+#ifdef SEEK_HOLE
+  case SEEK_HOLE:
+    if (offset < 0 || static_cast<uint64_t>(offset) >= in->size)
+      return -CEPHFS_ENXIO;
+    pos = in->size;
+    break;
+#endif
+
+  default:
+    ldout(cct, 1) << __func__ << ": invalid whence value " << whence << dendl;
+    return -CEPHFS_EINVAL;
+  }
+
+  if (pos < 0) {
+    return -CEPHFS_EINVAL;
+  } else {
+    f->pos = pos;
+  }
+
+  ldout(cct, 8) << "_lseek(" << f << ", " << offset << ", " << whence << ") = " << f->pos << dendl;
+  return f->pos;
+}
+
+
+void Client::lock_fh_pos(Fh *f)
+{
+  ldout(cct, 10) << __func__ << " " << f << dendl;
+
+  if (f->pos_locked || !f->pos_waiters.empty()) {
+    ceph::condition_variable cond;
+    f->pos_waiters.push_back(&cond);
+    ldout(cct, 10) << __func__ << " BLOCKING on " << f << dendl;
+    std::unique_lock l{client_lock, std::adopt_lock};
+    cond.wait(l, [f, me=&cond] {
+      return !f->pos_locked && f->pos_waiters.front() == me;
+    });
+    l.release();
+    ldout(cct, 10) << __func__ << " UNBLOCKING on " << f << dendl;
+    ceph_assert(f->pos_waiters.front() == &cond);
+    f->pos_waiters.pop_front();
+  }
+
+  f->pos_locked = true;
+}
+
+void Client::unlock_fh_pos(Fh *f)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  ldout(cct, 10) << __func__ << " " << f << dendl;
+  f->pos_locked = false;
+  if (!f->pos_waiters.empty()) {
+    // only wake up the oldest waiter
+    auto cond = f->pos_waiters.front();
+    cond->notify_one();
+  }
+}
+
+int Client::uninline_data(Inode *in, Context *onfinish)
+{
+  if (!in->inline_data.length()) {
+    onfinish->complete(0);
+    return 0;
+  }
+
+  char oid_buf[32];
+  snprintf(oid_buf, sizeof(oid_buf), "%llx.00000000", (long long unsigned)in->ino);
+  object_t oid = oid_buf;
+
+  ObjectOperation create_ops;
+  create_ops.create(false);
+
+  objecter->mutate(oid,
+		   OSDMap::file_to_object_locator(in->layout),
+		   create_ops,
+		   in->snaprealm->get_snap_context(),
+		   ceph::real_clock::now(),
+		   0,
+		   NULL);
+
+  bufferlist inline_version_bl;
+  encode(in->inline_version, inline_version_bl);
+
+  ObjectOperation uninline_ops;
+  uninline_ops.cmpxattr("inline_version",
+                        CEPH_OSD_CMPXATTR_OP_GT,
+                        CEPH_OSD_CMPXATTR_MODE_U64,
+                        inline_version_bl);
+  bufferlist inline_data = in->inline_data;
+  uninline_ops.write(0, inline_data, in->truncate_size, in->truncate_seq);
+  uninline_ops.setxattr("inline_version", stringify(in->inline_version));
+
+  objecter->mutate(oid,
+		   OSDMap::file_to_object_locator(in->layout),
+		   uninline_ops,
+		   in->snaprealm->get_snap_context(),
+		   ceph::real_clock::now(),
+		   0,
+		   onfinish);
+
+  return 0;
+}
+
+//
+
+// blocking osd interface
+
+int Client::read(int fd, char *buf, loff_t size, loff_t offset)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "read" << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << size << std::endl;
+  tout(cct) << offset << std::endl;
+
+  std::unique_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  bufferlist bl;
+  /* We can't return bytes written larger than INT_MAX, clamp size to that */
+  size = std::min(size, (loff_t)INT_MAX);
+  int r = _read(f, offset, size, &bl);
+  ldout(cct, 3) << "read(" << fd << ", " << (void*)buf << ", " << size << ", " << offset << ") = " << r << dendl;
+  if (r >= 0) {
+    lock.unlock();
+    bl.begin().copy(bl.length(), buf);
+    r = bl.length();
+  }
+  return r;
+}
+
+int Client::preadv(int fd, const struct iovec *iov, int iovcnt, loff_t offset)
+{
+  if (iovcnt < 0)
+    return -CEPHFS_EINVAL;
+  return _preadv_pwritev(fd, iov, iovcnt, offset, false);
+}
+
+int64_t Client::_read(Fh *f, int64_t offset, uint64_t size, bufferlist *bl)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  int want, have = 0;
+  bool movepos = false;
+  int64_t rc = 0;
+  const auto& conf = cct->_conf;
+  Inode *in = f->inode.get();
+  utime_t lat;
+  utime_t start = ceph_clock_now(); 
+
+  if ((f->mode & CEPH_FILE_MODE_RD) == 0)
+    return -CEPHFS_EBADF;
+  //bool lazy = f->mode == CEPH_FILE_MODE_LAZY;
+
+  if (offset < 0) {
+    lock_fh_pos(f);
+    offset = f->pos;
+    movepos = true;
+  }
+  loff_t start_pos = offset;
+
+  if (in->inline_version == 0) {
+    auto r = _getattr(in, CEPH_STAT_CAP_INLINE_DATA, f->actor_perms, true);
+    if (r < 0) {
+      rc = r;
+      goto done;
+    }
+    ceph_assert(in->inline_version > 0);
+  }
+
+retry:
+  if (f->mode & CEPH_FILE_MODE_LAZY)
+    want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
+  else
+    want = CEPH_CAP_FILE_CACHE;
+  {
+    auto r = get_caps(f, CEPH_CAP_FILE_RD, want, &have, -1);
+    if (r < 0) {
+      rc = r;
+      goto done;
+    }
+  }
+  if (f->flags & O_DIRECT)
+    have &= ~(CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO);
+
+  if (in->inline_version < CEPH_INLINE_NONE) {
+    uint32_t len = in->inline_data.length();
+    uint64_t endoff = offset + size;
+    if (endoff > in->size)
+      endoff = in->size;
+
+    if (offset < len) {
+      if (endoff <= len) {
+        bl->substr_of(in->inline_data, offset, endoff - offset);
+      } else {
+        bl->substr_of(in->inline_data, offset, len - offset);
+        bl->append_zero(endoff - len);
+      }
+      rc = endoff - offset;
+    } else if ((uint64_t)offset < endoff) {
+      bl->append_zero(endoff - offset);
+      rc = endoff - offset;
+    } else {
+      rc = 0;
+    }
+    goto success;
+  }
+
+  if (!conf->client_debug_force_sync_read &&
+      conf->client_oc &&
+      (have & (CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO))) {
+
+    if (f->flags & O_RSYNC) {
+      _flush_range(in, offset, size);
+    }
+    rc = _read_async(f, offset, size, bl);
+    if (rc < 0)
+      goto done;
+  } else {
+    if (f->flags & O_DIRECT)
+      _flush_range(in, offset, size);
+
+    bool checkeof = false;
+    rc = _read_sync(f, offset, size, bl, &checkeof);
+    if (rc < 0)
+      goto done;
+    if (checkeof) {
+      offset += rc;
+      size -= rc;
+
+      put_cap_ref(in, CEPH_CAP_FILE_RD);
+      have = 0;
+      // reverify size
+      {
+        auto r = _getattr(in, CEPH_STAT_CAP_SIZE, f->actor_perms);
+        if (r < 0) {
+          rc = r;
+          goto done;
+        }
+      }
+
+      // eof?  short read.
+      if ((uint64_t)offset < in->size)
+	goto retry;
+    }
+  }
+
+success:
+  ceph_assert(rc >= 0);
+  update_read_io_size(bl->length());
+  if (movepos) {
+    // adjust fd pos
+    f->pos = start_pos + rc;
+  }
+  
+  lat = ceph_clock_now();
+  lat -= start;
+
+  ++nr_read_request;
+  update_io_stat_read(lat);
+
+done:
+  // done!
+  if (have) {
+    put_cap_ref(in, CEPH_CAP_FILE_RD);
+  }
+  if (movepos) {
+    unlock_fh_pos(f);
+  }
+  return rc;
+}
+
+Client::C_Readahead::C_Readahead(Client *c, Fh *f) :
+    client(c), f(f) {
+  f->get();
+  f->readahead.inc_pending();
+}
+
+Client::C_Readahead::~C_Readahead() {
+  f->readahead.dec_pending();
+  client->_put_fh(f);
+}
+
+void Client::C_Readahead::finish(int r) {
+  lgeneric_subdout(client->cct, client, 20) << "client." << client->get_nodeid() << " " << "C_Readahead on " << f->inode << dendl;
+  client->put_cap_ref(f->inode.get(), CEPH_CAP_FILE_RD | CEPH_CAP_FILE_CACHE);
+  if (r > 0) {
+    client->update_read_io_size(r);
+  }
+}
+
+int Client::_read_async(Fh *f, uint64_t off, uint64_t len, bufferlist *bl)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  const auto& conf = cct->_conf;
+  Inode *in = f->inode.get();
+
+  ldout(cct, 10) << __func__ << " " << *in << " " << off << "~" << len << dendl;
+
+  // trim read based on file size?
+  if (off >= in->size)
+    return 0;
+  if (len == 0)
+    return 0;
+  if (off + len > in->size) {
+    len = in->size - off;    
+  }
+
+  ldout(cct, 10) << " min_bytes=" << f->readahead.get_min_readahead_size()
+                 << " max_bytes=" << f->readahead.get_max_readahead_size()
+                 << " max_periods=" << conf->client_readahead_max_periods << dendl;
+
+  // read (and possibly block)
+  int r = 0;
+  C_SaferCond onfinish("Client::_read_async flock");
+  r = objectcacher->file_read(&in->oset, &in->layout, in->snapid,
+			      off, len, bl, 0, &onfinish);
+  if (r == 0) {
+    get_cap_ref(in, CEPH_CAP_FILE_CACHE);
+    client_lock.unlock();
+    r = onfinish.wait();
+    client_lock.lock();
+    put_cap_ref(in, CEPH_CAP_FILE_CACHE);
+    update_read_io_size(bl->length());
+  }
+
+  if(f->readahead.get_min_readahead_size() > 0) {
+    pair<uint64_t, uint64_t> readahead_extent = f->readahead.update(off, len, in->size);
+    if (readahead_extent.second > 0) {
+      ldout(cct, 20) << "readahead " << readahead_extent.first << "~" << readahead_extent.second
+		     << " (caller wants " << off << "~" << len << ")" << dendl;
+      Context *onfinish2 = new C_Readahead(this, f);
+      int r2 = objectcacher->file_read(&in->oset, &in->layout, in->snapid,
+				       readahead_extent.first, readahead_extent.second,
+				       NULL, 0, onfinish2);
+      if (r2 == 0) {
+	ldout(cct, 20) << "readahead initiated, c " << onfinish2 << dendl;
+	get_cap_ref(in, CEPH_CAP_FILE_RD | CEPH_CAP_FILE_CACHE);
+      } else {
+	ldout(cct, 20) << "readahead was no-op, already cached" << dendl;
+	delete onfinish2;
+      }
+    }
+  }
+
+  return r;
+}
+
+int Client::_read_sync(Fh *f, uint64_t off, uint64_t len, bufferlist *bl,
+		       bool *checkeof)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  Inode *in = f->inode.get();
+  uint64_t pos = off;
+  int left = len;
+  int read = 0;
+
+  ldout(cct, 10) << __func__ << " " << *in << " " << off << "~" << len << dendl;
+
+  // 0 success, 1 continue and < 0 error happen.
+  auto wait_and_copy = [&](C_SaferCond &onfinish, bufferlist &tbl, int wanted) {
+    int r = onfinish.wait();
+
+    // if we get ENOENT from OSD, assume 0 bytes returned
+    if (r == -CEPHFS_ENOENT)
+      r = 0;
+    if (r < 0)
+      return r;
+
+    if (tbl.length()) {
+      r = tbl.length();
+
+      read += r;
+      pos += r;
+      left -= r;
+      bl->claim_append(tbl);
+    }
+    // short read?
+    if (r >= 0 && r < wanted) {
+      if (pos < in->size) {
+	// zero up to known EOF
+	int64_t some = in->size - pos;
+	if (some > left)
+	  some = left;
+	auto z = buffer::ptr_node::create(some);
+	z->zero();
+	bl->push_back(std::move(z));
+	read += some;
+	pos += some;
+	left -= some;
+	if (left == 0)
+          return 0;
+      }
+
+      *checkeof = true;
+      return 0;
+    }
+    return 1;
+  };
+
+  while (left > 0) {
+    C_SaferCond onfinish("Client::_read_sync flock");
+    bufferlist tbl;
+
+    int wanted = left;
+    filer->read_trunc(in->ino, &in->layout, in->snapid,
+		      pos, left, &tbl, 0,
+		      in->truncate_size, in->truncate_seq,
+		      &onfinish);
+    client_lock.unlock();
+    int r = wait_and_copy(onfinish, tbl, wanted);
+    client_lock.lock();
+    if (!r)
+      return read;
+    if (r < 0)
+      return r;
+  }
+  return read;
+}
+
+int Client::write(int fd, const char *buf, loff_t size, loff_t offset) 
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "write" << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << size << std::endl;
+  tout(cct) << offset << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *fh = get_filehandle(fd);
+  if (!fh)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (fh->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  /* We can't return bytes written larger than INT_MAX, clamp size to that */
+  size = std::min(size, (loff_t)INT_MAX);
+  int r = _write(fh, offset, size, buf, NULL, false);
+  ldout(cct, 3) << "write(" << fd << ", \"...\", " << size << ", " << offset << ") = " << r << dendl;
+  return r;
+}
+
+int Client::pwritev(int fd, const struct iovec *iov, int iovcnt, int64_t offset)
+{
+  if (iovcnt < 0)
+    return -CEPHFS_EINVAL;
+  return _preadv_pwritev(fd, iov, iovcnt, offset, true);
+}
+
+int64_t Client::_preadv_pwritev_locked(Fh *fh, const struct iovec *iov,
+				   unsigned iovcnt, int64_t offset, bool write,
+				   bool clamp_to_int, std::unique_lock<ceph::mutex> &cl)
+{
+#if defined(__linux__) && defined(O_PATH)
+    if (fh->flags & O_PATH)
+        return -CEPHFS_EBADF;
+#endif
+    loff_t totallen = 0;
+    for (unsigned i = 0; i < iovcnt; i++) {
+        totallen += iov[i].iov_len;
+    }
+
+    /*
+     * Some of the API functions take 64-bit size values, but only return
+     * 32-bit signed integers. Clamp the I/O sizes in those functions so that
+     * we don't do I/Os larger than the values we can return.
+     */
+    if (clamp_to_int) {
+      totallen = std::min(totallen, (loff_t)INT_MAX);
+    }
+    if (write) {
+        int64_t w = _write(fh, offset, totallen, NULL, iov, iovcnt);
+        ldout(cct, 3) << "pwritev(" << fh << ", \"...\", " << totallen << ", " << offset << ") = " << w << dendl;
+        return w;
+    } else {
+        bufferlist bl;
+        int64_t r = _read(fh, offset, totallen, &bl);
+        ldout(cct, 3) << "preadv(" << fh << ", " <<  offset << ") = " << r << dendl;
+        if (r <= 0)
+          return r;
+
+        cl.unlock();
+        auto iter = bl.cbegin();
+        for (unsigned j = 0, resid = r; j < iovcnt && resid > 0; j++) {
+               /*
+                * This piece of code aims to handle the case that bufferlist
+                * does not have enough data to fill in the iov
+                */
+               const auto round_size = std::min<unsigned>(resid, iov[j].iov_len);
+               iter.copy(round_size, reinterpret_cast<char*>(iov[j].iov_base));
+               resid -= round_size;
+               /* iter is self-updating */
+        }
+        cl.lock();
+        return r;
+    }
+}
+
+int Client::_preadv_pwritev(int fd, const struct iovec *iov, unsigned iovcnt, int64_t offset, bool write)
+{
+    RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+    if (!mref_reader.is_state_satisfied())
+      return -CEPHFS_ENOTCONN;
+
+    tout(cct) << fd << std::endl;
+    tout(cct) << offset << std::endl;
+
+    std::unique_lock cl(client_lock);
+    Fh *fh = get_filehandle(fd);
+    if (!fh)
+      return -CEPHFS_EBADF;
+    return _preadv_pwritev_locked(fh, iov, iovcnt, offset, write, true, cl);
+}
+
+int64_t Client::_write(Fh *f, int64_t offset, uint64_t size, const char *buf,
+	                const struct iovec *iov, int iovcnt)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  uint64_t fpos = 0;
+  Inode *in = f->inode.get();
+
+  if ( (uint64_t)(offset+size) > mdsmap->get_max_filesize() && //exceeds config
+       (uint64_t)(offset+size) > in->size ) { //exceeds filesize 
+      return -CEPHFS_EFBIG;              
+	}
+  //ldout(cct, 7) << "write fh " << fh << " size " << size << " offset " << offset << dendl;
+
+  if (objecter->osdmap_pool_full(in->layout.pool_id)) {
+    return -CEPHFS_ENOSPC;
+  }
+
+  ceph_assert(in->snapid == CEPH_NOSNAP);
+
+  // was Fh opened as writeable?
+  if ((f->mode & CEPH_FILE_MODE_WR) == 0)
+    return -CEPHFS_EBADF;
+
+  // use/adjust fd pos?
+  if (offset < 0) {
+    lock_fh_pos(f);
+    /*
+     * FIXME: this is racy in that we may block _after_ this point waiting for caps, and size may
+     * change out from under us.
+     */
+    if (f->flags & O_APPEND) {
+      auto r = _lseek(f, 0, SEEK_END);
+      if (r < 0) {
+        unlock_fh_pos(f);
+        return r;
+      }
+    }
+    offset = f->pos;
+    fpos = offset+size;
+    unlock_fh_pos(f);
+  }
+
+  // check quota
+  uint64_t endoff = offset + size;
+  if (endoff > in->size && is_quota_bytes_exceeded(in, endoff - in->size,
+						   f->actor_perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  //bool lazy = f->mode == CEPH_FILE_MODE_LAZY;
+
+  ldout(cct, 10) << "cur file size is " << in->size << dendl;
+
+  // time it.
+  utime_t start = ceph_clock_now();
+
+  if (in->inline_version == 0) {
+    int r = _getattr(in, CEPH_STAT_CAP_INLINE_DATA, f->actor_perms, true);
+    if (r < 0)
+      return r;
+    ceph_assert(in->inline_version > 0);
+  }
+
+  // copy into fresh buffer (since our write may be resub, async)
+  bufferlist bl;
+  if (buf) {
+    if (size > 0)
+      bl.append(buf, size);
+  } else if (iov){
+    for (int i = 0; i < iovcnt; i++) {
+      if (iov[i].iov_len > 0) {
+        bl.append((const char *)iov[i].iov_base, iov[i].iov_len);
+      }
+    }
+  }
+
+  utime_t lat;
+  uint64_t totalwritten;
+  int want, have;
+  if (f->mode & CEPH_FILE_MODE_LAZY)
+    want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
+  else
+    want = CEPH_CAP_FILE_BUFFER;
+  int r = get_caps(f, CEPH_CAP_FILE_WR|CEPH_CAP_AUTH_SHARED, want, &have, endoff);
+  if (r < 0)
+    return r;
+
+  /* clear the setuid/setgid bits, if any */
+  if (unlikely(in->mode & (S_ISUID|S_ISGID)) && size > 0) {
+    struct ceph_statx stx = { 0 };
+
+    put_cap_ref(in, CEPH_CAP_AUTH_SHARED);
+    r = __setattrx(in, &stx, CEPH_SETATTR_KILL_SGUID, f->actor_perms);
+    if (r < 0)
+      return r;
+  } else {
+    put_cap_ref(in, CEPH_CAP_AUTH_SHARED);
+  }
+
+  if (f->flags & O_DIRECT)
+    have &= ~(CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO);
+
+  ldout(cct, 10) << " snaprealm " << *in->snaprealm << dendl;
+
+  std::unique_ptr<C_SaferCond> onuninline = nullptr;
+  
+  if (in->inline_version < CEPH_INLINE_NONE) {
+    if (endoff > cct->_conf->client_max_inline_size ||
+        endoff > CEPH_INLINE_MAX_SIZE ||
+        !(have & CEPH_CAP_FILE_BUFFER)) {
+      onuninline.reset(new C_SaferCond("Client::_write_uninline_data flock"));
+      uninline_data(in, onuninline.get());
+    } else {
+      get_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+      uint32_t len = in->inline_data.length();
+
+      if (endoff < len)
+        in->inline_data.begin(endoff).copy(len - endoff, bl); // XXX
+
+      if (offset < len)
+        in->inline_data.splice(offset, len - offset);
+      else if (offset > len)
+        in->inline_data.append_zero(offset - len);
+
+      in->inline_data.append(bl);
+      in->inline_version++;
+
+      put_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+      goto success;
+    }
+  }
+
+  if (cct->_conf->client_oc &&
+      (have & (CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO))) {
+    // do buffered write
+    if (!in->oset.dirty_or_tx)
+      get_cap_ref(in, CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_BUFFER);
+
+    get_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+    // async, caching, non-blocking.
+    r = objectcacher->file_write(&in->oset, &in->layout,
+				 in->snaprealm->get_snap_context(),
+				 offset, size, bl, ceph::real_clock::now(),
+				 0);
+    put_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+    if (r < 0)
+      goto done;
+
+    // flush cached write if O_SYNC is set on file fh
+    // O_DSYNC == O_SYNC on linux < 2.6.33
+    // O_SYNC = __O_SYNC | O_DSYNC on linux >= 2.6.33
+    if ((f->flags & O_SYNC) || (f->flags & O_DSYNC)) {
+      _flush_range(in, offset, size);
+    }
+  } else {
+    if (f->flags & O_DIRECT)
+      _flush_range(in, offset, size);
+
+    // simple, non-atomic sync write
+    C_SaferCond onfinish("Client::_write flock");
+    get_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+    filer->write_trunc(in->ino, &in->layout, in->snaprealm->get_snap_context(),
+		       offset, size, bl, ceph::real_clock::now(), 0,
+		       in->truncate_size, in->truncate_seq,
+		       &onfinish);
+    client_lock.unlock();
+    r = onfinish.wait();
+    client_lock.lock();
+    put_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+    if (r < 0)
+      goto done;
+  }
+
+  // if we get here, write was successful, update client metadata
+success:
+  update_write_io_size(size);
+  // time
+  lat = ceph_clock_now();
+  lat -= start;
+
+  ++nr_write_request;
+  update_io_stat_write(lat);
+
+  if (fpos) {
+    lock_fh_pos(f);
+    f->pos = fpos;
+    unlock_fh_pos(f);
+  }
+  totalwritten = size;
+  r = (int64_t)totalwritten;
+
+  // extend file?
+  if (totalwritten + offset > in->size) {
+    in->size = totalwritten + offset;
+    in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+
+    if (is_quota_bytes_approaching(in, f->actor_perms)) {
+      check_caps(in, CHECK_CAPS_NODELAY);
+    } else if (is_max_size_approaching(in)) {
+      check_caps(in, 0);
+    }
+
+    ldout(cct, 7) << "wrote to " << totalwritten+offset << ", extending file size" << dendl;
+  } else {
+    ldout(cct, 7) << "wrote to " << totalwritten+offset << ", leaving file size at " << in->size << dendl;
+  }
+
+  // mtime
+  in->mtime = in->ctime = ceph_clock_now();
+  in->change_attr++;
+  in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+
+done:
+
+  if (nullptr != onuninline) {
+    client_lock.unlock();
+    int uninline_ret = onuninline->wait();
+    client_lock.lock();
+
+    if (uninline_ret >= 0 || uninline_ret == -CEPHFS_ECANCELED) {
+      in->inline_data.clear();
+      in->inline_version = CEPH_INLINE_NONE;
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+      check_caps(in, 0);
+    } else
+      r = uninline_ret;
+  }
+
+  put_cap_ref(in, CEPH_CAP_FILE_WR);
+  return r;
+}
+
+int Client::_flush(Fh *f)
+{
+  Inode *in = f->inode.get();
+  int err = f->take_async_err();
+  if (err != 0) {
+    ldout(cct, 1) << __func__ << ": " << f << " on inode " << *in << " caught async_err = "
+                  << cpp_strerror(err) << dendl;
+  } else {
+    ldout(cct, 10) << __func__ << ": " << f << " on inode " << *in << " no async_err state" << dendl;
+  }
+
+  return err;
+}
+
+int Client::truncate(const char *relpath, loff_t length, const UserPerm& perms) 
+{
+  struct ceph_statx stx;
+  stx.stx_size = length;
+  return setattrx(relpath, &stx, CEPH_SETATTR_SIZE, perms);
+}
+
+int Client::ftruncate(int fd, loff_t length, const UserPerm& perms) 
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << length << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  if ((f->mode & CEPH_FILE_MODE_WR) == 0)
+    return -CEPHFS_EBADF;
+  struct stat attr;
+  attr.st_size = length;
+  return _setattr(f->inode, &attr, CEPH_SETATTR_SIZE, perms);
+}
+
+int Client::fsync(int fd, bool syncdataonly) 
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "fsync" << std::endl;
+  tout(cct) << fd << std::endl;
+  tout(cct) << syncdataonly << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (f->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  int r = _fsync(f, syncdataonly);
+  if (r == 0) {
+    // The IOs in this fsync were okay, but maybe something happened
+    // in the background that we shoudl be reporting?
+    r = f->take_async_err();
+    ldout(cct, 5) << "fsync(" << fd << ", " << syncdataonly
+                  << ") = 0, async_err = " << r << dendl;
+  } else {
+    // Assume that an error we encountered during fsync, even reported
+    // synchronously, would also have applied the error to the Fh, and we
+    // should clear it here to avoid returning the same error again on next
+    // call.
+    ldout(cct, 5) << "fsync(" << fd << ", " << syncdataonly << ") = "
+                  << r << dendl;
+    f->take_async_err();
+  }
+  return r;
+}
+
+int Client::_fsync(Inode *in, bool syncdataonly)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  int r = 0;
+  std::unique_ptr<C_SaferCond> object_cacher_completion = nullptr;
+  ceph_tid_t flush_tid = 0;
+  InodeRef tmp_ref;
+  utime_t lat;
+  utime_t start = ceph_clock_now(); 
+
+  ldout(cct, 8) << "_fsync on " << *in << " " << (syncdataonly ? "(dataonly)":"(data+metadata)") << dendl;
+  
+  if (cct->_conf->client_oc) {
+    object_cacher_completion.reset(new C_SaferCond("Client::_fsync::lock"));
+    tmp_ref = in; // take a reference; C_SaferCond doesn't and _flush won't either
+    _flush(in, object_cacher_completion.get());
+    ldout(cct, 15) << "using return-valued form of _fsync" << dendl;
+  }
+  
+  if (!syncdataonly && in->dirty_caps) {
+    check_caps(in, CHECK_CAPS_NODELAY|CHECK_CAPS_SYNCHRONOUS);
+    if (in->flushing_caps)
+      flush_tid = last_flush_tid;
+  } else ldout(cct, 10) << "no metadata needs to commit" << dendl;
+
+  if (!syncdataonly && !in->unsafe_ops.empty()) {
+    flush_mdlog_sync(in);
+
+    MetaRequest *req = in->unsafe_ops.back();
+    ldout(cct, 15) << "waiting on unsafe requests, last tid " << req->get_tid() <<  dendl;
+
+    req->get();
+    wait_on_list(req->waitfor_safe);
+    put_request(req);
+  }
+
+  if (nullptr != object_cacher_completion) { // wait on a real reply instead of guessing
+    client_lock.unlock();
+    ldout(cct, 15) << "waiting on data to flush" << dendl;
+    r = object_cacher_completion->wait();
+    client_lock.lock();
+    ldout(cct, 15) << "got " << r << " from flush writeback" << dendl;
+  } else {
+    // FIXME: this can starve
+    while (in->cap_refs[CEPH_CAP_FILE_BUFFER] > 0) {
+      ldout(cct, 10) << "ino " << in->ino << " has " << in->cap_refs[CEPH_CAP_FILE_BUFFER]
+		     << " uncommitted, waiting" << dendl;
+      wait_on_list(in->waitfor_commit);
+    }
+  }
+
+  if (!r) {
+    if (flush_tid > 0)
+      wait_sync_caps(in, flush_tid);
+
+    ldout(cct, 10) << "ino " << in->ino << " has no uncommitted writes" << dendl;
+  } else {
+    ldout(cct, 8) << "ino " << in->ino << " failed to commit to disk! "
+		  << cpp_strerror(-r) << dendl;
+  }
+   
+  lat = ceph_clock_now();
+  lat -= start;
+  logger->tinc(l_c_fsync, lat);
+
+  return r;
+}
+
+int Client::_fsync(Fh *f, bool syncdataonly)
+{
+  ldout(cct, 8) << "_fsync(" << f << ", " << (syncdataonly ? "dataonly)":"data+metadata)") << dendl;
+  return _fsync(f->inode.get(), syncdataonly);
+}
+
+int Client::fstat(int fd, struct stat *stbuf, const UserPerm& perms, int mask)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "fstat mask " << hex << mask << dec << std::endl;
+  tout(cct) << fd << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  int r = _getattr(f->inode, mask, perms);
+  if (r < 0)
+    return r;
+  fill_stat(f->inode, stbuf, NULL);
+  ldout(cct, 5) << "fstat(" << fd << ", " << stbuf << ") = " << r << dendl;
+  return r;
+}
+
+int Client::fstatx(int fd, struct ceph_statx *stx, const UserPerm& perms,
+		   unsigned int want, unsigned int flags)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "fstatx flags " << hex << flags << " want " << want << dec << std::endl;
+  tout(cct) << fd << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+
+  unsigned mask = statx_to_mask(flags, want);
+
+  int r = 0;
+  if (mask) {
+    r = _getattr(f->inode, mask, perms);
+    if (r < 0) {
+      ldout(cct, 3) << "fstatx exit on error!" << dendl;
+      return r;
+    }
+  }
+
+  fill_statx(f->inode, mask, stx);
+  ldout(cct, 3) << "fstatx(" << fd << ", " << stx << ") = " << r << dendl;
+  return r;
+}
+
+int Client::statxat(int dirfd, const char *relpath,
+                    struct ceph_statx *stx, const UserPerm& perms,
+                    unsigned int want, unsigned int flags) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  tout(cct) << __func__ << " flags " << hex << flags << " want " << want << dec << std::endl;
+  tout(cct) << dirfd << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  unsigned mask = statx_to_mask(flags, want);
+
+  InodeRef dirinode;
+  std::scoped_lock lock(client_lock);
+  int r = get_fd_inode(dirfd, &dirinode);
+  if (r < 0) {
+    return r;
+  }
+
+  InodeRef in;
+  filepath path(relpath);
+  r = path_walk(path, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW), mask, dirinode);
+  if (r < 0) {
+    return r;
+  }
+  r = _getattr(in, mask, perms);
+  if (r < 0) {
+    ldout(cct, 3) << __func__ << " exit on error!" << dendl;
+    return r;
+  }
+
+  fill_statx(in, mask, stx);
+  ldout(cct, 3) << __func__ << " dirfd" << dirfd << ", r= " << r << dendl;
+  return r;
+}
+
+// not written yet, but i want to link!
+
+int Client::chdir(const char *relpath, std::string &new_cwd,
+		  const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "chdir" << std::endl;
+  tout(cct) << relpath << std::endl;
+
+  filepath path(relpath);
+  InodeRef in;
+
+  std::scoped_lock lock(client_lock);
+  int r = path_walk(path, &in, perms);
+  if (r < 0)
+    return r;
+
+  if (!(in.get()->is_dir()))
+    return -CEPHFS_ENOTDIR;
+
+  if (cwd != in)
+    cwd.swap(in);
+  ldout(cct, 3) << "chdir(" << relpath << ")  cwd now " << cwd->ino << dendl;
+
+  _getcwd(new_cwd, perms);
+  return 0;
+}
+
+void Client::_getcwd(string& dir, const UserPerm& perms)
+{
+  filepath path;
+  ldout(cct, 10) << __func__ << " " << *cwd << dendl;
+
+  Inode *in = cwd.get();
+  while (in != root.get()) {
+    ceph_assert(in->dentries.size() < 2); // dirs can't be hard-linked
+
+    // A cwd or ancester is unlinked
+    if (in->dentries.empty()) {
+      return;
+    }
+
+    Dentry *dn = in->get_first_parent();
+
+
+    if (!dn) {
+      // look it up
+      ldout(cct, 10) << __func__ << " looking up parent for " << *in << dendl;
+      MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LOOKUPNAME);
+      filepath path(in->ino);
+      req->set_filepath(path);
+      req->set_inode(in);
+      int res = make_request(req, perms);
+      if (res < 0)
+	break;
+
+      // start over
+      path = filepath();
+      in = cwd.get();
+      continue;
+    }
+    path.push_front_dentry(dn->name);
+    in = dn->dir->parent_inode;
+  }
+  dir = "/";
+  dir += path.get_path();
+}
+
+void Client::getcwd(string& dir, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return;
+
+  std::scoped_lock l(client_lock);
+
+  _getcwd(dir, perms);
+}
+
+int Client::statfs(const char *path, struct statvfs *stbuf,
+		   const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << std::endl;
+  unsigned long int total_files_on_fs;
+
+  ceph_statfs stats;
+  C_SaferCond cond;
+
+  std::unique_lock lock(client_lock);
+  const vector<int64_t> &data_pools = mdsmap->get_data_pools();
+  if (data_pools.size() == 1) {
+    objecter->get_fs_stats(stats, data_pools[0], &cond);
+  } else {
+    objecter->get_fs_stats(stats, boost::optional<int64_t>(), &cond);
+  }
+
+  lock.unlock();
+  int rval = cond.wait();
+  lock.lock();
+
+  assert(root);
+  total_files_on_fs = root->rstat.rfiles + root->rstat.rsubdirs;
+
+  if (rval < 0) {
+    ldout(cct, 1) << "underlying call to statfs returned error: "
+                  << cpp_strerror(rval)
+                  << dendl;
+    return rval;
+  }
+
+  memset(stbuf, 0, sizeof(*stbuf));
+
+  /*
+   * we're going to set a block size of 4MB so we can represent larger
+   * FSes without overflowing. Additionally convert the space
+   * measurements from KB to bytes while making them in terms of
+   * blocks.  We use 4MB only because it is big enough, and because it
+   * actually *is* the (ceph) default block size.
+   */
+  const int CEPH_BLOCK_SHIFT = 22;
+  stbuf->f_frsize = 1 << CEPH_BLOCK_SHIFT;
+  stbuf->f_bsize = 1 << CEPH_BLOCK_SHIFT;
+  stbuf->f_files = total_files_on_fs;
+  stbuf->f_ffree = -1;
+  stbuf->f_favail = -1;
+  stbuf->f_fsid = -1;       // ??
+  stbuf->f_flag = 0;        // ??
+  stbuf->f_namemax = NAME_MAX;
+
+  // Usually quota_root will == root_ancestor, but if the mount root has no
+  // quota but we can see a parent of it that does have a quota, we'll
+  // respect that one instead.
+  ceph_assert(root != nullptr);
+  InodeRef quota_root = root->quota.is_enable() ? root : get_quota_root(root.get(), perms);
+
+  // get_quota_root should always give us something if client quotas are
+  // enabled
+  ceph_assert(cct->_conf.get_val<bool>("client_quota") == false || quota_root != nullptr);
+
+  if (quota_root && cct->_conf->client_quota_df && quota_root->quota.max_bytes) {
+
+    // Skip the getattr if any sessions are stale, as we don't want to
+    // block `df` if this client has e.g. been evicted, or if the MDS cluster
+    // is unhealthy.
+    if (!_any_stale_sessions()) {
+      int r = _getattr(quota_root, 0, perms, true);
+      if (r != 0) {
+        // Ignore return value: error getting latest inode metadata is not a good
+        // reason to break "df".
+        lderr(cct) << "Error in getattr on quota root 0x"
+                   << std::hex << quota_root->ino << std::dec
+                   << " statfs result may be outdated" << dendl;
+      }
+    }
+
+    // Special case: if there is a size quota set on the Inode acting
+    // as the root for this client mount, then report the quota status
+    // as the filesystem statistics.
+    const fsblkcnt_t total = quota_root->quota.max_bytes >> CEPH_BLOCK_SHIFT;
+    const fsblkcnt_t used = quota_root->rstat.rbytes >> CEPH_BLOCK_SHIFT;
+    // It is possible for a quota to be exceeded: arithmetic here must
+    // handle case where used > total.
+    const fsblkcnt_t free = total > used ? total - used : 0;
+
+    stbuf->f_blocks = total;
+    stbuf->f_bfree = free;
+    stbuf->f_bavail = free;
+  } else {
+    // General case: report the cluster statistics returned from RADOS. Because
+    // multiple pools may be used without one filesystem namespace via
+    // layouts, this is the most correct thing we can do.
+    stbuf->f_blocks = stats.kb >> (CEPH_BLOCK_SHIFT - 10);
+    stbuf->f_bfree = stats.kb_avail >> (CEPH_BLOCK_SHIFT - 10);
+    stbuf->f_bavail = stats.kb_avail >> (CEPH_BLOCK_SHIFT - 10);
+  }
+
+  return rval;
+}
+
+int Client::_do_filelock(Inode *in, Fh *fh, int lock_type, int op, int sleep,
+			 struct flock *fl, uint64_t owner, bool removing)
+{
+  ldout(cct, 10) << __func__ << " ino " << in->ino
+		 << (lock_type == CEPH_LOCK_FCNTL ? " fcntl" : " flock")
+		 << " type " << fl->l_type << " owner " << owner
+		 << " " << fl->l_start << "~" << fl->l_len << dendl;
+
+  if (in->flags & I_ERROR_FILELOCK)
+    return -CEPHFS_EIO;
+
+  int lock_cmd;
+  if (F_RDLCK == fl->l_type)
+    lock_cmd = CEPH_LOCK_SHARED;
+  else if (F_WRLCK == fl->l_type)
+    lock_cmd = CEPH_LOCK_EXCL;
+  else if (F_UNLCK == fl->l_type)
+    lock_cmd = CEPH_LOCK_UNLOCK;
+  else
+    return -CEPHFS_EIO;
+
+  if (op != CEPH_MDS_OP_SETFILELOCK || lock_cmd == CEPH_LOCK_UNLOCK)
+    sleep = 0;
+
+  /*
+   * Set the most significant bit, so that MDS knows the 'owner'
+   * is sufficient to identify the owner of lock. (old code uses
+   * both 'owner' and 'pid')
+   */
+  owner |= (1ULL << 63);
+
+  MetaRequest *req = new MetaRequest(op);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_inode(in);
+
+  req->head.args.filelock_change.rule = lock_type;
+  req->head.args.filelock_change.type = lock_cmd;
+  req->head.args.filelock_change.owner = owner;
+  req->head.args.filelock_change.pid = fl->l_pid;
+  req->head.args.filelock_change.start = fl->l_start;
+  req->head.args.filelock_change.length = fl->l_len;
+  req->head.args.filelock_change.wait = sleep;
+
+  int ret;
+  bufferlist bl;
+
+  if (sleep && switch_interrupt_cb) {
+    // enable interrupt
+    switch_interrupt_cb(callback_handle, req->get());
+    ret = make_request(req, fh->actor_perms, NULL, NULL, -1, &bl);
+    // disable interrupt
+    switch_interrupt_cb(callback_handle, NULL);
+    if (ret == 0 && req->aborted()) {
+      // effect of this lock request has been revoked by the 'lock intr' request
+      ret = req->get_abort_code();
+    }
+    put_request(req);
+  } else {
+    ret = make_request(req, fh->actor_perms, NULL, NULL, -1, &bl);
+  }
+
+  if (ret == 0) {
+    if (op == CEPH_MDS_OP_GETFILELOCK) {
+      ceph_filelock filelock;
+      auto p = bl.cbegin();
+      decode(filelock, p);
+
+      if (CEPH_LOCK_SHARED == filelock.type)
+	fl->l_type = F_RDLCK;
+      else if (CEPH_LOCK_EXCL == filelock.type)
+	fl->l_type = F_WRLCK;
+      else
+	fl->l_type = F_UNLCK;
+
+      fl->l_whence = SEEK_SET;
+      fl->l_start = filelock.start;
+      fl->l_len = filelock.length;
+      fl->l_pid = filelock.pid;
+    } else if (op == CEPH_MDS_OP_SETFILELOCK) {
+      ceph_lock_state_t *lock_state;
+      if (lock_type == CEPH_LOCK_FCNTL) {
+	if (!in->fcntl_locks)
+	  in->fcntl_locks.reset(new ceph_lock_state_t(cct, CEPH_LOCK_FCNTL));
+	lock_state = in->fcntl_locks.get();
+      } else if (lock_type == CEPH_LOCK_FLOCK) {
+	if (!in->flock_locks)
+	  in->flock_locks.reset(new ceph_lock_state_t(cct, CEPH_LOCK_FLOCK));
+	lock_state = in->flock_locks.get();
+      } else {
+	ceph_abort();
+	return -CEPHFS_EINVAL;
+      }
+      _update_lock_state(fl, owner, lock_state);
+
+      if (!removing) {
+	if (lock_type == CEPH_LOCK_FCNTL) {
+	  if (!fh->fcntl_locks)
+	    fh->fcntl_locks.reset(new ceph_lock_state_t(cct, CEPH_LOCK_FCNTL));
+	  lock_state = fh->fcntl_locks.get();
+	} else {
+	  if (!fh->flock_locks)
+	    fh->flock_locks.reset(new ceph_lock_state_t(cct, CEPH_LOCK_FLOCK));
+	  lock_state = fh->flock_locks.get();
+	}
+	_update_lock_state(fl, owner, lock_state);
+      }
+    } else
+      ceph_abort();
+  }
+  return ret;
+}
+
+int Client::_interrupt_filelock(MetaRequest *req)
+{
+  // Set abort code, but do not kick. The abort code prevents the request
+  // from being re-sent.
+  req->abort(-CEPHFS_EINTR);
+  if (req->mds < 0)
+    return 0; // haven't sent the request
+
+  Inode *in = req->inode();
+
+  int lock_type;
+  if (req->head.args.filelock_change.rule == CEPH_LOCK_FLOCK)
+    lock_type = CEPH_LOCK_FLOCK_INTR;
+  else if (req->head.args.filelock_change.rule == CEPH_LOCK_FCNTL)
+    lock_type = CEPH_LOCK_FCNTL_INTR;
+  else {
+    ceph_abort();
+    return -CEPHFS_EINVAL;
+  }
+
+  MetaRequest *intr_req = new MetaRequest(CEPH_MDS_OP_SETFILELOCK);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  intr_req->set_filepath(path);
+  intr_req->set_inode(in);
+  intr_req->head.args.filelock_change = req->head.args.filelock_change;
+  intr_req->head.args.filelock_change.rule = lock_type;
+  intr_req->head.args.filelock_change.type = CEPH_LOCK_UNLOCK;
+
+  UserPerm perms(req->get_uid(), req->get_gid());
+  return make_request(intr_req, perms, NULL, NULL, -1);
+}
+
+void Client::_encode_filelocks(Inode *in, bufferlist& bl)
+{
+  if (!in->fcntl_locks && !in->flock_locks)
+    return;
+
+  unsigned nr_fcntl_locks = in->fcntl_locks ? in->fcntl_locks->held_locks.size() : 0;
+  encode(nr_fcntl_locks, bl);
+  if (nr_fcntl_locks) {
+    auto &lock_state = in->fcntl_locks;
+    for(multimap<uint64_t, ceph_filelock>::iterator p = lock_state->held_locks.begin();
+	p != lock_state->held_locks.end();
+	++p)
+      encode(p->second, bl);
+  }
+
+  unsigned nr_flock_locks = in->flock_locks ? in->flock_locks->held_locks.size() : 0;
+  encode(nr_flock_locks, bl);
+  if (nr_flock_locks) {
+    auto &lock_state = in->flock_locks;
+    for(multimap<uint64_t, ceph_filelock>::iterator p = lock_state->held_locks.begin();
+	p != lock_state->held_locks.end();
+	++p)
+      encode(p->second, bl);
+  }
+
+  ldout(cct, 10) << __func__ << " ino " << in->ino << ", " << nr_fcntl_locks
+		 << " fcntl locks, " << nr_flock_locks << " flock locks" <<  dendl;
+}
+
+void Client::_release_filelocks(Fh *fh)
+{
+  if (!fh->fcntl_locks && !fh->flock_locks)
+    return;
+
+  Inode *in = fh->inode.get();
+  ldout(cct, 10) << __func__ << " " << fh << " ino " << in->ino << dendl;
+
+  list<ceph_filelock> activated_locks;
+
+  list<pair<int, ceph_filelock> > to_release;
+
+  if (fh->fcntl_locks) {
+    auto &lock_state = fh->fcntl_locks;
+    for(auto p = lock_state->held_locks.begin(); p != lock_state->held_locks.end(); ) {
+      auto q = p++;
+      if (in->flags & I_ERROR_FILELOCK) {
+	lock_state->remove_lock(q->second, activated_locks);
+      } else {
+	to_release.push_back(pair<int, ceph_filelock>(CEPH_LOCK_FCNTL, q->second));
+      }
+    }
+    lock_state.reset();
+  }
+  if (fh->flock_locks) {
+    auto &lock_state = fh->flock_locks;
+    for(auto p = lock_state->held_locks.begin(); p != lock_state->held_locks.end(); ) {
+      auto q = p++;
+      if (in->flags & I_ERROR_FILELOCK) {
+	lock_state->remove_lock(q->second, activated_locks);
+      } else {
+	to_release.push_back(pair<int, ceph_filelock>(CEPH_LOCK_FLOCK, q->second));
+      }
+    }
+    lock_state.reset();
+  }
+
+  if ((in->flags & I_ERROR_FILELOCK) && !in->has_any_filelocks())
+    in->flags &= ~I_ERROR_FILELOCK;
+
+  if (to_release.empty())
+    return;
+
+  struct flock fl;
+  memset(&fl, 0, sizeof(fl));
+  fl.l_whence = SEEK_SET;
+  fl.l_type = F_UNLCK;
+
+  for (list<pair<int, ceph_filelock> >::iterator p = to_release.begin();
+       p != to_release.end();
+       ++p) {
+    fl.l_start = p->second.start;
+    fl.l_len = p->second.length;
+    fl.l_pid = p->second.pid;
+    _do_filelock(in, fh, p->first, CEPH_MDS_OP_SETFILELOCK, 0, &fl,
+		 p->second.owner, true);
+  }
+}
+
+void Client::_update_lock_state(struct flock *fl, uint64_t owner,
+				ceph_lock_state_t *lock_state)
+{
+  int lock_cmd;
+  if (F_RDLCK == fl->l_type)
+    lock_cmd = CEPH_LOCK_SHARED;
+  else if (F_WRLCK == fl->l_type)
+    lock_cmd = CEPH_LOCK_EXCL;
+  else
+    lock_cmd = CEPH_LOCK_UNLOCK;;
+
+  ceph_filelock filelock;
+  filelock.start = fl->l_start;
+  filelock.length = fl->l_len;
+  filelock.client = 0;
+  // see comment in _do_filelock()
+  filelock.owner = owner | (1ULL << 63);
+  filelock.pid = fl->l_pid;
+  filelock.type = lock_cmd;
+
+  if (filelock.type == CEPH_LOCK_UNLOCK) {
+    list<ceph_filelock> activated_locks;
+    lock_state->remove_lock(filelock, activated_locks);
+  } else {
+    bool r = lock_state->add_lock(filelock, false, false, NULL);
+    ceph_assert(r);
+  }
+}
+
+int Client::_getlk(Fh *fh, struct flock *fl, uint64_t owner)
+{
+  Inode *in = fh->inode.get();
+  ldout(cct, 10) << "_getlk " << fh << " ino " << in->ino << dendl;
+  int ret = _do_filelock(in, fh, CEPH_LOCK_FCNTL, CEPH_MDS_OP_GETFILELOCK, 0, fl, owner);
+  return ret;
+}
+
+int Client::_setlk(Fh *fh, struct flock *fl, uint64_t owner, int sleep)
+{
+  Inode *in = fh->inode.get();
+  ldout(cct, 10) << "_setlk " << fh << " ino " << in->ino << dendl;
+  int ret =  _do_filelock(in, fh, CEPH_LOCK_FCNTL, CEPH_MDS_OP_SETFILELOCK, sleep, fl, owner);
+  ldout(cct, 10) << "_setlk " << fh << " ino " << in->ino << " result=" << ret << dendl;
+  return ret;
+}
+
+int Client::_flock(Fh *fh, int cmd, uint64_t owner)
+{
+  Inode *in = fh->inode.get();
+  ldout(cct, 10) << "_flock " << fh << " ino " << in->ino << dendl;
+
+  int sleep = !(cmd & LOCK_NB);
+  cmd &= ~LOCK_NB;
+
+  int type;
+  switch (cmd) {
+    case LOCK_SH:
+      type = F_RDLCK;
+      break;
+    case LOCK_EX:
+      type = F_WRLCK;
+      break;
+    case LOCK_UN:
+      type = F_UNLCK;
+      break;
+    default:
+      return -CEPHFS_EINVAL;
+  }
+
+  struct flock fl;
+  memset(&fl, 0, sizeof(fl));
+  fl.l_type = type;
+  fl.l_whence = SEEK_SET;
+
+  int ret =  _do_filelock(in, fh, CEPH_LOCK_FLOCK, CEPH_MDS_OP_SETFILELOCK, sleep, &fl, owner);
+  ldout(cct, 10) << "_flock " << fh << " ino " << in->ino << " result=" << ret << dendl;
+  return ret;
+}
+
+int Client::get_snap_info(const char *path, const UserPerm &perms, SnapInfo *snap_info) {
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied()) {
+    return -CEPHFS_ENOTCONN;
+  }
+
+  std::unique_lock locker(client_lock);
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, true);
+  if (r < 0) {
+    return r;
+  }
+
+  if (in->snapid == CEPH_NOSNAP) {
+    return -CEPHFS_EINVAL;
+  }
+
+  snap_info->id = in->snapid;
+  snap_info->metadata = in->snap_metadata;
+  return 0;
+}
+
+int Client::ll_statfs(Inode *in, struct statvfs *stbuf, const UserPerm& perms)
+{
+  /* Since the only thing this does is wrap a call to statfs, and
+     statfs takes a lock, it doesn't seem we have a need to split it
+     out. */
+  return statfs(0, stbuf, perms);
+}
+
+void Client::ll_register_callbacks(struct ceph_client_callback_args *args)
+{
+  if (!args)
+    return;
+  std::scoped_lock l(client_lock);
+  ldout(cct, 10) << __func__ << " cb " << args->handle
+		 << " invalidate_ino_cb " << args->ino_cb
+		 << " invalidate_dentry_cb " << args->dentry_cb
+		 << " switch_interrupt_cb " << args->switch_intr_cb
+		 << " remount_cb " << args->remount_cb
+		 << dendl;
+  callback_handle = args->handle;
+  if (args->ino_cb) {
+    ino_invalidate_cb = args->ino_cb;
+    async_ino_invalidator.start();
+  }
+  if (args->dentry_cb) {
+    dentry_invalidate_cb = args->dentry_cb;
+    async_dentry_invalidator.start();
+  }
+  if (args->switch_intr_cb) {
+    switch_interrupt_cb = args->switch_intr_cb;
+    interrupt_finisher.start();
+  }
+  if (args->remount_cb) {
+    remount_cb = args->remount_cb;
+    remount_finisher.start();
+  }
+  if (args->ino_release_cb) {
+    ino_release_cb = args->ino_release_cb;
+    async_ino_releasor.start();
+  }
+  if (args->umask_cb)
+    umask_cb = args->umask_cb;
+}
+
+std::pair<int, bool> Client::test_dentry_handling(bool can_invalidate)
+{
+  std::pair <int, bool> r(0, false);
+
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  if (!iref_reader.is_state_satisfied())
+    return std::make_pair(-CEPHFS_ENOTCONN, false);
+
+  can_invalidate_dentries = can_invalidate;
+
+  if (can_invalidate_dentries) {
+    ceph_assert(dentry_invalidate_cb);
+    ldout(cct, 1) << "using dentry_invalidate_cb" << dendl;
+  } else {
+    ceph_assert(remount_cb);
+    ldout(cct, 1) << "using remount_cb" << dendl;
+    r = _do_remount(false);
+  }
+
+  return r;
+}
+
+int Client::_sync_fs()
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  ldout(cct, 10) << __func__ << dendl;
+
+  // flush file data
+  std::unique_ptr<C_SaferCond> cond = nullptr; 
+  if (cct->_conf->client_oc) {
+    cond.reset(new C_SaferCond("Client::_sync_fs:lock"));
+    objectcacher->flush_all(cond.get());
+  }
+
+  // flush caps
+  flush_caps_sync();
+  ceph_tid_t flush_tid = last_flush_tid;
+
+  // wait for unsafe mds requests
+  wait_unsafe_requests();
+
+  wait_sync_caps(flush_tid);
+
+  if (nullptr != cond) {
+    client_lock.unlock();
+    ldout(cct, 15) << __func__ << " waiting on data to flush" << dendl;
+    cond->wait();
+    ldout(cct, 15) << __func__ << " flush finished" << dendl;
+    client_lock.lock();
+  }
+
+  return 0;
+}
+
+int Client::sync_fs()
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock l(client_lock);
+
+  return _sync_fs();
+}
+
+int64_t Client::drop_caches()
+{
+  std::scoped_lock l(client_lock);
+  return objectcacher->release_all();
+}
+
+int Client::_lazyio(Fh *fh, int enable)
+{
+  Inode *in = fh->inode.get();
+  ldout(cct, 20) << __func__ << " " << *in << " " << !!enable << dendl;
+
+  if (!!(fh->mode & CEPH_FILE_MODE_LAZY) == !!enable)
+    return 0;
+
+  int orig_mode = fh->mode;
+  if (enable) {
+    fh->mode |= CEPH_FILE_MODE_LAZY;
+    in->get_open_ref(fh->mode);
+    in->put_open_ref(orig_mode);
+    check_caps(in, CHECK_CAPS_NODELAY);
+  } else {
+    fh->mode &= ~CEPH_FILE_MODE_LAZY;
+    in->get_open_ref(fh->mode);
+    in->put_open_ref(orig_mode);
+    check_caps(in, 0);
+  }
+
+  return 0;
+}
+
+int Client::lazyio(int fd, int enable)
+{
+  std::scoped_lock l(client_lock);
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+
+  return _lazyio(f, enable);
+}
+
+int Client::ll_lazyio(Fh *fh, int enable)
+{
+  ldout(cct, 3) << __func__ << " " << fh << " " << fh->inode->ino << " " << !!enable << dendl;
+  tout(cct) << __func__ << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _lazyio(fh, enable);
+}
+
+int Client::lazyio_propagate(int fd, loff_t offset, size_t count)
+{
+  std::scoped_lock l(client_lock);
+  ldout(cct, 3) << "op: client->lazyio_propagate(" << fd
+          << ", " << offset << ", " << count << ")" << dendl;
+  
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+
+  // for now
+  _fsync(f, true);
+
+  return 0;
+}
+
+int Client::lazyio_synchronize(int fd, loff_t offset, size_t count)
+{
+  std::scoped_lock l(client_lock);
+  ldout(cct, 3) << "op: client->lazyio_synchronize(" << fd
+          << ", " << offset << ", " << count << ")" << dendl;
+  
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  Inode *in = f->inode.get();
+  
+  _fsync(f, true);
+  if (_release(in)) {
+    int r =_getattr(in, CEPH_STAT_CAP_SIZE, f->actor_perms);
+    if (r < 0) 
+      return r;
+  }
+  return 0;
+}
+
+
+// =============================
+// snaps
+
+int Client::mksnap(const char *relpath, const char *name, const UserPerm& perm,
+                   mode_t mode, const std::map<std::string, std::string> &metadata)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock l(client_lock);
+
+  filepath path(relpath);
+  InodeRef in;
+  int r = path_walk(path, &in, perm);
+  if (r < 0)
+    return r;
+  if (cct->_conf->client_permissions) {
+    r = may_create(in.get(), perm);
+    if (r < 0)
+      return r;
+  }
+  Inode *snapdir = open_snapdir(in.get());
+  return _mkdir(snapdir, name, mode, perm, nullptr, metadata);
+}
+
+int Client::rmsnap(const char *relpath, const char *name, const UserPerm& perms, bool check_perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock l(client_lock);
+
+  filepath path(relpath);
+  InodeRef in;
+  int r = path_walk(path, &in, perms);
+  if (r < 0)
+    return r;
+  Inode *snapdir = open_snapdir(in.get());
+  if (cct->_conf->client_permissions) {
+    r = may_delete(snapdir, check_perms ? name : NULL, perms);
+    if (r < 0)
+      return r;
+  }
+  return _rmdir(snapdir, name, perms);
+}
+
+// =============================
+// expose caps
+
+int Client::get_caps_issued(int fd)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+
+  return f->inode->caps_issued();
+}
+
+int Client::get_caps_issued(const char *path, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  filepath p(path);
+  InodeRef in;
+  int r = path_walk(p, &in, perms, true);
+  if (r < 0)
+    return r;
+  return in->caps_issued();
+}
+
+// =========================================
+// low level
+
+void Client::refresh_snapdir_attrs(Inode *in, Inode *diri) {
+  ldout(cct, 10) << __func__ << ": snapdir inode=" << *in
+                 << ", inode=" << *diri << dendl;
+  in->ino = diri->ino;
+  in->snapid = CEPH_SNAPDIR;
+  in->mode = diri->mode;
+  in->uid = diri->uid;
+  in->gid = diri->gid;
+  in->nlink = 1;
+  in->mtime = diri->mtime;
+  in->ctime = diri->ctime;
+  in->btime = diri->btime;
+  in->atime = diri->atime;
+  in->size = diri->size;
+  in->change_attr = diri->change_attr;
+
+  in->dirfragtree.clear();
+  in->snapdir_parent = diri;
+  // copy posix acls to snapshotted inode
+  in->xattrs.clear();
+  for (auto &[xattr_key, xattr_value] : diri->xattrs) {
+    if (xattr_key.rfind("system.", 0) == 0) {
+      in->xattrs[xattr_key] = xattr_value;
+    }
+  }
+}
+
+Inode *Client::open_snapdir(Inode *diri)
+{
+  Inode *in;
+  vinodeno_t vino(diri->ino, CEPH_SNAPDIR);
+  if (!inode_map.count(vino)) {
+    in = new Inode(this, vino, &diri->layout);
+    refresh_snapdir_attrs(in, diri);
+    diri->flags |= I_SNAPDIR_OPEN;
+    inode_map[vino] = in;
+    if (use_faked_inos())
+      _assign_faked_ino(in);
+    ldout(cct, 10) << "open_snapdir created snapshot inode " << *in << dendl;
+  } else {
+    in = inode_map[vino];
+    ldout(cct, 10) << "open_snapdir had snapshot inode " << *in << dendl;
+  }
+  return in;
+}
+
+int Client::ll_lookup(Inode *parent, const char *name, struct stat *attr,
+		      Inode **out, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+  ldout(cct, 3) << __func__ << " " << vparent << " " << name << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  int r = 0;
+  if (!fuse_default_permissions) {
+    if (strcmp(name, ".") && strcmp(name, "..")) {
+      r = may_lookup(parent, perms);
+      if (r < 0)
+	return r;
+    }
+  }
+
+  string dname(name);
+  InodeRef in;
+
+  r = _lookup(parent, dname, CEPH_STAT_CAP_INODE_ALL, &in, perms);
+  if (r < 0) {
+    attr->st_ino = 0;
+    goto out;
+  }
+
+  ceph_assert(in);
+  fill_stat(in, attr);
+  _ll_get(in.get());
+
+ out:
+  ldout(cct, 3) << __func__ << " " << vparent << " " << name
+	  << " -> " << r << " (" << hex << attr->st_ino << dec << ")" << dendl;
+  tout(cct) << attr->st_ino << std::endl;
+  *out = in.get();
+  return r;
+}
+
+int Client::ll_lookup_vino(
+    vinodeno_t vino,
+    const UserPerm& perms,
+    Inode **inode)
+{
+  ceph_assert(inode != NULL);
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  if (is_reserved_vino(vino))
+    return -CEPHFS_ESTALE;
+
+  std::scoped_lock lock(client_lock);
+  ldout(cct, 3) << __func__ << " " << vino << dendl;
+
+  // Check the cache first
+  unordered_map<vinodeno_t,Inode*>::iterator p = inode_map.find(vino);
+  if (p != inode_map.end()) {
+    *inode = p->second;
+    _ll_get(*inode);
+    return 0;
+  }
+
+  uint64_t snapid = vino.snapid;
+
+  // for snapdir, find the non-snapped dir inode
+  if (snapid == CEPH_SNAPDIR)
+    vino.snapid = CEPH_NOSNAP;
+
+  int r = _lookup_vino(vino, perms, inode);
+  if (r)
+    return r;
+  ceph_assert(*inode != NULL);
+
+  if (snapid == CEPH_SNAPDIR) {
+    Inode *tmp = *inode;
+
+    // open the snapdir and put the inode ref
+    *inode = open_snapdir(tmp);
+    _ll_forget(tmp, 1);
+    _ll_get(*inode);
+  }
+  return 0;
+}
+
+int Client::ll_lookup_inode(
+    struct inodeno_t ino,
+    const UserPerm& perms,
+    Inode **inode)
+{
+  vinodeno_t vino(ino, CEPH_NOSNAP);
+  return ll_lookup_vino(vino, perms, inode);
+}
+
+int Client::ll_lookupx(Inode *parent, const char *name, Inode **out,
+		       struct ceph_statx *stx, unsigned want, unsigned flags,
+		       const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+  ldout(cct, 3) << __func__ << " " << vparent << " " << name << dendl;
+  tout(cct) << "ll_lookupx" << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  int r = 0;
+  if (!fuse_default_permissions) {
+    r = may_lookup(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  string dname(name);
+  InodeRef in;
+
+  unsigned mask = statx_to_mask(flags, want);
+  r = _lookup(parent, dname, mask, &in, perms);
+  if (r < 0) {
+    stx->stx_ino = 0;
+    stx->stx_mask = 0;
+  } else {
+    ceph_assert(in);
+    fill_statx(in, mask, stx);
+    _ll_get(in.get());
+  }
+
+  ldout(cct, 3) << __func__ << " " << vparent << " " << name
+	  << " -> " << r << " (" << hex << stx->stx_ino << dec << ")" << dendl;
+  tout(cct) << stx->stx_ino << std::endl;
+  *out = in.get();
+  return r;
+}
+
+int Client::ll_walk(const char* name, Inode **out, struct ceph_statx *stx,
+		    unsigned int want, unsigned int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  filepath fp(name, 0);
+  InodeRef in;
+  int rc;
+  unsigned mask = statx_to_mask(flags, want);
+
+  ldout(cct, 3) << __func__ << " " << name << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  rc = path_walk(fp, &in, perms, !(flags & AT_SYMLINK_NOFOLLOW), mask);
+  if (rc < 0) {
+    /* zero out mask, just in case... */
+    stx->stx_mask = 0;
+    stx->stx_ino = 0;
+    *out = NULL;
+    return rc;
+  } else {
+    ceph_assert(in);
+    fill_statx(in, mask, stx);
+    _ll_get(in.get());
+    *out = in.get();
+    return 0;
+  }
+}
+
+void Client::_ll_get(Inode *in)
+{
+  if (in->ll_ref == 0) {
+    in->iget();
+    if (in->is_dir() && !in->dentries.empty()) {
+      ceph_assert(in->dentries.size() == 1); // dirs can't be hard-linked
+      in->get_first_parent()->get(); // pin dentry
+    }
+    if (in->snapid != CEPH_NOSNAP)
+      ll_snap_ref[in->snapid]++;
+  }
+  in->ll_get();
+  ldout(cct, 20) << __func__ << " " << in << " " << in->ino << " -> " << in->ll_ref << dendl;
+}
+
+int Client::_ll_put(Inode *in, uint64_t num)
+{
+  in->ll_put(num);
+  ldout(cct, 20) << __func__ << " " << in << " " << in->ino << " " << num << " -> " << in->ll_ref << dendl;
+  if (in->ll_ref == 0) {
+    if (in->is_dir() && !in->dentries.empty()) {
+      ceph_assert(in->dentries.size() == 1); // dirs can't be hard-linked
+      in->get_first_parent()->put(); // unpin dentry
+    }
+    if (in->snapid != CEPH_NOSNAP) {
+      auto p = ll_snap_ref.find(in->snapid);
+      ceph_assert(p != ll_snap_ref.end());
+      ceph_assert(p->second > 0);
+      if (--p->second == 0)
+	ll_snap_ref.erase(p);
+    }
+    put_inode(in);
+    return 0;
+  } else {
+    return in->ll_ref;
+  }
+}
+
+void Client::_ll_drop_pins()
+{
+  ldout(cct, 10) << __func__ << dendl;
+  std::set<InodeRef> to_be_put; //this set will be deconstructed item by item when exit
+  ceph::unordered_map<vinodeno_t, Inode*>::iterator next;
+  for (ceph::unordered_map<vinodeno_t, Inode*>::iterator it = inode_map.begin();
+       it != inode_map.end();
+       it = next) {
+    Inode *in = it->second;
+    next = it;
+    ++next;
+    if (in->ll_ref){
+      to_be_put.insert(in);
+      _ll_put(in, in->ll_ref);
+    }
+  }
+}
+
+bool Client::_ll_forget(Inode *in, uint64_t count)
+{
+  inodeno_t ino = in->ino;
+
+  ldout(cct, 8) << __func__ << " " << ino << " " << count << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << ino.val << std::endl;
+  tout(cct) << count << std::endl;
+
+  // Ignore forget if we're no longer mounted
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return true;
+
+  if (ino == 1) return true;  // ignore forget on root.
+
+  bool last = false;
+  if (in->ll_ref < count) {
+    ldout(cct, 1) << "WARNING: ll_forget on " << ino << " " << count
+		  << ", which only has ll_ref=" << in->ll_ref << dendl;
+    _ll_put(in, in->ll_ref);
+    last = true;
+  } else {
+    if (_ll_put(in, count) == 0)
+      last = true;
+  }
+
+  return last;
+}
+
+bool Client::ll_forget(Inode *in, uint64_t count)
+{
+  std::scoped_lock lock(client_lock);
+  return _ll_forget(in, count);
+}
+
+bool Client::ll_put(Inode *in)
+{
+  /* ll_forget already takes the lock */
+  return ll_forget(in, 1);
+}
+
+int Client::ll_get_snap_ref(snapid_t snap)
+{
+  std::scoped_lock lock(client_lock);
+  auto p = ll_snap_ref.find(snap);
+  if (p != ll_snap_ref.end())
+    return p->second;
+  return 0;
+}
+
+snapid_t Client::ll_get_snapid(Inode *in)
+{
+  std::scoped_lock lock(client_lock);
+  return in->snapid;
+}
+
+Inode *Client::ll_get_inode(ino_t ino)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return NULL;
+
+  std::scoped_lock lock(client_lock);
+
+  vinodeno_t vino = _map_faked_ino(ino);
+  unordered_map<vinodeno_t,Inode*>::iterator p = inode_map.find(vino);
+  if (p == inode_map.end())
+    return NULL;
+  Inode *in = p->second;
+  _ll_get(in);
+  return in;
+}
+
+Inode *Client::ll_get_inode(vinodeno_t vino)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return NULL;
+
+  if (is_reserved_vino(vino))
+    return NULL;
+
+  std::scoped_lock lock(client_lock);
+
+  unordered_map<vinodeno_t,Inode*>::iterator p = inode_map.find(vino);
+  if (p == inode_map.end())
+    return NULL;
+  Inode *in = p->second;
+  _ll_get(in);
+  return in;
+}
+
+int Client::_ll_getattr(Inode *in, int caps, const UserPerm& perms)
+{
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 8) << __func__ << " " << vino << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+
+  if (vino.snapid < CEPH_NOSNAP)
+    return 0;
+  else
+    return _getattr(in, caps, perms);
+}
+
+int Client::ll_getattr(Inode *in, struct stat *attr, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  int res = _ll_getattr(in, CEPH_STAT_CAP_INODE_ALL, perms);
+
+  if (res == 0)
+    fill_stat(in, attr);
+  ldout(cct, 3) << __func__ << " " << _get_vino(in) << " = " << res << dendl;
+  return res;
+}
+
+int Client::ll_getattrx(Inode *in, struct ceph_statx *stx, unsigned int want,
+			unsigned int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  int res = 0;
+  unsigned mask = statx_to_mask(flags, want);
+
+  if (mask && !in->caps_issued_mask(mask, true))
+    res = _ll_getattr(in, mask, perms);
+
+  if (res == 0)
+    fill_statx(in, mask, stx);
+  ldout(cct, 3) << __func__ << " " << _get_vino(in) << " = " << res << dendl;
+  return res;
+}
+
+int Client::_ll_setattrx(Inode *in, struct ceph_statx *stx, int mask,
+			 const UserPerm& perms, InodeRef *inp)
+{
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 8) << __func__ << " " << vino << " mask " << hex << mask << dec
+		<< dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << stx->stx_mode << std::endl;
+  tout(cct) << stx->stx_uid << std::endl;
+  tout(cct) << stx->stx_gid << std::endl;
+  tout(cct) << stx->stx_size << std::endl;
+  tout(cct) << stx->stx_mtime << std::endl;
+  tout(cct) << stx->stx_atime << std::endl;
+  tout(cct) << stx->stx_btime << std::endl;
+  tout(cct) << mask << std::endl;
+
+  if (!fuse_default_permissions) {
+    int res = may_setattr(in, stx, mask, perms);
+    if (res < 0)
+      return res;
+  }
+
+  mask &= ~(CEPH_SETATTR_MTIME_NOW | CEPH_SETATTR_ATIME_NOW);
+
+  return __setattrx(in, stx, mask, perms, inp);
+}
+
+int Client::ll_setattrx(Inode *in, struct ceph_statx *stx, int mask,
+			const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef target(in);
+  int res = _ll_setattrx(in, stx, mask, perms, &target);
+  if (res == 0) {
+    ceph_assert(in == target.get());
+    fill_statx(in, in->caps_issued(), stx);
+  }
+
+  ldout(cct, 3) << __func__ << " " << _get_vino(in) << " = " << res << dendl;
+  return res;
+}
+
+int Client::ll_setattr(Inode *in, struct stat *attr, int mask,
+		       const UserPerm& perms)
+{
+  struct ceph_statx stx;
+  stat_to_statx(attr, &stx);
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef target(in);
+  int res = _ll_setattrx(in, &stx, mask, perms, &target);
+  if (res == 0) {
+    ceph_assert(in == target.get());
+    fill_stat(in, attr);
+  }
+
+  ldout(cct, 3) << __func__ << " " << _get_vino(in) << " = " << res << dendl;
+  return res;
+}
+
+
+// ----------
+// xattrs
+
+int Client::getxattr(const char *path, const char *name, void *value, size_t size,
+		     const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, true, CEPH_STAT_CAP_XATTR);
+  if (r < 0)
+    return r;
+  return _getxattr(in, name, value, size, perms);
+}
+
+int Client::lgetxattr(const char *path, const char *name, void *value, size_t size,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, false, CEPH_STAT_CAP_XATTR);
+  if (r < 0)
+    return r;
+  return _getxattr(in, name, value, size, perms);
+}
+
+int Client::fgetxattr(int fd, const char *name, void *value, size_t size,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  return _getxattr(f->inode, name, value, size, perms);
+}
+
+int Client::listxattr(const char *path, char *list, size_t size,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, true, CEPH_STAT_CAP_XATTR);
+  if (r < 0)
+    return r;
+  return Client::_listxattr(in.get(), list, size, perms);
+}
+
+int Client::llistxattr(const char *path, char *list, size_t size,
+		       const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, false, CEPH_STAT_CAP_XATTR);
+  if (r < 0)
+    return r;
+  return Client::_listxattr(in.get(), list, size, perms);
+}
+
+int Client::flistxattr(int fd, char *list, size_t size, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  return Client::_listxattr(f->inode.get(), list, size, perms);
+}
+
+int Client::removexattr(const char *path, const char *name,
+			const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, true);
+  if (r < 0)
+    return r;
+  return _removexattr(in, name, perms);
+}
+
+int Client::lremovexattr(const char *path, const char *name,
+			 const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, false);
+  if (r < 0)
+    return r;
+  return _removexattr(in, name, perms);
+}
+
+int Client::fremovexattr(int fd, const char *name, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  return _removexattr(f->inode, name, perms);
+}
+
+int Client::setxattr(const char *path, const char *name, const void *value,
+		     size_t size, int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  _setxattr_maybe_wait_for_osdmap(name, value, size);
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, true);
+  if (r < 0)
+    return r;
+  return _setxattr(in, name, value, size, flags, perms);
+}
+
+int Client::lsetxattr(const char *path, const char *name, const void *value,
+		      size_t size, int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  _setxattr_maybe_wait_for_osdmap(name, value, size);
+
+  std::scoped_lock lock(client_lock);
+
+  InodeRef in;
+  int r = Client::path_walk(path, &in, perms, false);
+  if (r < 0)
+    return r;
+  return _setxattr(in, name, value, size, flags, perms);
+}
+
+int Client::fsetxattr(int fd, const char *name, const void *value, size_t size,
+		      int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  _setxattr_maybe_wait_for_osdmap(name, value, size);
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  return _setxattr(f->inode, name, value, size, flags, perms);
+}
+
+int Client::_getxattr(Inode *in, const char *name, void *value, size_t size,
+		      const UserPerm& perms)
+{
+  int r;
+  const VXattr *vxattr = nullptr;
+
+  vxattr = _match_vxattr(in, name);
+  if (vxattr) {
+    r = -CEPHFS_ENODATA;
+
+    // Do a force getattr to get the latest quota before returning
+    // a value to userspace.
+    int flags = 0;
+    if (vxattr->flags & VXATTR_RSTAT) {
+      flags |= CEPH_STAT_RSTAT;
+    }
+    if (vxattr->flags & VXATTR_DIRSTAT) {
+      flags |= CEPH_CAP_FILE_SHARED;
+    }
+    r = _getattr(in, flags | CEPH_STAT_CAP_XATTR, perms, true);
+    if (r != 0) {
+      // Error from getattr!
+      return r;
+    }
+
+    // call pointer-to-member function
+    char buf[256];
+    if (!(vxattr->exists_cb && !(this->*(vxattr->exists_cb))(in))) {
+      r = (this->*(vxattr->getxattr_cb))(in, buf, sizeof(buf));
+    } else {
+      r = -CEPHFS_ENODATA;
+    }
+
+    if (size != 0) {
+      if (r > (int)size) {
+	r = -CEPHFS_ERANGE;
+      } else if (r > 0) {
+	memcpy(value, buf, r);
+      }
+    }
+    goto out;
+  }
+
+  if (!strncmp(name, "ceph.", 5)) {
+    r = _getvxattr(in, perms, name, size, value, MDS_RANK_NONE);
+    goto out;
+  }
+
+  if (acl_type == NO_ACL && !strncmp(name, "system.", 7)) {
+    r = -CEPHFS_EOPNOTSUPP;
+    goto out;
+  }
+
+  r = _getattr(in, CEPH_STAT_CAP_XATTR, perms, in->xattr_version == 0);
+  if (r == 0) {
+    string n(name);
+    r = -CEPHFS_ENODATA;
+    if (in->xattrs.count(n)) {
+      r = in->xattrs[n].length();
+      if (r > 0 && size != 0) {
+	if (size >= (unsigned)r)
+	  memcpy(value, in->xattrs[n].c_str(), r);
+	else
+	  r = -CEPHFS_ERANGE;
+      }
+    }
+  }
+ out:
+  ldout(cct, 8) << "_getxattr(" << in->ino << ", \"" << name << "\", " << size << ") = " << r << dendl;
+  return r;
+}
+
+int Client::_getxattr(InodeRef &in, const char *name, void *value, size_t size,
+		      const UserPerm& perms)
+{
+  if (cct->_conf->client_permissions) {
+    int r = xattr_permission(in.get(), name, MAY_READ, perms);
+    if (r < 0)
+      return r;
+  }
+  return _getxattr(in.get(), name, value, size, perms);
+}
+
+int Client::ll_getxattr(Inode *in, const char *name, void *value,
+			size_t size, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << __func__ << " " << vino << " " << name << " size " << size << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  if (!fuse_default_permissions) {
+    int r = xattr_permission(in, name, MAY_READ, perms);
+    if (r < 0)
+      return r;
+  }
+
+  return _getxattr(in, name, value, size, perms);
+}
+
+int Client::_listxattr(Inode *in, char *name, size_t size,
+		       const UserPerm& perms)
+{
+  bool len_only = (size == 0);
+  int r = _getattr(in, CEPH_STAT_CAP_XATTR, perms, in->xattr_version == 0);
+  if (r != 0) {
+    goto out;
+  }
+
+  r = 0;
+  for ([[maybe_unused]] const auto &[xattr_name, xattr_value_bl] : in->xattrs) {
+    if (xattr_name.rfind("ceph.", 0) == 0) {
+      continue;
+    }
+
+    size_t this_len = xattr_name.length() + 1;
+    r += this_len;
+    if (len_only)
+      continue;
+
+    if (this_len > size) {
+      r = -CEPHFS_ERANGE;
+      goto out;
+    }
+
+    memcpy(name, xattr_name.c_str(), this_len);
+    name += this_len;
+    size -= this_len;
+  }
+out:
+  ldout(cct, 8) << __func__ << "(" << in->ino << ", " << size << ") = " << r << dendl;
+  return r;
+}
+
+int Client::ll_listxattr(Inode *in, char *names, size_t size,
+			 const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << __func__ << " " << vino << " size " << size << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << size << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _listxattr(in, names, size, perms);
+}
+
+int Client::_do_setxattr(Inode *in, const char *name, const void *value,
+			 size_t size, int flags, const UserPerm& perms)
+{
+
+  int xattr_flags = 0;
+  if (!value)
+    xattr_flags |= CEPH_XATTR_REMOVE;
+  if (flags & XATTR_CREATE)
+    xattr_flags |= CEPH_XATTR_CREATE;
+  if (flags & XATTR_REPLACE)
+    xattr_flags |= CEPH_XATTR_REPLACE;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_SETXATTR);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_string2(name);
+  req->set_inode(in);
+  req->head.args.setxattr.flags = xattr_flags;
+
+  bufferlist bl;
+  assert (value || size == 0);
+  bl.append((const char*)value, size);
+  req->set_data(bl);
+
+  int res = make_request(req, perms);
+
+  trim_cache();
+  ldout(cct, 3) << __func__ << "(" << in->ino << ", \"" << name << "\") = " <<
+    res << dendl;
+  return res;
+}
+
+int Client::_setxattr(Inode *in, const char *name, const void *value,
+		      size_t size, int flags, const UserPerm& perms)
+{
+  if (in->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+
+  if (size == 0) {
+    value = "";
+  } else if (value == NULL) {
+      return -CEPHFS_EINVAL;
+  }
+
+  bool posix_acl_xattr = false;
+  if (acl_type == POSIX_ACL)
+    posix_acl_xattr = !strncmp(name, "system.", 7);
+
+  if (strncmp(name, "user.", 5) &&
+      strncmp(name, "security.", 9) &&
+      strncmp(name, "trusted.", 8) &&
+      strncmp(name, "ceph.", 5) &&
+      !posix_acl_xattr)
+    return -CEPHFS_EOPNOTSUPP;
+
+  bool check_realm = false;
+
+  if (posix_acl_xattr) {
+    if (!strcmp(name, ACL_EA_ACCESS)) {
+      mode_t new_mode = in->mode;
+      if (value) {
+	int ret = posix_acl_equiv_mode(value, size, &new_mode);
+	if (ret < 0)
+	  return ret;
+	if (ret == 0) {
+	  value = NULL;
+	  size = 0;
+	}
+	if (new_mode != in->mode) {
+	  struct ceph_statx stx;
+	  stx.stx_mode = new_mode;
+	  ret = _do_setattr(in, &stx, CEPH_SETATTR_MODE, perms, NULL);
+	  if (ret < 0)
+	    return ret;
+	}
+      }
+    } else if (!strcmp(name, ACL_EA_DEFAULT)) {
+      if (value) {
+	if (!S_ISDIR(in->mode))
+	  return -CEPHFS_EACCES;
+	int ret = posix_acl_check(value, size);
+	if (ret < 0)
+	  return -CEPHFS_EINVAL;
+	if (ret == 0) {
+	  value = NULL;
+	  size = 0;
+	}
+      }
+    } else {
+      return -CEPHFS_EOPNOTSUPP;
+    }
+  } else {
+    const VXattr *vxattr = _match_vxattr(in, name);
+    if (vxattr) {
+      if (vxattr->readonly)
+	return -CEPHFS_EOPNOTSUPP;
+      if (vxattr->name.compare(0, 10, "ceph.quota") == 0 && value)
+	check_realm = true;
+    }
+  }
+
+  int ret = _do_setxattr(in, name, value, size, flags, perms);
+  if (ret >= 0 && check_realm) {
+    // check if snaprealm was created for quota inode
+    if (in->quota.is_enable() &&
+	!(in->snaprealm && in->snaprealm->ino == in->ino))
+      ret = -CEPHFS_EOPNOTSUPP;
+  }
+
+  return ret;
+}
+
+int Client::_setxattr(InodeRef &in, const char *name, const void *value,
+		      size_t size, int flags, const UserPerm& perms)
+{
+  if (cct->_conf->client_permissions) {
+    int r = xattr_permission(in.get(), name, MAY_WRITE, perms);
+    if (r < 0)
+      return r;
+  }
+  return _setxattr(in.get(), name, value, size, flags, perms);
+}
+
+int Client::_setxattr_check_data_pool(string& name, string& value, const OSDMap *osdmap)
+{
+  string tmp;
+  if (name == "layout") {
+    string::iterator begin = value.begin();
+    string::iterator end = value.end();
+    keys_and_values<string::iterator> p;    // create instance of parser
+    std::map<string, string> m;             // map to receive results
+    if (!qi::parse(begin, end, p, m)) {     // returns true if successful
+      return -CEPHFS_EINVAL;
+    }
+    if (begin != end)
+      return -CEPHFS_EINVAL;
+    for (map<string,string>::iterator q = m.begin(); q != m.end(); ++q) {
+      if (q->first == "pool") {
+	tmp = q->second;
+	break;
+      }
+    }
+  } else if (name == "layout.pool") {
+    tmp = value;
+  }
+
+  if (tmp.length()) {
+    int64_t pool;
+    try {
+      pool = boost::lexical_cast<unsigned>(tmp);
+      if (!osdmap->have_pg_pool(pool))
+	return -CEPHFS_ENOENT;
+    } catch (boost::bad_lexical_cast const&) {
+      pool = osdmap->lookup_pg_pool_name(tmp);
+      if (pool < 0) {
+	return -CEPHFS_ENOENT;
+      }
+    }
+  }
+
+  return 0;
+}
+
+void Client::_setxattr_maybe_wait_for_osdmap(const char *name, const void *value, size_t size)
+{
+  // For setting pool of layout, MetaRequest need osdmap epoch.
+  // There is a race which create a new data pool but client and mds both don't have.
+  // Make client got the latest osdmap which make mds quickly judge whether get newer osdmap.
+  ldout(cct, 15) << __func__ << ": name = " << name << dendl;
+  if (strcmp(name, "ceph.file.layout.pool") == 0 || strcmp(name, "ceph.dir.layout.pool") == 0 ||
+      strcmp(name, "ceph.file.layout") == 0 || strcmp(name, "ceph.dir.layout") == 0) {
+    string rest(strstr(name, "layout"));
+    string v((const char*)value, size);
+    int r = objecter->with_osdmap([&](const OSDMap& o) {
+      return _setxattr_check_data_pool(rest, v, &o);
+    });
+
+    if (r == -CEPHFS_ENOENT) {
+      bs::error_code ec;
+      ldout(cct, 20) << __func__ << ": waiting for latest osdmap" << dendl;
+      objecter->wait_for_latest_osdmap(ca::use_blocked[ec]);
+      ldout(cct, 20) << __func__ << ": got latest osdmap: " << ec << dendl;
+    }
+  }
+}
+
+int Client::ll_setxattr(Inode *in, const char *name, const void *value,
+			size_t size, int flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  _setxattr_maybe_wait_for_osdmap(name, value, size);
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << __func__ << " " << vino << " " << name << " size " << size << dendl;
+  tout(cct) << __func__ << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  if (!fuse_default_permissions) {
+    int r = xattr_permission(in, name, MAY_WRITE, perms);
+    if (r < 0)
+      return r;
+  }
+  return _setxattr(in, name, value, size, flags, perms);
+}
+
+int Client::_removexattr(Inode *in, const char *name, const UserPerm& perms)
+{
+  if (in->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+
+  // same xattrs supported by kernel client
+  if (strncmp(name, "user.", 5) &&
+      strncmp(name, "system.", 7) &&
+      strncmp(name, "security.", 9) &&
+      strncmp(name, "trusted.", 8) &&
+      strncmp(name, "ceph.", 5))
+    return -CEPHFS_EOPNOTSUPP;
+
+  const VXattr *vxattr = _match_vxattr(in, name);
+  if (vxattr && vxattr->readonly)
+    return -CEPHFS_EOPNOTSUPP;
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_RMXATTR);
+  filepath path;
+  in->make_nosnap_relative_path(path);
+  req->set_filepath(path);
+  req->set_filepath2(name);
+  req->set_inode(in);
+ 
+  int res = make_request(req, perms);
+
+  trim_cache();
+  ldout(cct, 8) << "_removexattr(" << in->ino << ", \"" << name << "\") = " << res << dendl;
+  return res;
+}
+
+int Client::_removexattr(InodeRef &in, const char *name, const UserPerm& perms)
+{
+  if (cct->_conf->client_permissions) {
+    int r = xattr_permission(in.get(), name, MAY_WRITE, perms);
+    if (r < 0)
+      return r;
+  }
+  return _removexattr(in.get(), name, perms);
+}
+
+int Client::ll_removexattr(Inode *in, const char *name, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_removexattr " << vino << " " << name << dendl;
+  tout(cct) << "ll_removexattr" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  if (!fuse_default_permissions) {
+    int r = xattr_permission(in, name, MAY_WRITE, perms);
+    if (r < 0)
+      return r;
+  }
+
+  return _removexattr(in, name, perms);
+}
+
+bool Client::_vxattrcb_quota_exists(Inode *in)
+{
+  return in->quota.is_enable() &&
+   (in->snapid != CEPH_NOSNAP ||
+    (in->snaprealm && in->snaprealm->ino == in->ino));
+}
+size_t Client::_vxattrcb_quota(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size,
+                  "max_bytes=%lld max_files=%lld",
+                  (long long int)in->quota.max_bytes,
+                  (long long int)in->quota.max_files);
+}
+size_t Client::_vxattrcb_quota_max_bytes(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%lld", (long long int)in->quota.max_bytes);
+}
+size_t Client::_vxattrcb_quota_max_files(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%lld", (long long int)in->quota.max_files);
+}
+
+bool Client::_vxattrcb_layout_exists(Inode *in)
+{
+  return in->layout != file_layout_t();
+}
+size_t Client::_vxattrcb_layout(Inode *in, char *val, size_t size)
+{
+  int r = snprintf(val, size,
+      "stripe_unit=%llu stripe_count=%llu object_size=%llu pool=",
+      (unsigned long long)in->layout.stripe_unit,
+      (unsigned long long)in->layout.stripe_count,
+      (unsigned long long)in->layout.object_size);
+  objecter->with_osdmap([&](const OSDMap& o) {
+      if (o.have_pg_pool(in->layout.pool_id))
+	r += snprintf(val + r, size - r, "%s",
+		      o.get_pool_name(in->layout.pool_id).c_str());
+      else
+	r += snprintf(val + r, size - r, "%" PRIu64,
+		      (uint64_t)in->layout.pool_id);
+    });
+  if (in->layout.pool_ns.length())
+    r += snprintf(val + r, size - r, " pool_namespace=%s",
+		  in->layout.pool_ns.c_str());
+  return r;
+}
+size_t Client::_vxattrcb_layout_stripe_unit(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->layout.stripe_unit);
+}
+size_t Client::_vxattrcb_layout_stripe_count(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->layout.stripe_count);
+}
+size_t Client::_vxattrcb_layout_object_size(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->layout.object_size);
+}
+size_t Client::_vxattrcb_layout_pool(Inode *in, char *val, size_t size)
+{
+  size_t r;
+  objecter->with_osdmap([&](const OSDMap& o) {
+      if (o.have_pg_pool(in->layout.pool_id))
+	r = snprintf(val, size, "%s", o.get_pool_name(
+		       in->layout.pool_id).c_str());
+      else
+	r = snprintf(val, size, "%" PRIu64, (uint64_t)in->layout.pool_id);
+    });
+  return r;
+}
+size_t Client::_vxattrcb_layout_pool_namespace(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%s", in->layout.pool_ns.c_str());
+}
+size_t Client::_vxattrcb_dir_entries(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)(in->dirstat.nfiles + in->dirstat.nsubdirs));
+}
+size_t Client::_vxattrcb_dir_files(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->dirstat.nfiles);
+}
+size_t Client::_vxattrcb_dir_subdirs(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->dirstat.nsubdirs);
+}
+size_t Client::_vxattrcb_dir_rentries(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)(in->rstat.rfiles + in->rstat.rsubdirs));
+}
+size_t Client::_vxattrcb_dir_rfiles(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->rstat.rfiles);
+}
+size_t Client::_vxattrcb_dir_rsubdirs(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->rstat.rsubdirs);
+}
+size_t Client::_vxattrcb_dir_rsnaps(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->rstat.rsnaps);
+}
+size_t Client::_vxattrcb_dir_rbytes(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu", (unsigned long long)in->rstat.rbytes);
+}
+size_t Client::_vxattrcb_dir_rctime(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%ld.%09ld", (long)in->rstat.rctime.sec(),
+      (long)in->rstat.rctime.nsec());
+}
+bool Client::_vxattrcb_dir_pin_exists(Inode *in)
+{
+  return in->dir_pin != -CEPHFS_ENODATA;
+}
+size_t Client::_vxattrcb_dir_pin(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%ld", (long)in->dir_pin);
+}
+
+bool Client::_vxattrcb_snap_btime_exists(Inode *in)
+{
+  return !in->snap_btime.is_zero();
+}
+
+size_t Client::_vxattrcb_snap_btime(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%llu.%09lu",
+      (long long unsigned)in->snap_btime.sec(),
+      (long unsigned)in->snap_btime.nsec());
+}
+
+bool Client::_vxattrcb_mirror_info_exists(Inode *in)
+{
+  // checking one of the xattrs would suffice
+  return in->xattrs.count("ceph.mirror.info.cluster_id") != 0;
+}
+
+size_t Client::_vxattrcb_mirror_info(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "cluster_id=%.*s fs_id=%.*s",
+                  in->xattrs["ceph.mirror.info.cluster_id"].length(),
+                  in->xattrs["ceph.mirror.info.cluster_id"].c_str(),
+                  in->xattrs["ceph.mirror.info.fs_id"].length(),
+                  in->xattrs["ceph.mirror.info.fs_id"].c_str());
+}
+
+size_t Client::_vxattrcb_cluster_fsid(Inode *in, char *val, size_t size)
+{
+  return snprintf(val, size, "%s", monclient->get_fsid().to_string().c_str());
+}
+
+size_t Client::_vxattrcb_client_id(Inode *in, char *val, size_t size)
+{
+  auto name = messenger->get_myname();
+  return snprintf(val, size, "%s%ld", name.type_str(), name.num());
+}
+
+#define CEPH_XATTR_NAME(_type, _name) "ceph." #_type "." #_name
+#define CEPH_XATTR_NAME2(_type, _name, _name2) "ceph." #_type "." #_name "." #_name2
+
+#define XATTR_NAME_CEPH(_type, _name, _flags)                 \
+{                                                              \
+  name: CEPH_XATTR_NAME(_type, _name),                         \
+  getxattr_cb: &Client::_vxattrcb_ ## _type ## _ ## _name,     \
+  readonly: true,                                              \
+  exists_cb: NULL,                                             \
+  flags: _flags,                                               \
+}
+#define XATTR_LAYOUT_FIELD(_type, _name, _field)		\
+{								\
+  name: CEPH_XATTR_NAME2(_type, _name, _field),			\
+  getxattr_cb: &Client::_vxattrcb_ ## _name ## _ ## _field,	\
+  readonly: false,						\
+  exists_cb: &Client::_vxattrcb_layout_exists,			\
+  flags: 0,                                                     \
+}
+#define XATTR_QUOTA_FIELD(_type, _name)		                \
+{								\
+  name: CEPH_XATTR_NAME(_type, _name),			        \
+  getxattr_cb: &Client::_vxattrcb_ ## _type ## _ ## _name,	\
+  readonly: false,						\
+  exists_cb: &Client::_vxattrcb_quota_exists,			\
+  flags: 0,                                                     \
+}
+
+const Client::VXattr Client::_dir_vxattrs[] = {
+  {
+    name: "ceph.dir.layout",
+    getxattr_cb: &Client::_vxattrcb_layout,
+    readonly: false,
+    exists_cb: &Client::_vxattrcb_layout_exists,
+    flags: 0,
+  },
+  // FIXME
+  // Delete the following dir layout field definitions for release "S"
+  XATTR_LAYOUT_FIELD(dir, layout, stripe_unit),
+  XATTR_LAYOUT_FIELD(dir, layout, stripe_count),
+  XATTR_LAYOUT_FIELD(dir, layout, object_size),
+  XATTR_LAYOUT_FIELD(dir, layout, pool),
+  XATTR_LAYOUT_FIELD(dir, layout, pool_namespace),
+  XATTR_NAME_CEPH(dir, entries, VXATTR_DIRSTAT),
+  XATTR_NAME_CEPH(dir, files, VXATTR_DIRSTAT),
+  XATTR_NAME_CEPH(dir, subdirs, VXATTR_DIRSTAT),
+  XATTR_NAME_CEPH(dir, rentries, VXATTR_RSTAT),
+  XATTR_NAME_CEPH(dir, rfiles, VXATTR_RSTAT),
+  XATTR_NAME_CEPH(dir, rsubdirs, VXATTR_RSTAT),
+  XATTR_NAME_CEPH(dir, rsnaps, VXATTR_RSTAT),
+  XATTR_NAME_CEPH(dir, rbytes, VXATTR_RSTAT),
+  XATTR_NAME_CEPH(dir, rctime, VXATTR_RSTAT),
+  {
+    name: "ceph.quota",
+    getxattr_cb: &Client::_vxattrcb_quota,
+    readonly: false,
+    exists_cb: &Client::_vxattrcb_quota_exists,
+    flags: 0,
+  },
+  XATTR_QUOTA_FIELD(quota, max_bytes),
+  XATTR_QUOTA_FIELD(quota, max_files),
+  // FIXME
+  // Delete the following dir pin field definitions for release "S"
+  {
+    name: "ceph.dir.pin",
+    getxattr_cb: &Client::_vxattrcb_dir_pin,
+    readonly: false,
+    exists_cb: &Client::_vxattrcb_dir_pin_exists,
+    flags: 0,
+  },
+  {
+    name: "ceph.snap.btime",
+    getxattr_cb: &Client::_vxattrcb_snap_btime,
+    readonly: true,
+    exists_cb: &Client::_vxattrcb_snap_btime_exists,
+    flags: 0,
+  },
+  {
+    name: "ceph.mirror.info",
+    getxattr_cb: &Client::_vxattrcb_mirror_info,
+    readonly: false,
+    exists_cb: &Client::_vxattrcb_mirror_info_exists,
+    flags: 0,
+  },
+  { name: "" }     /* Required table terminator */
+};
+
+const Client::VXattr Client::_file_vxattrs[] = {
+  {
+    name: "ceph.file.layout",
+    getxattr_cb: &Client::_vxattrcb_layout,
+    readonly: false,
+    exists_cb: &Client::_vxattrcb_layout_exists,
+    flags: 0,
+  },
+  XATTR_LAYOUT_FIELD(file, layout, stripe_unit),
+  XATTR_LAYOUT_FIELD(file, layout, stripe_count),
+  XATTR_LAYOUT_FIELD(file, layout, object_size),
+  XATTR_LAYOUT_FIELD(file, layout, pool),
+  XATTR_LAYOUT_FIELD(file, layout, pool_namespace),
+  {
+    name: "ceph.snap.btime",
+    getxattr_cb: &Client::_vxattrcb_snap_btime,
+    readonly: true,
+    exists_cb: &Client::_vxattrcb_snap_btime_exists,
+    flags: 0,
+  },
+  { name: "" }     /* Required table terminator */
+};
+
+const Client::VXattr Client::_common_vxattrs[] = {
+  {
+    name: "ceph.cluster_fsid",
+    getxattr_cb: &Client::_vxattrcb_cluster_fsid,
+    readonly: true,
+    exists_cb: nullptr,
+    flags: 0,
+  },
+  {
+    name: "ceph.client_id",
+    getxattr_cb: &Client::_vxattrcb_client_id,
+    readonly: true,
+    exists_cb: nullptr,
+    flags: 0,
+  },
+  { name: "" }     /* Required table terminator */
+};
+
+const Client::VXattr *Client::_get_vxattrs(Inode *in)
+{
+  if (in->is_dir())
+    return _dir_vxattrs;
+  else if (in->is_file())
+    return _file_vxattrs;
+  return NULL;
+}
+
+const Client::VXattr *Client::_match_vxattr(Inode *in, const char *name)
+{
+  if (strncmp(name, "ceph.", 5) == 0) {
+    const VXattr *vxattr = _get_vxattrs(in);
+    if (vxattr) {
+      while (!vxattr->name.empty()) {
+	if (vxattr->name == name)
+	  return vxattr;
+	vxattr++;
+      }
+    }
+
+    // for common vxattrs
+    vxattr = _common_vxattrs;
+    while (!vxattr->name.empty()) {
+      if (vxattr->name == name)
+        return vxattr;
+      vxattr++;
+    }
+  }
+
+  return NULL;
+}
+
+int Client::ll_readlink(Inode *in, char *buf, size_t buflen, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_readlink " << vino << dendl;
+  tout(cct) << "ll_readlink" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  for (auto dn : in->dentries) {
+    touch_dn(dn);
+  }
+
+  int r = _readlink(in, buf, buflen); // FIXME: no permission checking!
+  ldout(cct, 3) << "ll_readlink " << vino << " = " << r << dendl;
+  return r;
+}
+
+int Client::_mknod(Inode *dir, const char *name, mode_t mode, dev_t rdev,
+		   const UserPerm& perms, InodeRef *inp)
+{
+  ldout(cct, 8) << "_mknod(" << dir->ino << " " << name << ", 0" << oct
+		<< mode << dec << ", " << rdev << ", uid " << perms.uid()
+		<< ", gid " << perms.gid() << ")" << dendl;
+
+  if (strlen(name) > NAME_MAX)
+    return -CEPHFS_ENAMETOOLONG;
+
+  if (dir->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+  if (is_quota_files_exceeded(dir, perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_MKNOD);
+
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path); 
+  req->set_inode(dir);
+  req->head.args.mknod.rdev = rdev;
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+
+  bufferlist xattrs_bl;
+  int res = _posix_acl_create(dir, &mode, xattrs_bl, perms);
+  if (res < 0)
+    goto fail;
+  req->head.args.mknod.mode = mode;
+  if (xattrs_bl.length() > 0)
+    req->set_data(xattrs_bl);
+
+  Dentry *de;
+  res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+
+  res = make_request(req, perms, inp);
+
+  trim_cache();
+
+  ldout(cct, 8) << "mknod(" << path << ", 0" << oct << mode << dec << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_mknod(Inode *parent, const char *name, mode_t mode,
+		     dev_t rdev, struct stat *attr, Inode **out,
+		     const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_mknod " << vparent << " " << name << dendl;
+  tout(cct) << "ll_mknod" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << mode << std::endl;
+  tout(cct) << rdev << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _mknod(parent, name, mode, rdev, perms, &in);
+  if (r == 0) {
+    fill_stat(in, attr);
+    _ll_get(in.get());
+  }
+  tout(cct) << attr->st_ino << std::endl;
+  ldout(cct, 3) << "ll_mknod " << vparent << " " << name
+	  << " = " << r << " (" << hex << attr->st_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::ll_mknodx(Inode *parent, const char *name, mode_t mode,
+		      dev_t rdev, Inode **out,
+		      struct ceph_statx *stx, unsigned want, unsigned flags,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  unsigned caps = statx_to_mask(flags, want);
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_mknodx " << vparent << " " << name << dendl;
+  tout(cct) << "ll_mknodx" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << mode << std::endl;
+  tout(cct) << rdev << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _mknod(parent, name, mode, rdev, perms, &in);
+  if (r == 0) {
+    fill_statx(in, caps, stx);
+    _ll_get(in.get());
+  }
+  tout(cct) << stx->stx_ino << std::endl;
+  ldout(cct, 3) << "ll_mknodx " << vparent << " " << name
+	  << " = " << r << " (" << hex << stx->stx_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::_create(Inode *dir, const char *name, int flags, mode_t mode,
+		    InodeRef *inp, Fh **fhp, int stripe_unit, int stripe_count,
+		    int object_size, const char *data_pool, bool *created,
+		    const UserPerm& perms, std::string alternate_name)
+{
+  ldout(cct, 8) << "_create(" << dir->ino << " " << name << ", 0" << oct <<
+    mode << dec << ")" << dendl;
+
+  if (strlen(name) > NAME_MAX)
+    return -CEPHFS_ENAMETOOLONG;
+  if (dir->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+  if (is_quota_files_exceeded(dir, perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  // use normalized flags to generate cmode
+  int cflags = ceph_flags_sys2wire(flags);
+  if (cct->_conf.get_val<bool>("client_force_lazyio"))
+    cflags |= CEPH_O_LAZY;
+
+  int cmode = ceph_flags_to_mode(cflags);
+
+  int64_t pool_id = -1;
+  if (data_pool && *data_pool) {
+    pool_id = objecter->with_osdmap(
+      std::mem_fn(&OSDMap::lookup_pg_pool_name), data_pool);
+    if (pool_id < 0)
+      return -CEPHFS_EINVAL;
+    if (pool_id > 0xffffffffll)
+      return -CEPHFS_ERANGE;  // bummer!
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_CREATE);
+
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+  req->set_alternate_name(std::move(alternate_name));
+  req->set_inode(dir);
+  req->head.args.open.flags = cflags | CEPH_O_CREAT;
+
+  req->head.args.open.stripe_unit = stripe_unit;
+  req->head.args.open.stripe_count = stripe_count;
+  req->head.args.open.object_size = object_size;
+  if (cct->_conf->client_debug_getattr_caps)
+    req->head.args.open.mask = DEBUG_GETATTR_CAPS;
+  else
+    req->head.args.open.mask = 0;
+  req->head.args.open.pool = pool_id;
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+
+  mode |= S_IFREG;
+  bufferlist xattrs_bl;
+  int res = _posix_acl_create(dir, &mode, xattrs_bl, perms);
+  if (res < 0)
+    goto fail;
+  req->head.args.open.mode = mode;
+  if (xattrs_bl.length() > 0)
+    req->set_data(xattrs_bl);
+
+  Dentry *de;
+  res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+
+  res = make_request(req, perms, inp, created);
+  if (res < 0) {
+    goto reply_error;
+  }
+
+  /* If the caller passed a value in fhp, do the open */
+  if(fhp) {
+    (*inp)->get_open_ref(cmode);
+    *fhp = _create_fh(inp->get(), flags, cmode, perms);
+  }
+
+ reply_error:
+  trim_cache();
+
+  ldout(cct, 8) << "create(" << path << ", 0" << oct << mode << dec
+		<< " layout " << stripe_unit
+		<< ' ' << stripe_count
+		<< ' ' << object_size
+		<<") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::_mkdir(Inode *dir, const char *name, mode_t mode, const UserPerm& perm,
+		   InodeRef *inp, const std::map<std::string, std::string> &metadata,
+                   std::string alternate_name)
+{
+  ldout(cct, 8) << "_mkdir(" << dir->ino << " " << name << ", 0" << oct
+		<< mode << dec << ", uid " << perm.uid()
+		<< ", gid " << perm.gid() << ")" << dendl;
+
+  if (strlen(name) > NAME_MAX)
+    return -CEPHFS_ENAMETOOLONG;
+
+  if (dir->snapid != CEPH_NOSNAP && dir->snapid != CEPH_SNAPDIR) {
+    return -CEPHFS_EROFS;
+  }
+  if (is_quota_files_exceeded(dir, perm)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  bool is_snap_op = dir->snapid == CEPH_SNAPDIR;
+  MetaRequest *req = new MetaRequest(is_snap_op ?
+				     CEPH_MDS_OP_MKSNAP : CEPH_MDS_OP_MKDIR);
+
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+  req->set_inode(dir);
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+  req->set_alternate_name(std::move(alternate_name));
+
+  mode |= S_IFDIR;
+  bufferlist bl;
+  int res = _posix_acl_create(dir, &mode, bl, perm);
+  if (res < 0)
+    goto fail;
+  req->head.args.mkdir.mode = mode;
+  if (is_snap_op) {
+    SnapPayload payload;
+    // clear the bufferlist that may have been populated by the call
+    // to _posix_acl_create(). MDS mksnap does not make use of it.
+    // So, reuse it to pass metadata payload.
+    bl.clear();
+    payload.metadata = metadata;
+    encode(payload, bl);
+  }
+  if (bl.length() > 0) {
+    req->set_data(bl);
+  }
+
+  Dentry *de;
+  res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+  
+  ldout(cct, 10) << "_mkdir: making request" << dendl;
+  res = make_request(req, perm, inp);
+  ldout(cct, 10) << "_mkdir result is " << res << dendl;
+
+  trim_cache();
+
+  ldout(cct, 8) << "_mkdir(" << path << ", 0" << oct << mode << dec << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_mkdir(Inode *parent, const char *name, mode_t mode,
+		     struct stat *attr, Inode **out, const UserPerm& perm)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_mkdir " << vparent << " " << name << dendl;
+  tout(cct) << "ll_mkdir" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << mode << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perm);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _mkdir(parent, name, mode, perm, &in);
+  if (r == 0) {
+    fill_stat(in, attr);
+    _ll_get(in.get());
+  }
+  tout(cct) << attr->st_ino << std::endl;
+  ldout(cct, 3) << "ll_mkdir " << vparent << " " << name
+	  << " = " << r << " (" << hex << attr->st_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::ll_mkdirx(Inode *parent, const char *name, mode_t mode, Inode **out,
+		      struct ceph_statx *stx, unsigned want, unsigned flags,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_mkdirx " << vparent << " " << name << dendl;
+  tout(cct) << "ll_mkdirx" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << mode << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _mkdir(parent, name, mode, perms, &in);
+  if (r == 0) {
+    fill_statx(in, statx_to_mask(flags, want), stx);
+    _ll_get(in.get());
+  } else {
+    stx->stx_ino = 0;
+    stx->stx_mask = 0;
+  }
+  tout(cct) << stx->stx_ino << std::endl;
+  ldout(cct, 3) << "ll_mkdirx " << vparent << " " << name
+	  << " = " << r << " (" << hex << stx->stx_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::_symlink(Inode *dir, const char *name, const char *target,
+		     const UserPerm& perms, std::string alternate_name, InodeRef *inp)
+{
+  ldout(cct, 8) << "_symlink(" << dir->ino << " " << name << ", " << target
+		<< ", uid " << perms.uid() << ", gid " << perms.gid() << ")"
+		<< dendl;
+
+  if (strlen(name) > NAME_MAX)
+    return -CEPHFS_ENAMETOOLONG;
+
+  if (dir->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+  if (is_quota_files_exceeded(dir, perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_SYMLINK);
+
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+  req->set_alternate_name(std::move(alternate_name));
+  req->set_inode(dir);
+  req->set_string2(target); 
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+
+  Dentry *de;
+  int res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+
+  res = make_request(req, perms, inp);
+
+  trim_cache();
+  ldout(cct, 8) << "_symlink(\"" << path << "\", \"" << target << "\") = " <<
+    res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_symlink(Inode *parent, const char *name, const char *value,
+		       struct stat *attr, Inode **out, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_symlink " << vparent << " " << name << " -> " << value
+		<< dendl;
+  tout(cct) << "ll_symlink" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << value << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _symlink(parent, name, value, perms, "", &in);
+  if (r == 0) {
+    fill_stat(in, attr);
+    _ll_get(in.get());
+  }
+  tout(cct) << attr->st_ino << std::endl;
+  ldout(cct, 3) << "ll_symlink " << vparent << " " << name
+	  << " = " << r << " (" << hex << attr->st_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::ll_symlinkx(Inode *parent, const char *name, const char *value,
+			Inode **out, struct ceph_statx *stx, unsigned want,
+			unsigned flags, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 3) << "ll_symlinkx " << vparent << " " << name << " -> " << value
+		<< dendl;
+  tout(cct) << "ll_symlinkx" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << value << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_create(parent, perms);
+    if (r < 0)
+      return r;
+  }
+
+  InodeRef in;
+  int r = _symlink(parent, name, value, perms, "", &in);
+  if (r == 0) {
+    fill_statx(in, statx_to_mask(flags, want), stx);
+    _ll_get(in.get());
+  }
+  tout(cct) << stx->stx_ino << std::endl;
+  ldout(cct, 3) << "ll_symlinkx " << vparent << " " << name
+	  << " = " << r << " (" << hex << stx->stx_ino << dec << ")" << dendl;
+  *out = in.get();
+  return r;
+}
+
+int Client::_unlink(Inode *dir, const char *name, const UserPerm& perm)
+{
+  ldout(cct, 8) << "_unlink(" << dir->ino << " " << name
+		<< " uid " << perm.uid() << " gid " << perm.gid()
+		<< ")" << dendl;
+
+  if (dir->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_UNLINK);
+
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+
+  InodeRef otherin;
+  Inode *in;
+  Dentry *de;
+
+  int res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+
+  res = _lookup(dir, name, 0, &otherin, perm);
+  if (res < 0)
+    goto fail;
+
+  in = otherin.get();
+  req->set_other_inode(in);
+  in->break_all_delegs();
+  req->other_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
+
+  req->set_inode(dir);
+
+  res = make_request(req, perm);
+
+  trim_cache();
+  ldout(cct, 8) << "unlink(" << path << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_unlink(Inode *in, const char *name, const UserPerm& perm)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_unlink " << vino << " " << name << dendl;
+  tout(cct) << "ll_unlink" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_delete(in, name, perm);
+    if (r < 0)
+      return r;
+  }
+  return _unlink(in, name, perm);
+}
+
+int Client::_rmdir(Inode *dir, const char *name, const UserPerm& perms)
+{
+  ldout(cct, 8) << "_rmdir(" << dir->ino << " " << name << " uid "
+		<< perms.uid() << " gid " << perms.gid() << ")" << dendl;
+
+  if (dir->snapid != CEPH_NOSNAP && dir->snapid != CEPH_SNAPDIR) {
+    return -CEPHFS_EROFS;
+  }
+  
+  int op = dir->snapid == CEPH_SNAPDIR ? CEPH_MDS_OP_RMSNAP : CEPH_MDS_OP_RMDIR;
+  MetaRequest *req = new MetaRequest(op);
+  filepath path;
+  dir->make_nosnap_relative_path(path);
+  path.push_dentry(name);
+  req->set_filepath(path);
+  req->set_inode(dir);
+
+  req->dentry_drop = CEPH_CAP_FILE_SHARED;
+  req->dentry_unless = CEPH_CAP_FILE_EXCL;
+  req->other_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
+
+  InodeRef in;
+
+  Dentry *de;
+  int res = get_or_create(dir, name, &de);
+  if (res < 0)
+    goto fail;
+  if (op == CEPH_MDS_OP_RMDIR) 
+    req->set_dentry(de);
+  else
+    de->get();
+
+  res = _lookup(dir, name, 0, &in, perms);
+  if (res < 0)
+    goto fail;
+
+  if (op == CEPH_MDS_OP_RMSNAP) {
+    unlink(de, true, true);
+    de->put();
+  }
+  req->set_other_inode(in.get());
+
+  res = make_request(req, perms);
+
+  trim_cache();
+  ldout(cct, 8) << "rmdir(" << path << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_rmdir(Inode *in, const char *name, const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_rmdir " << vino << " " << name << dendl;
+  tout(cct) << "ll_rmdir" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_delete(in, name, perms);
+    if (r < 0)
+      return r;
+  }
+
+  return _rmdir(in, name, perms);
+}
+
+int Client::_rename(Inode *fromdir, const char *fromname, Inode *todir, const char *toname, const UserPerm& perm, std::string alternate_name)
+{
+  ldout(cct, 8) << "_rename(" << fromdir->ino << " " << fromname << " to "
+		<< todir->ino << " " << toname
+		<< " uid " << perm.uid() << " gid " << perm.gid() << ")"
+		<< dendl;
+
+  if (fromdir->snapid != todir->snapid)
+    return -CEPHFS_EXDEV;
+
+  int op = CEPH_MDS_OP_RENAME;
+  if (fromdir->snapid != CEPH_NOSNAP) {
+    if (fromdir == todir && fromdir->snapid == CEPH_SNAPDIR)
+      op = CEPH_MDS_OP_RENAMESNAP;
+    else
+      return -CEPHFS_EROFS;
+  }
+  if (cct->_conf.get_val<bool>("client_quota") && fromdir != todir) {
+    Inode *fromdir_root =
+      fromdir->quota.is_enable() ? fromdir : get_quota_root(fromdir, perm);
+    Inode *todir_root =
+      todir->quota.is_enable() ? todir : get_quota_root(todir, perm);
+    if (fromdir_root != todir_root) {
+      return -CEPHFS_EXDEV;
+    }
+  }
+
+  InodeRef target;
+  MetaRequest *req = new MetaRequest(op);
+
+  filepath from;
+  fromdir->make_nosnap_relative_path(from);
+  from.push_dentry(fromname);
+  filepath to;
+  todir->make_nosnap_relative_path(to);
+  to.push_dentry(toname);
+  req->set_filepath(to);
+  req->set_filepath2(from);
+  req->set_alternate_name(std::move(alternate_name));
+
+  Dentry *oldde;
+  int res = get_or_create(fromdir, fromname, &oldde);
+  if (res < 0)
+    goto fail;
+  Dentry *de;
+  res = get_or_create(todir, toname, &de);
+  if (res < 0)
+    goto fail;
+
+  if (op == CEPH_MDS_OP_RENAME) {
+    req->set_old_dentry(oldde);
+    req->old_dentry_drop = CEPH_CAP_FILE_SHARED;
+    req->old_dentry_unless = CEPH_CAP_FILE_EXCL;
+
+    req->set_dentry(de);
+    req->dentry_drop = CEPH_CAP_FILE_SHARED;
+    req->dentry_unless = CEPH_CAP_FILE_EXCL;
+
+    InodeRef oldin, otherin;
+    res = _lookup(fromdir, fromname, 0, &oldin, perm);
+    if (res < 0)
+      goto fail;
+
+    Inode *oldinode = oldin.get();
+    oldinode->break_all_delegs();
+    req->set_old_inode(oldinode);
+    req->old_inode_drop = CEPH_CAP_LINK_SHARED;
+
+    res = _lookup(todir, toname, 0, &otherin, perm);
+    switch (res) {
+    case 0:
+      {
+	Inode *in = otherin.get();
+	req->set_other_inode(in);
+	in->break_all_delegs();
+      }
+      req->other_inode_drop = CEPH_CAP_LINK_SHARED | CEPH_CAP_LINK_EXCL;
+      break;
+    case -CEPHFS_ENOENT:
+      break;
+    default:
+      goto fail;
+    }
+
+    req->set_inode(todir);
+  } else {
+    // renamesnap reply contains no tracedn, so we need to invalidate
+    // dentry manually
+    unlink(oldde, true, true);
+    unlink(de, true, true);
+
+    req->set_inode(todir);
+  }
+
+  res = make_request(req, perm, &target);
+  ldout(cct, 10) << "rename result is " << res << dendl;
+
+  // renamed item from our cache
+
+  trim_cache();
+  ldout(cct, 8) << "_rename(" << from << ", " << to << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_rename(Inode *parent, const char *name, Inode *newparent,
+		      const char *newname, const UserPerm& perm)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vparent = _get_vino(parent);
+  vinodeno_t vnewparent = _get_vino(newparent);
+
+  ldout(cct, 3) << "ll_rename " << vparent << " " << name << " to "
+	  << vnewparent << " " << newname << dendl;
+  tout(cct) << "ll_rename" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << vnewparent.ino.val << std::endl;
+  tout(cct) << newname << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_delete(parent, name, perm);
+    if (r < 0)
+      return r;
+    r = may_delete(newparent, newname, perm);
+    if (r < 0 && r != -CEPHFS_ENOENT)
+      return r;
+  }
+
+  return _rename(parent, name, newparent, newname, perm, "");
+}
+
+int Client::_link(Inode *in, Inode *dir, const char *newname, const UserPerm& perm, std::string alternate_name, InodeRef *inp)
+{
+  ldout(cct, 8) << "_link(" << in->ino << " to " << dir->ino << " " << newname
+		<< " uid " << perm.uid() << " gid " << perm.gid() << ")" << dendl;
+
+  if (strlen(newname) > NAME_MAX)
+    return -CEPHFS_ENAMETOOLONG;
+
+  if (in->snapid != CEPH_NOSNAP || dir->snapid != CEPH_NOSNAP) {
+    return -CEPHFS_EROFS;
+  }
+  if (is_quota_files_exceeded(dir, perm)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  in->break_all_delegs();
+  MetaRequest *req = new MetaRequest(CEPH_MDS_OP_LINK);
+
+  filepath path(newname, dir->ino);
+  req->set_filepath(path);
+  req->set_alternate_name(std::move(alternate_name));
+  filepath existing(in->ino);
+  req->set_filepath2(existing);
+
+  req->set_inode(dir);
+  req->inode_drop = CEPH_CAP_FILE_SHARED;
+  req->inode_unless = CEPH_CAP_FILE_EXCL;
+
+  Dentry *de;
+  int res = get_or_create(dir, newname, &de);
+  if (res < 0)
+    goto fail;
+  req->set_dentry(de);
+  
+  res = make_request(req, perm, inp);
+  ldout(cct, 10) << "link result is " << res << dendl;
+
+  trim_cache();
+  ldout(cct, 8) << "link(" << existing << ", " << path << ") = " << res << dendl;
+  return res;
+
+ fail:
+  put_request(req);
+  return res;
+}
+
+int Client::ll_link(Inode *in, Inode *newparent, const char *newname,
+		    const UserPerm& perm)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+  vinodeno_t vnewparent = _get_vino(newparent);
+
+  ldout(cct, 3) << "ll_link " << vino << " to " << vnewparent << " " <<
+    newname << dendl;
+  tout(cct) << "ll_link" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << vnewparent << std::endl;
+  tout(cct) << newname << std::endl;
+
+  InodeRef target;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    if (S_ISDIR(in->mode))
+      return -CEPHFS_EPERM;
+
+    int r = may_hardlink(in, perm);
+    if (r < 0)
+      return r;
+
+    r = may_create(newparent, perm);
+    if (r < 0)
+      return r;
+  }
+
+  return _link(in, newparent, newname, perm, "", &target);
+}
+
+int Client::ll_num_osds(void)
+{
+  std::scoped_lock lock(client_lock);
+  return objecter->with_osdmap(std::mem_fn(&OSDMap::get_num_osds));
+}
+
+int Client::ll_osdaddr(int osd, uint32_t *addr)
+{
+  std::scoped_lock lock(client_lock);
+
+  entity_addr_t g;
+  bool exists = objecter->with_osdmap([&](const OSDMap& o) {
+      if (!o.exists(osd))
+	return false;
+      g = o.get_addrs(osd).front();
+      return true;
+    });
+  if (!exists)
+    return -1;
+  uint32_t nb_addr = (g.in4_addr()).sin_addr.s_addr;
+  *addr = ntohl(nb_addr);
+  return 0;
+}
+
+uint32_t Client::ll_stripe_unit(Inode *in)
+{
+  std::scoped_lock lock(client_lock);
+  return in->layout.stripe_unit;
+}
+
+uint64_t Client::ll_snap_seq(Inode *in)
+{
+  std::scoped_lock lock(client_lock);
+  return in->snaprealm->seq;
+}
+
+int Client::ll_file_layout(Inode *in, file_layout_t *layout)
+{
+  std::scoped_lock lock(client_lock);
+  *layout = in->layout;
+  return 0;
+}
+
+int Client::ll_file_layout(Fh *fh, file_layout_t *layout)
+{
+  return ll_file_layout(fh->inode.get(), layout);
+}
+
+/* Currently we cannot take advantage of redundancy in reads, since we
+   would have to go through all possible placement groups (a
+   potentially quite large number determined by a hash), and use CRUSH
+   to calculate the appropriate set of OSDs for each placement group,
+   then index into that.  An array with one entry per OSD is much more
+   tractable and works for demonstration purposes. */
+
+int Client::ll_get_stripe_osd(Inode *in, uint64_t blockno,
+			      file_layout_t* layout)
+{
+  std::scoped_lock lock(client_lock);
+
+  inodeno_t ino = in->ino;
+  uint32_t object_size = layout->object_size;
+  uint32_t su = layout->stripe_unit;
+  uint32_t stripe_count = layout->stripe_count;
+  uint64_t stripes_per_object = object_size / su;
+  uint64_t stripeno = 0, stripepos = 0;
+
+  if(stripe_count) {
+      stripeno = blockno / stripe_count;    // which horizontal stripe        (Y)
+      stripepos = blockno % stripe_count;   // which object in the object set (X)
+  }
+  uint64_t objectsetno = stripeno / stripes_per_object;       // which object set
+  uint64_t objectno = objectsetno * stripe_count + stripepos;  // object id
+
+  object_t oid = file_object_t(ino, objectno);
+  return objecter->with_osdmap([&](const OSDMap& o) {
+      ceph_object_layout olayout =
+	o.file_to_object_layout(oid, *layout);
+      pg_t pg = (pg_t)olayout.ol_pgid;
+      vector<int> osds;
+      int primary;
+      o.pg_to_acting_osds(pg, &osds, &primary);
+      return primary;
+    });
+}
+
+/* Return the offset of the block, internal to the object */
+
+uint64_t Client::ll_get_internal_offset(Inode *in, uint64_t blockno)
+{
+  std::scoped_lock lock(client_lock);
+  file_layout_t *layout=&(in->layout);
+  uint32_t object_size = layout->object_size;
+  uint32_t su = layout->stripe_unit;
+  uint64_t stripes_per_object = object_size / su;
+
+  return (blockno % stripes_per_object) * su;
+}
+
+int Client::ll_opendir(Inode *in, int flags, dir_result_t** dirpp,
+		       const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_opendir " << vino << dendl;
+  tout(cct) << "ll_opendir" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (!fuse_default_permissions) {
+    int r = may_open(in, flags, perms);
+    if (r < 0)
+      return r;
+  }
+
+  int r = _opendir(in, dirpp, perms);
+  tout(cct) << (uintptr_t)*dirpp << std::endl;
+
+  ldout(cct, 3) << "ll_opendir " << vino << " = " << r << " (" << *dirpp << ")"
+		<< dendl;
+  return r;
+}
+
+int Client::ll_releasedir(dir_result_t *dirp)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_releasedir " << dirp << dendl;
+  tout(cct) << "ll_releasedir" << std::endl;
+  tout(cct) << (uintptr_t)dirp << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  _closedir(dirp);
+  return 0;
+}
+
+int Client::ll_fsyncdir(dir_result_t *dirp)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_fsyncdir " << dirp << dendl;
+  tout(cct) << "ll_fsyncdir" << std::endl;
+  tout(cct) << (uintptr_t)dirp << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _fsync(dirp->inode.get(), false);
+}
+
+int Client::ll_open(Inode *in, int flags, Fh **fhp, const UserPerm& perms)
+{
+  ceph_assert(!(flags & O_CREAT));
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+
+  ldout(cct, 3) << "ll_open " << vino << " " << ceph_flags_sys2wire(flags) << dendl;
+  tout(cct) << "ll_open" << std::endl;
+  tout(cct) << vino.ino.val << std::endl;
+  tout(cct) << ceph_flags_sys2wire(flags) << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  int r;
+  if (!fuse_default_permissions) {
+    r = may_open(in, flags, perms);
+    if (r < 0)
+      goto out;
+  }
+
+  r = _open(in, flags, 0, fhp /* may be NULL */, perms);
+
+ out:
+  Fh *fhptr = fhp ? *fhp : NULL;
+  if (fhptr) {
+    ll_unclosed_fh_set.insert(fhptr);
+  }
+  tout(cct) << (uintptr_t)fhptr << std::endl;
+  ldout(cct, 3) << "ll_open " << vino << " " << ceph_flags_sys2wire(flags) <<
+      " = " << r << " (" << fhptr << ")" << dendl;
+  return r;
+}
+
+int Client::_ll_create(Inode *parent, const char *name, mode_t mode,
+		      int flags, InodeRef *in, int caps, Fh **fhp,
+		      const UserPerm& perms)
+{
+  *fhp = NULL;
+
+  vinodeno_t vparent = _get_vino(parent);
+
+  ldout(cct, 8) << "_ll_create " << vparent << " " << name << " 0" << oct <<
+    mode << dec << " " << ceph_flags_sys2wire(flags) << ", uid " << perms.uid()
+		<< ", gid " << perms.gid() << dendl;
+  tout(cct) << "ll_create" << std::endl;
+  tout(cct) << vparent.ino.val << std::endl;
+  tout(cct) << name << std::endl;
+  tout(cct) << mode << std::endl;
+  tout(cct) << ceph_flags_sys2wire(flags) << std::endl;
+
+  bool created = false;
+  int r = _lookup(parent, name, caps, in, perms);
+
+  if (r == 0 && (flags & O_CREAT) && (flags & O_EXCL))
+    return -CEPHFS_EEXIST;
+
+  if (r == -CEPHFS_ENOENT && (flags & O_CREAT)) {
+    if (!fuse_default_permissions) {
+      r = may_create(parent, perms);
+      if (r < 0)
+	goto out;
+    }
+    r = _create(parent, name, flags, mode, in, fhp, 0, 0, 0, NULL, &created,
+		perms, "");
+    if (r < 0)
+      goto out;
+  }
+
+  if (r < 0)
+    goto out;
+
+  ceph_assert(*in);
+
+  ldout(cct, 20) << "_ll_create created = " << created << dendl;
+  if (!created) {
+    if (!fuse_default_permissions) {
+      r = may_open(in->get(), flags, perms);
+      if (r < 0) {
+	if (*fhp) {
+	  int release_r = _release_fh(*fhp);
+	  ceph_assert(release_r == 0);  // during create, no async data ops should have happened
+	}
+	goto out;
+      }
+    }
+    if (*fhp == NULL) {
+      r = _open(in->get(), flags, mode, fhp, perms);
+      if (r < 0)
+	goto out;
+    }
+  }
+
+out:
+  if (*fhp) {
+    ll_unclosed_fh_set.insert(*fhp);
+  }
+
+  ino_t ino = 0;
+  if (r >= 0) {
+    Inode *inode = in->get();
+    if (use_faked_inos())
+      ino = inode->faked_ino;
+    else
+      ino = inode->ino;
+  }
+
+  tout(cct) << (uintptr_t)*fhp << std::endl;
+  tout(cct) << ino << std::endl;
+  ldout(cct, 8) << "_ll_create " << vparent << " " << name << " 0" << oct <<
+    mode << dec << " " << ceph_flags_sys2wire(flags) << " = " << r << " (" <<
+    *fhp << " " << hex << ino << dec << ")" << dendl;
+
+  return r;
+}
+
+int Client::ll_create(Inode *parent, const char *name, mode_t mode,
+		      int flags, struct stat *attr, Inode **outp, Fh **fhp,
+		      const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+  InodeRef in;
+
+  int r = _ll_create(parent, name, mode, flags, &in, CEPH_STAT_CAP_INODE_ALL,
+		      fhp, perms);
+  if (r >= 0) {
+    ceph_assert(in);
+
+    // passing an Inode in outp requires an additional ref
+    if (outp) {
+      _ll_get(in.get());
+      *outp = in.get();
+    }
+    fill_stat(in, attr);
+  } else {
+    attr->st_ino = 0;
+  }
+
+  return r;
+}
+
+int Client::ll_createx(Inode *parent, const char *name, mode_t mode,
+			int oflags, Inode **outp, Fh **fhp,
+			struct ceph_statx *stx, unsigned want, unsigned lflags,
+			const UserPerm& perms)
+{
+  unsigned caps = statx_to_mask(lflags, want);
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+  InodeRef in;
+
+  int r = _ll_create(parent, name, mode, oflags, &in, caps, fhp, perms);
+  if (r >= 0) {
+    ceph_assert(in);
+
+    // passing an Inode in outp requires an additional ref
+    if (outp) {
+      _ll_get(in.get());
+      *outp = in.get();
+    }
+    fill_statx(in, caps, stx);
+  } else {
+    stx->stx_ino = 0;
+    stx->stx_mask = 0;
+  }
+
+  return r;
+}
+
+loff_t Client::ll_lseek(Fh *fh, loff_t offset, int whence)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << "ll_lseek" << std::endl;
+  tout(cct) << offset << std::endl;
+  tout(cct) << whence << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _lseek(fh, offset, whence);
+}
+
+int Client::ll_read(Fh *fh, loff_t off, loff_t len, bufferlist *bl)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_read " << fh << " " << fh->inode->ino << " " << " " << off << "~" << len << dendl;
+  tout(cct) << "ll_read" << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+  tout(cct) << off << std::endl;
+  tout(cct) << len << std::endl;
+
+  /* We can't return bytes written larger than INT_MAX, clamp len to that */
+  len = std::min(len, (loff_t)INT_MAX);
+  std::scoped_lock lock(client_lock);
+
+  int r = _read(fh, off, len, bl);
+  ldout(cct, 3) << "ll_read " << fh << " " << off << "~" << len << " = " << r
+		<< dendl;
+  return r;
+}
+
+int Client::ll_read_block(Inode *in, uint64_t blockid,
+			  char *buf,
+			  uint64_t offset,
+			  uint64_t length,
+			  file_layout_t* layout)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  vinodeno_t vino = _get_vino(in);
+  object_t oid = file_object_t(vino.ino, blockid);
+  C_SaferCond onfinish;
+  bufferlist bl;
+
+  objecter->read(oid,
+		 object_locator_t(layout->pool_id),
+		 offset,
+		 length,
+		 vino.snapid,
+		 &bl,
+		 CEPH_OSD_FLAG_READ,
+                 &onfinish);
+
+  int r = onfinish.wait();
+  if (r >= 0) {
+      bl.begin().copy(bl.length(), buf);
+      r = bl.length();
+  }
+
+  return r;
+}
+
+/* It appears that the OSD doesn't return success unless the entire
+   buffer was written, return the write length on success. */
+
+int Client::ll_write_block(Inode *in, uint64_t blockid,
+			   char* buf, uint64_t offset,
+			   uint64_t length, file_layout_t* layout,
+			   uint64_t snapseq, uint32_t sync)
+{
+  vinodeno_t vino = ll_get_vino(in);
+  int r = 0;
+  std::unique_ptr<C_SaferCond> onsafe = nullptr;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  if (length == 0) {
+    return -CEPHFS_EINVAL;
+  }
+  if (true || sync) {
+    /* if write is stable, the epilogue is waiting on
+     * flock */
+    onsafe.reset(new C_SaferCond("Client::ll_write_block flock"));
+  }
+  object_t oid = file_object_t(vino.ino, blockid);
+  SnapContext fakesnap;
+  ceph::bufferlist bl;
+  if (length > 0) {
+    bl.push_back(buffer::copy(buf, length));
+  }
+
+  ldout(cct, 1) << "ll_block_write for " << vino.ino << "." << blockid
+		<< dendl;
+
+  fakesnap.seq = snapseq;
+
+  /* lock just in time */
+  objecter->write(oid,
+		  object_locator_t(layout->pool_id),
+		  offset,
+		  length,
+		  fakesnap,
+		  bl,
+		  ceph::real_clock::now(),
+		  0,
+		  onsafe.get());
+
+  if (nullptr != onsafe) {
+    r = onsafe->wait();
+  }
+
+  if (r < 0) {
+    return r;
+  } else {
+    return length;
+  }
+}
+
+int Client::ll_commit_blocks(Inode *in,
+			     uint64_t offset,
+			     uint64_t length)
+{
+    /*
+    BarrierContext *bctx;
+    vinodeno_t vino = _get_vino(in);
+    uint64_t ino = vino.ino;
+
+    ldout(cct, 1) << "ll_commit_blocks for " << vino.ino << " from "
+		  << offset << " to " << length << dendl;
+
+    if (length == 0) {
+      return -CEPHFS_EINVAL;
+    }
+
+    std::scoped_lock lock(client_lock);
+    map<uint64_t, BarrierContext*>::iterator p = barriers.find(ino);
+    if (p != barriers.end()) {
+      barrier_interval civ(offset, offset + length);
+      p->second->commit_barrier(civ);
+    }
+    */
+    return 0;
+}
+
+int Client::ll_write(Fh *fh, loff_t off, loff_t len, const char *data)
+{
+  ldout(cct, 3) << "ll_write " << fh << " " << fh->inode->ino << " " << off <<
+    "~" << len << dendl;
+  tout(cct) << "ll_write" << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+  tout(cct) << off << std::endl;
+  tout(cct) << len << std::endl;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  /* We can't return bytes written larger than INT_MAX, clamp len to that */
+  len = std::min(len, (loff_t)INT_MAX);
+  std::scoped_lock lock(client_lock);
+
+  int r = _write(fh, off, len, data, NULL, 0);
+  ldout(cct, 3) << "ll_write " << fh << " " << off << "~" << len << " = " << r
+		<< dendl;
+  return r;
+}
+
+int64_t Client::ll_writev(struct Fh *fh, const struct iovec *iov, int iovcnt, int64_t off)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::unique_lock cl(client_lock);
+  return _preadv_pwritev_locked(fh, iov, iovcnt, off, true, false, cl);
+}
+
+int64_t Client::ll_readv(struct Fh *fh, const struct iovec *iov, int iovcnt, int64_t off)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::unique_lock cl(client_lock);
+  return _preadv_pwritev_locked(fh, iov, iovcnt, off, false, false, cl);
+}
+
+int Client::ll_flush(Fh *fh)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_flush " << fh << " " << fh->inode->ino << " " << dendl;
+  tout(cct) << "ll_flush" << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _flush(fh);
+}
+
+int Client::ll_fsync(Fh *fh, bool syncdataonly)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_fsync " << fh << " " << fh->inode->ino << " " << dendl;
+  tout(cct) << "ll_fsync" << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  int r = _fsync(fh, syncdataonly);
+  if (r) {
+    // If we're returning an error, clear it from the FH
+    fh->take_async_err();
+  }
+  return r;
+}
+
+int Client::ll_sync_inode(Inode *in, bool syncdataonly)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_sync_inode " << *in << " " << dendl;
+  tout(cct) << "ll_sync_inode" << std::endl;
+  tout(cct) << (uintptr_t)in << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _fsync(in, syncdataonly);
+}
+
+#ifdef FALLOC_FL_PUNCH_HOLE
+
+int Client::_fallocate(Fh *fh, int mode, int64_t offset, int64_t length)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  if (offset < 0 || length <= 0)
+    return -CEPHFS_EINVAL;
+
+  if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
+    return -CEPHFS_EOPNOTSUPP;
+
+  if ((mode & FALLOC_FL_PUNCH_HOLE) && !(mode & FALLOC_FL_KEEP_SIZE))
+    return -CEPHFS_EOPNOTSUPP;
+
+  Inode *in = fh->inode.get();
+
+  if (objecter->osdmap_pool_full(in->layout.pool_id) &&
+      !(mode & FALLOC_FL_PUNCH_HOLE)) {
+    return -CEPHFS_ENOSPC;
+  }
+
+  if (in->snapid != CEPH_NOSNAP)
+    return -CEPHFS_EROFS;
+
+  if ((fh->mode & CEPH_FILE_MODE_WR) == 0)
+    return -CEPHFS_EBADF;
+
+  uint64_t size = offset + length;
+  if (!(mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE)) &&
+      size > in->size &&
+      is_quota_bytes_exceeded(in, size - in->size, fh->actor_perms)) {
+    return -CEPHFS_EDQUOT;
+  }
+
+  int have;
+  int r = get_caps(fh, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER, &have, -1);
+  if (r < 0)
+    return r;
+
+  std::unique_ptr<C_SaferCond> onuninline = nullptr;
+  if (mode & FALLOC_FL_PUNCH_HOLE) {
+    if (in->inline_version < CEPH_INLINE_NONE &&
+        (have & CEPH_CAP_FILE_BUFFER)) {
+      bufferlist bl;
+      auto inline_iter = in->inline_data.cbegin();
+      int len = in->inline_data.length();
+      if (offset < len) {
+        if (offset > 0)
+          inline_iter.copy(offset, bl);
+        int size = length;
+        if (offset + size > len)
+          size = len - offset;
+        if (size > 0)
+          bl.append_zero(size);
+        if (offset + size < len) {
+          inline_iter += size;
+          inline_iter.copy(len - offset - size, bl);
+        }
+        in->inline_data = bl;
+        in->inline_version++;
+      }
+      in->mtime = in->ctime = ceph_clock_now();
+      in->change_attr++;
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+    } else {
+      if (in->inline_version < CEPH_INLINE_NONE) {
+        onuninline.reset(new C_SaferCond("Client::_fallocate_uninline_data flock"));
+        uninline_data(in, onuninline.get());
+      }
+
+      C_SaferCond onfinish("Client::_punch_hole flock");
+
+      get_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+
+      _invalidate_inode_cache(in, offset, length);
+      filer->zero(in->ino, &in->layout,
+		  in->snaprealm->get_snap_context(),
+		  offset, length,
+		  ceph::real_clock::now(),
+		  0, true, &onfinish);
+      in->mtime = in->ctime = ceph_clock_now();
+      in->change_attr++;
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+
+      client_lock.unlock();
+      onfinish.wait();
+      client_lock.lock();
+      put_cap_ref(in, CEPH_CAP_FILE_BUFFER);
+    }
+  } else if (!(mode & FALLOC_FL_KEEP_SIZE)) {
+    uint64_t size = offset + length;
+    if (size > in->size) {
+      in->size = size;
+      in->mtime = in->ctime = ceph_clock_now();
+      in->change_attr++;
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+
+      if (is_quota_bytes_approaching(in, fh->actor_perms)) {
+        check_caps(in, CHECK_CAPS_NODELAY);
+      } else if (is_max_size_approaching(in)) {
+	check_caps(in, 0);
+      }
+    }
+  }
+
+  if (nullptr != onuninline) {
+    client_lock.unlock();
+    int ret = onuninline->wait();
+    client_lock.lock();
+
+    if (ret >= 0 || ret == -CEPHFS_ECANCELED) {
+      in->inline_data.clear();
+      in->inline_version = CEPH_INLINE_NONE;
+      in->mark_caps_dirty(CEPH_CAP_FILE_WR);
+      check_caps(in, 0);
+    } else
+      r = ret;
+  }
+
+  put_cap_ref(in, CEPH_CAP_FILE_WR);
+  return r;
+}
+#else
+
+int Client::_fallocate(Fh *fh, int mode, int64_t offset, int64_t length)
+{
+  return -CEPHFS_EOPNOTSUPP;
+}
+
+#endif
+
+
+int Client::ll_fallocate(Fh *fh, int mode, int64_t offset, int64_t length)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << " " << fh << " " << fh->inode->ino << " " << dendl;
+  tout(cct) << __func__ << " " << mode << " " << offset << " " << length << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _fallocate(fh, mode, offset, length);
+}
+
+int Client::fallocate(int fd, int mode, loff_t offset, loff_t length)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  tout(cct) << __func__ << " " << " " << fd << mode << " " << offset << " " << length << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  Fh *fh = get_filehandle(fd);
+  if (!fh)
+    return -CEPHFS_EBADF;
+#if defined(__linux__) && defined(O_PATH)
+  if (fh->flags & O_PATH)
+    return -CEPHFS_EBADF;
+#endif
+  return _fallocate(fh, mode, offset, length);
+}
+
+int Client::ll_release(Fh *fh)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << " (fh)" << fh << " " << fh->inode->ino << " " <<
+    dendl;
+  tout(cct) << __func__ << " (fh)" << std::endl;
+  tout(cct) << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+
+  if (ll_unclosed_fh_set.count(fh))
+    ll_unclosed_fh_set.erase(fh);
+  return _release_fh(fh);
+}
+
+int Client::ll_getlk(Fh *fh, struct flock *fl, uint64_t owner)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << "ll_getlk (fh)" << fh << " " << fh->inode->ino << dendl;
+  tout(cct) << "ll_getk (fh)" << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _getlk(fh, fl, owner);
+}
+
+int Client::ll_setlk(Fh *fh, struct flock *fl, uint64_t owner, int sleep)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << "  (fh) " << fh << " " << fh->inode->ino << dendl;
+  tout(cct) << __func__ << " (fh)" << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _setlk(fh, fl, owner, sleep);
+}
+
+int Client::ll_flock(Fh *fh, int cmd, uint64_t owner)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  ldout(cct, 3) << __func__ << "  (fh) " << fh << " " << fh->inode->ino << dendl;
+  tout(cct) << __func__ << " (fh)" << (uintptr_t)fh << std::endl;
+
+  std::scoped_lock lock(client_lock);
+  return _flock(fh, cmd, owner);
+}
+
+int Client::set_deleg_timeout(uint32_t timeout)
+{
+  std::scoped_lock lock(client_lock);
+
+  /*
+   * The whole point is to prevent blocklisting so we must time out the
+   * delegation before the session autoclose timeout kicks in.
+   */
+  if (timeout >= mdsmap->get_session_autoclose())
+    return -CEPHFS_EINVAL;
+
+  deleg_timeout = timeout;
+  return 0;
+}
+
+int Client::ll_delegation(Fh *fh, unsigned cmd, ceph_deleg_cb_t cb, void *priv)
+{
+  int ret = -CEPHFS_EINVAL;
+
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Inode *inode = fh->inode.get();
+
+  switch(cmd) {
+  case CEPH_DELEGATION_NONE:
+    inode->unset_deleg(fh);
+    ret = 0;
+    break;
+  default:
+    try {
+      ret = inode->set_deleg(fh, cmd, cb, priv);
+    } catch (std::bad_alloc&) {
+      ret = -CEPHFS_ENOMEM;
+    }
+    break;
+  }
+  return ret;
+}
+
+class C_Client_RequestInterrupt : public Context  {
+private:
+  Client *client;
+  MetaRequest *req;
+public:
+  C_Client_RequestInterrupt(Client *c, MetaRequest *r) : client(c), req(r) {
+    req->get();
+  }
+  void finish(int r) override {
+    std::scoped_lock l(client->client_lock);
+    ceph_assert(req->head.op == CEPH_MDS_OP_SETFILELOCK);
+    client->_interrupt_filelock(req);
+    client->put_request(req);
+  }
+};
+
+void Client::ll_interrupt(void *d)
+{
+  MetaRequest *req = static_cast<MetaRequest*>(d);
+  ldout(cct, 3) << __func__ << " tid " << req->get_tid() << dendl;
+  tout(cct) << __func__ << " tid " << req->get_tid() << std::endl;
+  interrupt_finisher.queue(new C_Client_RequestInterrupt(this, req));
+}
+
+// =========================================
+// layout
+
+// expose file layouts
+
+int Client::describe_layout(const char *relpath, file_layout_t *lp,
+			    const UserPerm& perms)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  filepath path(relpath);
+  InodeRef in;
+  int r = path_walk(path, &in, perms);
+  if (r < 0)
+    return r;
+
+  *lp = in->layout;
+
+  ldout(cct, 3) << __func__ << "(" << relpath << ") = 0" << dendl;
+  return 0;
+}
+
+int Client::fdescribe_layout(int fd, file_layout_t *lp)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  Inode *in = f->inode.get();
+
+  *lp = in->layout;
+
+  ldout(cct, 3) << __func__ << "(" << fd << ") = 0" << dendl;
+  return 0;
+}
+
+int64_t Client::get_default_pool_id()
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  /* first data pool is the default */ 
+  return mdsmap->get_first_data_pool(); 
+}
+
+// expose osdmap
+
+int64_t Client::get_pool_id(const char *pool_name)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  return objecter->with_osdmap(std::mem_fn(&OSDMap::lookup_pg_pool_name),
+			       pool_name);
+}
+
+string Client::get_pool_name(int64_t pool)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return string();
+
+  std::scoped_lock lock(client_lock);
+
+  return objecter->with_osdmap([pool](const OSDMap& o) {
+      return o.have_pg_pool(pool) ? o.get_pool_name(pool) : string();
+    });
+}
+
+int Client::get_pool_replication(int64_t pool)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  return objecter->with_osdmap([pool](const OSDMap& o) {
+      return o.have_pg_pool(pool) ? o.get_pg_pool(pool)->get_size() : -CEPHFS_ENOENT;
+    });
+}
+
+int Client::get_file_extent_osds(int fd, loff_t off, loff_t *len, vector<int>& osds)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  Inode *in = f->inode.get();
+
+  vector<ObjectExtent> extents;
+  Striper::file_to_extents(cct, in->ino, &in->layout, off, 1, in->truncate_size, extents);
+  ceph_assert(extents.size() == 1);
+
+  objecter->with_osdmap([&](const OSDMap& o) {
+      pg_t pg = o.object_locator_to_pg(extents[0].oid, extents[0].oloc);
+      o.pg_to_acting_osds(pg, osds);
+    });
+
+  if (osds.empty())
+    return -CEPHFS_EINVAL;
+
+  /*
+   * Return the remainder of the extent (stripe unit)
+   *
+   * If length = 1 is passed to Striper::file_to_extents we get a single
+   * extent back, but its length is one so we still need to compute the length
+   * to the end of the stripe unit.
+   *
+   * If length = su then we may get 1 or 2 objects back in the extents vector
+   * which would have to be examined. Even then, the offsets are local to the
+   * object, so matching up to the file offset is extra work.
+   *
+   * It seems simpler to stick with length = 1 and manually compute the
+   * remainder.
+   */
+  if (len) {
+    uint64_t su = in->layout.stripe_unit;
+    *len = su - (off % su);
+  }
+
+  return 0;
+}
+
+int Client::get_osd_crush_location(int id, vector<pair<string, string> >& path)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  if (id < 0)
+    return -CEPHFS_EINVAL;
+  return objecter->with_osdmap([&](const OSDMap& o) {
+      return o.crush->get_full_location_ordered(id, path);
+    });
+}
+
+int Client::get_file_stripe_address(int fd, loff_t offset,
+				    vector<entity_addr_t>& address)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  Inode *in = f->inode.get();
+
+  // which object?
+  vector<ObjectExtent> extents;
+  Striper::file_to_extents(cct, in->ino, &in->layout, offset, 1,
+			   in->truncate_size, extents);
+  ceph_assert(extents.size() == 1);
+
+  // now we have the object and its 'layout'
+  return objecter->with_osdmap([&](const OSDMap& o) {
+      pg_t pg = o.object_locator_to_pg(extents[0].oid, extents[0].oloc);
+      vector<int> osds;
+      o.pg_to_acting_osds(pg, osds);
+      if (osds.empty())
+	return -CEPHFS_EINVAL;
+      for (unsigned i = 0; i < osds.size(); i++) {
+	entity_addr_t addr = o.get_addrs(osds[i]).front();
+	address.push_back(addr);
+      }
+      return 0;
+    });
+}
+
+int Client::get_osd_addr(int osd, entity_addr_t& addr)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  return objecter->with_osdmap([&](const OSDMap& o) {
+      if (!o.exists(osd))
+	return -CEPHFS_ENOENT;
+
+      addr = o.get_addrs(osd).front();
+      return 0;
+    });
+}
+
+int Client::enumerate_layout(int fd, vector<ObjectExtent>& result,
+			     loff_t length, loff_t offset)
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  Fh *f = get_filehandle(fd);
+  if (!f)
+    return -CEPHFS_EBADF;
+  Inode *in = f->inode.get();
+
+  // map to a list of extents
+  Striper::file_to_extents(cct, in->ino, &in->layout, offset, length, in->truncate_size, result);
+
+  ldout(cct, 3) << __func__ << "(" << fd << ", " << length << ", " << offset << ") = 0" << dendl;
+  return 0;
+}
+
+
+/* find an osd with the same ip.  -CEPHFS_ENXIO if none. */
+int Client::get_local_osd()
+{
+  RWRef_t mref_reader(mount_state, CLIENT_MOUNTING);
+  if (!mref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  std::scoped_lock lock(client_lock);
+
+  objecter->with_osdmap([this](const OSDMap& o) {
+      if (o.get_epoch() != local_osd_epoch) {
+	local_osd = o.find_osd_on_ip(messenger->get_myaddrs().front());
+	local_osd_epoch = o.get_epoch();
+      }
+    });
+  return local_osd;
+}
+
+
+
+
+
+
+// ===============================
+
+void Client::ms_handle_connect(Connection *con)
+{
+  ldout(cct, 10) << __func__ << " on " << con->get_peer_addr() << dendl;
+}
+
+bool Client::ms_handle_reset(Connection *con)
+{
+  ldout(cct, 0) << __func__ << " on " << con->get_peer_addr() << dendl;
+  return false;
+}
+
+void Client::ms_handle_remote_reset(Connection *con)
+{
+  std::scoped_lock lock(client_lock);
+  ldout(cct, 0) << __func__ << " on " << con->get_peer_addr() << dendl;
+  switch (con->get_peer_type()) {
+  case CEPH_ENTITY_TYPE_MDS:
+    {
+      // kludge to figure out which mds this is; fixme with a Connection* state
+      mds_rank_t mds = MDS_RANK_NONE;
+      MetaSession *s = NULL;
+      for (auto &p : mds_sessions) {
+	if (mdsmap->have_inst(p.first) && mdsmap->get_addrs(p.first) == con->get_peer_addrs()) {
+	  mds = p.first;
+	  s = &p.second;
+	}
+      }
+      if (mds >= 0) {
+	assert (s != NULL);
+	switch (s->state) {
+	case MetaSession::STATE_CLOSING:
+	  ldout(cct, 1) << "reset from mds we were closing; we'll call that closed" << dendl;
+	  _closed_mds_session(s);
+	  break;
+
+	case MetaSession::STATE_OPENING:
+	  {
+	    ldout(cct, 1) << "reset from mds we were opening; retrying" << dendl;
+	    list<Context*> waiters;
+	    waiters.swap(s->waiting_for_open);
+	    _closed_mds_session(s);
+	    MetaSession *news = _get_or_open_mds_session(mds);
+	    news->waiting_for_open.swap(waiters);
+	  }
+	  break;
+
+	case MetaSession::STATE_OPEN:
+	  {
+	    objecter->maybe_request_map(); /* to check if we are blocklisted */
+	    if (cct->_conf.get_val<bool>("client_reconnect_stale")) {
+	      ldout(cct, 1) << "reset from mds we were open; close mds session for reconnect" << dendl;
+	      _closed_mds_session(s);
+	    } else {
+	      ldout(cct, 1) << "reset from mds we were open; mark session as stale" << dendl;
+	      s->state = MetaSession::STATE_STALE;
+	    }
+	  }
+	  break;
+
+	case MetaSession::STATE_NEW:
+	case MetaSession::STATE_CLOSED:
+	default:
+	  break;
+	}
+      }
+    }
+    break;
+  }
+}
+
+bool Client::ms_handle_refused(Connection *con)
+{
+  ldout(cct, 1) << __func__ << " on " << con->get_peer_addr() << dendl;
+  return false;
+}
+
+Inode *Client::get_quota_root(Inode *in, const UserPerm& perms)
+{
+  Inode *quota_in = root_ancestor;
+  SnapRealm *realm = in->snaprealm;
+
+  if (!cct->_conf.get_val<bool>("client_quota"))
+    return NULL;
+
+  while (realm) {
+    ldout(cct, 10) << __func__ << " realm " << realm->ino << dendl;
+    if (realm->ino != in->ino) {
+      auto p = inode_map.find(vinodeno_t(realm->ino, CEPH_NOSNAP));
+      if (p == inode_map.end())
+	break;
+
+      if (p->second->quota.is_enable()) {
+	quota_in = p->second;
+	break;
+      }
+    }
+    realm = realm->pparent;
+  }
+  ldout(cct, 10) << __func__ << " " << in->vino() << " -> " << quota_in->vino() << dendl;
+  return quota_in;
+}
+
+/**
+ * Traverse quota ancestors of the Inode, return true
+ * if any of them passes the passed function
+ */
+bool Client::check_quota_condition(Inode *in, const UserPerm& perms,
+				   std::function<bool (const Inode &in)> test)
+{
+  if (!cct->_conf.get_val<bool>("client_quota"))
+    return false;
+
+  while (true) {
+    ceph_assert(in != NULL);
+    if (test(*in)) {
+      return true;
+    }
+
+    if (in == root_ancestor) {
+      // We're done traversing, drop out
+      return false;
+    } else {
+      // Continue up the tree
+      in = get_quota_root(in, perms);
+    }
+  }
+
+  return false;
+}
+
+bool Client::is_quota_files_exceeded(Inode *in, const UserPerm& perms)
+{
+  return check_quota_condition(in, perms,
+      [](const Inode &in) {
+        return in.quota.max_files && in.rstat.rsize() >= in.quota.max_files;
+      });
+}
+
+bool Client::is_quota_bytes_exceeded(Inode *in, int64_t new_bytes,
+				     const UserPerm& perms)
+{
+  return check_quota_condition(in, perms,
+      [&new_bytes](const Inode &in) {
+        return in.quota.max_bytes && (in.rstat.rbytes + new_bytes)
+               > in.quota.max_bytes;
+      });
+}
+
+bool Client::is_quota_bytes_approaching(Inode *in, const UserPerm& perms)
+{
+  ceph_assert(in->size >= in->reported_size);
+  const uint64_t size = in->size - in->reported_size;
+  return check_quota_condition(in, perms,
+      [&size](const Inode &in) {
+        if (in.quota.max_bytes) {
+          if (in.rstat.rbytes >= in.quota.max_bytes) {
+            return true;
+          }
+
+          const uint64_t space = in.quota.max_bytes - in.rstat.rbytes;
+          return (space >> 4) < size;
+        } else {
+          return false;
+        }
+      });
+}
+
+enum {
+  POOL_CHECKED = 1,
+  POOL_CHECKING = 2,
+  POOL_READ = 4,
+  POOL_WRITE = 8,
+};
+
+int Client::check_pool_perm(Inode *in, int need)
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  if (!cct->_conf->client_check_pool_perm)
+    return 0;
+
+  /* Only need to do this for regular files */
+  if (!in->is_file())
+    return 0;
+
+  int64_t pool_id = in->layout.pool_id;
+  std::string pool_ns = in->layout.pool_ns;
+  std::pair<int64_t, std::string> perm_key(pool_id, pool_ns);
+  int have = 0;
+  while (true) {
+    auto it = pool_perms.find(perm_key);
+    if (it == pool_perms.end())
+      break;
+    if (it->second == POOL_CHECKING) {
+      // avoid concurrent checkings
+      wait_on_list(waiting_for_pool_perm);
+    } else {
+      have = it->second;
+      ceph_assert(have & POOL_CHECKED);
+      break;
+    }
+  }
+
+  if (!have) {
+    if (in->snapid != CEPH_NOSNAP) {
+      // pool permission check needs to write to the first object. But for snapshot,
+      // head of the first object may have alread been deleted. To avoid creating
+      // orphan object, skip the check for now.
+      return 0;
+    }
+
+    pool_perms[perm_key] = POOL_CHECKING;
+
+    char oid_buf[32];
+    snprintf(oid_buf, sizeof(oid_buf), "%llx.00000000", (unsigned long long)in->ino);
+    object_t oid = oid_buf;
+
+    SnapContext nullsnapc;
+
+    C_SaferCond rd_cond;
+    ObjectOperation rd_op;
+    rd_op.stat(nullptr, nullptr, nullptr);
+
+    objecter->mutate(oid, OSDMap::file_to_object_locator(in->layout), rd_op,
+		     nullsnapc, ceph::real_clock::now(), 0, &rd_cond);
+
+    C_SaferCond wr_cond;
+    ObjectOperation wr_op;
+    wr_op.create(true);
+
+    objecter->mutate(oid, OSDMap::file_to_object_locator(in->layout), wr_op,
+		     nullsnapc, ceph::real_clock::now(), 0, &wr_cond);
+
+    client_lock.unlock();
+    int rd_ret = rd_cond.wait();
+    int wr_ret = wr_cond.wait();
+    client_lock.lock();
+
+    bool errored = false;
+
+    if (rd_ret == 0 || rd_ret == -CEPHFS_ENOENT)
+      have |= POOL_READ;
+    else if (rd_ret != -CEPHFS_EPERM) {
+      ldout(cct, 10) << __func__ << " on pool " << pool_id << " ns " << pool_ns
+		     << " rd_err = " << rd_ret << " wr_err = " << wr_ret << dendl;
+      errored = true;
+    }
+
+    if (wr_ret == 0 || wr_ret == -CEPHFS_EEXIST)
+      have |= POOL_WRITE;
+    else if (wr_ret != -CEPHFS_EPERM) {
+      ldout(cct, 10) << __func__ << " on pool " << pool_id << " ns " << pool_ns
+		     << " rd_err = " << rd_ret << " wr_err = " << wr_ret << dendl;
+      errored = true;
+    }
+
+    if (errored) {
+      // Indeterminate: erase CHECKING state so that subsequent calls re-check.
+      // Raise EIO because actual error code might be misleading for
+      // userspace filesystem user.
+      pool_perms.erase(perm_key);
+      signal_cond_list(waiting_for_pool_perm);
+      return -CEPHFS_EIO;
+    }
+
+    pool_perms[perm_key] = have | POOL_CHECKED;
+    signal_cond_list(waiting_for_pool_perm);
+  }
+
+  if ((need & CEPH_CAP_FILE_RD) && !(have & POOL_READ)) {
+    ldout(cct, 10) << __func__ << " on pool " << pool_id << " ns " << pool_ns
+		   << " need " << ccap_string(need) << ", but no read perm" << dendl;
+    return -CEPHFS_EPERM;
+  }
+  if ((need & CEPH_CAP_FILE_WR) && !(have & POOL_WRITE)) {
+    ldout(cct, 10) << __func__ << " on pool " << pool_id << " ns " << pool_ns
+		   << " need " << ccap_string(need) << ", but no write perm" << dendl;
+    return -CEPHFS_EPERM;
+  }
+
+  return 0;
+}
+
+int Client::_posix_acl_permission(Inode *in, const UserPerm& perms, unsigned want)
+{
+  if (acl_type == POSIX_ACL) {
+    if (in->xattrs.count(ACL_EA_ACCESS)) {
+      const bufferptr& access_acl = in->xattrs[ACL_EA_ACCESS];
+
+      return posix_acl_permits(access_acl, in->uid, in->gid, perms, want);
+    }
+  }
+  return -CEPHFS_EAGAIN;
+}
+
+int Client::_posix_acl_chmod(Inode *in, mode_t mode, const UserPerm& perms)
+{
+  if (acl_type == NO_ACL)
+    return 0;
+
+  int r = _getattr(in, CEPH_STAT_CAP_XATTR, perms, in->xattr_version == 0);
+  if (r < 0)
+    goto out;
+
+  if (acl_type == POSIX_ACL) {
+    if (in->xattrs.count(ACL_EA_ACCESS)) {
+      const bufferptr& access_acl = in->xattrs[ACL_EA_ACCESS];
+      bufferptr acl(access_acl.c_str(), access_acl.length());
+      r = posix_acl_access_chmod(acl, mode);
+      if (r < 0)
+	goto out;
+      r = _do_setxattr(in, ACL_EA_ACCESS, acl.c_str(), acl.length(), 0, perms);
+    } else {
+      r = 0;
+    }
+  }
+out:
+  ldout(cct, 10) << __func__ << " ino " << in->ino << " result=" << r << dendl;
+  return r;
+}
+
+int Client::_posix_acl_create(Inode *dir, mode_t *mode, bufferlist& xattrs_bl,
+			      const UserPerm& perms)
+{
+  if (acl_type == NO_ACL)
+    return 0;
+
+  if (S_ISLNK(*mode))
+    return 0;
+
+  int r = _getattr(dir, CEPH_STAT_CAP_XATTR, perms, dir->xattr_version == 0);
+  if (r < 0)
+    goto out;
+
+  if (acl_type == POSIX_ACL) {
+    if (dir->xattrs.count(ACL_EA_DEFAULT)) {
+      map<string, bufferptr> xattrs;
+
+      const bufferptr& default_acl = dir->xattrs[ACL_EA_DEFAULT];
+      bufferptr acl(default_acl.c_str(), default_acl.length());
+      r = posix_acl_inherit_mode(acl, mode);
+      if (r < 0)
+	goto out;
+
+      if (r > 0) {
+	r = posix_acl_equiv_mode(acl.c_str(), acl.length(), mode);
+	if (r < 0)
+	  goto out;
+	if (r > 0)
+	  xattrs[ACL_EA_ACCESS] = acl;
+      }
+
+      if (S_ISDIR(*mode))
+	xattrs[ACL_EA_DEFAULT] = dir->xattrs[ACL_EA_DEFAULT];
+
+      r = xattrs.size();
+      if (r > 0)
+	encode(xattrs, xattrs_bl);
+    } else {
+      if (umask_cb)
+	*mode &= ~umask_cb(callback_handle);
+      r = 0;
+    }
+  }
+out:
+  ldout(cct, 10) << __func__ << " dir ino " << dir->ino << " result=" << r << dendl;
+  return r;
+}
+
+void Client::set_filer_flags(int flags)
+{
+  std::scoped_lock l(client_lock);
+  ceph_assert(flags == 0 ||
+	 flags == CEPH_OSD_FLAG_LOCALIZE_READS);
+  objecter->add_global_op_flags(flags);
+}
+
+void Client::clear_filer_flags(int flags)
+{
+  std::scoped_lock l(client_lock);
+  ceph_assert(flags == CEPH_OSD_FLAG_LOCALIZE_READS);
+  objecter->clear_global_op_flag(flags);
+}
+
+// called before mount
+void Client::set_uuid(const std::string& uuid)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  ceph_assert(iref_reader.is_state_satisfied());
+
+  std::scoped_lock l(client_lock);
+  assert(!uuid.empty());
+
+  metadata["uuid"] = uuid;
+  _close_sessions();
+}
+
+// called before mount. 0 means infinite
+void Client::set_session_timeout(unsigned timeout)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  ceph_assert(iref_reader.is_state_satisfied());
+
+  std::scoped_lock l(client_lock);
+
+  metadata["timeout"] = stringify(timeout);
+}
+
+// called before mount
+int Client::start_reclaim(const std::string& uuid, unsigned flags,
+			  const std::string& fs_name)
+{
+  RWRef_t iref_reader(initialize_state, CLIENT_INITIALIZED);
+  if (!iref_reader.is_state_satisfied())
+    return -CEPHFS_ENOTCONN;
+
+  if (uuid.empty())
+    return -CEPHFS_EINVAL;
+
+  std::unique_lock l(client_lock);
+  {
+    auto it = metadata.find("uuid");
+    if (it != metadata.end() && it->second == uuid)
+      return -CEPHFS_EINVAL;
+  }
+
+  int r = subscribe_mdsmap(fs_name);
+  if (r < 0) {
+    lderr(cct) << "mdsmap subscription failed: " << cpp_strerror(r) << dendl;
+    return r;
+  }
+
+  if (metadata.empty())
+    populate_metadata("");
+
+  while (mdsmap->get_epoch() == 0)
+    wait_on_list(waiting_for_mdsmap);
+
+  reclaim_errno = 0;
+  for (unsigned mds = 0; mds < mdsmap->get_num_in_mds(); ) {
+    if (!mdsmap->is_up(mds)) {
+      ldout(cct, 10) << "mds." << mds << " not active, waiting for new mdsmap" << dendl;
+      wait_on_list(waiting_for_mdsmap);
+      continue;
+    }
+
+    MetaSession *session;
+    if (!have_open_session(mds)) {
+      session = _get_or_open_mds_session(mds);
+      if (session->state == MetaSession::STATE_REJECTED)
+	return -CEPHFS_EPERM;
+      if (session->state != MetaSession::STATE_OPENING) {
+	// umounting?
+	return -CEPHFS_EINVAL;
+      }
+      ldout(cct, 10) << "waiting for session to mds." << mds << " to open" << dendl;
+      wait_on_context_list(session->waiting_for_open);
+      continue;
+    }
+
+    session = &mds_sessions.at(mds);
+    if (!session->mds_features.test(CEPHFS_FEATURE_RECLAIM_CLIENT))
+      return -CEPHFS_EOPNOTSUPP;
+
+    if (session->reclaim_state == MetaSession::RECLAIM_NULL ||
+	session->reclaim_state == MetaSession::RECLAIMING) {
+      session->reclaim_state = MetaSession::RECLAIMING;
+      auto m = make_message<MClientReclaim>(uuid, flags);
+      session->con->send_message2(std::move(m));
+      wait_on_list(waiting_for_reclaim);
+    } else if (session->reclaim_state == MetaSession::RECLAIM_FAIL) {
+      return reclaim_errno ? : -CEPHFS_ENOTRECOVERABLE;
+    } else {
+      mds++;
+    }
+  }
+
+  // didn't find target session in any mds
+  if (reclaim_target_addrs.empty()) {
+    if (flags & CEPH_RECLAIM_RESET)
+      return -CEPHFS_ENOENT;
+    return -CEPHFS_ENOTRECOVERABLE;
+  }
+
+  if (flags & CEPH_RECLAIM_RESET)
+    return 0;
+
+  // use blocklist to check if target session was killed
+  // (config option mds_session_blocklist_on_evict needs to be true)
+  ldout(cct, 10) << __func__ << ": waiting for OSD epoch " << reclaim_osd_epoch << dendl;
+  bs::error_code ec;
+  l.unlock();
+  objecter->wait_for_map(reclaim_osd_epoch, ca::use_blocked[ec]);
+  l.lock();
+
+  if (ec)
+    return ceph::from_error_code(ec);
+
+  bool blocklisted = objecter->with_osdmap(
+      [this](const OSDMap &osd_map) -> bool {
+	return osd_map.is_blocklisted(reclaim_target_addrs);
+      });
+  if (blocklisted)
+    return -CEPHFS_ENOTRECOVERABLE;
+
+  metadata["reclaiming_uuid"] = uuid;
+  return 0;
+}
+
+void Client::finish_reclaim()
+{
+  auto it = metadata.find("reclaiming_uuid");
+  if (it == metadata.end()) {
+    for (auto &p : mds_sessions)
+      p.second.reclaim_state = MetaSession::RECLAIM_NULL;
+    return;
+  }
+
+  for (auto &p : mds_sessions) {
+    p.second.reclaim_state = MetaSession::RECLAIM_NULL;
+    auto m = make_message<MClientReclaim>("", MClientReclaim::FLAG_FINISH);
+    p.second.con->send_message2(std::move(m));
+  }
+
+  metadata["uuid"] = it->second;
+  metadata.erase(it);
+}
+
+void Client::handle_client_reclaim_reply(const MConstRef<MClientReclaimReply>& reply)
+{
+  mds_rank_t from = mds_rank_t(reply->get_source().num());
+  ldout(cct, 10) << __func__ << " " << *reply << " from mds." << from << dendl;
+
+  std::scoped_lock cl(client_lock);
+  MetaSession *session = _get_mds_session(from, reply->get_connection().get());
+  if (!session) {
+    ldout(cct, 10) << " discarding reclaim reply from sessionless mds." <<  from << dendl;
+    return;
+  }
+
+  if (reply->get_result() >= 0) {
+    session->reclaim_state = MetaSession::RECLAIM_OK;
+    if (reply->get_epoch() > reclaim_osd_epoch)
+      reclaim_osd_epoch = reply->get_epoch();
+    if (!reply->get_addrs().empty())
+      reclaim_target_addrs = reply->get_addrs();
+  } else {
+    session->reclaim_state = MetaSession::RECLAIM_FAIL;
+    reclaim_errno = reply->get_result();
+  }
+
+  signal_cond_list(waiting_for_reclaim);
+}
+
+/**
+ * This is included in cap release messages, to cause
+ * the MDS to wait until this OSD map epoch.  It is necessary
+ * in corner cases where we cancel RADOS ops, so that
+ * nobody else tries to do IO to the same objects in
+ * the same epoch as the cancelled ops.
+ */
+void Client::set_cap_epoch_barrier(epoch_t e)
+{
+  ldout(cct, 5) << __func__ << " epoch = " << e << dendl;
+  cap_epoch_barrier = e;
+}
+
+const char** Client::get_tracked_conf_keys() const
+{
+  static const char* keys[] = {
+    "client_cache_size",
+    "client_cache_mid",
+    "client_acl_type",
+    "client_deleg_timeout",
+    "client_deleg_break_on_open",
+    "client_oc_size",
+    "client_oc_max_objects",
+    "client_oc_max_dirty",
+    "client_oc_target_dirty",
+    "client_oc_max_dirty_age",
+    "client_caps_release_delay",
+    "client_mount_timeout",
+    NULL
+  };
+  return keys;
+}
+
+void Client::handle_conf_change(const ConfigProxy& conf,
+				const std::set <std::string> &changed)
+{
+  std::scoped_lock lock(client_lock);
+
+  if (changed.count("client_cache_mid")) {
+    lru.lru_set_midpoint(cct->_conf->client_cache_mid);
+  }
+  if (changed.count("client_acl_type")) {
+    acl_type = NO_ACL;
+    if (cct->_conf->client_acl_type == "posix_acl")
+      acl_type = POSIX_ACL;
+  }
+  if (changed.count("client_oc_size")) {
+    objectcacher->set_max_size(cct->_conf->client_oc_size);
+  }
+  if (changed.count("client_oc_max_objects")) {
+    objectcacher->set_max_objects(cct->_conf->client_oc_max_objects);
+  }
+  if (changed.count("client_oc_max_dirty")) {
+    objectcacher->set_max_dirty(cct->_conf->client_oc_max_dirty);
+  }
+  if (changed.count("client_oc_target_dirty")) {
+    objectcacher->set_target_dirty(cct->_conf->client_oc_target_dirty);
+  }
+  if (changed.count("client_oc_max_dirty_age")) {
+    objectcacher->set_max_dirty_age(cct->_conf->client_oc_max_dirty_age);
+  }
+  if (changed.count("client_collect_and_send_global_metrics")) {
+    _collect_and_send_global_metrics = cct->_conf.get_val<bool>(
+      "client_collect_and_send_global_metrics");
+  }    
+  if (changed.count("client_caps_release_delay")) {
+    caps_release_delay = cct->_conf.get_val<std::chrono::seconds>(
+      "client_caps_release_delay");
+  }
+  if (changed.count("client_mount_timeout")) {
+    mount_timeout = cct->_conf.get_val<std::chrono::seconds>(
+      "client_mount_timeout");
+  }
+}
+
+void intrusive_ptr_add_ref(Inode *in)
+{
+  in->iget();
+}
+
+void intrusive_ptr_release(Inode *in)
+{
+  in->client->put_inode(in);
+}
+
+mds_rank_t Client::_get_random_up_mds() const
+{
+  ceph_assert(ceph_mutex_is_locked_by_me(client_lock));
+
+  std::set<mds_rank_t> up;
+  mdsmap->get_up_mds_set(up);
+
+  if (up.empty())
+    return MDS_RANK_NONE;
+  std::set<mds_rank_t>::const_iterator p = up.begin();
+  for (int n = rand() % up.size(); n; n--)
+    ++p;
+  return *p;
+}
+
+
+StandaloneClient::StandaloneClient(Messenger *m, MonClient *mc,
+				   boost::asio::io_context& ictx)
+  : Client(m, mc, new Objecter(m->cct, m, mc, ictx))
+{
+  monclient->set_messenger(m);
+  objecter->set_client_incarnation(0);
+}
+
+StandaloneClient::~StandaloneClient()
+{
+  delete objecter;
+  objecter = nullptr;
+}
+
+int StandaloneClient::init()
+{
+  RWRef_t iref_writer(initialize_state, CLIENT_INITIALIZING, false);
+  ceph_assert(iref_writer.is_first_writer());
+
+  _pre_init();
+  objecter->init();
+
+  client_lock.lock();
+
+  messenger->add_dispatcher_tail(objecter);
+  messenger->add_dispatcher_tail(this);
+
+  monclient->set_want_keys(CEPH_ENTITY_TYPE_MDS | CEPH_ENTITY_TYPE_OSD);
+  int r = monclient->init();
+  if (r < 0) {
+    // need to do cleanup because we're in an intermediate init state
+    {
+      std::scoped_lock l(timer_lock);
+      timer.shutdown();
+    }
+
+    client_lock.unlock();
+    objecter->shutdown();
+    objectcacher->stop();
+    monclient->shutdown();
+    return r;
+  }
+  objecter->start();
+
+  client_lock.unlock();
+  _finish_init();
+  iref_writer.update_state(CLIENT_INITIALIZED);
+
+  return 0;
+}
+
+void StandaloneClient::shutdown()
+{
+  Client::shutdown();
+  objecter->shutdown();
+  monclient->shutdown();
+}
diff --git a/src/client/Client.h b/src/client/Client.h
new file mode 100644
index 000000000..497f016da
--- /dev/null
+++ b/src/client/Client.h
@@ -0,0 +1,1613 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+
+#ifndef CEPH_CLIENT_H
+#define CEPH_CLIENT_H
+
+#include "common/CommandTable.h"
+#include "common/Finisher.h"
+#include "common/Timer.h"
+#include "common/ceph_mutex.h"
+#include "common/cmdparse.h"
+#include "common/compiler_extensions.h"
+#include "include/common_fwd.h"
+#include "include/cephfs/ceph_ll_client.h"
+#include "include/filepath.h"
+#include "include/interval_set.h"
+#include "include/lru.h"
+#include "include/types.h"
+#include "include/unordered_map.h"
+#include "include/unordered_set.h"
+#include "include/cephfs/metrics/Types.h"
+#include "mds/mdstypes.h"
+#include "msg/Dispatcher.h"
+#include "msg/MessageRef.h"
+#include "msg/Messenger.h"
+#include "osdc/ObjectCacher.h"
+
+#include "RWRef.h"
+#include "InodeRef.h"
+#include "MetaSession.h"
+#include "UserPerm.h"
+
+#include <fstream>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <thread>
+
+using std::set;
+using std::map;
+using std::fstream;
+
+class FSMap;
+class FSMapUser;
+class MonClient;
+
+
+struct DirStat;
+struct LeaseStat;
+struct InodeStat;
+
+class Filer;
+class Objecter;
+class WritebackHandler;
+
+class MDSMap;
+class Message;
+class destructive_lock_ref_t;
+
+enum {
+  l_c_first = 20000,
+  l_c_reply,
+  l_c_lat,
+  l_c_wrlat,
+  l_c_read,
+  l_c_fsync,
+  l_c_md_avg,
+  l_c_md_sqsum,
+  l_c_md_ops,
+  l_c_rd_avg,
+  l_c_rd_sqsum,
+  l_c_rd_ops,
+  l_c_wr_avg,
+  l_c_wr_sqsum,
+  l_c_wr_ops,
+  l_c_last,
+};
+
+
+class MDSCommandOp : public CommandOp
+{
+  public:
+  mds_gid_t     mds_gid;
+
+  explicit MDSCommandOp(ceph_tid_t t) : CommandOp(t) {}
+};
+
+/* error code for ceph_fuse */
+#define CEPH_FUSE_NO_MDS_UP    -((1<<16)+0) /* no mds up deteced in ceph_fuse */
+#define CEPH_FUSE_LAST         -((1<<16)+1) /* (unused) */
+
+// ============================================
+// types for my local metadata cache
+/* basic structure:
+   
+ - Dentries live in an LRU loop.  they get expired based on last access.
+      see include/lru.h.  items can be bumped to "mid" or "top" of list, etc.
+ - Inode has ref count for each Fh, Dir, or Dentry that points to it.
+ - when Inode ref goes to 0, it's expired.
+ - when Dir is empty, it's removed (and it's Inode ref--)
+ 
+*/
+
+/* getdir result */
+struct DirEntry {
+  explicit DirEntry(const string &s) : d_name(s), stmask(0) {}
+  DirEntry(const string &n, struct stat& s, int stm) : d_name(n), st(s), stmask(stm) {}
+
+  string d_name;
+  struct stat st;
+  int stmask;
+};
+
+struct Cap;
+class Dir;
+class Dentry;
+struct SnapRealm;
+struct Fh;
+struct CapSnap;
+
+struct MetaRequest;
+class ceph_lock_state_t;
+
+// ========================================================
+// client interface
+
+struct dir_result_t {
+  static const int SHIFT = 28;
+  static const int64_t MASK = (1 << SHIFT) - 1;
+  static const int64_t HASH = 0xFFULL << (SHIFT + 24); // impossible frag bits
+  static const loff_t END = 1ULL << (SHIFT + 32);
+
+  struct dentry {
+    int64_t offset;
+    std::string name;
+    std::string alternate_name;
+    InodeRef inode;
+    explicit dentry(int64_t o) : offset(o) {}
+    dentry(int64_t o, std::string n, std::string an, InodeRef in) :
+      offset(o), name(std::move(n)), alternate_name(std::move(an)), inode(std::move(in)) {}
+  };
+  struct dentry_off_lt {
+    bool operator()(const dentry& d, int64_t off) const {
+      return dir_result_t::fpos_cmp(d.offset, off) < 0;
+    }
+  };
+
+
+  explicit dir_result_t(Inode *in, const UserPerm& perms);
+
+
+  static uint64_t make_fpos(unsigned h, unsigned l, bool hash) {
+    uint64_t v =  ((uint64_t)h<< SHIFT) | (uint64_t)l;
+    if (hash)
+      v |= HASH;
+    else
+      ceph_assert((v & HASH) != HASH);
+    return v;
+  }
+  static unsigned fpos_high(uint64_t p) {
+    unsigned v = (p & (END-1)) >> SHIFT;
+    if ((p & HASH) == HASH)
+      return ceph_frag_value(v);
+    return v;
+  }
+  static unsigned fpos_low(uint64_t p) {
+    return p & MASK;
+  }
+  static int fpos_cmp(uint64_t l, uint64_t r) {
+    int c = ceph_frag_compare(fpos_high(l), fpos_high(r));
+    if (c)
+      return c;
+    if (fpos_low(l) == fpos_low(r))
+      return 0;
+    return fpos_low(l) < fpos_low(r) ? -1 : 1;
+  }
+
+  unsigned offset_high() { return fpos_high(offset); }
+  unsigned offset_low() { return fpos_low(offset); }
+
+  void set_end() { offset |= END; }
+  bool at_end() { return (offset & END); }
+
+  void set_hash_order() { offset |= HASH; }
+  bool hash_order() { return (offset & HASH) == HASH; }
+
+  bool is_cached() {
+    if (buffer.empty())
+      return false;
+    if (hash_order()) {
+      return buffer_frag.contains(offset_high());
+    } else {
+      return buffer_frag == frag_t(offset_high());
+    }
+  }
+
+  void reset() {
+    last_name.clear();
+    next_offset = 2;
+    offset = 0;
+    ordered_count = 0;
+    cache_index = 0;
+    buffer.clear();
+  }
+
+  InodeRef inode;
+  int64_t offset;        // hash order:
+			 //   (0xff << 52) | ((24 bits hash) << 28) |
+			 //   (the nth entry has hash collision);
+			 // frag+name order;
+			 //   ((frag value) << 28) | (the nth entry in frag);
+
+  unsigned next_offset;  // offset of next chunk (last_name's + 1)
+  string last_name;      // last entry in previous chunk
+
+  uint64_t release_count;
+  uint64_t ordered_count;
+  unsigned cache_index;
+  int start_shared_gen;  // dir shared_gen at start of readdir
+  UserPerm perms;
+
+  frag_t buffer_frag;
+
+  vector<dentry> buffer;
+  struct dirent de;
+};
+
+class Client : public Dispatcher, public md_config_obs_t {
+public:
+  friend class C_Block_Sync; // Calls block map and protected helpers
+  friend class C_Client_CacheInvalidate;  // calls ino_invalidate_cb
+  friend class C_Client_DentryInvalidate;  // calls dentry_invalidate_cb
+  friend class C_Client_FlushComplete; // calls put_inode()
+  friend class C_Client_Remount;
+  friend class C_Client_RequestInterrupt;
+  friend class C_Deleg_Timeout; // Asserts on client_lock, called when a delegation is unreturned
+  friend class C_Client_CacheRelease; // Asserts on client_lock
+  friend class SyntheticClient;
+  friend void intrusive_ptr_release(Inode *in);
+  template <typename T> friend struct RWRefState;
+  template <typename T> friend class RWRef;
+
+  using Dispatcher::cct;
+  using clock = ceph::coarse_mono_clock;
+
+  typedef int (*add_dirent_cb_t)(void *p, struct dirent *de, struct ceph_statx *stx, off_t off, Inode *in);
+
+  struct walk_dentry_result {
+    InodeRef in;
+    std::string alternate_name;
+  };
+
+  class CommandHook : public AdminSocketHook {
+  public:
+    explicit CommandHook(Client *client);
+    int call(std::string_view command, const cmdmap_t& cmdmap,
+	     Formatter *f,
+	     std::ostream& errss,
+	     bufferlist& out) override;
+  private:
+    Client *m_client;
+  };
+
+  // snapshot info returned via get_snap_info(). nothing to do
+  // with SnapInfo on the MDS.
+  struct SnapInfo {
+    snapid_t id;
+    std::map<std::string, std::string> metadata;
+  };
+
+  Client(Messenger *m, MonClient *mc, Objecter *objecter_);
+  Client(const Client&) = delete;
+  Client(const Client&&) = delete;
+  virtual ~Client() override;
+
+  static UserPerm pick_my_perms(CephContext *c) {
+    uid_t uid = c->_conf->client_mount_uid >= 0 ? c->_conf->client_mount_uid : -1;
+    gid_t gid = c->_conf->client_mount_gid >= 0 ? c->_conf->client_mount_gid : -1;
+    return UserPerm(uid, gid);
+  }
+  UserPerm pick_my_perms() {
+    uid_t uid = user_id >= 0 ? user_id : -1;
+    gid_t gid = group_id >= 0 ? group_id : -1;
+    return UserPerm(uid, gid);
+  }
+
+  int mount(const std::string &mount_root, const UserPerm& perms,
+	    bool require_mds=false, const std::string &fs_name="");
+  void unmount();
+  bool is_unmounting() const {
+    return mount_state.check_current_state(CLIENT_UNMOUNTING);
+  }
+  bool is_mounted() const {
+    return mount_state.check_current_state(CLIENT_MOUNTED);
+  }
+  bool is_mounting() const {
+    return mount_state.check_current_state(CLIENT_MOUNTING);
+  }
+  bool is_initialized() const {
+    return initialize_state.check_current_state(CLIENT_INITIALIZED);
+  }
+  void abort_conn();
+
+  void set_uuid(const std::string& uuid);
+  void set_session_timeout(unsigned timeout);
+  int start_reclaim(const std::string& uuid, unsigned flags,
+		    const std::string& fs_name);
+  void finish_reclaim();
+
+  fs_cluster_id_t get_fs_cid() {
+    return fscid;
+  }
+
+  int mds_command(
+    const std::string &mds_spec,
+    const std::vector<std::string>& cmd,
+    const bufferlist& inbl,
+    bufferlist *poutbl, std::string *prs, Context *onfinish);
+
+  // these should (more or less) mirror the actual system calls.
+  int statfs(const char *path, struct statvfs *stbuf, const UserPerm& perms);
+
+  // crap
+  int chdir(const char *s, std::string &new_cwd, const UserPerm& perms);
+  void _getcwd(std::string& cwd, const UserPerm& perms);
+  void getcwd(std::string& cwd, const UserPerm& perms);
+
+  // namespace ops
+  int opendir(const char *name, dir_result_t **dirpp, const UserPerm& perms);
+  int fdopendir(int dirfd, dir_result_t **dirpp, const UserPerm& perms);
+  int closedir(dir_result_t *dirp);
+
+  /**
+   * Fill a directory listing from dirp, invoking cb for each entry
+   * with the given pointer, the dirent, the struct stat, the stmask,
+   * and the offset.
+   *
+   * Returns 0 if it reached the end of the directory.
+   * If @a cb returns a negative error code, stop and return that.
+   */
+  int readdir_r_cb(dir_result_t *dirp, add_dirent_cb_t cb, void *p,
+		   unsigned want=0, unsigned flags=AT_STATX_DONT_SYNC,
+		   bool getref=false);
+
+  struct dirent * readdir(dir_result_t *d);
+  int readdir_r(dir_result_t *dirp, struct dirent *de);
+  int readdirplus_r(dir_result_t *dirp, struct dirent *de, struct ceph_statx *stx, unsigned want, unsigned flags, Inode **out);
+
+  int getdir(const char *relpath, list<string>& names,
+	     const UserPerm& perms);  // get the whole dir at once.
+
+  /**
+   * Returns the length of the buffer that got filled in, or -errno.
+   * If it returns -CEPHFS_ERANGE you just need to increase the size of the
+   * buffer and try again.
+   */
+  int _getdents(dir_result_t *dirp, char *buf, int buflen, bool ful);  // get a bunch of dentries at once
+  int getdents(dir_result_t *dirp, char *buf, int buflen) {
+    return _getdents(dirp, buf, buflen, true);
+  }
+  int getdnames(dir_result_t *dirp, char *buf, int buflen) {
+    return _getdents(dirp, buf, buflen, false);
+  }
+
+  void rewinddir(dir_result_t *dirp);
+  loff_t telldir(dir_result_t *dirp);
+  void seekdir(dir_result_t *dirp, loff_t offset);
+
+  int may_delete(const char *relpath, const UserPerm& perms);
+  int link(const char *existing, const char *newname, const UserPerm& perm, std::string alternate_name="");
+  int unlink(const char *path, const UserPerm& perm);
+  int unlinkat(int dirfd, const char *relpath, int flags, const UserPerm& perm);
+  int rename(const char *from, const char *to, const UserPerm& perm, std::string alternate_name="");
+
+  // dirs
+  int mkdir(const char *path, mode_t mode, const UserPerm& perm, std::string alternate_name="");
+  int mkdirat(int dirfd, const char *relpath, mode_t mode, const UserPerm& perm,
+              std::string alternate_name="");
+  int mkdirs(const char *path, mode_t mode, const UserPerm& perms);
+  int rmdir(const char *path, const UserPerm& perms);
+
+  // symlinks
+  int readlink(const char *path, char *buf, loff_t size, const UserPerm& perms);
+  int readlinkat(int dirfd, const char *relpath, char *buf, loff_t size, const UserPerm& perms);
+
+  int symlink(const char *existing, const char *newname, const UserPerm& perms, std::string alternate_name="");
+  int symlinkat(const char *target, int dirfd, const char *relpath, const UserPerm& perms,
+                std::string alternate_name="");
+
+  // path traversal for high-level interface
+  int walk(std::string_view path, struct walk_dentry_result* result, const UserPerm& perms, bool followsym=true);
+
+  // inode stuff
+  unsigned statx_to_mask(unsigned int flags, unsigned int want);
+  int stat(const char *path, struct stat *stbuf, const UserPerm& perms,
+	   frag_info_t *dirstat=0, int mask=CEPH_STAT_CAP_INODE_ALL);
+  int statx(const char *path, struct ceph_statx *stx,
+	    const UserPerm& perms,
+	    unsigned int want, unsigned int flags);
+  int lstat(const char *path, struct stat *stbuf, const UserPerm& perms,
+	    frag_info_t *dirstat=0, int mask=CEPH_STAT_CAP_INODE_ALL);
+
+  int setattr(const char *relpath, struct stat *attr, int mask,
+	      const UserPerm& perms);
+  int setattrx(const char *relpath, struct ceph_statx *stx, int mask,
+	       const UserPerm& perms, int flags=0);
+  int fsetattr(int fd, struct stat *attr, int mask, const UserPerm& perms);
+  int fsetattrx(int fd, struct ceph_statx *stx, int mask, const UserPerm& perms);
+  int chmod(const char *path, mode_t mode, const UserPerm& perms);
+  int fchmod(int fd, mode_t mode, const UserPerm& perms);
+  int chmodat(int dirfd, const char *relpath, mode_t mode, int flags, const UserPerm& perms);
+  int lchmod(const char *path, mode_t mode, const UserPerm& perms);
+  int chown(const char *path, uid_t new_uid, gid_t new_gid,
+	    const UserPerm& perms);
+  int fchown(int fd, uid_t new_uid, gid_t new_gid, const UserPerm& perms);
+  int lchown(const char *path, uid_t new_uid, gid_t new_gid,
+	     const UserPerm& perms);
+  int chownat(int dirfd, const char *relpath, uid_t new_uid, gid_t new_gid,
+              int flags, const UserPerm& perms);
+  int utime(const char *path, struct utimbuf *buf, const UserPerm& perms);
+  int lutime(const char *path, struct utimbuf *buf, const UserPerm& perms);
+  int futime(int fd, struct utimbuf *buf, const UserPerm& perms);
+  int utimes(const char *relpath, struct timeval times[2], const UserPerm& perms);
+  int lutimes(const char *relpath, struct timeval times[2], const UserPerm& perms);
+  int futimes(int fd, struct timeval times[2], const UserPerm& perms);
+  int futimens(int fd, struct timespec times[2], const UserPerm& perms);
+  int utimensat(int dirfd, const char *relpath, struct timespec times[2], int flags,
+                const UserPerm& perms);
+  int flock(int fd, int operation, uint64_t owner);
+  int truncate(const char *path, loff_t size, const UserPerm& perms);
+
+  // file ops
+  int mknod(const char *path, mode_t mode, const UserPerm& perms, dev_t rdev=0);
+
+  int create_and_open(std::optional<int> dirfd, const char *relpath, int flags, const UserPerm& perms,
+                      mode_t mode, int stripe_unit, int stripe_count, int object_size, const char *data_pool,
+                      std::string alternate_name);
+  int open(const char *path, int flags, const UserPerm& perms, mode_t mode=0, std::string alternate_name="") {
+    return open(path, flags, perms, mode, 0, 0, 0, NULL, alternate_name);
+  }
+  int open(const char *path, int flags, const UserPerm& perms,
+	   mode_t mode, int stripe_unit, int stripe_count, int object_size,
+	   const char *data_pool, std::string alternate_name="");
+  int _openat(int dirfd, const char *relpath, int flags, const UserPerm& perms,
+              mode_t mode=0, std::string alternate_name="");
+  int openat(int dirfd, const char *relpath, int flags, const UserPerm& perms,
+             mode_t mode, int stripe_unit, int stripe_count,
+             int object_size, const char *data_pool, std::string alternate_name);
+  int openat(int dirfd, const char *path, int flags, const UserPerm& perms, mode_t mode=0,
+             std::string alternate_name="") {
+    return openat(dirfd, path, flags, perms, mode, 0, 0, 0, NULL, alternate_name);
+  }
+
+  int lookup_hash(inodeno_t ino, inodeno_t dirino, const char *name,
+		  const UserPerm& perms);
+  int lookup_ino(inodeno_t ino, const UserPerm& perms, Inode **inode=NULL);
+  int lookup_name(Inode *in, Inode *parent, const UserPerm& perms);
+  int _close(int fd);
+  int close(int fd);
+  loff_t lseek(int fd, loff_t offset, int whence);
+  int read(int fd, char *buf, loff_t size, loff_t offset=-1);
+  int preadv(int fd, const struct iovec *iov, int iovcnt, loff_t offset=-1);
+  int write(int fd, const char *buf, loff_t size, loff_t offset=-1);
+  int pwritev(int fd, const struct iovec *iov, int iovcnt, loff_t offset=-1);
+  int fake_write_size(int fd, loff_t size);
+  int ftruncate(int fd, loff_t size, const UserPerm& perms);
+  int fsync(int fd, bool syncdataonly);
+  int fstat(int fd, struct stat *stbuf, const UserPerm& perms,
+	    int mask=CEPH_STAT_CAP_INODE_ALL);
+  int fstatx(int fd, struct ceph_statx *stx, const UserPerm& perms,
+	     unsigned int want, unsigned int flags);
+  int statxat(int dirfd, const char *relpath,
+              struct ceph_statx *stx, const UserPerm& perms,
+              unsigned int want, unsigned int flags);
+  int fallocate(int fd, int mode, loff_t offset, loff_t length);
+
+  // full path xattr ops
+  int getxattr(const char *path, const char *name, void *value, size_t size,
+	       const UserPerm& perms);
+  int lgetxattr(const char *path, const char *name, void *value, size_t size,
+		const UserPerm& perms);
+  int fgetxattr(int fd, const char *name, void *value, size_t size,
+		const UserPerm& perms);
+  int listxattr(const char *path, char *list, size_t size, const UserPerm& perms);
+  int llistxattr(const char *path, char *list, size_t size, const UserPerm& perms);
+  int flistxattr(int fd, char *list, size_t size, const UserPerm& perms);
+  int removexattr(const char *path, const char *name, const UserPerm& perms);
+  int lremovexattr(const char *path, const char *name, const UserPerm& perms);
+  int fremovexattr(int fd, const char *name, const UserPerm& perms);
+  int setxattr(const char *path, const char *name, const void *value,
+	       size_t size, int flags, const UserPerm& perms);
+  int lsetxattr(const char *path, const char *name, const void *value,
+		size_t size, int flags, const UserPerm& perms);
+  int fsetxattr(int fd, const char *name, const void *value, size_t size,
+		int flags, const UserPerm& perms);
+
+  int sync_fs();
+  int64_t drop_caches();
+
+  int get_snap_info(const char *path, const UserPerm &perms, SnapInfo *snap_info);
+
+  // hpc lazyio
+  int lazyio(int fd, int enable);
+  int lazyio_propagate(int fd, loff_t offset, size_t count);
+  int lazyio_synchronize(int fd, loff_t offset, size_t count);
+
+  // expose file layout
+  int describe_layout(const char *path, file_layout_t* layout,
+		      const UserPerm& perms);
+  int fdescribe_layout(int fd, file_layout_t* layout);
+  int get_file_stripe_address(int fd, loff_t offset, vector<entity_addr_t>& address);
+  int get_file_extent_osds(int fd, loff_t off, loff_t *len, vector<int>& osds);
+  int get_osd_addr(int osd, entity_addr_t& addr);
+
+  // expose mdsmap
+  int64_t get_default_pool_id();
+
+  // expose osdmap
+  int get_local_osd();
+  int get_pool_replication(int64_t pool);
+  int64_t get_pool_id(const char *pool_name);
+  string get_pool_name(int64_t pool);
+  int get_osd_crush_location(int id, vector<pair<string, string> >& path);
+
+  int enumerate_layout(int fd, vector<ObjectExtent>& result,
+		       loff_t length, loff_t offset);
+
+  int mksnap(const char *path, const char *name, const UserPerm& perm,
+             mode_t mode=0, const std::map<std::string, std::string> &metadata={});
+  int rmsnap(const char *path, const char *name, const UserPerm& perm, bool check_perms=false);
+
+  // Inode permission checking
+  int inode_permission(Inode *in, const UserPerm& perms, unsigned want);
+
+  // expose caps
+  int get_caps_issued(int fd);
+  int get_caps_issued(const char *path, const UserPerm& perms);
+
+  snapid_t ll_get_snapid(Inode *in);
+  vinodeno_t ll_get_vino(Inode *in) {
+    std::lock_guard lock(client_lock);
+    return _get_vino(in);
+  }
+  // get inode from faked ino
+  Inode *ll_get_inode(ino_t ino);
+  Inode *ll_get_inode(vinodeno_t vino);
+  int ll_lookup(Inode *parent, const char *name, struct stat *attr,
+		Inode **out, const UserPerm& perms);
+  int ll_lookup_inode(struct inodeno_t ino, const UserPerm& perms, Inode **inode);
+  int ll_lookup_vino(vinodeno_t vino, const UserPerm& perms, Inode **inode);
+  int ll_lookupx(Inode *parent, const char *name, Inode **out,
+			struct ceph_statx *stx, unsigned want, unsigned flags,
+			const UserPerm& perms);
+  bool ll_forget(Inode *in, uint64_t count);
+  bool ll_put(Inode *in);
+  int ll_get_snap_ref(snapid_t snap);
+
+  int ll_getattr(Inode *in, struct stat *st, const UserPerm& perms);
+  int ll_getattrx(Inode *in, struct ceph_statx *stx, unsigned int want,
+		  unsigned int flags, const UserPerm& perms);
+  int ll_setattrx(Inode *in, struct ceph_statx *stx, int mask,
+		  const UserPerm& perms);
+  int ll_setattr(Inode *in, struct stat *st, int mask,
+		 const UserPerm& perms);
+  int ll_getxattr(Inode *in, const char *name, void *value, size_t size,
+		  const UserPerm& perms);
+  int ll_setxattr(Inode *in, const char *name, const void *value, size_t size,
+		  int flags, const UserPerm& perms);
+  int ll_removexattr(Inode *in, const char *name, const UserPerm& perms);
+  int ll_listxattr(Inode *in, char *list, size_t size, const UserPerm& perms);
+  int ll_opendir(Inode *in, int flags, dir_result_t **dirpp,
+		 const UserPerm& perms);
+  int ll_releasedir(dir_result_t* dirp);
+  int ll_fsyncdir(dir_result_t* dirp);
+  int ll_readlink(Inode *in, char *buf, size_t bufsize, const UserPerm& perms);
+  int ll_mknod(Inode *in, const char *name, mode_t mode, dev_t rdev,
+	       struct stat *attr, Inode **out, const UserPerm& perms);
+  int ll_mknodx(Inode *parent, const char *name, mode_t mode, dev_t rdev,
+	        Inode **out, struct ceph_statx *stx, unsigned want,
+		unsigned flags, const UserPerm& perms);
+  int ll_mkdir(Inode *in, const char *name, mode_t mode, struct stat *attr,
+	       Inode **out, const UserPerm& perm);
+  int ll_mkdirx(Inode *parent, const char *name, mode_t mode, Inode **out,
+		struct ceph_statx *stx, unsigned want, unsigned flags,
+		const UserPerm& perms);
+  int ll_symlink(Inode *in, const char *name, const char *value,
+		 struct stat *attr, Inode **out, const UserPerm& perms);
+  int ll_symlinkx(Inode *parent, const char *name, const char *value,
+		  Inode **out, struct ceph_statx *stx, unsigned want,
+		  unsigned flags, const UserPerm& perms);
+  int ll_unlink(Inode *in, const char *name, const UserPerm& perm);
+  int ll_rmdir(Inode *in, const char *name, const UserPerm& perms);
+  int ll_rename(Inode *parent, const char *name, Inode *newparent,
+		const char *newname, const UserPerm& perm);
+  int ll_link(Inode *in, Inode *newparent, const char *newname,
+	      const UserPerm& perm);
+  int ll_open(Inode *in, int flags, Fh **fh, const UserPerm& perms);
+  int _ll_create(Inode *parent, const char *name, mode_t mode,
+	      int flags, InodeRef *in, int caps, Fh **fhp,
+	      const UserPerm& perms);
+  int ll_create(Inode *parent, const char *name, mode_t mode, int flags,
+		struct stat *attr, Inode **out, Fh **fhp,
+		const UserPerm& perms);
+  int ll_createx(Inode *parent, const char *name, mode_t mode,
+		int oflags, Inode **outp, Fh **fhp,
+		struct ceph_statx *stx, unsigned want, unsigned lflags,
+		const UserPerm& perms);
+  int ll_read_block(Inode *in, uint64_t blockid, char *buf,  uint64_t offset,
+		    uint64_t length, file_layout_t* layout);
+
+  int ll_write_block(Inode *in, uint64_t blockid,
+		     char* buf, uint64_t offset,
+		     uint64_t length, file_layout_t* layout,
+		     uint64_t snapseq, uint32_t sync);
+  int ll_commit_blocks(Inode *in, uint64_t offset, uint64_t length);
+
+  int ll_statfs(Inode *in, struct statvfs *stbuf, const UserPerm& perms);
+  int ll_walk(const char* name, Inode **i, struct ceph_statx *stx,
+	       unsigned int want, unsigned int flags, const UserPerm& perms);
+  uint32_t ll_stripe_unit(Inode *in);
+  int ll_file_layout(Inode *in, file_layout_t *layout);
+  uint64_t ll_snap_seq(Inode *in);
+
+  int ll_read(Fh *fh, loff_t off, loff_t len, bufferlist *bl);
+  int ll_write(Fh *fh, loff_t off, loff_t len, const char *data);
+  int64_t ll_readv(struct Fh *fh, const struct iovec *iov, int iovcnt, int64_t off);
+  int64_t ll_writev(struct Fh *fh, const struct iovec *iov, int iovcnt, int64_t off);
+  loff_t ll_lseek(Fh *fh, loff_t offset, int whence);
+  int ll_flush(Fh *fh);
+  int ll_fsync(Fh *fh, bool syncdataonly);
+  int ll_sync_inode(Inode *in, bool syncdataonly);
+  int ll_fallocate(Fh *fh, int mode, int64_t offset, int64_t length);
+  int ll_release(Fh *fh);
+  int ll_getlk(Fh *fh, struct flock *fl, uint64_t owner);
+  int ll_setlk(Fh *fh, struct flock *fl, uint64_t owner, int sleep);
+  int ll_flock(Fh *fh, int cmd, uint64_t owner);
+  int ll_lazyio(Fh *fh, int enable);
+  int ll_file_layout(Fh *fh, file_layout_t *layout);
+  void ll_interrupt(void *d);
+  bool ll_handle_umask() {
+    return acl_type != NO_ACL;
+  }
+
+  int ll_get_stripe_osd(struct Inode *in, uint64_t blockno,
+			file_layout_t* layout);
+  uint64_t ll_get_internal_offset(struct Inode *in, uint64_t blockno);
+
+  int ll_num_osds(void);
+  int ll_osdaddr(int osd, uint32_t *addr);
+  int ll_osdaddr(int osd, char* buf, size_t size);
+
+  void ll_register_callbacks(struct ceph_client_callback_args *args);
+  std::pair<int, bool> test_dentry_handling(bool can_invalidate);
+
+  const char** get_tracked_conf_keys() const override;
+  void handle_conf_change(const ConfigProxy& conf,
+	                          const std::set <std::string> &changed) override;
+  uint32_t get_deleg_timeout() { return deleg_timeout; }
+  int set_deleg_timeout(uint32_t timeout);
+  int ll_delegation(Fh *fh, unsigned cmd, ceph_deleg_cb_t cb, void *priv);
+
+  entity_name_t get_myname() { return messenger->get_myname(); }
+  void wait_on_list(std::list<ceph::condition_variable*>& ls);
+  void signal_cond_list(std::list<ceph::condition_variable*>& ls);
+
+  void set_filer_flags(int flags);
+  void clear_filer_flags(int flags);
+
+  void tear_down_cache();
+
+  void update_metadata(std::string const &k, std::string const &v);
+
+  client_t get_nodeid() { return whoami; }
+
+  inodeno_t get_root_ino();
+  Inode *get_root();
+
+  virtual int init();
+  virtual void shutdown();
+
+  // messaging
+  void cancel_commands(const MDSMap& newmap);
+  void handle_mds_map(const MConstRef<MMDSMap>& m);
+  void handle_fs_map(const MConstRef<MFSMap>& m);
+  void handle_fs_map_user(const MConstRef<MFSMapUser>& m);
+  void handle_osd_map(const MConstRef<MOSDMap>& m);
+
+  void handle_lease(const MConstRef<MClientLease>& m);
+
+  // inline data
+  int uninline_data(Inode *in, Context *onfinish);
+
+  // file caps
+  void check_cap_issue(Inode *in, unsigned issued);
+  void add_update_cap(Inode *in, MetaSession *session, uint64_t cap_id,
+		      unsigned issued, unsigned wanted, unsigned seq, unsigned mseq,
+		      inodeno_t realm, int flags, const UserPerm& perms);
+  void remove_cap(Cap *cap, bool queue_release);
+  void remove_all_caps(Inode *in);
+  void remove_session_caps(MetaSession *session, int err);
+  int mark_caps_flushing(Inode *in, ceph_tid_t *ptid);
+  void adjust_session_flushing_caps(Inode *in, MetaSession *old_s, MetaSession *new_s);
+  void flush_caps_sync();
+  void kick_flushing_caps(Inode *in, MetaSession *session);
+  void kick_flushing_caps(MetaSession *session);
+  void early_kick_flushing_caps(MetaSession *session);
+  int get_caps(Fh *fh, int need, int want, int *have, loff_t endoff);
+  int get_caps_used(Inode *in);
+
+  void maybe_update_snaprealm(SnapRealm *realm, snapid_t snap_created, snapid_t snap_highwater,
+			      vector<snapid_t>& snaps);
+
+  void handle_quota(const MConstRef<MClientQuota>& m);
+  void handle_snap(const MConstRef<MClientSnap>& m);
+  void handle_caps(const MConstRef<MClientCaps>& m);
+  void handle_cap_import(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m);
+  void handle_cap_export(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m);
+  void handle_cap_trunc(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m);
+  void handle_cap_flush_ack(MetaSession *session, Inode *in, Cap *cap, const MConstRef<MClientCaps>& m);
+  void handle_cap_flushsnap_ack(MetaSession *session, Inode *in, const MConstRef<MClientCaps>& m);
+  void handle_cap_grant(MetaSession *session, Inode *in, Cap *cap, const MConstRef<MClientCaps>& m);
+  void cap_delay_requeue(Inode *in);
+
+  void send_cap(Inode *in, MetaSession *session, Cap *cap, int flags,
+		int used, int want, int retain, int flush,
+		ceph_tid_t flush_tid);
+
+  void send_flush_snap(Inode *in, MetaSession *session, snapid_t follows, CapSnap& capsnap);
+
+  void flush_snaps(Inode *in);
+  void get_cap_ref(Inode *in, int cap);
+  void put_cap_ref(Inode *in, int cap);
+  void wait_sync_caps(Inode *in, ceph_tid_t want);
+  void wait_sync_caps(ceph_tid_t want);
+  void queue_cap_snap(Inode *in, SnapContext &old_snapc);
+  void finish_cap_snap(Inode *in, CapSnap &capsnap, int used);
+
+  void _schedule_invalidate_dentry_callback(Dentry *dn, bool del);
+  void _async_dentry_invalidate(vinodeno_t dirino, vinodeno_t ino, string& name);
+  void _try_to_trim_inode(Inode *in, bool sched_inval);
+
+  void _schedule_invalidate_callback(Inode *in, int64_t off, int64_t len);
+  void _invalidate_inode_cache(Inode *in);
+  void _invalidate_inode_cache(Inode *in, int64_t off, int64_t len);
+  void _async_invalidate(vinodeno_t ino, int64_t off, int64_t len);
+
+  void _schedule_ino_release_callback(Inode *in);
+  void _async_inode_release(vinodeno_t ino);
+
+  bool _release(Inode *in);
+
+  /**
+   * Initiate a flush of the data associated with the given inode.
+   * If you specify a Context, you are responsible for holding an inode
+   * reference for the duration of the flush. If not, _flush() will
+   * take the reference for you.
+   * @param in The Inode whose data you wish to flush.
+   * @param c The Context you wish us to complete once the data is
+   * flushed. If already flushed, this will be called in-line.
+   *
+   * @returns true if the data was already flushed, false otherwise.
+   */
+  bool _flush(Inode *in, Context *c);
+  void _flush_range(Inode *in, int64_t off, uint64_t size);
+  void _flushed(Inode *in);
+  void flush_set_callback(ObjectCacher::ObjectSet *oset);
+
+  void close_release(Inode *in);
+  void close_safe(Inode *in);
+
+  void lock_fh_pos(Fh *f);
+  void unlock_fh_pos(Fh *f);
+
+  // metadata cache
+  void update_dir_dist(Inode *in, DirStat *st, mds_rank_t from);
+
+  void clear_dir_complete_and_ordered(Inode *diri, bool complete);
+  void insert_readdir_results(MetaRequest *request, MetaSession *session, Inode *diri);
+  Inode* insert_trace(MetaRequest *request, MetaSession *session);
+  void update_inode_file_size(Inode *in, int issued, uint64_t size,
+			      uint64_t truncate_seq, uint64_t truncate_size);
+  void update_inode_file_time(Inode *in, int issued, uint64_t time_warp_seq,
+			      utime_t ctime, utime_t mtime, utime_t atime);
+
+  Inode *add_update_inode(InodeStat *st, utime_t ttl, MetaSession *session,
+			  const UserPerm& request_perms);
+  Dentry *insert_dentry_inode(Dir *dir, const string& dname, LeaseStat *dlease,
+			      Inode *in, utime_t from, MetaSession *session,
+			      Dentry *old_dentry = NULL);
+  void update_dentry_lease(Dentry *dn, LeaseStat *dlease, utime_t from, MetaSession *session);
+
+  bool use_faked_inos() { return _use_faked_inos; }
+  vinodeno_t map_faked_ino(ino_t ino);
+
+  //notify the mds to flush the mdlog
+  void flush_mdlog_sync(Inode *in);
+  void flush_mdlog_sync();
+  void flush_mdlog(MetaSession *session);
+
+  void renew_caps();
+  void renew_caps(MetaSession *session);
+  void flush_cap_releases();
+  void renew_and_flush_cap_releases();
+  void tick();
+  void start_tick_thread();
+
+  void update_read_io_size(size_t size) {
+    total_read_ops++;
+    total_read_size += size;
+  }
+
+  void update_write_io_size(size_t size) {
+    total_write_ops++;
+    total_write_size += size;
+  }
+
+  void inc_dentry_nr() {
+    ++dentry_nr;
+  }
+  void dec_dentry_nr() {
+    --dentry_nr;
+  }
+  void dlease_hit() {
+    ++dlease_hits;
+  }
+  void dlease_miss() {
+    ++dlease_misses;
+  }
+  std::tuple<uint64_t, uint64_t, uint64_t> get_dlease_hit_rates() {
+    return std::make_tuple(dlease_hits, dlease_misses, dentry_nr);
+  }
+
+  void cap_hit() {
+    ++cap_hits;
+  }
+  void cap_miss() {
+    ++cap_misses;
+  }
+  std::pair<uint64_t, uint64_t> get_cap_hit_rates() {
+    return std::make_pair(cap_hits, cap_misses);
+  }
+
+  void inc_opened_files() {
+    ++opened_files;
+  }
+  void dec_opened_files() {
+    --opened_files;
+  }
+  std::pair<uint64_t, uint64_t> get_opened_files_rates() {
+    return std::make_pair(opened_files, inode_map.size());
+  }
+
+  void inc_pinned_icaps() {
+    ++pinned_icaps;
+  }
+  void dec_pinned_icaps(uint64_t nr=1) {
+    pinned_icaps -= nr;
+  }
+  std::pair<uint64_t, uint64_t> get_pinned_icaps_rates() {
+    return std::make_pair(pinned_icaps, inode_map.size());
+  }
+
+  void inc_opened_inodes() {
+    ++opened_inodes;
+  }
+  void dec_opened_inodes() {
+    --opened_inodes;
+  }
+  std::pair<uint64_t, uint64_t> get_opened_inodes_rates() {
+    return std::make_pair(opened_inodes, inode_map.size());
+  }
+
+  xlist<Inode*> &get_dirty_list() { return dirty_list; }
+
+  /* timer_lock for 'timer' */
+  ceph::mutex timer_lock = ceph::make_mutex("Client::timer_lock");
+  SafeTimer timer;
+
+  /* tick thread */
+  std::thread upkeeper;
+  ceph::condition_variable upkeep_cond;
+  bool tick_thread_stopped = false;
+
+  std::unique_ptr<PerfCounters> logger;
+  std::unique_ptr<MDSMap> mdsmap;
+
+  bool fuse_default_permissions;
+  bool _collect_and_send_global_metrics;
+
+protected:
+  std::list<ceph::condition_variable*> waiting_for_reclaim;
+  /* Flags for check_caps() */
+  static const unsigned CHECK_CAPS_NODELAY = 0x1;
+  static const unsigned CHECK_CAPS_SYNCHRONOUS = 0x2;
+
+  void check_caps(Inode *in, unsigned flags);
+
+  void set_cap_epoch_barrier(epoch_t e);
+
+  void handle_command_reply(const MConstRef<MCommandReply>& m);
+  int fetch_fsmap(bool user);
+  int resolve_mds(
+      const std::string &mds_spec,
+      std::vector<mds_gid_t> *targets);
+
+  void get_session_metadata(std::map<std::string, std::string> *meta) const;
+  bool have_open_session(mds_rank_t mds);
+  void got_mds_push(MetaSession *s);
+  MetaSession *_get_mds_session(mds_rank_t mds, Connection *con);  ///< return session for mds *and* con; null otherwise
+  MetaSession *_get_or_open_mds_session(mds_rank_t mds);
+  MetaSession *_open_mds_session(mds_rank_t mds);
+  void _close_mds_session(MetaSession *s);
+  void _closed_mds_session(MetaSession *s, int err=0, bool rejected=false);
+  bool _any_stale_sessions() const;
+  void _kick_stale_sessions();
+  void handle_client_session(const MConstRef<MClientSession>& m);
+  void send_reconnect(MetaSession *s);
+  void resend_unsafe_requests(MetaSession *s);
+  void wait_unsafe_requests();
+
+  void dump_mds_requests(Formatter *f);
+  void dump_mds_sessions(Formatter *f, bool cap_dump=false);
+
+  int make_request(MetaRequest *req, const UserPerm& perms,
+                   InodeRef *ptarget = 0, bool *pcreated = 0,
+                   mds_rank_t use_mds=-1, bufferlist *pdirbl=0,
+                   size_t feature_needed=ULONG_MAX);
+  void put_request(MetaRequest *request);
+  void unregister_request(MetaRequest *request);
+
+  int verify_reply_trace(int r, MetaSession *session, MetaRequest *request,
+			 const MConstRef<MClientReply>& reply,
+			 InodeRef *ptarget, bool *pcreated,
+			 const UserPerm& perms);
+  void encode_cap_releases(MetaRequest *request, mds_rank_t mds);
+  int encode_inode_release(Inode *in, MetaRequest *req,
+			   mds_rank_t mds, int drop,
+			   int unless,int force=0);
+  void encode_dentry_release(Dentry *dn, MetaRequest *req,
+			     mds_rank_t mds, int drop, int unless);
+  mds_rank_t choose_target_mds(MetaRequest *req, Inode** phash_diri=NULL);
+  void connect_mds_targets(mds_rank_t mds);
+  void send_request(MetaRequest *request, MetaSession *session,
+		    bool drop_cap_releases=false);
+  MRef<MClientRequest> build_client_request(MetaRequest *request);
+  void kick_requests(MetaSession *session);
+  void kick_requests_closed(MetaSession *session);
+  void handle_client_request_forward(const MConstRef<MClientRequestForward>& reply);
+  void handle_client_reply(const MConstRef<MClientReply>& reply);
+  bool is_dir_operation(MetaRequest *request);
+
+  int path_walk(const filepath& fp, struct walk_dentry_result* result, const UserPerm& perms, bool followsym=true, int mask=0,
+                InodeRef dirinode=nullptr);
+  int path_walk(const filepath& fp, InodeRef *end, const UserPerm& perms,
+                bool followsym=true, int mask=0, InodeRef dirinode=nullptr);
+
+  // fake inode number for 32-bits ino_t
+  void _assign_faked_ino(Inode *in);
+  void _assign_faked_root(Inode *in);
+  void _release_faked_ino(Inode *in);
+  void _reset_faked_inos();
+  vinodeno_t _map_faked_ino(ino_t ino);
+
+  // Optional extra metadata about me to send to the MDS
+  void populate_metadata(const std::string &mount_root);
+
+  SnapRealm *get_snap_realm(inodeno_t r);
+  SnapRealm *get_snap_realm_maybe(inodeno_t r);
+  void put_snap_realm(SnapRealm *realm);
+  bool adjust_realm_parent(SnapRealm *realm, inodeno_t parent);
+  void update_snap_trace(const bufferlist& bl, SnapRealm **realm_ret, bool must_flush=true);
+  void invalidate_snaprealm_and_children(SnapRealm *realm);
+
+  void refresh_snapdir_attrs(Inode *in, Inode *diri);
+  Inode *open_snapdir(Inode *diri);
+
+  int get_fd() {
+    int fd = free_fd_set.range_start();
+    free_fd_set.erase(fd, 1);
+    return fd;
+  }
+  void put_fd(int fd) {
+    free_fd_set.insert(fd, 1);
+  }
+
+  /*
+   * Resolve file descriptor, or return NULL.
+   */
+  Fh *get_filehandle(int fd) {
+    auto it = fd_map.find(fd);
+    if (it == fd_map.end())
+      return NULL;
+    return it->second;
+  }
+  int get_fd_inode(int fd, InodeRef *in);
+
+  // helpers
+  void wake_up_session_caps(MetaSession *s, bool reconnect);
+
+  void wait_on_context_list(list<Context*>& ls);
+  void signal_context_list(list<Context*>& ls);
+
+  // -- metadata cache stuff
+
+  // decrease inode ref.  delete if dangling.
+  void _put_inode(Inode *in, int n);
+  void delay_put_inodes(bool wakeup=false);
+  void put_inode(Inode *in, int n=1);
+  void close_dir(Dir *dir);
+
+  int subscribe_mdsmap(const std::string &fs_name="");
+
+  void _abort_mds_sessions(int err);
+
+  // same as unmount() but for when the client_lock is already held
+  void _unmount(bool abort);
+
+  //int get_cache_size() { return lru.lru_get_size(); }
+
+  /**
+   * Don't call this with in==NULL, use get_or_create for that
+   * leave dn set to default NULL unless you're trying to add
+   * a new inode to a pre-created Dentry
+   */
+  Dentry* link(Dir *dir, const string& name, Inode *in, Dentry *dn);
+  void unlink(Dentry *dn, bool keepdir, bool keepdentry);
+
+  int fill_stat(Inode *in, struct stat *st, frag_info_t *dirstat=0, nest_info_t *rstat=0);
+  int fill_stat(InodeRef& in, struct stat *st, frag_info_t *dirstat=0, nest_info_t *rstat=0) {
+    return fill_stat(in.get(), st, dirstat, rstat);
+  }
+
+  void fill_statx(Inode *in, unsigned int mask, struct ceph_statx *stx);
+  void fill_statx(InodeRef& in, unsigned int mask, struct ceph_statx *stx) {
+    return fill_statx(in.get(), mask, stx);
+  }
+
+  void touch_dn(Dentry *dn);
+
+  // trim cache.
+  void trim_cache(bool trim_kernel_dcache=false);
+  void trim_cache_for_reconnect(MetaSession *s);
+  void trim_dentry(Dentry *dn);
+  void trim_caps(MetaSession *s, uint64_t max);
+  void _invalidate_kernel_dcache();
+  void _trim_negative_child_dentries(InodeRef& in);
+
+  void dump_inode(Formatter *f, Inode *in, set<Inode*>& did, bool disconnected);
+  void dump_cache(Formatter *f);  // debug
+
+  // force read-only
+  void force_session_readonly(MetaSession *s);
+
+  void dump_status(Formatter *f);  // debug
+
+  bool ms_dispatch2(const MessageRef& m) override;
+
+  void ms_handle_connect(Connection *con) override;
+  bool ms_handle_reset(Connection *con) override;
+  void ms_handle_remote_reset(Connection *con) override;
+  bool ms_handle_refused(Connection *con) override;
+
+  int authenticate();
+
+  Inode* get_quota_root(Inode *in, const UserPerm& perms);
+  bool check_quota_condition(Inode *in, const UserPerm& perms,
+			     std::function<bool (const Inode &)> test);
+  bool is_quota_files_exceeded(Inode *in, const UserPerm& perms);
+  bool is_quota_bytes_exceeded(Inode *in, int64_t new_bytes,
+			       const UserPerm& perms);
+  bool is_quota_bytes_approaching(Inode *in, const UserPerm& perms);
+
+  int check_pool_perm(Inode *in, int need);
+
+  void handle_client_reclaim_reply(const MConstRef<MClientReclaimReply>& reply);
+
+  /**
+   * Call this when an OSDMap is seen with a full flag (global or per pool)
+   * set.
+   *
+   * @param pool the pool ID affected, or -1 if all.
+   */
+  void _handle_full_flag(int64_t pool);
+
+  void _close_sessions();
+
+  void _pre_init();
+
+  /**
+   * The basic housekeeping parts of init (perf counters, admin socket)
+   * that is independent of how objecters/monclient/messengers are
+   * being set up.
+   */
+  void _finish_init();
+
+  // global client lock
+  //  - protects Client and buffer cache both!
+  ceph::mutex client_lock = ceph::make_mutex("Client::client_lock");
+
+  std::map<snapid_t, int> ll_snap_ref;
+
+  InodeRef               root = nullptr;
+  map<Inode*, InodeRef>  root_parents;
+  Inode*                 root_ancestor = nullptr;
+  LRU                    lru;    // lru list of Dentry's in our local metadata cache.
+
+  InodeRef cwd;
+
+  std::unique_ptr<Filer>             filer;
+  std::unique_ptr<ObjectCacher>      objectcacher;
+  std::unique_ptr<WritebackHandler>  writeback_handler;
+
+  Messenger *messenger;
+  MonClient *monclient;
+  Objecter  *objecter;
+
+  client_t whoami;
+
+  /* The state migration mechanism */
+  enum _state {
+    /* For the initialize_state */
+    CLIENT_NEW, // The initial state for the initialize_state or after Client::shutdown()
+    CLIENT_INITIALIZING, // At the beginning of the Client::init()
+    CLIENT_INITIALIZED, // At the end of CLient::init()
+
+    /* For the mount_state */
+    CLIENT_UNMOUNTED, // The initial state for the mount_state or after unmounted
+    CLIENT_MOUNTING, // At the beginning of Client::mount()
+    CLIENT_MOUNTED, // At the end of Client::mount()
+    CLIENT_UNMOUNTING, // At the beginning of the Client::_unmout()
+  };
+
+  typedef enum _state state_t;
+  using RWRef_t = RWRef<state_t>;
+
+  struct mount_state_t : public RWRefState<state_t> {
+    public:
+      bool is_valid_state(state_t state) const override {
+        switch (state) {
+	  case Client::CLIENT_MOUNTING:
+	  case Client::CLIENT_MOUNTED:
+	  case Client::CLIENT_UNMOUNTING:
+	  case Client::CLIENT_UNMOUNTED:
+            return true;
+          default:
+            return false;
+        }
+      }
+
+      int check_reader_state(state_t require) const override {
+        if (require == Client::CLIENT_MOUNTING &&
+            (state == Client::CLIENT_MOUNTING || state == Client::CLIENT_MOUNTED))
+          return true;
+        else
+          return false;
+      }
+
+      /* The state migration check */
+      int check_writer_state(state_t require) const override {
+        if (require == Client::CLIENT_MOUNTING &&
+            state == Client::CLIENT_UNMOUNTED)
+          return true;
+	else if (require == Client::CLIENT_MOUNTED &&
+            state == Client::CLIENT_MOUNTING)
+          return true;
+	else if (require == Client::CLIENT_UNMOUNTING &&
+            state == Client::CLIENT_MOUNTED)
+          return true;
+	else if (require == Client::CLIENT_UNMOUNTED &&
+            state == Client::CLIENT_UNMOUNTING)
+          return true;
+        else
+          return false;
+      }
+
+      mount_state_t(state_t state, const char *lockname, uint64_t reader_cnt=0)
+        : RWRefState (state, lockname, reader_cnt) {}
+      ~mount_state_t() {}
+  };
+
+  struct initialize_state_t : public RWRefState<state_t> {
+    public:
+      bool is_valid_state(state_t state) const override {
+        switch (state) {
+	  case Client::CLIENT_NEW:
+          case Client::CLIENT_INITIALIZING:
+	  case Client::CLIENT_INITIALIZED:
+            return true;
+          default:
+            return false;
+        }
+      }
+
+      int check_reader_state(state_t require) const override {
+        if (require == Client::CLIENT_INITIALIZED &&
+            state >= Client::CLIENT_INITIALIZED)
+          return true;
+        else
+          return false;
+      }
+
+      /* The state migration check */
+      int check_writer_state(state_t require) const override {
+        if (require == Client::CLIENT_INITIALIZING &&
+            (state == Client::CLIENT_NEW))
+          return true;
+	else if (require == Client::CLIENT_INITIALIZED &&
+            (state == Client::CLIENT_INITIALIZING))
+          return true;
+	else if (require == Client::CLIENT_NEW &&
+            (state == Client::CLIENT_INITIALIZED))
+          return true;
+        else
+          return false;
+      }
+
+      initialize_state_t(state_t state, const char *lockname, uint64_t reader_cnt=0)
+        : RWRefState (state, lockname, reader_cnt) {}
+      ~initialize_state_t() {}
+  };
+
+  struct mount_state_t mount_state;
+  struct initialize_state_t initialize_state;
+
+private:
+  struct C_Readahead : public Context {
+    C_Readahead(Client *c, Fh *f);
+    ~C_Readahead() override;
+    void finish(int r) override;
+
+    Client *client;
+    Fh *f;
+  };
+
+  /*
+   * These define virtual xattrs exposing the recursive directory
+   * statistics and layout metadata.
+   */
+  struct VXattr {
+	  const string name;
+	  size_t (Client::*getxattr_cb)(Inode *in, char *val, size_t size);
+	  bool readonly;
+	  bool (Client::*exists_cb)(Inode *in);
+	  unsigned int flags;
+  };
+
+  enum {
+    NO_ACL = 0,
+    POSIX_ACL,
+  };
+
+  enum {
+    MAY_EXEC = 1,
+    MAY_WRITE = 2,
+    MAY_READ = 4,
+  };
+
+  std::unique_ptr<CephContext, std::function<void(CephContext*)>> cct_deleter;
+
+  /* Flags for VXattr */
+  static const unsigned VXATTR_RSTAT = 0x1;
+  static const unsigned VXATTR_DIRSTAT = 0x2;
+
+  static const VXattr _dir_vxattrs[];
+  static const VXattr _file_vxattrs[];
+  static const VXattr _common_vxattrs[];
+
+
+  bool is_reserved_vino(vinodeno_t &vino);
+
+  void fill_dirent(struct dirent *de, const char *name, int type, uint64_t ino, loff_t next_off);
+
+  int _opendir(Inode *in, dir_result_t **dirpp, const UserPerm& perms);
+  void _readdir_drop_dirp_buffer(dir_result_t *dirp);
+  bool _readdir_have_frag(dir_result_t *dirp);
+  void _readdir_next_frag(dir_result_t *dirp);
+  void _readdir_rechoose_frag(dir_result_t *dirp);
+  int _readdir_get_frag(dir_result_t *dirp);
+  int _readdir_cache_cb(dir_result_t *dirp, add_dirent_cb_t cb, void *p, int caps, bool getref);
+  void _closedir(dir_result_t *dirp);
+
+  // other helpers
+  void _fragmap_remove_non_leaves(Inode *in);
+  void _fragmap_remove_stopped_mds(Inode *in, mds_rank_t mds);
+
+  void _ll_get(Inode *in);
+  int _ll_put(Inode *in, uint64_t num);
+  void _ll_drop_pins();
+
+  Fh *_create_fh(Inode *in, int flags, int cmode, const UserPerm& perms);
+  int _release_fh(Fh *fh);
+  void _put_fh(Fh *fh);
+
+  std::pair<int, bool> _do_remount(bool retry_on_error);
+
+  int _read_sync(Fh *f, uint64_t off, uint64_t len, bufferlist *bl, bool *checkeof);
+  int _read_async(Fh *f, uint64_t off, uint64_t len, bufferlist *bl);
+
+  bool _dentry_valid(const Dentry *dn);
+
+  // internal interface
+  //   call these with client_lock held!
+  int _do_lookup(Inode *dir, const string& name, int mask, InodeRef *target,
+		 const UserPerm& perms);
+
+  int _lookup(Inode *dir, const string& dname, int mask, InodeRef *target,
+	      const UserPerm& perm, std::string* alternate_name=nullptr);
+
+  int _link(Inode *in, Inode *dir, const char *name, const UserPerm& perm, std::string alternate_name,
+	    InodeRef *inp = 0);
+  int _unlink(Inode *dir, const char *name, const UserPerm& perm);
+  int _rename(Inode *olddir, const char *oname, Inode *ndir, const char *nname, const UserPerm& perm, std::string alternate_name);
+  int _mkdir(Inode *dir, const char *name, mode_t mode, const UserPerm& perm,
+	     InodeRef *inp = 0, const std::map<std::string, std::string> &metadata={},
+             std::string alternate_name="");
+  int _rmdir(Inode *dir, const char *name, const UserPerm& perms);
+  int _symlink(Inode *dir, const char *name, const char *target,
+	       const UserPerm& perms, std::string alternate_name, InodeRef *inp = 0);
+  int _mknod(Inode *dir, const char *name, mode_t mode, dev_t rdev,
+	     const UserPerm& perms, InodeRef *inp = 0);
+  int _do_setattr(Inode *in, struct ceph_statx *stx, int mask,
+		  const UserPerm& perms, InodeRef *inp);
+  void stat_to_statx(struct stat *st, struct ceph_statx *stx);
+  int __setattrx(Inode *in, struct ceph_statx *stx, int mask,
+		 const UserPerm& perms, InodeRef *inp = 0);
+  int _setattrx(InodeRef &in, struct ceph_statx *stx, int mask,
+		const UserPerm& perms);
+  int _setattr(InodeRef &in, struct stat *attr, int mask,
+	       const UserPerm& perms);
+  int _ll_setattrx(Inode *in, struct ceph_statx *stx, int mask,
+		   const UserPerm& perms, InodeRef *inp = 0);
+  int _getattr(Inode *in, int mask, const UserPerm& perms, bool force=false);
+  int _getattr(InodeRef &in, int mask, const UserPerm& perms, bool force=false) {
+    return _getattr(in.get(), mask, perms, force);
+  }
+  int _readlink(Inode *in, char *buf, size_t size);
+  int _getxattr(Inode *in, const char *name, void *value, size_t len,
+		const UserPerm& perms);
+  int _getxattr(InodeRef &in, const char *name, void *value, size_t len,
+		const UserPerm& perms);
+  int _getvxattr(Inode *in, const UserPerm& perms, const char *attr_name,
+                 ssize_t size, void *value, mds_rank_t rank);
+  int _listxattr(Inode *in, char *names, size_t len, const UserPerm& perms);
+  int _do_setxattr(Inode *in, const char *name, const void *value, size_t len,
+		   int flags, const UserPerm& perms);
+  int _setxattr(Inode *in, const char *name, const void *value, size_t len,
+		int flags, const UserPerm& perms);
+  int _setxattr(InodeRef &in, const char *name, const void *value, size_t len,
+		int flags, const UserPerm& perms);
+  int _setxattr_check_data_pool(string& name, string& value, const OSDMap *osdmap);
+  void _setxattr_maybe_wait_for_osdmap(const char *name, const void *value, size_t len);
+  int _removexattr(Inode *in, const char *nm, const UserPerm& perms);
+  int _removexattr(InodeRef &in, const char *nm, const UserPerm& perms);
+  int _open(Inode *in, int flags, mode_t mode, Fh **fhp,
+	    const UserPerm& perms);
+  int _renew_caps(Inode *in);
+  int _create(Inode *in, const char *name, int flags, mode_t mode, InodeRef *inp,
+	      Fh **fhp, int stripe_unit, int stripe_count, int object_size,
+	      const char *data_pool, bool *created, const UserPerm &perms,
+              std::string alternate_name);
+
+  loff_t _lseek(Fh *fh, loff_t offset, int whence);
+  int64_t _read(Fh *fh, int64_t offset, uint64_t size, bufferlist *bl);
+  int64_t _write(Fh *fh, int64_t offset, uint64_t size, const char *buf,
+          const struct iovec *iov, int iovcnt);
+  int64_t _preadv_pwritev_locked(Fh *fh, const struct iovec *iov,
+                                 unsigned iovcnt, int64_t offset,
+                                 bool write, bool clamp_to_int,
+                                 std::unique_lock<ceph::mutex> &cl);
+  int _preadv_pwritev(int fd, const struct iovec *iov, unsigned iovcnt, int64_t offset, bool write);
+  int _flush(Fh *fh);
+  int _fsync(Fh *fh, bool syncdataonly);
+  int _fsync(Inode *in, bool syncdataonly);
+  int _sync_fs();
+  int _fallocate(Fh *fh, int mode, int64_t offset, int64_t length);
+  int _getlk(Fh *fh, struct flock *fl, uint64_t owner);
+  int _setlk(Fh *fh, struct flock *fl, uint64_t owner, int sleep);
+  int _flock(Fh *fh, int cmd, uint64_t owner);
+  int _lazyio(Fh *fh, int enable);
+
+  int get_or_create(Inode *dir, const char* name,
+		    Dentry **pdn, bool expect_null=false);
+
+  int xattr_permission(Inode *in, const char *name, unsigned want,
+		       const UserPerm& perms);
+  int may_setattr(Inode *in, struct ceph_statx *stx, int mask,
+		  const UserPerm& perms);
+  int may_open(Inode *in, int flags, const UserPerm& perms);
+  int may_lookup(Inode *dir, const UserPerm& perms);
+  int may_create(Inode *dir, const UserPerm& perms);
+  int may_delete(Inode *dir, const char *name, const UserPerm& perms);
+  int may_hardlink(Inode *in, const UserPerm& perms);
+
+  int _getattr_for_perm(Inode *in, const UserPerm& perms);
+
+  vinodeno_t _get_vino(Inode *in);
+
+  bool _vxattrcb_quota_exists(Inode *in);
+  size_t _vxattrcb_quota(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_quota_max_bytes(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_quota_max_files(Inode *in, char *val, size_t size);
+
+  bool _vxattrcb_layout_exists(Inode *in);
+  size_t _vxattrcb_layout(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_layout_stripe_unit(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_layout_stripe_count(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_layout_object_size(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_layout_pool(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_layout_pool_namespace(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_entries(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_files(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_subdirs(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rentries(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rfiles(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rsubdirs(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rsnaps(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rbytes(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_dir_rctime(Inode *in, char *val, size_t size);
+
+  bool _vxattrcb_dir_pin_exists(Inode *in);
+  size_t _vxattrcb_dir_pin(Inode *in, char *val, size_t size);
+
+  bool _vxattrcb_snap_btime_exists(Inode *in);
+  size_t _vxattrcb_snap_btime(Inode *in, char *val, size_t size);
+
+  bool _vxattrcb_mirror_info_exists(Inode *in);
+  size_t _vxattrcb_mirror_info(Inode *in, char *val, size_t size);
+
+  size_t _vxattrcb_cluster_fsid(Inode *in, char *val, size_t size);
+  size_t _vxattrcb_client_id(Inode *in, char *val, size_t size);
+
+  static const VXattr *_get_vxattrs(Inode *in);
+  static const VXattr *_match_vxattr(Inode *in, const char *name);
+
+  int _do_filelock(Inode *in, Fh *fh, int lock_type, int op, int sleep,
+		   struct flock *fl, uint64_t owner, bool removing=false);
+  int _interrupt_filelock(MetaRequest *req);
+  void _encode_filelocks(Inode *in, bufferlist& bl);
+  void _release_filelocks(Fh *fh);
+  void _update_lock_state(struct flock *fl, uint64_t owner, ceph_lock_state_t *lock_state);
+
+  int _posix_acl_create(Inode *dir, mode_t *mode, bufferlist& xattrs_bl,
+			const UserPerm& perms);
+  int _posix_acl_chmod(Inode *in, mode_t mode, const UserPerm& perms);
+  int _posix_acl_permission(Inode *in, const UserPerm& perms, unsigned want);
+
+  mds_rank_t _get_random_up_mds() const;
+
+  int _ll_getattr(Inode *in, int caps, const UserPerm& perms);
+  int _lookup_parent(Inode *in, const UserPerm& perms, Inode **parent=NULL);
+  int _lookup_name(Inode *in, Inode *parent, const UserPerm& perms);
+  int _lookup_vino(vinodeno_t ino, const UserPerm& perms, Inode **inode=NULL);
+  bool _ll_forget(Inode *in, uint64_t count);
+
+  void collect_and_send_metrics();
+  void collect_and_send_global_metrics();
+
+  void update_io_stat_metadata(utime_t latency);
+  void update_io_stat_read(utime_t latency);
+  void update_io_stat_write(utime_t latency);
+
+  uint32_t deleg_timeout = 0;
+
+  client_switch_interrupt_callback_t switch_interrupt_cb = nullptr;
+  client_remount_callback_t remount_cb = nullptr;
+  client_ino_callback_t ino_invalidate_cb = nullptr;
+  client_dentry_callback_t dentry_invalidate_cb = nullptr;
+  client_umask_callback_t umask_cb = nullptr;
+  client_ino_release_t ino_release_cb = nullptr;
+  void *callback_handle = nullptr;
+  bool can_invalidate_dentries = false;
+
+  Finisher async_ino_invalidator;
+  Finisher async_dentry_invalidator;
+  Finisher interrupt_finisher;
+  Finisher remount_finisher;
+  Finisher async_ino_releasor;
+  Finisher objecter_finisher;
+
+  ceph::coarse_mono_time last_cap_renew;
+
+  CommandHook m_command_hook;
+
+  int user_id, group_id;
+  int acl_type = NO_ACL;
+
+  epoch_t cap_epoch_barrier = 0;
+
+  // mds sessions
+  map<mds_rank_t, MetaSession> mds_sessions;  // mds -> push seq
+  std::set<mds_rank_t> mds_ranks_closing;  // mds ranks currently tearing down sessions
+  std::list<ceph::condition_variable*> waiting_for_mdsmap;
+
+  // FSMap, for when using mds_command
+  std::list<ceph::condition_variable*> waiting_for_fsmap;
+  std::unique_ptr<FSMap> fsmap;
+  std::unique_ptr<FSMapUser> fsmap_user;
+
+  // This mutex only protects command_table
+  ceph::mutex command_lock = ceph::make_mutex("Client::command_lock");
+  // MDS command state
+  CommandTable<MDSCommandOp> command_table;
+
+  bool _use_faked_inos;
+
+  // Cluster fsid
+  fs_cluster_id_t fscid;
+
+  // file handles, etc.
+  interval_set<int> free_fd_set;  // unused fds
+  ceph::unordered_map<int, Fh*> fd_map;
+  set<Fh*> ll_unclosed_fh_set;
+  ceph::unordered_set<dir_result_t*> opened_dirs;
+  uint64_t fd_gen = 1;
+
+  bool   mount_aborted = false;
+  bool   blocklisted = false;
+
+  ceph::unordered_map<vinodeno_t, Inode*> inode_map;
+  ceph::unordered_map<ino_t, vinodeno_t> faked_ino_map;
+  interval_set<ino_t> free_faked_inos;
+  ino_t last_used_faked_ino;
+  ino_t last_used_faked_root;
+
+  int local_osd = -CEPHFS_ENXIO;
+  epoch_t local_osd_epoch = 0;
+
+  // mds requests
+  ceph_tid_t last_tid = 0;
+  ceph_tid_t oldest_tid = 0; // oldest incomplete mds request, excluding setfilelock requests
+  map<ceph_tid_t, MetaRequest*> mds_requests;
+
+  // cap flushing
+  ceph_tid_t last_flush_tid = 1;
+
+  // dirty_list keeps all the dirty inodes before flushing.
+  xlist<Inode*> delayed_list, dirty_list;
+  int num_flushing_caps = 0;
+  ceph::unordered_map<inodeno_t,SnapRealm*> snap_realms;
+  std::map<std::string, std::string> metadata;
+
+  ceph::coarse_mono_time last_auto_reconnect;
+  std::chrono::seconds caps_release_delay, mount_timeout;
+  // trace generation
+  ofstream traceout;
+
+  ceph::condition_variable mount_cond, sync_cond;
+
+  std::map<std::pair<int64_t,std::string>, int> pool_perms;
+  std::list<ceph::condition_variable*> waiting_for_pool_perm;
+
+  uint64_t retries_on_invalidate = 0;
+
+  // state reclaim
+  int reclaim_errno = 0;
+  epoch_t reclaim_osd_epoch = 0;
+  entity_addrvec_t reclaim_target_addrs;
+
+  // dentry lease metrics
+  uint64_t dentry_nr = 0;
+  uint64_t dlease_hits = 0;
+  uint64_t dlease_misses = 0;
+
+  uint64_t cap_hits = 0;
+  uint64_t cap_misses = 0;
+
+  uint64_t opened_files = 0;
+  uint64_t pinned_icaps = 0;
+  uint64_t opened_inodes = 0;
+
+  uint64_t total_read_ops = 0;
+  uint64_t total_read_size = 0;
+
+  uint64_t total_write_ops = 0;
+  uint64_t total_write_size = 0;
+
+  ceph::spinlock delay_i_lock;
+  std::map<Inode*,int> delay_i_release;
+
+  uint64_t nr_metadata_request = 0;
+  uint64_t nr_read_request = 0;
+  uint64_t nr_write_request = 0;
+};
+
+/**
+ * Specialization of Client that manages its own Objecter instance
+ * and handles init/shutdown of messenger/monclient
+ */
+class StandaloneClient : public Client
+{
+public:
+  StandaloneClient(Messenger *m, MonClient *mc, boost::asio::io_context& ictx);
+
+  ~StandaloneClient() override;
+
+  int init() override;
+  void shutdown() override;
+};
+
+#endif
diff --git a/src/client/ClientSnapRealm.cc b/src/client/ClientSnapRealm.cc
new file mode 100644
index 000000000..3656fbdf5
--- /dev/null
+++ b/src/client/ClientSnapRealm.cc
@@ -0,0 +1,29 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "ClientSnapRealm.h"
+#include "common/Formatter.h"
+
+void SnapRealm::dump(Formatter *f) const
+{
+  f->dump_stream("ino") << ino;
+  f->dump_int("nref", nref);
+  f->dump_stream("created") << created;
+  f->dump_stream("seq") << seq;
+  f->dump_stream("parent_ino") << parent;
+  f->dump_stream("parent_since") << parent_since;
+
+  f->open_array_section("prior_parent_snaps");
+  for (vector<snapid_t>::const_iterator p = prior_parent_snaps.begin(); p != prior_parent_snaps.end(); ++p)
+    f->dump_stream("snapid") << *p;
+  f->close_section();
+  f->open_array_section("my_snaps");
+  for (vector<snapid_t>::const_iterator p = my_snaps.begin(); p != my_snaps.end(); ++p)
+    f->dump_stream("snapid") << *p;
+  f->close_section();
+
+  f->open_array_section("children");
+  for (set<SnapRealm*>::const_iterator p = pchildren.begin(); p != pchildren.end(); ++p)
+    f->dump_stream("child") << (*p)->ino;
+  f->close_section();
+}
diff --git a/src/client/ClientSnapRealm.h b/src/client/ClientSnapRealm.h
new file mode 100644
index 000000000..ccb129d78
--- /dev/null
+++ b/src/client/ClientSnapRealm.h
@@ -0,0 +1,60 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#ifndef CEPH_CLIENT_SNAPREALM_H
+#define CEPH_CLIENT_SNAPREALM_H
+
+#include "include/types.h"
+#include "common/snap_types.h"
+#include "include/xlist.h"
+
+struct Inode;
+
+struct SnapRealm {
+  inodeno_t ino;
+  int nref;
+  snapid_t created;
+  snapid_t seq;
+  
+  inodeno_t parent;
+  snapid_t parent_since;
+  vector<snapid_t> prior_parent_snaps;  // snaps prior to parent_since
+  vector<snapid_t> my_snaps;
+
+  SnapRealm *pparent;
+  set<SnapRealm*> pchildren;
+
+private:
+  SnapContext cached_snap_context;  // my_snaps + parent snaps + past_parent_snaps
+  friend ostream& operator<<(ostream& out, const SnapRealm& r);
+
+public:
+  xlist<Inode*> inodes_with_caps;
+
+  explicit SnapRealm(inodeno_t i) :
+    ino(i), nref(0), created(0), seq(0),
+    pparent(NULL) { }
+
+  void build_snap_context();
+  void invalidate_cache() {
+    cached_snap_context.clear();
+  }
+
+  const SnapContext& get_snap_context() {
+    if (cached_snap_context.seq == 0)
+      build_snap_context();
+    return cached_snap_context;
+  }
+
+  void dump(Formatter *f) const;
+};
+
+inline ostream& operator<<(ostream& out, const SnapRealm& r) {
+  return out << "snaprealm(" << r.ino << " nref=" << r.nref << " c=" << r.created << " seq=" << r.seq
+	     << " parent=" << r.parent
+	     << " my_snaps=" << r.my_snaps
+	     << " cached_snapc=" << r.cached_snap_context
+	     << ")";
+}
+
+#endif
diff --git a/src/client/Delegation.cc b/src/client/Delegation.cc
new file mode 100644
index 000000000..0c0ac1b35
--- /dev/null
+++ b/src/client/Delegation.cc
@@ -0,0 +1,131 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+#include "common/Clock.h"
+#include "common/Timer.h"
+
+#include "Client.h"
+#include "Inode.h"
+#include "Fh.h"
+#include "Delegation.h"
+
+class C_Deleg_Timeout : public Context {
+  Delegation *deleg;
+public:
+  explicit C_Deleg_Timeout(Delegation *d) : deleg(d) {}
+  void finish(int r) override {
+    Inode *in = deleg->get_fh()->inode.get();
+    Client *client = in->client;
+
+    lsubdout(client->cct, client, 0) << __func__ <<
+	  ": delegation return timeout for inode 0x" <<
+	  std::hex << in->ino << ". Forcibly unmounting client. "<<
+	  client << std::dec << dendl;
+    client->_unmount(false);
+  }
+};
+
+/**
+ * ceph_deleg_caps_for_type - what caps are necessary for a delegation?
+ * @type: delegation request type
+ *
+ * Determine what caps are necessary in order to grant a delegation of a given
+ * type. For read delegations, we need whatever we require in order to do
+ * cached reads, plus AsLs to cover metadata changes that should trigger a
+ * recall. We also grab Xs since changing xattrs usually alters the mtime and
+ * so would trigger a recall.
+ *
+ * For write delegations, we need whatever read delegations need plus the
+ * caps to allow writing to the file (Fbwx).
+ */
+int ceph_deleg_caps_for_type(unsigned type)
+{
+	int caps = CEPH_CAP_PIN;
+
+	switch (type) {
+	case CEPH_DELEGATION_WR:
+		caps |= CEPH_CAP_FILE_EXCL |
+			CEPH_CAP_FILE_WR | CEPH_CAP_FILE_BUFFER;
+		/* Fallthrough */
+	case CEPH_DELEGATION_RD:
+		caps |= CEPH_CAP_FILE_SHARED |
+			CEPH_CAP_FILE_RD | CEPH_CAP_FILE_CACHE |
+			CEPH_CAP_XATTR_SHARED |
+			CEPH_CAP_LINK_SHARED | CEPH_CAP_AUTH_SHARED;
+		break;
+	default:
+		// Should never happen
+		ceph_abort();
+	}
+	return caps;
+}
+
+/*
+ * A delegation is a container for holding caps on behalf of a client that
+ * wants to be able to rely on them until recalled.
+ */
+Delegation::Delegation(Fh *_fh, unsigned _type, ceph_deleg_cb_t _cb, void *_priv)
+	: fh(_fh), priv(_priv), type(_type), recall_cb(_cb),
+	  recall_time(utime_t()), timeout_event(nullptr)
+{
+  Inode *inode = _fh->inode.get();
+  inode->client->get_cap_ref(inode, ceph_deleg_caps_for_type(_type));
+};
+
+Delegation::~Delegation()
+{
+  disarm_timeout();
+  Inode *inode = fh->inode.get();
+  inode->client->put_cap_ref(inode, ceph_deleg_caps_for_type(type));
+}
+
+void Delegation::reinit(unsigned _type, ceph_deleg_cb_t _recall_cb, void *_priv)
+{
+  /* update cap refs -- note that we do a get first to avoid any going to 0 */
+  if (type != _type) {
+    Inode *inode = fh->inode.get();
+
+    inode->client->get_cap_ref(inode, ceph_deleg_caps_for_type(_type));
+    inode->client->put_cap_ref(inode, ceph_deleg_caps_for_type(type));
+    type = _type;
+  }
+
+  recall_cb = _recall_cb;
+  priv = _priv;
+}
+
+void Delegation::arm_timeout()
+{
+  Client *client = fh->inode.get()->client;
+
+  std::scoped_lock l(client->timer_lock);
+  if (timeout_event)
+    return;
+
+  timeout_event = new C_Deleg_Timeout(this);
+  client->timer.add_event_after(client->get_deleg_timeout(), timeout_event);
+}
+
+void Delegation::disarm_timeout()
+{
+  Client *client = fh->inode.get()->client;
+
+  std::scoped_lock l(client->timer_lock);
+  if (!timeout_event)
+    return;
+
+  client->timer.cancel_event(timeout_event);
+  timeout_event = nullptr;
+}
+
+void Delegation::recall(bool skip_read)
+{
+  /* If skip_read is true, don't break read delegations */
+  if (skip_read && type == CEPH_DELEGATION_RD)
+    return;
+
+  if (!is_recalled()) {
+    recall_cb(fh, priv);
+    recall_time = ceph_clock_now();
+    arm_timeout();
+  }
+}
diff --git a/src/client/Delegation.h b/src/client/Delegation.h
new file mode 100644
index 000000000..d24a02487
--- /dev/null
+++ b/src/client/Delegation.h
@@ -0,0 +1,57 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+#ifndef _CEPH_CLIENT_DELEGATION_H
+#define _CEPH_CLIENT_DELEGATION_H
+
+#include "common/Clock.h"
+#include "common/Timer.h"
+#include "include/cephfs/ceph_ll_client.h"
+
+/* Commands for manipulating delegation state */
+#ifndef CEPH_DELEGATION_NONE
+# define CEPH_DELEGATION_NONE	0
+# define CEPH_DELEGATION_RD	1
+# define CEPH_DELEGATION_WR	2
+#endif
+
+/* Converts CEPH_DELEGATION_* to cap mask */
+int ceph_deleg_caps_for_type(unsigned type);
+
+/*
+ * A delegation is a container for holding caps on behalf of a client that
+ * wants to be able to rely on them until recalled.
+ */
+class Delegation {
+public:
+  Delegation(Fh *_fh, unsigned _type, ceph_deleg_cb_t _cb, void *_priv);
+  ~Delegation();
+  Fh *get_fh() { return fh; }
+  unsigned get_type() { return type; }
+  bool is_recalled() { return !recall_time.is_zero(); }
+
+  void reinit(unsigned _type, ceph_deleg_cb_t _recall_cb, void *_priv);
+  void recall(bool skip_read);
+private:
+  // Filehandle against which it was acquired
+  Fh				*fh;
+
+  // opaque token that will be passed to the callback
+  void				*priv;
+
+  // CEPH_DELEGATION_* type
+  unsigned			type;
+
+  // callback into application to recall delegation
+  ceph_deleg_cb_t		recall_cb;
+
+  // time of first recall
+  utime_t			recall_time;
+
+  // timer for unreturned delegations
+  Context			*timeout_event;
+
+  void arm_timeout();
+  void disarm_timeout();
+};
+
+#endif /* _CEPH_CLIENT_DELEGATION_H */
diff --git a/src/client/Dentry.cc b/src/client/Dentry.cc
new file mode 100644
index 000000000..f8741050a
--- /dev/null
+++ b/src/client/Dentry.cc
@@ -0,0 +1,33 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "include/types.h"
+#include "include/utime.h"
+
+#include "Dentry.h"
+#include "Dir.h"
+#include "Inode.h"
+
+#include "common/Formatter.h"
+
+void Dentry::dump(Formatter *f) const
+{
+  f->dump_string("name", name);
+  f->dump_stream("dir") << dir->parent_inode->ino;
+  if (inode)
+    f->dump_stream("ino") << inode->ino;
+  f->dump_int("ref", ref);
+  f->dump_int("offset", offset);
+  if (lease_mds >= 0) {
+    f->dump_int("lease_mds", lease_mds);
+    f->dump_stream("lease_ttl") << lease_ttl;
+    f->dump_unsigned("lease_gen", lease_gen);
+    f->dump_unsigned("lease_seq", lease_seq);
+  }
+  f->dump_int("cap_shared_gen", cap_shared_gen);
+}
+
+std::ostream &operator<<(std::ostream &oss, const Dentry &dn)
+{
+  return oss << dn.dir->parent_inode->vino() << "[\"" << dn.name << "\"]";
+}
diff --git a/src/client/Dentry.h b/src/client/Dentry.h
new file mode 100644
index 000000000..483a31ecc
--- /dev/null
+++ b/src/client/Dentry.h
@@ -0,0 +1,99 @@
+#ifndef CEPH_CLIENT_DENTRY_H
+#define CEPH_CLIENT_DENTRY_H
+
+#include "include/lru.h"
+#include "include/xlist.h"
+
+#include "mds/mdstypes.h"
+#include "Inode.h"
+#include "InodeRef.h"
+#include "Dir.h"
+
+class Dentry : public LRUObject {
+public:
+  explicit Dentry(Dir *_dir, const std::string &_name) :
+    dir(_dir), name(_name), inode_xlist_link(this)
+  {
+    auto r = dir->dentries.insert(make_pair(name, this));
+    ceph_assert(r.second);
+    dir->num_null_dentries++;
+  }
+  ~Dentry() {
+    ceph_assert(ref == 0);
+    ceph_assert(dir == nullptr);
+  }
+
+  /*
+   * ref==1 -> cached, unused
+   * ref >1 -> pinned in lru
+   */
+  void get() {
+    ceph_assert(ref > 0);
+    if (++ref == 2)
+      lru_pin();
+    //cout << "dentry.get on " << this << " " << name << " now " << ref << std::endl;
+  }
+  void put() {
+    ceph_assert(ref > 0);
+    if (--ref == 1)
+      lru_unpin();
+    //cout << "dentry.put on " << this << " " << name << " now " << ref << std::endl;
+    if (ref == 0)
+      delete this;
+  }
+  void link(InodeRef in) {
+    inode = in;
+    inode->dentries.push_back(&inode_xlist_link);
+    if (inode->is_dir()) {
+      if (inode->dir)
+        get(); // dir -> dn pin
+      if (inode->ll_ref)
+        get(); // ll_ref -> dn pin
+    }
+    dir->num_null_dentries--;
+  }
+  void unlink(void) {
+    if (inode->is_dir()) {
+      if (inode->dir)
+        put(); // dir -> dn pin
+      if (inode->ll_ref)
+        put(); // ll_ref -> dn pin
+    }
+    ceph_assert(inode_xlist_link.get_list() == &inode->dentries);
+    inode_xlist_link.remove_myself();
+    inode.reset();
+    dir->num_null_dentries++;
+  }
+  void mark_primary() {
+    if (inode && inode->dentries.front() != this)
+      inode->dentries.push_front(&inode_xlist_link);
+  }
+  void detach(void) {
+    ceph_assert(!inode);
+    auto p = dir->dentries.find(name);
+    ceph_assert(p != dir->dentries.end());
+    dir->dentries.erase(p);
+    dir->num_null_dentries--;
+    dir = nullptr;
+  }
+
+  void dump(Formatter *f) const;
+  friend std::ostream &operator<<(std::ostream &oss, const Dentry &Dentry);
+
+  Dir	   *dir;
+  const string name;
+  InodeRef inode;
+  int	   ref = 1; // 1 if there's a dir beneath me.
+  int64_t offset = 0;
+  mds_rank_t lease_mds = -1;
+  utime_t lease_ttl;
+  uint64_t lease_gen = 0;
+  ceph_seq_t lease_seq = 0;
+  int cap_shared_gen = 0;
+  std::string alternate_name;
+
+private:
+  xlist<Dentry *>::item inode_xlist_link;
+};
+
+#endif
diff --git a/src/client/Dir.h b/src/client/Dir.h
new file mode 100644
index 000000000..731a2038e
--- /dev/null
+++ b/src/client/Dir.h
@@ -0,0 +1,19 @@
+#ifndef CEPH_CLIENT_DIR_H
+#define CEPH_CLIENT_DIR_H
+
+struct Inode;
+
+class Dir {
+ public:
+  Inode    *parent_inode;  // my inode
+  ceph::unordered_map<string, Dentry*> dentries;
+  unsigned num_null_dentries = 0;
+
+  vector<Dentry*> readdir_cache;
+
+  explicit Dir(Inode* in) { parent_inode = in; }
+
+  bool is_empty() {  return dentries.empty(); }
+};
+
+#endif
diff --git a/src/client/Fh.cc b/src/client/Fh.cc
new file mode 100644
index 000000000..62bd26140
--- /dev/null
+++ b/src/client/Fh.cc
@@ -0,0 +1,32 @@
+
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2017 Red Hat Inc
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+
+#include "Inode.h"
+
+#include "Fh.h"
+
+Fh::Fh(InodeRef in, int flags, int cmode, uint64_t _gen, const UserPerm &perms) :
+    inode(in), _ref(1), pos(0), mds(0), mode(cmode), gen(_gen), flags(flags),
+    pos_locked(false), actor_perms(perms), readahead()
+{
+  inode->add_fh(this);
+}
+
+Fh::~Fh()
+{
+  inode->rm_fh(this);
+}
+
diff --git a/src/client/Fh.h b/src/client/Fh.h
new file mode 100644
index 000000000..c3355ba6c
--- /dev/null
+++ b/src/client/Fh.h
@@ -0,0 +1,62 @@
+#ifndef CEPH_CLIENT_FH_H
+#define CEPH_CLIENT_FH_H
+
+#include "common/Readahead.h"
+#include "include/types.h"
+#include "InodeRef.h"
+#include "UserPerm.h"
+#include "mds/flock.h"
+
+class Inode;
+
+// file handle for any open file state
+
+struct Fh {
+  InodeRef  inode;
+  int	    _ref;
+  loff_t    pos;
+  int       mds;        // have to talk to mds we opened with (for now)
+  int       mode;       // the mode i opened the file with
+  uint64_t  gen;
+
+  int flags;
+  bool pos_locked;           // pos is currently in use
+  std::list<ceph::condition_variable*> pos_waiters;   // waiters for pos
+
+  UserPerm actor_perms; // perms I opened the file with
+
+  Readahead readahead;
+
+  // file lock
+  std::unique_ptr<ceph_lock_state_t> fcntl_locks;
+  std::unique_ptr<ceph_lock_state_t> flock_locks;
+
+  bool has_any_filelocks() {
+    return
+      (fcntl_locks && !fcntl_locks->empty()) ||
+      (flock_locks && !flock_locks->empty());
+  }
+
+  // IO error encountered by any writeback on this Inode while
+  // this Fh existed (i.e. an fsync on another Fh will still show
+  // up as an async_err here because it could have been the same
+  // bytes we wrote via this Fh).
+  int async_err = {0};
+
+  int take_async_err()
+  {
+      int e = async_err;
+      async_err = 0;
+      return e;
+  }
+
+  Fh() = delete;
+  Fh(InodeRef in, int flags, int cmode, uint64_t gen, const UserPerm &perms);
+  ~Fh();
+
+  void get() { ++_ref; }
+  int put() { return --_ref; }
+};
+
+
+#endif
diff --git a/src/client/Inode.cc b/src/client/Inode.cc
new file mode 100644
index 000000000..5b1b93d9f
--- /dev/null
+++ b/src/client/Inode.cc
@@ -0,0 +1,794 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "Client.h"
+#include "Inode.h"
+#include "Dentry.h"
+#include "Dir.h"
+#include "Fh.h"
+#include "MetaSession.h"
+#include "ClientSnapRealm.h"
+#include "Delegation.h"
+
+#include "mds/flock.h"
+
+Inode::~Inode()
+{
+  delay_cap_item.remove_myself();
+  dirty_cap_item.remove_myself(); 
+  snaprealm_item.remove_myself();
+
+  if (snapdir_parent) {
+    snapdir_parent->flags &= ~I_SNAPDIR_OPEN;
+    snapdir_parent.reset();
+  }
+
+  if (!oset.objects.empty()) {
+    lsubdout(client->cct, client, 0) << __func__ << ": leftover objects on inode 0x"
+      << std::hex << ino << std::dec << dendl;
+    ceph_assert(oset.objects.empty());
+  }
+
+  if (!delegations.empty()) {
+    lsubdout(client->cct, client, 0) << __func__ << ": leftover delegations on inode 0x"
+      << std::hex << ino << std::dec << dendl;
+    ceph_assert(delegations.empty());
+  }
+}
+
+ostream& operator<<(ostream &out, const Inode &in)
+{
+  out << in.vino() << "("
+      << "faked_ino=" << in.faked_ino
+      << " nref=" << in.get_nref()
+      << " ll_ref=" << in.ll_ref
+      << " cap_refs=" << in.cap_refs
+      << " open=" << in.open_by_mode
+      << " mode=" << oct << in.mode << dec
+      << " size=" << in.size << "/" << in.max_size
+      << " nlink=" << in.nlink
+      << " btime=" << in.btime
+      << " mtime=" << in.mtime
+      << " ctime=" << in.ctime
+      << " change_attr=" << in.change_attr
+      << " caps=" << ccap_string(in.caps_issued());
+  if (!in.caps.empty()) {
+    out << "(";
+    bool first = true;
+    for (const auto &pair : in.caps) {
+      if (!first)
+        out << ',';
+      out << pair.first << '=' << ccap_string(pair.second.issued);
+      first = false;
+    }
+    out << ")";
+  }
+  if (in.dirty_caps)
+    out << " dirty_caps=" << ccap_string(in.dirty_caps);
+  if (in.flushing_caps)
+    out << " flushing_caps=" << ccap_string(in.flushing_caps);
+
+  if (in.flags & I_COMPLETE)
+    out << " COMPLETE";
+
+  if (in.is_file())
+    out << " " << in.oset;
+
+  if (!in.dentries.empty())
+    out << " parents=" << in.dentries;
+
+  if (in.is_dir() && in.has_dir_layout())
+    out << " has_dir_layout";
+
+  if (in.quota.is_enable())
+    out << " " << in.quota;
+
+  out << ' ' << &in << ")";
+  return out;
+}
+
+
+void Inode::make_long_path(filepath& p)
+{
+  if (!dentries.empty()) {
+    Dentry *dn = get_first_parent();
+    ceph_assert(dn->dir && dn->dir->parent_inode);
+    dn->dir->parent_inode->make_long_path(p);
+    p.push_dentry(dn->name);
+  } else if (snapdir_parent) {
+    make_nosnap_relative_path(p);
+  } else
+    p = filepath(ino);
+}
+
+void Inode::make_short_path(filepath& p)
+{
+  if (!dentries.empty()) {
+    Dentry *dn = get_first_parent();
+    ceph_assert(dn->dir && dn->dir->parent_inode);
+    p = filepath(dn->name, dn->dir->parent_inode->ino);
+  } else if (snapdir_parent) {
+    make_nosnap_relative_path(p);
+  } else
+    p = filepath(ino);
+}
+
+/*
+ * make a filepath suitable for an mds request:
+ *  - if we are non-snapped/live, the ino is sufficient, e.g. #1234
+ *  - if we are snapped, make filepath relative to first non-snapped parent.
+ */
+void Inode::make_nosnap_relative_path(filepath& p)
+{
+  if (snapid == CEPH_NOSNAP) {
+    p = filepath(ino);
+  } else if (snapdir_parent) {
+    snapdir_parent->make_nosnap_relative_path(p);
+    string empty;
+    p.push_dentry(empty);
+  } else if (!dentries.empty()) {
+    Dentry *dn = get_first_parent();
+    ceph_assert(dn->dir && dn->dir->parent_inode);
+    dn->dir->parent_inode->make_nosnap_relative_path(p);
+    p.push_dentry(dn->name);
+  } else {
+    p = filepath(ino);
+  }
+}
+
+void Inode::get_open_ref(int mode)
+{
+  client->inc_opened_files();
+  if (open_by_mode[mode] == 0) {
+    client->inc_opened_inodes();
+  }
+  open_by_mode[mode]++;
+  break_deleg(!(mode & CEPH_FILE_MODE_WR));
+}
+
+bool Inode::put_open_ref(int mode)
+{
+  //cout << "open_by_mode[" << mode << "] " << open_by_mode[mode] << " -> " << (open_by_mode[mode]-1) << std::endl;
+  auto& ref = open_by_mode.at(mode);
+  ceph_assert(ref > 0);
+  client->dec_opened_files();
+  if (--ref == 0) {
+    client->dec_opened_inodes();
+    return true;
+  }
+  return false;
+}
+
+void Inode::get_cap_ref(int cap)
+{
+  int n = 0;
+  while (cap) {
+    if (cap & 1) {
+      int c = 1 << n;
+      cap_refs[c]++;
+      //cout << "inode " << *this << " get " << cap_string(c) << " " << (cap_refs[c]-1) << " -> " << cap_refs[c] << std::endl;
+    }
+    cap >>= 1;
+    n++;
+  }
+}
+
+int Inode::put_cap_ref(int cap)
+{
+  int last = 0;
+  int n = 0;
+  while (cap) {
+    if (cap & 1) {
+      int c = 1 << n;
+      if (cap_refs[c] <= 0) {
+	lderr(client->cct) << "put_cap_ref " << ccap_string(c) << " went negative on " << *this << dendl;
+	ceph_assert(cap_refs[c] > 0);
+      }
+      if (--cap_refs[c] == 0)
+        last |= c;
+      //cout << "inode " << *this << " put " << cap_string(c) << " " << (cap_refs[c]+1) << " -> " << cap_refs[c] << std::endl;
+    }
+    cap >>= 1;
+    n++;
+  }
+  return last;
+}
+
+bool Inode::is_any_caps()
+{
+  return !caps.empty() || snap_caps;
+}
+
+bool Inode::cap_is_valid(const Cap &cap) const
+{
+  /*cout << "cap_gen     " << cap->session-> cap_gen << std::endl
+    << "session gen " << cap->gen << std::endl
+    << "cap expire  " << cap->session->cap_ttl << std::endl
+    << "cur time    " << ceph_clock_now(cct) << std::endl;*/
+  if ((cap.session->cap_gen <= cap.gen)
+      && (ceph_clock_now() < cap.session->cap_ttl)) {
+    return true;
+  }
+  return false;
+}
+
+int Inode::caps_issued(int *implemented) const
+{
+  int c = snap_caps;
+  int i = 0;
+  for (const auto &[mds, cap] : caps) {
+    if (cap_is_valid(cap)) {
+      c |= cap.issued;
+      i |= cap.implemented;
+    }
+  }
+  // exclude caps issued by non-auth MDS, but are been revoking by
+  // the auth MDS. The non-auth MDS should be revoking/exporting
+  // these caps, but the message is delayed.
+  if (auth_cap)
+    c &= ~auth_cap->implemented | auth_cap->issued;
+
+  if (implemented)
+    *implemented = i;
+  return c;
+}
+
+void Inode::try_touch_cap(mds_rank_t mds)
+{
+  auto it = caps.find(mds);
+  if (it != caps.end()) {
+    it->second.touch();
+  }
+}
+
+/**
+ * caps_issued_mask - check whether we have all of the caps in the mask
+ * @mask: mask to check against
+ * @allow_impl: whether the caller can also use caps that are implemented but not issued
+ *
+ * This is the bog standard "check whether we have the required caps" operation.
+ * Typically, we only check against the capset that is currently "issued".
+ * In other words, we ignore caps that have been revoked but not yet released.
+ * Also account capability hit/miss stats.
+ *
+ * Some callers (particularly those doing attribute retrieval) can also make
+ * use of the full set of "implemented" caps to satisfy requests from the
+ * cache.
+ *
+ * Those callers should refrain from taking new references to implemented
+ * caps!
+ */
+bool Inode::caps_issued_mask(unsigned mask, bool allow_impl)
+{
+  int c = snap_caps;
+  int i = 0;
+
+  if ((c & mask) == mask)
+    return true;
+  // prefer auth cap
+  if (auth_cap &&
+      cap_is_valid(*auth_cap) &&
+      (auth_cap->issued & mask) == mask) {
+    auth_cap->touch();
+    client->cap_hit();
+    return true;
+  }
+  // try any cap
+  for (auto &pair : caps) {
+    Cap &cap = pair.second;
+    if (cap_is_valid(cap)) {
+      if ((cap.issued & mask) == mask) {
+        cap.touch();
+	client->cap_hit();
+	return true;
+      }
+      c |= cap.issued;
+      i |= cap.implemented;
+    }
+  }
+
+  if (allow_impl)
+    c |= i;
+
+  if ((c & mask) == mask) {
+    // bah.. touch them all
+    for (auto &pair : caps) {
+      pair.second.touch();
+    }
+    client->cap_hit();
+    return true;
+  }
+
+  client->cap_miss();
+  return false;
+}
+
+int Inode::caps_used()
+{
+  int w = 0;
+  for (const auto &[cap, cnt] : cap_refs)
+    if (cnt)
+      w |= cap;
+  return w;
+}
+
+int Inode::caps_file_wanted()
+{
+  int want = 0;
+  for (const auto &[mode, cnt] : open_by_mode)
+    if (cnt)
+      want |= ceph_caps_for_mode(mode);
+  return want;
+}
+
+int Inode::caps_wanted()
+{
+  int want = caps_file_wanted() | caps_used();
+  if (want & CEPH_CAP_FILE_BUFFER)
+    want |= CEPH_CAP_FILE_EXCL;
+  return want;
+}
+
+int Inode::caps_mds_wanted()
+{
+  int want = 0;
+  for (const auto &pair : caps) {
+    want |= pair.second.wanted;
+  }
+  return want;
+}
+
+int Inode::caps_dirty()
+{
+  return dirty_caps | flushing_caps;
+}
+
+const UserPerm* Inode::get_best_perms()
+{
+  const UserPerm *perms = NULL;
+  for (const auto &pair : caps) {
+    const UserPerm& iperm = pair.second.latest_perms;
+    if (!perms) { // we don't have any, take what's present
+      perms = &iperm;
+    } else if (iperm.uid() == uid) {
+      if (iperm.gid() == gid) { // we have the best possible, return
+	return &iperm;
+      }
+      if (perms->uid() != uid) { // take uid > gid every time
+	perms = &iperm;
+      }
+    } else if (perms->uid() != uid && iperm.gid() == gid) {
+      perms = &iperm; // a matching gid is better than nothing
+    }
+  }
+  return perms;
+}
+
+bool Inode::have_valid_size()
+{
+  // RD+RDCACHE or WR+WRBUFFER => valid size
+  if (caps_issued() & (CEPH_CAP_FILE_SHARED | CEPH_CAP_FILE_EXCL))
+    return true;
+  return false;
+}
+
+// open Dir for an inode.  if it's not open, allocated it (and pin dentry in memory).
+Dir *Inode::open_dir()
+{
+  if (!dir) {
+    dir = new Dir(this);
+    lsubdout(client->cct, client, 15) << "open_dir " << dir << " on " << this << dendl;
+    ceph_assert(dentries.size() < 2); // dirs can't be hard-linked
+    if (!dentries.empty())
+      get_first_parent()->get();      // pin dentry
+    iget();                  // pin inode
+  }
+  return dir;
+}
+
+bool Inode::check_mode(const UserPerm& perms, unsigned want)
+{
+  if (uid == perms.uid()) {
+    // if uid is owner, owner entry determines access
+    want = want << 6;
+  } else if (perms.gid_in_groups(gid)) {
+    // if a gid or sgid matches the owning group, group entry determines access
+    want = want << 3;
+  }
+
+  return (mode & want) == want;
+}
+
+void Inode::dump(Formatter *f) const
+{
+  f->dump_stream("ino") << ino;
+  f->dump_stream("snapid") << snapid;
+  if (rdev)
+    f->dump_unsigned("rdev", rdev);
+  f->dump_stream("ctime") << ctime;
+  f->dump_stream("btime") << btime;
+  f->dump_stream("mode") << '0' << std::oct << mode << std::dec;
+  f->dump_unsigned("uid", uid);
+  f->dump_unsigned("gid", gid);
+  f->dump_int("nlink", nlink);
+
+  f->dump_unsigned("size", size);
+  f->dump_unsigned("max_size", max_size);
+  f->dump_unsigned("truncate_seq", truncate_seq);
+  f->dump_unsigned("truncate_size", truncate_size);
+  f->dump_stream("mtime") << mtime;
+  f->dump_stream("atime") << atime;
+  f->dump_unsigned("time_warp_seq", time_warp_seq);
+  f->dump_unsigned("change_attr", change_attr);
+
+  f->dump_object("layout", layout);
+  if (is_dir()) {
+    f->open_object_section("dir_layout");
+    ::dump(dir_layout, f);
+    f->close_section();
+
+    f->dump_bool("complete", flags & I_COMPLETE);
+    f->dump_bool("ordered", flags & I_DIR_ORDERED);
+
+    /* FIXME when wip-mds-encoding is merged ***
+    f->open_object_section("dir_stat");
+    dirstat.dump(f);
+    f->close_section();
+
+    f->open_object_section("rstat");
+    rstat.dump(f);
+    f->close_section();
+    */
+  }
+
+  f->dump_unsigned("version", version);
+  f->dump_unsigned("xattr_version", xattr_version);
+  f->dump_unsigned("flags", flags);
+
+  if (is_dir()) {
+    f->dump_int("dir_hashed", (int)dir_hashed);
+    f->dump_int("dir_replicated", (int)dir_replicated);
+    if (dir_replicated) {
+      f->open_array_section("dirfrags");
+      for (const auto &frag : frag_repmap) {
+        f->open_object_section("frags");
+        CachedStackStringStream css;
+        *css << std::hex << frag.first.value() << "/" << std::dec << frag.first.bits();
+        f->dump_string("frag", css->strv());
+
+        f->open_array_section("repmap");
+        for (const auto &mds : frag.second) {
+          f->dump_int("mds", mds);
+        }
+        f->close_section();
+
+        f->close_section();
+      }
+      f->close_section();
+    }
+  }
+
+  f->open_array_section("caps");
+  for (const auto &pair : caps) {
+    f->open_object_section("cap");
+    f->dump_int("mds", pair.first);
+    if (&pair.second == auth_cap)
+      f->dump_int("auth", 1);
+    pair.second.dump(f);
+    f->close_section();
+  }
+  f->close_section();
+  if (auth_cap)
+    f->dump_int("auth_cap", auth_cap->session->mds_num);
+
+  f->dump_stream("dirty_caps") << ccap_string(dirty_caps);
+  if (flushing_caps) {
+    f->dump_stream("flushings_caps") << ccap_string(flushing_caps);
+    f->open_object_section("flushing_cap_tid");
+    for (map<ceph_tid_t, int>::const_iterator p = flushing_cap_tids.begin();
+	 p != flushing_cap_tids.end();
+	 ++p) {
+      string n(ccap_string(p->second));
+      f->dump_unsigned(n.c_str(), p->first);
+    }
+    f->close_section();
+  }
+  f->dump_int("shared_gen", shared_gen);
+  f->dump_int("cache_gen", cache_gen);
+  if (snap_caps) {
+    f->dump_int("snap_caps", snap_caps);
+    f->dump_int("snap_cap_refs", snap_cap_refs);
+  }
+
+  f->dump_stream("hold_caps_until") << hold_caps_until;
+
+  if (snaprealm) {
+    f->open_object_section("snaprealm");
+    snaprealm->dump(f);
+    f->close_section();
+  }
+  if (!cap_snaps.empty()) {
+    for (const auto &p : cap_snaps) {
+      f->open_object_section("cap_snap");
+      f->dump_stream("follows") << p.first;
+      p.second.dump(f);
+      f->close_section();
+    }
+  }
+
+  // open
+  if (!open_by_mode.empty()) {
+    f->open_array_section("open_by_mode");
+    for (map<int,int>::const_iterator p = open_by_mode.begin(); p != open_by_mode.end(); ++p) {
+      f->open_object_section("ref");
+      f->dump_int("mode", p->first);
+      f->dump_int("refs", p->second);
+      f->close_section();
+    }
+    f->close_section();
+  }
+  if (!cap_refs.empty()) {
+    f->open_array_section("cap_refs");
+    for (map<int,int>::const_iterator p = cap_refs.begin(); p != cap_refs.end(); ++p) {
+      f->open_object_section("cap_ref");
+      f->dump_stream("cap") << ccap_string(p->first);
+      f->dump_int("refs", p->second);
+      f->close_section();
+    }
+    f->close_section();
+  }
+
+  f->dump_unsigned("reported_size", reported_size);
+  if (wanted_max_size != max_size)
+    f->dump_unsigned("wanted_max_size", wanted_max_size);
+  if (requested_max_size != max_size)
+    f->dump_unsigned("requested_max_size", requested_max_size);
+
+  f->dump_int("nref", get_nref());
+  f->dump_int("ll_ref", ll_ref);
+
+  if (!dentries.empty()) {
+    f->open_array_section("parents");
+    for (const auto &&dn : dentries) {
+      f->open_object_section("dentry");
+      f->dump_stream("dir_ino") << dn->dir->parent_inode->ino;
+      f->dump_string("name", dn->name);
+      f->close_section();
+    }
+    f->close_section();
+  }
+}
+
+void Cap::dump(Formatter *f) const
+{
+  f->dump_int("mds", session->mds_num);
+  f->dump_stream("ino") << inode.ino;
+  f->dump_unsigned("cap_id", cap_id);
+  f->dump_stream("issued") << ccap_string(issued);
+  if (implemented != issued)
+    f->dump_stream("implemented") << ccap_string(implemented);
+  f->dump_stream("wanted") << ccap_string(wanted);
+  f->dump_unsigned("seq", seq);
+  f->dump_unsigned("issue_seq", issue_seq);
+  f->dump_unsigned("mseq", mseq);
+  f->dump_unsigned("gen", gen);
+}
+
+void CapSnap::dump(Formatter *f) const
+{
+  f->dump_stream("ino") << in->ino;
+  f->dump_stream("issued") << ccap_string(issued);
+  f->dump_stream("dirty") << ccap_string(dirty);
+  f->dump_unsigned("size", size);
+  f->dump_stream("ctime") << ctime;
+  f->dump_stream("mtime") << mtime;
+  f->dump_stream("atime") << atime;
+  f->dump_int("time_warp_seq", time_warp_seq);
+  f->dump_stream("mode") << '0' << std::oct << mode << std::dec;
+  f->dump_unsigned("uid", uid);
+  f->dump_unsigned("gid", gid);
+  if (!xattrs.empty()) {
+    f->open_object_section("xattr_lens");
+    for (map<string,bufferptr>::const_iterator p = xattrs.begin(); p != xattrs.end(); ++p)
+      f->dump_int(p->first.c_str(), p->second.length());
+    f->close_section();
+  }
+  f->dump_unsigned("xattr_version", xattr_version);
+  f->dump_int("writing", (int)writing);
+  f->dump_int("dirty_data", (int)dirty_data);
+  f->dump_unsigned("flush_tid", flush_tid);
+}
+
+void Inode::set_async_err(int r)
+{
+  for (const auto &fh : fhs) {
+    fh->async_err = r;
+  }
+}
+
+bool Inode::has_recalled_deleg()
+{
+  if (delegations.empty())
+    return false;
+
+  // Either all delegations are recalled or none are. Just check the first.
+  Delegation& deleg = delegations.front();
+  return deleg.is_recalled();
+}
+
+void Inode::recall_deleg(bool skip_read)
+{
+  if (delegations.empty())
+    return;
+
+  // Issue any recalls
+  for (list<Delegation>::iterator d = delegations.begin();
+       d != delegations.end(); ++d) {
+
+    Delegation& deleg = *d;
+    deleg.recall(skip_read);
+  }
+}
+
+bool Inode::delegations_broken(bool skip_read)
+{
+  if (delegations.empty()) {
+    lsubdout(client->cct, client, 10) <<
+	  __func__ << ": delegations empty on " << *this << dendl;
+    return true;
+  }
+
+  if (skip_read) {
+    Delegation& deleg = delegations.front();
+    lsubdout(client->cct, client, 10) <<
+	__func__ << ": read delegs only on " << *this << dendl;
+    if (deleg.get_type() == CEPH_FILE_MODE_RD) {
+	return true;
+    }
+  }
+  lsubdout(client->cct, client, 10) <<
+	__func__ << ": not broken" << *this << dendl;
+  return false;
+}
+
+void Inode::break_deleg(bool skip_read)
+{
+  lsubdout(client->cct, client, 10) <<
+	  __func__ << ": breaking delegs on " << *this << dendl;
+
+  recall_deleg(skip_read);
+
+  while (!delegations_broken(skip_read))
+    client->wait_on_list(waitfor_deleg);
+}
+
+/**
+ * set_deleg: request a delegation on an open Fh
+ * @fh: filehandle on which to acquire it
+ * @type: delegation request type
+ * @cb: delegation recall callback function
+ * @priv: private pointer to be passed to callback
+ *
+ * Attempt to acquire a delegation on an open file handle. If there are no
+ * conflicts and we have the right caps, allocate a new delegation, fill it
+ * out and return 0. Return an error if we can't get one for any reason.
+ */
+int Inode::set_deleg(Fh *fh, unsigned type, ceph_deleg_cb_t cb, void *priv)
+{
+  lsubdout(client->cct, client, 10) <<
+	  __func__ << ": inode " << *this << dendl;
+
+  /*
+   * 0 deleg timeout means that they haven't been explicitly enabled. Don't
+   * allow it, with an unusual error to make it clear.
+   */
+  if (!client->get_deleg_timeout())
+    return -CEPHFS_ETIME;
+
+  // Just say no if we have any recalled delegs still outstanding
+  if (has_recalled_deleg()) {
+    lsubdout(client->cct, client, 10) << __func__ <<
+	  ": has_recalled_deleg" << dendl;
+    return -CEPHFS_EAGAIN;
+  }
+
+  // check vs. currently open files on this inode
+  switch (type) {
+  case CEPH_DELEGATION_RD:
+    if (open_count_for_write()) {
+      lsubdout(client->cct, client, 10) << __func__ <<
+	    ": open for write" << dendl;
+      return -CEPHFS_EAGAIN;
+    }
+    break;
+  case CEPH_DELEGATION_WR:
+    if (open_count() > 1) {
+      lsubdout(client->cct, client, 10) << __func__ << ": open" << dendl;
+      return -CEPHFS_EAGAIN;
+    }
+    break;
+  default:
+    return -CEPHFS_EINVAL;
+  }
+
+  /*
+   * A delegation is essentially a long-held container for cap references that
+   * we delegate to the client until recalled. The caps required depend on the
+   * type of delegation (read vs. rw). This is entirely an opportunistic thing.
+   * If we don't have the necessary caps for the delegation, then we just don't
+   * grant one.
+   *
+   * In principle we could request the caps from the MDS, but a delegation is
+   * usually requested just after an open. If we don't have the necessary caps
+   * already, then it's likely that there is some sort of conflicting access.
+   *
+   * In the future, we may need to add a way to have this request caps more
+   * aggressively -- for instance, to handle WANT_DELEGATION for NFSv4.1+.
+   */
+  int need = ceph_deleg_caps_for_type(type);
+  if (!caps_issued_mask(need)) {
+    lsubdout(client->cct, client, 10) << __func__ << ": cap mismatch, have="
+      << ccap_string(caps_issued()) << " need=" << ccap_string(need) << dendl;
+    return -CEPHFS_EAGAIN;
+  }
+
+  for (list<Delegation>::iterator d = delegations.begin();
+       d != delegations.end(); ++d) {
+    Delegation& deleg = *d;
+    if (deleg.get_fh() == fh) {
+      deleg.reinit(type, cb, priv);
+      return 0;
+    }
+  }
+
+  delegations.emplace_back(fh, type, cb, priv);
+  return 0;
+}
+
+/**
+ * unset_deleg - remove a delegation that was previously set
+ * @fh: file handle to clear delegation of
+ *
+ * Unlink delegation from the Inode (if there is one), put caps and free it.
+ */
+void Inode::unset_deleg(Fh *fh)
+{
+  for (list<Delegation>::iterator d = delegations.begin();
+       d != delegations.end(); ++d) {
+    Delegation& deleg = *d;
+    if (deleg.get_fh() == fh) {
+      delegations.erase(d);
+      client->signal_cond_list(waitfor_deleg);
+      break;
+    }
+  }
+}
+
+/**
+* mark_caps_dirty - mark some caps dirty
+* @caps: the dirty caps
+*
+* note that if there is no dirty and flushing caps before, we need to pin this inode.
+* it will be unpined by handle_cap_flush_ack when there are no dirty and flushing caps.
+*/
+void Inode::mark_caps_dirty(int caps)
+{
+  lsubdout(client->cct, client, 10) << __func__ << " " << *this << " " << ccap_string(dirty_caps) << " -> "
+           << ccap_string(dirty_caps | caps) << dendl;
+  if (caps && !caps_dirty())
+    iget();
+  dirty_caps |= caps;
+  client->get_dirty_list().push_back(&dirty_cap_item);
+}
+
+/**
+* mark_caps_clean - only clean the dirty_caps and caller should start flushing the dirty caps.
+*/
+void Inode::mark_caps_clean()
+{
+  lsubdout(client->cct, client, 10) << __func__ << " " << *this << dendl;
+  dirty_caps = 0;
+  dirty_cap_item.remove_myself();
+}
+
+
diff --git a/src/client/Inode.h b/src/client/Inode.h
new file mode 100644
index 000000000..1cb6285ba
--- /dev/null
+++ b/src/client/Inode.h
@@ -0,0 +1,361 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#ifndef CEPH_CLIENT_INODE_H
+#define CEPH_CLIENT_INODE_H
+
+#include <numeric>
+
+#include "include/compat.h"
+#include "include/ceph_assert.h"
+#include "include/types.h"
+#include "include/xlist.h"
+
+#include "mds/flock.h"
+#include "mds/mdstypes.h" // hrm
+
+#include "osdc/ObjectCacher.h"
+
+#include "InodeRef.h"
+#include "MetaSession.h"
+#include "UserPerm.h"
+#include "Delegation.h"
+
+class Client;
+class Dentry;
+class Dir;
+struct SnapRealm;
+struct Inode;
+class MetaRequest;
+class filepath;
+class Fh;
+
+class Cap {
+public:
+  Cap() = delete;
+  Cap(Inode &i, MetaSession *s) : inode(i),
+                                  session(s),
+                                  gen(s->cap_gen),
+                                  cap_item(this)
+  {
+    s->caps.push_back(&cap_item);
+  }
+  ~Cap() {
+    cap_item.remove_myself();
+  }
+
+  void touch(void) {
+    // move to back of LRU
+    session->caps.push_back(&cap_item);
+  }
+
+  void dump(Formatter *f) const;
+
+  Inode &inode;
+  MetaSession *session;
+  uint64_t cap_id = 0;
+  unsigned issued = 0;
+  unsigned implemented = 0;
+  unsigned wanted = 0;   // as known to mds.
+  uint64_t seq = 0;
+  uint64_t issue_seq = 0;
+  __u32 mseq = 0;  // migration seq
+  __u32 gen;
+  UserPerm latest_perms;
+
+private:
+  /* Note that this Cap will not move (see Inode::caps):
+   *
+   * Section 23.1.2#8
+   * The insert members shall not affect the validity of iterators and
+   * references to the container, and the erase members shall invalidate only
+   * iterators and references to the erased elements.
+   */
+  xlist<Cap *>::item cap_item;
+};
+
+struct CapSnap {
+  //snapid_t follows;  // map key
+  InodeRef in;
+  SnapContext context;
+  int issued = 0, dirty = 0;
+
+  uint64_t size = 0;
+  utime_t ctime, btime, mtime, atime;
+  version_t time_warp_seq = 0;
+  uint64_t change_attr = 0;
+  uint32_t   mode = 0;
+  uid_t      uid = 0;
+  gid_t      gid = 0;
+  map<string,bufferptr> xattrs;
+  version_t xattr_version = 0;
+
+  bufferlist inline_data;
+  version_t inline_version = 0;
+
+  bool writing = false, dirty_data = false;
+  uint64_t flush_tid = 0;
+
+  int64_t cap_dirtier_uid = -1;
+  int64_t cap_dirtier_gid = -1;
+
+  explicit CapSnap(Inode *i)
+    : in(i)
+  {}
+
+  void dump(Formatter *f) const;
+};
+
+// inode flags
+#define I_COMPLETE		(1 << 0)
+#define I_DIR_ORDERED		(1 << 1)
+#define I_SNAPDIR_OPEN		(1 << 2)
+#define I_KICK_FLUSH		(1 << 3)
+#define I_CAP_DROPPED		(1 << 4)
+#define I_ERROR_FILELOCK	(1 << 5)
+
+struct Inode : RefCountedObject {
+  ceph::coarse_mono_time hold_caps_until;
+  Client *client;
+
+  // -- the actual inode --
+  inodeno_t ino; // ORDER DEPENDENCY: oset
+  snapid_t  snapid;
+  ino_t faked_ino = 0;
+
+  uint32_t   rdev = 0;    // if special file
+
+  // affected by any inode change...
+  utime_t    ctime;   // inode change time
+  utime_t    btime;   // birth time
+
+  // perm (namespace permissions)
+  uint32_t   mode = 0;
+  uid_t      uid = 0;
+  gid_t      gid = 0;
+
+  // nlink
+  int32_t    nlink = 0;
+
+  // file (data access)
+  ceph_dir_layout dir_layout{};
+  file_layout_t layout;
+  uint64_t   size = 0;        // on directory, # dentries
+  uint32_t   truncate_seq = 1;
+  uint64_t   truncate_size = -1;
+  utime_t    mtime;   // file data modify time.
+  utime_t    atime;   // file data access time.
+  uint32_t   time_warp_seq = 0;  // count of (potential) mtime/atime timewarps (i.e., utimes())
+  uint64_t   change_attr = 0;
+
+  uint64_t max_size = 0;  // max size we can write to
+
+  // dirfrag, recursive accountin
+  frag_info_t dirstat;
+  nest_info_t rstat;
+
+  // special stuff
+  version_t version = 0;           // auth only
+  version_t xattr_version = 0;
+  utime_t   snap_btime;        // snapshot creation (birth) time
+  std::map<std::string, std::string> snap_metadata;
+
+  // inline data
+  version_t  inline_version = 0;
+  bufferlist inline_data;
+
+  bool fscrypt = false; // fscrypt enabled ?
+
+  bool is_root()    const { return ino == CEPH_INO_ROOT; }
+  bool is_symlink() const { return (mode & S_IFMT) == S_IFLNK; }
+  bool is_dir()     const { return (mode & S_IFMT) == S_IFDIR; }
+  bool is_file()    const { return (mode & S_IFMT) == S_IFREG; }
+
+  bool has_dir_layout() const {
+    return layout != file_layout_t();
+  }
+
+  __u32 hash_dentry_name(const string &dn) {
+    int which = dir_layout.dl_dir_hash;
+    if (!which)
+      which = CEPH_STR_HASH_LINUX;
+    ceph_assert(ceph_str_hash_valid(which));
+    return ceph_str_hash(which, dn.data(), dn.length());
+  }
+
+  unsigned flags = 0;
+
+  quota_info_t quota;
+
+  bool is_complete_and_ordered() {
+    static const unsigned wants = I_COMPLETE | I_DIR_ORDERED;
+    return (flags & wants) == wants;
+  }
+
+  // about the dir (if this is one!)
+  Dir       *dir = 0;     // if i'm a dir.
+  fragtree_t dirfragtree;
+  uint64_t dir_release_count = 1;
+  uint64_t dir_ordered_count = 1;
+  bool dir_hashed = false;
+  bool dir_replicated = false;
+
+  // per-mds caps
+  std::map<mds_rank_t, Cap> caps;            // mds -> Cap
+  Cap *auth_cap = 0;
+  int64_t cap_dirtier_uid = -1;
+  int64_t cap_dirtier_gid = -1;
+  unsigned dirty_caps = 0;
+  unsigned flushing_caps = 0;
+  std::map<ceph_tid_t, int> flushing_cap_tids;
+  int shared_gen = 0;
+  int cache_gen = 0;
+  int snap_caps = 0;
+  int snap_cap_refs = 0;
+  xlist<Inode*>::item delay_cap_item, dirty_cap_item, flushing_cap_item;
+
+  SnapRealm *snaprealm = 0;
+  xlist<Inode*>::item snaprealm_item;
+  InodeRef snapdir_parent;  // only if we are a snapdir inode
+  map<snapid_t,CapSnap> cap_snaps;   // pending flush to mds
+
+  //int open_by_mode[CEPH_FILE_MODE_NUM];
+  map<int,int> open_by_mode;
+  map<int,int> cap_refs;
+
+  ObjectCacher::ObjectSet oset; // ORDER DEPENDENCY: ino
+
+  uint64_t reported_size = 0;
+  uint64_t wanted_max_size = 0;
+  uint64_t requested_max_size = 0;
+
+  uint64_t  ll_ref = 0;   // separate ref count for ll client
+  xlist<Dentry *> dentries; // if i'm linked to a dentry.
+  string    symlink;  // symlink content, if it's a symlink
+  map<string,bufferptr> xattrs;
+  map<frag_t,int> fragmap;  // known frag -> mds mappings
+  map<frag_t, std::vector<mds_rank_t>> frag_repmap; // non-auth mds mappings
+
+  std::list<ceph::condition_variable*> waitfor_caps;
+  std::list<ceph::condition_variable*> waitfor_commit;
+  std::list<ceph::condition_variable*> waitfor_deleg;
+
+  Dentry *get_first_parent() {
+    ceph_assert(!dentries.empty());
+    return *dentries.begin();
+  }
+
+  void make_long_path(filepath& p);
+  void make_short_path(filepath& p);
+  void make_nosnap_relative_path(filepath& p);
+
+  // The ref count. 1 for each dentry, fh, inode_map,
+  // cwd that links to me.
+  void iget() { get(); }
+  void iput(int n=1) { ceph_assert(n >= 0); while (n--) put(); }
+
+  void ll_get() {
+    ll_ref++;
+  }
+  void ll_put(uint64_t n=1) {
+    ceph_assert(ll_ref >= n);
+    ll_ref -= n;
+  }
+
+  // file locks
+  std::unique_ptr<ceph_lock_state_t> fcntl_locks;
+  std::unique_ptr<ceph_lock_state_t> flock_locks;
+
+  bool has_any_filelocks() {
+    return
+      (fcntl_locks && !fcntl_locks->empty()) ||
+      (flock_locks && !flock_locks->empty());
+  }
+
+  list<Delegation> delegations;
+
+  xlist<MetaRequest*> unsafe_ops;
+
+  std::set<Fh*> fhs;
+
+  mds_rank_t dir_pin = MDS_RANK_NONE;
+
+  Inode() = delete;
+  Inode(Client *c, vinodeno_t vino, file_layout_t *newlayout)
+    : client(c), ino(vino.ino), snapid(vino.snapid), delay_cap_item(this),
+      dirty_cap_item(this), flushing_cap_item(this), snaprealm_item(this),
+      oset((void *)this, newlayout->pool_id, this->ino) {}
+  ~Inode();
+
+  vinodeno_t vino() const { return vinodeno_t(ino, snapid); }
+
+  struct Compare {
+    bool operator() (Inode* const & left, Inode* const & right) {
+      if (left->ino.val < right->ino.val) {
+	return (left->snapid.val < right->snapid.val);
+      }
+      return false;
+    }
+  };
+
+  bool check_mode(const UserPerm& perms, unsigned want);
+
+  // CAPS --------
+  void get_open_ref(int mode);
+  bool put_open_ref(int mode);
+
+  void get_cap_ref(int cap);
+  int put_cap_ref(int cap);
+  bool is_any_caps();
+  bool cap_is_valid(const Cap &cap) const;
+  int caps_issued(int *implemented = 0) const;
+  void try_touch_cap(mds_rank_t mds);
+  bool caps_issued_mask(unsigned mask, bool allow_impl=false);
+  int caps_used();
+  int caps_file_wanted();
+  int caps_wanted();
+  int caps_mds_wanted();
+  int caps_dirty();
+  const UserPerm *get_best_perms();
+
+  bool have_valid_size();
+  Dir *open_dir();
+
+  void add_fh(Fh *f) {fhs.insert(f);}
+  void rm_fh(Fh *f) {fhs.erase(f);}
+  void set_async_err(int r);
+  void dump(Formatter *f) const;
+
+  void break_all_delegs() { break_deleg(false); };
+
+  void recall_deleg(bool skip_read);
+  bool has_recalled_deleg();
+  int set_deleg(Fh *fh, unsigned type, ceph_deleg_cb_t cb, void *priv);
+  void unset_deleg(Fh *fh);
+
+  void mark_caps_dirty(int caps);
+  void mark_caps_clean();
+private:
+  // how many opens for write on this Inode?
+  long open_count_for_write()
+  {
+    return (long)(open_by_mode[CEPH_FILE_MODE_RDWR] +
+		  open_by_mode[CEPH_FILE_MODE_WR]);
+  };
+
+  // how many opens of any sort on this inode?
+  long open_count()
+  {
+    return (long) std::accumulate(open_by_mode.begin(), open_by_mode.end(), 0,
+				  [] (int value, const std::map<int, int>::value_type& p)
+                   { return value + p.second; });
+  };
+
+  void break_deleg(bool skip_read);
+  bool delegations_broken(bool skip_read);
+
+};
+
+ostream& operator<<(ostream &out, const Inode &in);
+
+#endif
diff --git a/src/client/InodeRef.h b/src/client/InodeRef.h
new file mode 100644
index 000000000..822ec0ffc
--- /dev/null
+++ b/src/client/InodeRef.h
@@ -0,0 +1,12 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#ifndef CEPH_CLIENT_INODEREF_H
+#define CEPH_CLIENT_INODEREF_H
+
+#include <boost/intrusive_ptr.hpp>
+class Inode;
+void intrusive_ptr_add_ref(Inode *in);
+void intrusive_ptr_release(Inode *in);
+typedef boost::intrusive_ptr<Inode> InodeRef;
+#endif
diff --git a/src/client/MetaRequest.cc b/src/client/MetaRequest.cc
new file mode 100644
index 000000000..2450a1f68
--- /dev/null
+++ b/src/client/MetaRequest.cc
@@ -0,0 +1,80 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "include/types.h"
+#include "client/MetaRequest.h"
+#include "client/Dentry.h"
+#include "client/Inode.h"
+#include "messages/MClientReply.h"
+#include "common/Formatter.h"
+
+void MetaRequest::dump(Formatter *f) const
+{
+  auto age = std::chrono::duration<double>(ceph_clock_now() - op_stamp);
+
+  f->dump_unsigned("tid", tid);
+  f->dump_string("op", ceph_mds_op_name(head.op));
+  f->dump_stream("path") << path;
+  f->dump_stream("path2") << path2;
+  if (_inode)
+    f->dump_stream("ino") << _inode->ino;
+  if (_old_inode)
+    f->dump_stream("old_ino") << _old_inode->ino;
+  if (_other_inode)
+    f->dump_stream("other_ino") << _other_inode->ino;
+  if (target)
+    f->dump_stream("target_ino") << target->ino;
+  if (_dentry)
+    f->dump_string("dentry", _dentry->name);
+  if (_old_dentry)
+    f->dump_string("old_dentry", _old_dentry->name);
+  f->dump_stream("hint_ino") << inodeno_t(head.ino);
+
+  f->dump_stream("sent_stamp") << sent_stamp;
+  f->dump_float("age", age.count());
+  f->dump_int("mds", mds);
+  f->dump_int("resend_mds", resend_mds);
+  f->dump_int("send_to_auth", send_to_auth);
+  f->dump_unsigned("sent_on_mseq", sent_on_mseq);
+  f->dump_int("retry_attempt", retry_attempt);
+
+  f->dump_int("got_unsafe", got_unsafe);
+
+  f->dump_unsigned("uid", head.caller_uid);
+  f->dump_unsigned("gid", head.caller_gid);
+
+  f->dump_unsigned("oldest_client_tid", head.oldest_client_tid);
+  f->dump_unsigned("mdsmap_epoch", head.mdsmap_epoch);
+  f->dump_unsigned("flags", head.flags);
+  f->dump_unsigned("num_retry", head.num_retry);
+  f->dump_unsigned("num_fwd", head.num_fwd);
+  f->dump_unsigned("num_releases", head.num_releases);
+
+  f->dump_int("abort_rc", abort_rc);
+}
+
+MetaRequest::~MetaRequest()
+{
+  if (_dentry)
+    _dentry->put();
+  if (_old_dentry)
+    _old_dentry->put();
+}
+
+void MetaRequest::set_dentry(Dentry *d) {
+  ceph_assert(_dentry == NULL);
+  _dentry = d;
+  _dentry->get();
+}
+Dentry *MetaRequest::dentry() {
+  return _dentry;
+}
+
+void MetaRequest::set_old_dentry(Dentry *d) {
+  ceph_assert(_old_dentry == NULL);
+  _old_dentry = d;
+  _old_dentry->get();
+}
+Dentry *MetaRequest::old_dentry() {
+  return _old_dentry;
+}
diff --git a/src/client/MetaRequest.h b/src/client/MetaRequest.h
new file mode 100644
index 000000000..03eae6522
--- /dev/null
+++ b/src/client/MetaRequest.h
@@ -0,0 +1,248 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#ifndef CEPH_CLIENT_METAREQUEST_H
+#define CEPH_CLIENT_METAREQUEST_H
+
+
+#include "include/types.h"
+#include "include/xlist.h"
+#include "include/filepath.h"
+#include "mds/mdstypes.h"
+#include "InodeRef.h"
+#include "UserPerm.h"
+
+#include "messages/MClientRequest.h"
+#include "messages/MClientReply.h"
+
+class Dentry;
+class dir_result_t;
+
+struct MetaRequest {
+private:
+  InodeRef _inode, _old_inode, _other_inode;
+  Dentry *_dentry; //associated with path
+  Dentry *_old_dentry; //associated with path2
+  int abort_rc;
+public:
+  ceph::coarse_mono_time created = ceph::coarse_mono_clock::zero();
+  uint64_t tid;
+  utime_t  op_stamp;
+  ceph_mds_request_head head;
+  filepath path, path2;
+  std::string alternate_name;
+  bufferlist data;
+  int inode_drop; //the inode caps this operation will drop
+  int inode_unless; //unless we have these caps already
+  int old_inode_drop, old_inode_unless;
+  int dentry_drop, dentry_unless;
+  int old_dentry_drop, old_dentry_unless;
+  int other_inode_drop, other_inode_unless;
+  vector<MClientRequest::Release> cap_releases;
+
+  int regetattr_mask;          // getattr mask if i need to re-stat after a traceless reply
+ 
+  utime_t  sent_stamp;
+  mds_rank_t mds;                // who i am asking
+  mds_rank_t resend_mds;         // someone wants you to (re)send the request here
+  bool     send_to_auth;       // must send to auth mds
+  __u32    sent_on_mseq;       // mseq at last submission of this request
+  int      num_fwd;            // # of times i've been forwarded
+  int      retry_attempt;
+  std::atomic<uint64_t> ref = { 1 };
+  
+  ceph::cref_t<MClientReply> reply;         // the reply
+  bool kick;
+  bool success;
+  
+  // readdir result
+  dir_result_t *dirp;
+
+  //possible responses
+  bool got_unsafe;
+
+  xlist<MetaRequest*>::item item;
+  xlist<MetaRequest*>::item unsafe_item;
+  xlist<MetaRequest*>::item unsafe_dir_item;
+  xlist<MetaRequest*>::item unsafe_target_item;
+
+  ceph::condition_variable *caller_cond;          // who to take up
+  ceph::condition_variable *dispatch_cond;        // who to kick back
+  list<ceph::condition_variable*> waitfor_safe;
+
+  InodeRef target;
+  UserPerm perms;
+
+  explicit MetaRequest(int op) :
+    _dentry(NULL), _old_dentry(NULL), abort_rc(0),
+    tid(0),
+    inode_drop(0), inode_unless(0),
+    old_inode_drop(0), old_inode_unless(0),
+    dentry_drop(0), dentry_unless(0),
+    old_dentry_drop(0), old_dentry_unless(0),
+    other_inode_drop(0), other_inode_unless(0),
+    regetattr_mask(0),
+    mds(-1), resend_mds(-1), send_to_auth(false), sent_on_mseq(0),
+    num_fwd(0), retry_attempt(0),
+    reply(0),
+    kick(false), success(false), dirp(NULL),
+    got_unsafe(false), item(this), unsafe_item(this),
+    unsafe_dir_item(this), unsafe_target_item(this),
+    caller_cond(0), dispatch_cond(0) {
+    memset(&head, 0, sizeof(head));
+    head.op = op;
+  }
+  ~MetaRequest();
+
+  /**
+   * Prematurely terminate the request, such that callers
+   * to make_request will receive `rc` as their result.
+   */
+  void abort(int rc)
+  {
+    ceph_assert(rc != 0);
+    abort_rc = rc;
+  }
+
+  /**
+   * Whether abort() has been called for this request
+   */
+  inline bool aborted() const
+  {
+    return abort_rc != 0;
+  }
+
+  /**
+   * Given that abort() has been called for this request, what `rc` was
+   * passed into it?
+   */
+  int get_abort_code() const
+  {
+    return abort_rc;
+  }
+
+  void set_inode(Inode *in) {
+    _inode = in;
+  }
+  Inode *inode() {
+    return _inode.get();
+  }
+  void take_inode(InodeRef *out) {
+    out->swap(_inode);
+  }
+  void set_old_inode(Inode *in) {
+    _old_inode = in;
+  }
+  Inode *old_inode() {
+    return _old_inode.get();
+  }
+  void take_old_inode(InodeRef *out) {
+    out->swap(_old_inode);
+  }
+  void set_other_inode(Inode *in) {
+    _other_inode = in;
+  }
+  Inode *other_inode() {
+    return _other_inode.get();
+  }
+  void take_other_inode(InodeRef *out) {
+    out->swap(_other_inode);
+  }
+  void set_dentry(Dentry *d);
+  Dentry *dentry();
+  void set_old_dentry(Dentry *d);
+  Dentry *old_dentry();
+
+  MetaRequest* get() {
+    ref++;
+    return this;
+  }
+
+  /// psuedo-private put method; use Client::put_request()
+  bool _put() {
+    int v = --ref;
+    return v == 0;
+  }
+
+  // normal fields
+  void set_tid(ceph_tid_t t) { tid = t; }
+  void set_oldest_client_tid(ceph_tid_t t) { head.oldest_client_tid = t; }
+  void inc_num_fwd() { head.num_fwd = head.num_fwd + 1; }
+  void set_retry_attempt(int a) { head.num_retry = a; }
+  void set_filepath(const filepath& fp) { path = fp; }
+  void set_filepath2(const filepath& fp) { path2 = fp; }
+  void set_alternate_name(std::string an) { alternate_name = an; }
+  void set_string2(const char *s) { path2.set_path(std::string_view(s), 0); }
+  void set_caller_perms(const UserPerm& _perms) {
+    perms.shallow_copy(_perms);
+    head.caller_uid = perms.uid();
+    head.caller_gid = perms.gid();
+  }
+  uid_t get_uid() { return perms.uid(); }
+  uid_t get_gid() { return perms.gid(); }
+  void set_data(const bufferlist &d) { data = d; }
+  void set_dentry_wanted() {
+    head.flags = head.flags | CEPH_MDS_FLAG_WANT_DENTRY;
+  }
+  int get_op() { return head.op; }
+  ceph_tid_t get_tid() { return tid; }
+  filepath& get_filepath() { return path; }
+  filepath& get_filepath2() { return path2; }
+
+  bool is_write() {
+    return
+      (head.op & CEPH_MDS_OP_WRITE) || 
+      (head.op == CEPH_MDS_OP_OPEN && (head.args.open.flags & (O_CREAT|O_TRUNC)));
+  }
+  bool can_forward() {
+    if ((head.op & CEPH_MDS_OP_WRITE) ||
+	head.op == CEPH_MDS_OP_OPEN)   // do not forward _any_ open request.
+      return false;
+    return true;
+  }
+  bool auth_is_best(int issued) {
+    if (send_to_auth)
+      return true;
+
+    /* Any write op ? */
+    if (head.op & CEPH_MDS_OP_WRITE)
+      return true;
+
+    switch (head.op) {
+    case CEPH_MDS_OP_OPEN:
+    case CEPH_MDS_OP_READDIR:
+      return true;
+    case CEPH_MDS_OP_GETATTR:
+      /*
+       * If any 'x' caps is issued we can just choose the auth MDS
+       * instead of the random replica MDSes. Because only when the
+       * Locker is in LOCK_EXEC state will the loner client could
+       * get the 'x' caps. And if we send the getattr requests to
+       * any replica MDS it must auth pin and tries to rdlock from
+       * the auth MDS, and then the auth MDS need to do the Locker
+       * state transition to LOCK_SYNC. And after that the lock state
+       * will change back.
+       *
+       * This cost much when doing the Locker state transition and
+       * usually will need to revoke caps from clients.
+       *
+       * And for the 'Xs' caps for getxattr we will also choose the
+       * auth MDS, because the MDS side code is buggy due to setxattr
+       * won't notify the replica MDSes when the values changed and
+       * the replica MDS will return the old values. Though we will
+       * fix it in MDS code, but this still makes sense for old ceph.
+       */
+      if (((head.args.getattr.mask & CEPH_CAP_ANY_SHARED) &&
+	   (issued & CEPH_CAP_ANY_EXCL)) ||
+          (head.args.getattr.mask & (CEPH_STAT_RSTAT | CEPH_STAT_CAP_XATTR)))
+        return true;
+    default:
+      return false;
+    }
+  }
+
+  void dump(Formatter *f) const;
+
+};
+
+#endif
diff --git a/src/client/MetaSession.cc b/src/client/MetaSession.cc
new file mode 100644
index 000000000..b5160a843
--- /dev/null
+++ b/src/client/MetaSession.cc
@@ -0,0 +1,62 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#include "include/types.h"
+#include "messages/MClientCapRelease.h"
+
+#include "MetaSession.h"
+#include "Inode.h"
+
+#include "common/Formatter.h"
+
+const char *MetaSession::get_state_name() const
+{
+  switch (state) {
+  case STATE_NEW: return "new";
+  case STATE_OPENING: return "opening";
+  case STATE_OPEN: return "open";
+  case STATE_CLOSING: return "closing";
+  case STATE_CLOSED: return "closed";
+  case STATE_STALE: return "stale";
+  default: return "unknown";
+  }
+}
+
+void MetaSession::dump(Formatter *f, bool cap_dump) const
+{
+  f->dump_int("mds", mds_num);
+  f->dump_object("addrs", addrs);
+  f->dump_unsigned("seq", seq);
+  f->dump_unsigned("cap_gen", cap_gen);
+  f->dump_stream("cap_ttl") << cap_ttl;
+  f->dump_stream("last_cap_renew_request") << last_cap_renew_request;
+  f->dump_unsigned("cap_renew_seq", cap_renew_seq);
+  f->dump_int("num_caps", caps.size());
+  if (cap_dump) {
+    f->open_array_section("caps");
+    for (const auto& cap : caps) {
+      f->dump_object("cap", *cap);
+    }
+    f->close_section();
+  }
+  f->dump_string("state", get_state_name());
+}
+
+void MetaSession::enqueue_cap_release(inodeno_t ino, uint64_t cap_id, ceph_seq_t iseq,
+    ceph_seq_t mseq, epoch_t osd_barrier)
+{
+  if (!release) {
+    release = ceph::make_message<MClientCapRelease>();
+  }
+
+  if (osd_barrier > release->osd_epoch_barrier) {
+    release->osd_epoch_barrier = osd_barrier;
+  }
+
+  ceph_mds_cap_item i;
+  i.ino = ino;
+  i.cap_id = cap_id;
+  i.seq = iseq;
+  i.migrate_seq = mseq;
+  release->caps.push_back(i);
+}
diff --git a/src/client/MetaSession.h b/src/client/MetaSession.h
new file mode 100644
index 000000000..d9a1e481a
--- /dev/null
+++ b/src/client/MetaSession.h
@@ -0,0 +1,72 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+
+#ifndef CEPH_CLIENT_METASESSION_H
+#define CEPH_CLIENT_METASESSION_H
+
+#include "include/types.h"
+#include "include/utime.h"
+#include "include/xlist.h"
+#include "mds/MDSMap.h"
+#include "mds/mdstypes.h"
+#include "messages/MClientCapRelease.h"
+
+struct Cap;
+struct Inode;
+struct CapSnap;
+struct MetaRequest;
+
+struct MetaSession {
+  mds_rank_t mds_num;
+  ConnectionRef con;
+  version_t seq = 0;
+  uint64_t cap_gen = 0;
+  utime_t cap_ttl, last_cap_renew_request;
+  uint64_t cap_renew_seq = 0;
+  entity_addrvec_t addrs;
+  feature_bitset_t mds_features;
+  feature_bitset_t mds_metric_flags;
+
+  enum {
+    STATE_NEW, // Unused
+    STATE_OPENING,
+    STATE_OPEN,
+    STATE_CLOSING,
+    STATE_CLOSED,
+    STATE_STALE,
+    STATE_REJECTED,
+  } state = STATE_OPENING;
+
+  enum {
+    RECLAIM_NULL,
+    RECLAIMING,
+    RECLAIM_OK,
+    RECLAIM_FAIL,
+  } reclaim_state = RECLAIM_NULL;
+
+  int mds_state = MDSMap::STATE_NULL;
+  bool readonly = false;
+
+  list<Context*> waiting_for_open;
+
+  xlist<Cap*> caps;
+  xlist<Inode*> flushing_caps;
+  xlist<MetaRequest*> requests;
+  xlist<MetaRequest*> unsafe_requests;
+  std::set<ceph_tid_t> flushing_caps_tids;
+
+  ceph::ref_t<MClientCapRelease> release;
+
+  MetaSession(mds_rank_t mds_num, ConnectionRef con, const entity_addrvec_t& addrs)
+    : mds_num(mds_num), con(con), addrs(addrs) {
+  }
+
+  const char *get_state_name() const;
+
+  void dump(Formatter *f, bool cap_dump=false) const;
+
+  void enqueue_cap_release(inodeno_t ino, uint64_t cap_id, ceph_seq_t iseq,
+      ceph_seq_t mseq, epoch_t osd_barrier);
+};
+
+#endif
diff --git a/src/client/ObjecterWriteback.h b/src/client/ObjecterWriteback.h
new file mode 100644
index 000000000..7a159e80e
--- /dev/null
+++ b/src/client/ObjecterWriteback.h
@@ -0,0 +1,72 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+#ifndef CEPH_OSDC_OBJECTERWRITEBACKHANDLER_H
+#define CEPH_OSDC_OBJECTERWRITEBACKHANDLER_H
+
+#include "osdc/Objecter.h"
+#include "osdc/WritebackHandler.h"
+
+class ObjecterWriteback : public WritebackHandler {
+ public:
+  ObjecterWriteback(Objecter *o, Finisher *fin, ceph::mutex *lock)
+    : m_objecter(o),
+      m_finisher(fin),
+      m_lock(lock) { }
+  ~ObjecterWriteback() override {}
+
+  void read(const object_t& oid, uint64_t object_no,
+		    const object_locator_t& oloc, uint64_t off, uint64_t len,
+		    snapid_t snapid, bufferlist *pbl, uint64_t trunc_size,
+		    __u32 trunc_seq, int op_flags,
+                    const ZTracer::Trace &parent_trace,
+                    Context *onfinish) override {
+    m_objecter->read_trunc(oid, oloc, off, len, snapid, pbl, 0,
+			   trunc_size, trunc_seq,
+			   new C_OnFinisher(new C_Lock(m_lock, onfinish),
+					    m_finisher));
+  }
+
+  bool may_copy_on_write(const object_t& oid, uint64_t read_off,
+				 uint64_t read_len, snapid_t snapid) override {
+    return false;
+  }
+
+  ceph_tid_t write(const object_t& oid, const object_locator_t& oloc,
+			   uint64_t off, uint64_t len,
+			   const SnapContext& snapc, const bufferlist &bl,
+			   ceph::real_time mtime, uint64_t trunc_size,
+			   __u32 trunc_seq, ceph_tid_t journal_tid,
+                           const ZTracer::Trace &parent_trace,
+			   Context *oncommit) override {
+    return m_objecter->write_trunc(oid, oloc, off, len, snapc, bl, mtime, 0,
+				   trunc_size, trunc_seq,
+				   new C_OnFinisher(new C_Lock(m_lock,
+							       oncommit),
+						    m_finisher));
+  }
+
+  bool can_scattered_write() override { return true; }
+  using WritebackHandler::write;
+  ceph_tid_t write(const object_t& oid, const object_locator_t& oloc,
+                           vector<pair<uint64_t, bufferlist> >& io_vec,
+			   const SnapContext& snapc, ceph::real_time mtime,
+			   uint64_t trunc_size, __u32 trunc_seq,
+			   Context *oncommit) override {
+    ObjectOperation op;
+    for (vector<pair<uint64_t, bufferlist> >::iterator p = io_vec.begin();
+	 p != io_vec.end();
+	 ++p)
+      op.write(p->first, p->second, trunc_size, trunc_seq);
+
+    return m_objecter->mutate(oid, oloc, op, snapc, mtime, 0,
+			      new C_OnFinisher(new C_Lock(m_lock, oncommit),
+					       m_finisher));
+  }
+
+ private:
+  Objecter *m_objecter;
+  Finisher *m_finisher;
+  ceph::mutex *m_lock;
+};
+
+#endif
diff --git a/src/client/RWRef.h b/src/client/RWRef.h
new file mode 100644
index 000000000..9035a0937
--- /dev/null
+++ b/src/client/RWRef.h
@@ -0,0 +1,244 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2020 Red Hat, Inc.
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ * ============
+ *
+ * This is a common read/write reference framework, which will work
+ * simliarly to a RW lock, the difference here is that for the "readers"
+ * they won't hold any lock but will increase a reference instead when
+ * the "require" state is matched, or set a flag to tell the callers
+ * that the "require" state is not matched and also there is no any
+ * wait mechanism for "readers" to wait the state until it matches. It
+ * will let the callers determine what to do next.
+ *
+ * The usage, such as in libcephfs's client/Client.cc case:
+ *
+ * The Readers:
+ *
+ *   For the ll_read()/ll_write(), etc fucntions, they will work as
+ *   "readers", in the beginning they just need to define a RWRef
+ *   object and in RWRef constructor it will check if the state is
+ *   MOUNTED or MOUTING, if not it will fail and return directly with
+ *   doing nothing, or it will increase the reference and continue.
+ *   And when destructing the RWRef object, in the RWRef destructor
+ *   it will decrease the reference and notify the "writers" who maybe
+ *   waiting.
+ *
+ * The Writers:
+ *
+ *   And for the _unmount() function , as a "writer", in the beginning
+ *   it will also just need to define a RWRef object and in RWRef
+ *   constructor it will update the state to next stage first, which then
+ *   will fail all the new comming "readers", and then wait for all the
+ *   "readers" to finish.
+ *
+ * With this we can get rid of the locks for all the "readers" and they
+ * can run in parallel. And we won't have any potential deadlock issue
+ * with RWRef, such as:
+ *
+ * With RWLock:
+ *
+ *     ThreadA:                           ThreadB:
+ *
+ *     write_lock<RWLock1>.lock();        another_lock.lock();
+ *     state = NEXT_STATE;                ...
+ *     another_lock.lock();               read_lock<RWLock1>.lock();
+ *     ...                                if (state == STATE) {
+ *                                          ...
+ *                                        }
+ *                                        ...
+ *
+ * With RWRef:
+ *
+ *     ThreadA:                           ThreadB:
+ *
+ *     w = RWRef(myS, NEXT_STATE, false); another_lock.lock();
+ *     another_lock.lock();               r = RWRef(myS, STATE);
+ *     ...                                if (r.is_state_satisfied()) {
+ *                                          ...
+ *                                        }
+ *                                        ...
+ *
+ * And also in ThreadA, if it needs to do the cond.wait(&another_lock),
+ * it will goto sleep by holding the write_lock<RWLock1> for the RWLock
+ * case, if the ThreadBs are for some IOs, they may stuck for a very long
+ * time that may get timedout in the uplayer which may keep retrying.
+ * With the RWRef, the ThreadB will fail or continue directly without any
+ * stuck, and the uplayer will knew what next to do quickly.
+ */
+
+#ifndef CEPH_RWRef_Posix__H
+#define CEPH_RWRef_Posix__H
+
+#include <string>
+#include "include/ceph_assert.h"
+#include "common/ceph_mutex.h"
+
+/* The status mechanism info */
+template<typename T>
+struct RWRefState {
+  public:
+    template <typename T1> friend class RWRef;
+
+    /*
+     * This will be status mechanism. Currently you need to define
+     * it by yourself.
+     */
+    T state;
+
+    /*
+     * User defined method to check whether the "require" state
+     * is in the proper range we need.
+     *
+     * For example for the client/Client.cc:
+     * In some reader operation cases we need to make sure the
+     * client state is in mounting or mounted states, then it
+     * will set the "require = mounting" in class RWRef's constructor.
+     * Then the check_reader_state() should return truth if the
+     * state is already in mouting or mounted state.
+     */
+    virtual int check_reader_state(T require) const = 0;
+
+    /*
+     * User defined method to check whether the "require" state
+     * is in the proper range we need.
+     *
+     * This will usually be the state migration check.
+     */
+    virtual int check_writer_state(T require) const = 0;
+
+    /*
+     * User defined method to check whether the "require"
+     * state is valid or not.
+     */
+    virtual bool is_valid_state(T require) const = 0;
+
+    int64_t get_state() const {
+      std::scoped_lock l{lock};
+      return state;
+    }
+
+    bool check_current_state(T require) const {
+      ceph_assert(is_valid_state(require));
+
+      std::scoped_lock l{lock};
+      return state == require;
+    }
+
+    RWRefState(T init_state, const char *lockname, uint64_t _reader_cnt=0)
+      : state(init_state), lock(ceph::make_mutex(lockname)), reader_cnt(_reader_cnt) {}
+    virtual ~RWRefState() {}
+
+  private:
+    mutable ceph::mutex lock;
+    ceph::condition_variable cond;
+    uint64_t reader_cnt = 0;
+};
+
+template<typename T>
+class RWRef {
+public:
+  RWRef(const RWRef& other) = delete;
+  const RWRef& operator=(const RWRef& other) = delete;
+
+  RWRef(RWRefState<T> &s, T require, bool ir=true)
+    :S(s), is_reader(ir) {
+    ceph_assert(S.is_valid_state(require));
+
+    std::scoped_lock l{S.lock};
+    if (likely(is_reader)) { // Readers will update the reader_cnt
+      if (S.check_reader_state(require)) {
+        S.reader_cnt++;
+        satisfied = true;
+      }
+    } else { // Writers will update the state
+      is_reader = false;
+
+      /*
+       * If the current state is not the same as "require"
+       * then update the state and we are the first writer.
+       *
+       * Or if there already has one writer running or
+       * finished, it will let user to choose to continue
+       * or just break.
+       */
+      if (S.check_writer_state(require)) {
+        first_writer = true;
+        S.state = require;
+      }
+      satisfied = true;
+    }
+  }
+
+  /*
+   * Whether the "require" state is in the proper range of
+   * the states.
+   */
+  bool is_state_satisfied() const {
+    return satisfied;
+  }
+
+  /*
+   * Update the state, and only the writer could do the update.
+   */
+  void update_state(T new_state) {
+    ceph_assert(!is_reader);
+    ceph_assert(S.is_valid_state(new_state));
+
+    std::scoped_lock l{S.lock};
+    S.state = new_state;
+  }
+
+  /*
+   * For current state whether we are the first writer or not
+   */
+  bool is_first_writer() const {
+    return first_writer;
+  }
+
+  /*
+   * Will wait for all the in-flight "readers" to finish
+   */
+  void wait_readers_done() {
+    // Only writers can wait
+    ceph_assert(!is_reader);
+
+    std::unique_lock l{S.lock};
+
+    S.cond.wait(l, [this] {
+      return !S.reader_cnt;
+    });
+  }
+
+  ~RWRef() {
+    std::scoped_lock l{S.lock};
+    if (!is_reader)
+      return;
+
+    if (!satisfied)
+      return;
+
+    /*
+     * Decrease the refcnt and notify the waiters
+     */
+    if (--S.reader_cnt == 0)
+      S.cond.notify_all();
+  }
+
+private:
+  RWRefState<T> &S;
+  bool satisfied = false;
+  bool first_writer = false;
+  bool is_reader = true;
+};
+
+#endif // !CEPH_RWRef_Posix__H
diff --git a/src/client/SyntheticClient.cc b/src/client/SyntheticClient.cc
new file mode 100644
index 000000000..9ef5ccf19
--- /dev/null
+++ b/src/client/SyntheticClient.cc
@@ -0,0 +1,3437 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+#include "include/compat.h"
+
+#include <iostream>
+#include <sstream>
+
+
+#include "common/config.h"
+#include "SyntheticClient.h"
+#include "osdc/Objecter.h"
+#include "osdc/Filer.h"
+
+
+#include "include/filepath.h"
+#include "common/perf_counters.h"
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <utime.h>
+#include <math.h>
+#include <sys/statvfs.h>
+
+#include "common/errno.h"
+#include "include/ceph_assert.h"
+#include "include/cephfs/ceph_ll_client.h"
+
+#define dout_context g_ceph_context
+#define dout_subsys ceph_subsys_client
+#undef dout_prefix
+#define dout_prefix *_dout << "client." << (whoami >= 0 ? whoami:client->get_nodeid()) << " "
+
+// traces
+//void trace_include(SyntheticClient *syn, Client *cl, string& prefix);
+//void trace_openssh(SyntheticClient *syn, Client *cl, string& prefix);
+
+int num_client = 1;
+list<int> syn_modes;
+list<int> syn_iargs;
+list<string> syn_sargs;
+int syn_filer_flags = 0;
+
+void parse_syn_options(vector<const char*>& args)
+{
+  vector<const char*> nargs;
+
+  for (unsigned i=0; i<args.size(); i++) {
+    if (strcmp(args[i],"--num-client") == 0) {
+      num_client = atoi(args[++i]);
+      continue;
+    }
+    if (strcmp(args[i],"--syn") == 0) {
+      ++i;
+
+      if (strcmp(args[i], "mksnap") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_MKSNAP);
+	syn_sargs.push_back(args[++i]); // path
+	syn_sargs.push_back(args[++i]); // name
+      }
+      else if (strcmp(args[i], "rmsnap") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_RMSNAP);
+	syn_sargs.push_back(args[++i]); // path
+	syn_sargs.push_back(args[++i]); // name
+      } else if (strcmp(args[i], "mksnapfile") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_MKSNAPFILE);
+	syn_sargs.push_back(args[++i]); // path
+      } else if (strcmp(args[i],"rmfile") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_RMFILE );
+      } else if (strcmp(args[i],"writefile") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_WRITEFILE );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"wrshared") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_WRSHARED );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"writebatch") == 0) {
+          syn_modes.push_back( SYNCLIENT_MODE_WRITEBATCH );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"readfile") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_READFILE );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"readwriterandom") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_RDWRRANDOM );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"readwriterandom_ex") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_RDWRRANDOM_EX );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"overloadosd0") == 0) {
+	syn_modes.push_back( SYNCLIENT_MODE_OVERLOAD_OSD_0 );
+	syn_iargs.push_back( atoi(args[++i]) );
+	syn_iargs.push_back( atoi(args[++i]) );
+	syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"readshared") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_READSHARED );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"rw") == 0) {
+        int a = atoi(args[++i]);
+        int b = atoi(args[++i]);
+        syn_modes.push_back( SYNCLIENT_MODE_WRITEFILE );
+        syn_iargs.push_back( a );
+        syn_iargs.push_back( b );
+        syn_modes.push_back( SYNCLIENT_MODE_READFILE );
+        syn_iargs.push_back( a );
+        syn_iargs.push_back( b );
+      } else if (strcmp(args[i],"dumpplacement") == 0) {
+	syn_modes.push_back( SYNCLIENT_MODE_DUMP );
+	syn_sargs.push_back( args[++i] );   
+      } else if (strcmp(args[i],"dropcache") == 0) {
+	syn_modes.push_back( SYNCLIENT_MODE_DROPCACHE );
+      } else if (strcmp(args[i],"makedirs") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_MAKEDIRS );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"makedirmess") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_MAKEDIRMESS );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"statdirs") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_STATDIRS );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"readdirs") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_READDIRS );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"makefiles") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_MAKEFILES );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"makefiles2") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_MAKEFILES2 );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"linktest") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_LINKTEST );
+      } else if (strcmp(args[i],"createshared") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_CREATESHARED );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"openshared") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_OPENSHARED );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"createobjects") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_CREATEOBJECTS );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"objectrw") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_OBJECTRW );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"walk") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_FULLWALK );
+        //syn_sargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"randomwalk") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_RANDOMWALK );
+        syn_iargs.push_back( atoi(args[++i]) );       
+      } else if (strcmp(args[i],"trace") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_TRACE );
+        syn_sargs.push_back( args[++i] );
+        syn_iargs.push_back( atoi(args[++i]) );
+	syn_iargs.push_back(1);// data
+      } else if (strcmp(args[i],"mtrace") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_TRACE );
+        syn_sargs.push_back( args[++i] );
+        syn_iargs.push_back( atoi(args[++i]) );
+	syn_iargs.push_back(0);// no data
+      } else if (strcmp(args[i],"thrashlinks") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_THRASHLINKS );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"foo") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_FOO );
+      } else if (strcmp(args[i],"until") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_UNTIL );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"sleepuntil") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_SLEEPUNTIL );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"only") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_ONLY );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"onlyrange") == 0) {
+        syn_modes.push_back( SYNCLIENT_MODE_ONLYRANGE );
+        syn_iargs.push_back( atoi(args[++i]) );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"sleep") == 0) { 
+        syn_modes.push_back( SYNCLIENT_MODE_SLEEP );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"randomsleep") == 0) { 
+        syn_modes.push_back( SYNCLIENT_MODE_RANDOMSLEEP );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"opentest") == 0) { 
+        syn_modes.push_back( SYNCLIENT_MODE_OPENTEST );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"optest") == 0) {
+	syn_modes.push_back( SYNCLIENT_MODE_OPTEST );
+        syn_iargs.push_back( atoi(args[++i]) );
+      } else if (strcmp(args[i],"truncate") == 0) { 
+        syn_modes.push_back( SYNCLIENT_MODE_TRUNCATE );
+	syn_sargs.push_back(args[++i]);
+        syn_iargs.push_back(atoi(args[++i]));
+      } else if (strcmp(args[i],"importfind") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_IMPORTFIND);
+	syn_sargs.push_back(args[++i]);
+	syn_sargs.push_back(args[++i]);
+	syn_iargs.push_back(atoi(args[++i]));
+      } else if (strcmp(args[i], "lookuphash") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_LOOKUPHASH);
+	syn_sargs.push_back(args[++i]);
+	syn_sargs.push_back(args[++i]);
+	syn_sargs.push_back(args[++i]);
+      } else if (strcmp(args[i], "lookupino") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_LOOKUPINO);
+	syn_sargs.push_back(args[++i]);
+      } else if (strcmp(args[i], "chunkfile") == 0) {
+	syn_modes.push_back(SYNCLIENT_MODE_CHUNK);
+	syn_sargs.push_back(args[++i]);
+      } else {
+        cerr << "unknown syn arg " << args[i] << std::endl;
+        ceph_abort();
+      }
+    }
+    else if (strcmp(args[i], "localize_reads") == 0) {
+      cerr << "set CEPH_OSD_FLAG_LOCALIZE_READS" << std::endl;
+      syn_filer_flags |= CEPH_OSD_FLAG_LOCALIZE_READS;
+    }
+    else {
+      nargs.push_back(args[i]);
+    }
+  }
+
+  args = nargs;
+}
+
+
+SyntheticClient::SyntheticClient(StandaloneClient *client, int w)
+{
+  this->client = client;
+  whoami = w;
+  thread_id = 0;
+  
+  did_readdir = false;
+
+  run_only = -1;
+  exclude = -1;
+
+  this->modes = syn_modes;
+  this->iargs = syn_iargs;
+  this->sargs = syn_sargs;
+
+  run_start = ceph_clock_now();
+}
+
+
+
+
+#define DBL 2
+
+void *synthetic_client_thread_entry(void *ptr)
+{
+  SyntheticClient *sc = static_cast<SyntheticClient*>(ptr);
+  //int r = 
+  sc->run();
+  return 0;//(void*)r;
+}
+
+string SyntheticClient::get_sarg(int seq) 
+{
+  string a;
+  if (!sargs.empty()) {
+    a = sargs.front(); 
+    sargs.pop_front();
+  }
+  if (a.length() == 0 || a == "~") {
+    char s[30];
+    snprintf(s, sizeof(s), "syn.%lld.%d", (long long)client->whoami.v, seq);
+    a = s;
+  } 
+  return a;
+}
+
+int SyntheticClient::run()
+{
+  UserPerm perms = client->pick_my_perms();
+  dout(15) << "initing" << dendl;
+  int err = client->init();
+  if (err < 0) {
+    dout(0) << "failed to initialize: " << cpp_strerror(err) << dendl;
+    return -1;
+  }
+
+  dout(15) << "mounting" << dendl;
+  err = client->mount("", perms);
+  if (err < 0) {
+    dout(0) << "failed to mount: " << cpp_strerror(err) << dendl;
+    client->shutdown();
+    return -1;
+  }
+
+  //run_start = ceph_clock_now(client->cct);
+  run_until = utime_t(0,0);
+  dout(5) << "run" << dendl;
+
+  int seq = 0;
+
+  for (list<int>::iterator it = modes.begin();
+       it != modes.end();
+       ++it) {
+    int mode = *it;
+    dout(3) << "mode " << mode << dendl;
+
+    switch (mode) {
+
+
+      // WHO?
+
+    case SYNCLIENT_MODE_ONLY:
+      {
+        run_only = iargs.front();
+        iargs.pop_front();
+        if (run_only == client->get_nodeid())
+          dout(2) << "only " << run_only << dendl;
+      }
+      break;
+    case SYNCLIENT_MODE_ONLYRANGE:
+      {
+        int first = iargs.front();
+        iargs.pop_front();
+        int last = iargs.front();
+        iargs.pop_front();
+        if (first <= client->get_nodeid() &&
+	    last > client->get_nodeid()) {
+	  run_only = client->get_nodeid();
+          dout(2) << "onlyrange [" << first << ", " << last << ") includes me" << dendl;
+	} else
+	  run_only = client->get_nodeid().v+1;  // not me
+      }
+      break;
+    case SYNCLIENT_MODE_EXCLUDE:
+      {
+        exclude = iargs.front();
+        iargs.pop_front();
+        if (exclude == client->get_nodeid()) {
+	  run_only = client->get_nodeid().v + 1;
+          dout(2) << "not running " << exclude << dendl;
+	} else
+	  run_only = -1;
+      }
+      break;
+
+      // HOW LONG?
+
+    case SYNCLIENT_MODE_UNTIL:
+      {
+        int iarg1 = iargs.front();
+        iargs.pop_front();
+	if (run_me()) {
+	  if (iarg1) {
+	    dout(2) << "until " << iarg1 << dendl;
+	    utime_t dur(iarg1,0);
+	    run_until = run_start + dur;
+	  } else {
+	    dout(2) << "until " << iarg1 << " (no limit)" << dendl;
+	    run_until = utime_t(0,0);
+	  }
+	}
+      }
+      break;
+
+
+      // ...
+
+    case SYNCLIENT_MODE_FOO:
+      if (run_me()) {
+	foo();
+      }
+      did_run_me();
+      break;
+
+    case SYNCLIENT_MODE_RANDOMSLEEP:
+      {
+        int iarg1 = iargs.front();
+        iargs.pop_front();
+        if (run_me()) {
+          srand(time(0) + getpid() + client->whoami.v);
+          sleep(rand() % iarg1);
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_SLEEP:
+      {
+        int iarg1 = iargs.front();
+        iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "sleep " << iarg1 << dendl;
+          sleep(iarg1);
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_SLEEPUNTIL:
+      {
+        int iarg1 = iargs.front();
+        iargs.pop_front();
+        if (iarg1 && run_me()) {
+          dout(2) << "sleepuntil " << iarg1 << dendl;
+          utime_t at = ceph_clock_now() - run_start;
+          if (at.sec() < iarg1)
+            sleep(iarg1 - at.sec());
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_RANDOMWALK:
+      {
+        int iarg1 = iargs.front();
+        iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "randomwalk " << iarg1 << dendl;
+          random_walk(iarg1);
+        }
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_DROPCACHE:
+      {
+	client->sync_fs();
+	client->drop_caches();
+      }
+      break;
+
+    case SYNCLIENT_MODE_DUMP:
+      {
+	string sarg1 = get_sarg(0);
+	if (run_me()) {
+	  dout(2) << "placement dump " << sarg1 << dendl;
+	  dump_placement(sarg1);
+	}
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_MAKEDIRMESS:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "makedirmess " << sarg1 << " " << iarg1 << dendl;
+          make_dir_mess(sarg1.c_str(), iarg1);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_MAKEDIRS:
+      {
+        string sarg1 = get_sarg(seq++);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        int iarg3 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "makedirs " << sarg1 << " " << iarg1 << " " << iarg2 << " " << iarg3 << dendl;
+          make_dirs(sarg1.c_str(), iarg1, iarg2, iarg3);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_STATDIRS:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        int iarg3 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "statdirs " << sarg1 << " " << iarg1 << " " << iarg2 << " " << iarg3 << dendl;
+          stat_dirs(sarg1.c_str(), iarg1, iarg2, iarg3);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_READDIRS:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        int iarg3 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "readdirs " << sarg1 << " " << iarg1 << " " << iarg2 << " " << iarg3 << dendl;
+          read_dirs(sarg1.c_str(), iarg1, iarg2, iarg3);
+        }
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_THRASHLINKS:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        int iarg3 = iargs.front();  iargs.pop_front();
+        int iarg4 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "thrashlinks " << sarg1 << " " << iarg1 << " " << iarg2 << " " << iarg3 << dendl;
+          thrash_links(sarg1.c_str(), iarg1, iarg2, iarg3, iarg4);
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_LINKTEST:
+      {
+	if (run_me()) {
+	  link_test();
+	}
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_MAKEFILES:
+      {
+        int num = iargs.front();  iargs.pop_front();
+        int count = iargs.front();  iargs.pop_front();
+        int priv = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "makefiles " << num << " " << count << " " << priv << dendl;
+          make_files(num, count, priv, false);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_MAKEFILES2:
+      {
+        int num = iargs.front();  iargs.pop_front();
+        int count = iargs.front();  iargs.pop_front();
+        int priv = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "makefiles2 " << num << " " << count << " " << priv << dendl;
+          make_files(num, count, priv, true);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_CREATESHARED:
+      {
+        string sarg1 = get_sarg(0);
+        int num = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "createshared " << num << dendl;
+          create_shared(num);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_OPENSHARED:
+      {
+        string sarg1 = get_sarg(0);
+        int num = iargs.front();  iargs.pop_front();
+        int count = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "openshared " << num << dendl;
+          open_shared(num, count);
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_CREATEOBJECTS:
+      {
+        int count = iargs.front();  iargs.pop_front();
+        int size = iargs.front();  iargs.pop_front();
+        int inflight = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "createobjects " << count << " of " << size << " bytes"
+		  << ", " << inflight << " in flight" << dendl;
+          create_objects(count, size, inflight);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_OBJECTRW:
+      {
+        int count = iargs.front();  iargs.pop_front();
+        int size = iargs.front();  iargs.pop_front();
+        int wrpc = iargs.front();  iargs.pop_front();
+        int overlap = iargs.front();  iargs.pop_front();
+        int rskew = iargs.front();  iargs.pop_front();
+        int wskew = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          dout(2) << "objectrw " << count << " " << size << " " << wrpc 
+		  << " " << overlap << " " << rskew << " " << wskew << dendl;
+          object_rw(count, size, wrpc, overlap, rskew, wskew);
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_FULLWALK:
+      {
+        string sarg1;// = get_sarg(0);
+        if (run_me()) {
+          dout(2) << "fullwalk" << sarg1 << dendl;
+          full_walk(sarg1);
+        }
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_REPEATWALK:
+      {
+        string sarg1 = get_sarg(0);
+        if (run_me()) {
+          dout(2) << "repeatwalk " << sarg1 << dendl;
+          while (full_walk(sarg1) == 0) ;
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_RMFILE:
+      {
+        string sarg1 = get_sarg(0);
+        if (run_me()) {
+          rm_file(sarg1);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_WRITEFILE:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        dout(1) << "WRITING SYN CLIENT" << dendl;
+        if (run_me()) {
+          write_file(sarg1, iarg1, iarg2);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_CHUNK:
+      if (run_me()) {
+        string sarg1 = get_sarg(0);
+	chunk_file(sarg1);
+      }
+      did_run_me();
+      break;
+
+
+    case SYNCLIENT_MODE_OVERLOAD_OSD_0:
+      {
+	dout(1) << "OVERLOADING OSD 0" << dendl;
+	int iarg1 = iargs.front(); iargs.pop_front();
+	int iarg2 = iargs.front(); iargs.pop_front();
+	int iarg3 = iargs.front(); iargs.pop_front();
+	if (run_me()) {
+	  overload_osd_0(iarg1, iarg2, iarg3);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_WRSHARED:
+      {
+        string sarg1 = "shared";
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          write_file(sarg1, iarg1, iarg2);
+	}
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_READSHARED:
+      {
+        string sarg1 = "shared";
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          read_file(sarg1, iarg1, iarg2, true);
+	}
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_WRITEBATCH:
+      {
+	int iarg1 = iargs.front(); iargs.pop_front();
+        int iarg2 = iargs.front(); iargs.pop_front();
+        int iarg3 = iargs.front(); iargs.pop_front();
+
+        if (run_me()) {
+          write_batch(iarg1, iarg2, iarg3);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_READFILE:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+
+        dout(1) << "READING SYN CLIENT" << dendl;
+        if (run_me()) {
+          read_file(sarg1, iarg1, iarg2);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_RDWRRANDOM:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+
+        dout(1) << "RANDOM READ WRITE SYN CLIENT" << dendl;
+        if (run_me()) {
+          read_random(sarg1, iarg1, iarg2);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_RDWRRANDOM_EX:
+      {
+        string sarg1 = get_sarg(0);
+        int iarg1 = iargs.front();  iargs.pop_front();
+        int iarg2 = iargs.front();  iargs.pop_front();
+
+        dout(1) << "RANDOM READ WRITE SYN CLIENT" << dendl;
+        if (run_me()) {
+          read_random_ex(sarg1, iarg1, iarg2);
+	}
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_TRACE:
+      {
+        string tfile = get_sarg(0);
+        sargs.push_front(string("~"));
+        int iarg1 = iargs.front();  iargs.pop_front();
+	int playdata = iargs.front(); iargs.pop_front();
+        string prefix = get_sarg(0);
+	char realtfile[100];
+	snprintf(realtfile, sizeof(realtfile), tfile.c_str(), (int)client->get_nodeid().v);
+
+        if (run_me()) {
+          dout(0) << "trace " << tfile << " prefix=" << prefix << " count=" << iarg1 << " data=" << playdata << dendl;
+          
+          Trace t(realtfile);
+          
+	  if (iarg1 == 0) iarg1 = 1; // play trace at least once!
+
+          for (int i=0; i<iarg1; i++) {
+            utime_t start = ceph_clock_now();
+            
+            if (time_to_stop()) break;
+            play_trace(t, prefix, !playdata);
+            if (time_to_stop()) break;
+            if (iarg1 > 1) clean_dir(prefix);  // clean only if repeat
+            
+            utime_t lat = ceph_clock_now();
+            lat -= start;
+            
+            dout(0) << " trace " << tfile << " loop " << (i+1) << "/" << iarg1 << " done in " << (double)lat << " seconds" << dendl;
+            if (client->logger
+                && i > 0
+                && i < iarg1-1
+                ) {
+              //client->logger->finc("trsum", (double)lat);
+              //client->logger->inc("trnum");
+            }
+          }
+	  dout(1) << "done " << dendl;
+        }
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_OPENTEST:
+      {
+        int count = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+          for (int i=0; i<count; i++) {
+            int fd = client->open("test", (rand()%2) ?
+				  (O_WRONLY|O_CREAT) : O_RDONLY,
+				  perms);
+            if (fd > 0) client->close(fd);
+          }
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_OPTEST:
+      {
+        int count = iargs.front();  iargs.pop_front();
+        if (run_me()) {
+	  client->mknod("test", 0777, perms);
+	  struct stat st;
+	  for (int i=0; i<count; i++) {
+	    client->lstat("test", &st, perms);
+	    client->chmod("test", 0777, perms);
+          }
+        }
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_TRUNCATE:
+      {
+        string file = get_sarg(0);
+        sargs.push_front(file);
+        int iarg1 = iargs.front();  iargs.pop_front();
+	if (run_me()) {
+	  client->truncate(file.c_str(), iarg1, perms);
+	}
+	did_run_me();
+      }
+      break;
+
+
+    case SYNCLIENT_MODE_IMPORTFIND:
+      {
+	string base = get_sarg(0);
+	string find = get_sarg(0);
+	int data = get_iarg();
+	if (run_me()) {
+	  import_find(base.c_str(), find.c_str(), data);
+	}
+	did_run_me();
+      }
+      break;
+
+    case SYNCLIENT_MODE_LOOKUPHASH:
+      {
+	inodeno_t ino;
+	string iname = get_sarg(0);
+	sscanf(iname.c_str(), "%llx", (long long unsigned*)&ino.val);
+	inodeno_t dirino;
+	string diname = get_sarg(0);
+	sscanf(diname.c_str(), "%llx", (long long unsigned*)&dirino.val);
+	string name = get_sarg(0);
+	if (run_me()) {
+	  lookup_hash(ino, dirino, name.c_str(), perms);
+	}
+      }
+      break;
+    case SYNCLIENT_MODE_LOOKUPINO:
+      {
+	inodeno_t ino;
+	string iname = get_sarg(0);
+	sscanf(iname.c_str(), "%llx", (long long unsigned*)&ino.val);
+	if (run_me()) {
+	  lookup_ino(ino, perms);
+	}
+      }
+      break;
+      
+    case SYNCLIENT_MODE_MKSNAP:
+      {
+	string base = get_sarg(0);
+	string name = get_sarg(0);
+	if (run_me())
+	  mksnap(base.c_str(), name.c_str(), perms);
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_RMSNAP:
+      {
+	string base = get_sarg(0);
+	string name = get_sarg(0);
+	if (run_me())
+	  rmsnap(base.c_str(), name.c_str(), perms);
+	did_run_me();
+      }
+      break;
+    case SYNCLIENT_MODE_MKSNAPFILE:
+      {
+	string base = get_sarg(0);
+	if (run_me())
+	  mksnapfile(base.c_str());
+	did_run_me();
+      }
+      break;
+
+    default:
+      ceph_abort();
+    }
+  }
+  dout(1) << "syn done, unmounting " << dendl;
+
+  client->unmount();
+  client->shutdown();
+  return 0;
+}
+
+
+int SyntheticClient::start_thread()
+{
+  ceph_assert(!thread_id);
+
+  pthread_create(&thread_id, NULL, synthetic_client_thread_entry, this);
+  ceph_assert(thread_id);
+  ceph_pthread_setname(thread_id, "client");
+  return 0;
+}
+
+int SyntheticClient::join_thread()
+{
+  ceph_assert(thread_id);
+  void *rv;
+  pthread_join(thread_id, &rv);
+  return 0;
+}
+
+
+bool roll_die(float p) 
+{
+  float r = (float)(rand() % 100000) / 100000.0;
+  if (r < p) 
+    return true;
+  else 
+    return false;
+}
+
+void SyntheticClient::init_op_dist()
+{
+  op_dist.clear();
+#if 0
+  op_dist.add( CEPH_MDS_OP_STAT, 610 );
+  op_dist.add( CEPH_MDS_OP_UTIME, 0 );
+  op_dist.add( CEPH_MDS_OP_CHMOD, 1 );
+  op_dist.add( CEPH_MDS_OP_CHOWN, 1 );
+#endif
+
+  op_dist.add( CEPH_MDS_OP_READDIR, 2 );
+  op_dist.add( CEPH_MDS_OP_MKNOD, 30 );
+  op_dist.add( CEPH_MDS_OP_LINK, 0 );
+  op_dist.add( CEPH_MDS_OP_UNLINK, 20 );
+  op_dist.add( CEPH_MDS_OP_RENAME, 40 );
+
+  op_dist.add( CEPH_MDS_OP_MKDIR, 10 );
+  op_dist.add( CEPH_MDS_OP_RMDIR, 20 );
+  op_dist.add( CEPH_MDS_OP_SYMLINK, 20 );
+
+  op_dist.add( CEPH_MDS_OP_OPEN, 200 );
+  //op_dist.add( CEPH_MDS_OP_READ, 0 );
+  //op_dist.add( CEPH_MDS_OP_WRITE, 0 );
+  //op_dist.add( CEPH_MDS_OP_TRUNCATE, 0 );
+  //op_dist.add( CEPH_MDS_OP_FSYNC, 0 );
+  //op_dist.add( CEPH_MDS_OP_RELEASE, 200 );
+  op_dist.normalize();
+}
+
+void SyntheticClient::up()
+{
+  cwd = cwd.prefixpath(cwd.depth()-1);
+  dout(DBL) << "cd .. -> " << cwd << dendl;
+  clear_dir();
+}
+
+int SyntheticClient::play_trace(Trace& t, string& prefix, bool metadata_only)
+{
+  dout(4) << "play trace prefix '" << prefix << "'" << dendl;
+  UserPerm perms = client->pick_my_perms();
+  t.start();
+
+  string buf;
+  string buf2;
+
+  utime_t start = ceph_clock_now();
+
+  ceph::unordered_map<int64_t, int64_t> open_files;
+  ceph::unordered_map<int64_t, dir_result_t*> open_dirs;
+
+  ceph::unordered_map<int64_t, Fh*> ll_files;
+  ceph::unordered_map<int64_t, dir_result_t*> ll_dirs;
+  ceph::unordered_map<uint64_t, int64_t> ll_inos;
+
+  Inode *i1, *i2;
+
+  ll_inos[1] = 1; // root inode is known.
+
+  // prefix?
+  const char *p = prefix.c_str();
+  if (prefix.length()) {
+    client->mkdir(prefix.c_str(), 0755, perms);
+    struct ceph_statx stx;
+    i1 = client->ll_get_inode(vinodeno_t(1, CEPH_NOSNAP));
+    if (client->ll_lookupx(i1, prefix.c_str(), &i2, &stx, CEPH_STATX_INO, 0, perms) == 0) {
+      ll_inos[1] = stx.stx_ino;
+      dout(5) << "'root' ino is " << inodeno_t(stx.stx_ino) << dendl;
+      client->ll_put(i1);
+    } else {
+      dout(0) << "warning: play_trace couldn't lookup up my per-client directory" << dendl;
+    }
+  } else
+    (void) client->ll_get_inode(vinodeno_t(1, CEPH_NOSNAP));
+
+  utime_t last_status = start;
+
+  int n = 0;
+
+  // for object traces
+  ceph::mutex lock = ceph::make_mutex("synclient foo");
+  ceph::condition_variable cond;
+  bool ack;
+
+  while (!t.end()) {
+
+    if (++n == 100) {
+      n = 00;
+      utime_t now = last_status;
+      if (now - last_status > 1.0) {
+	last_status = now;
+	dout(1) << "play_trace at line " << t.get_line() << dendl;
+      }
+    }
+    
+    if (time_to_stop()) break;
+    
+    // op
+    const char *op = t.get_string(buf, 0);
+    dout(4) << (t.get_line()-1) << ": trace op " << op << dendl;
+    
+    if (op[0] == '@') {
+      // timestamp... ignore it!
+      t.get_int(); // sec
+      t.get_int(); // usec
+      op = t.get_string(buf, 0);
+    }
+
+    // high level ops ---------------------
+    UserPerm perms = client->pick_my_perms();
+    if (strcmp(op, "link") == 0) {
+      const char *a = t.get_string(buf, p);
+      const char *b = t.get_string(buf2, p);
+      client->link(a, b, perms);
+    } else if (strcmp(op, "unlink") == 0) {
+      const char *a = t.get_string(buf, p);
+      client->unlink(a, perms);
+    } else if (strcmp(op, "rename") == 0) {
+      const char *a = t.get_string(buf, p);
+      const char *b = t.get_string(buf2, p);
+      client->rename(a,b, perms);
+    } else if (strcmp(op, "mkdir") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      client->mkdir(a, b, perms);
+    } else if (strcmp(op, "rmdir") == 0) {
+      const char *a = t.get_string(buf, p);
+      client->rmdir(a, perms);
+    } else if (strcmp(op, "symlink") == 0) {
+      const char *a = t.get_string(buf, p);
+      const char *b = t.get_string(buf2, p);
+      client->symlink(a, b, perms);
+    } else if (strcmp(op, "readlink") == 0) {
+      const char *a = t.get_string(buf, p);
+      char buf[100];
+      client->readlink(a, buf, 100, perms);
+    } else if (strcmp(op, "lstat") == 0) {
+      struct stat st;
+      const char *a = t.get_string(buf, p);
+      if (strcmp(a, p) != 0 &&
+	  strcmp(a, "/") != 0 &&
+	  strcmp(a, "/lib") != 0 && // or /lib.. that would be a lookup. hack.
+	  a[0] != 0)  // stop stating the root directory already
+	client->lstat(a, &st, perms);
+    } else if (strcmp(op, "chmod") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      client->chmod(a, b, perms);
+    } else if (strcmp(op, "chown") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      int64_t c = t.get_int();
+      client->chown(a, b, c, perms);
+    } else if (strcmp(op, "utime") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      int64_t c = t.get_int();
+      struct utimbuf u;
+      u.actime = b;
+      u.modtime = c;
+      client->utime(a, &u, perms);
+    } else if (strcmp(op, "mknod") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      int64_t c = t.get_int();
+      client->mknod(a, b, perms, c);
+    } else if (strcmp(op, "oldmknod") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      client->mknod(a, b, perms, 0);
+    } else if (strcmp(op, "getdir") == 0) {
+      const char *a = t.get_string(buf, p);
+      list<string> contents;
+      int r = client->getdir(a, contents, perms);
+      if (r < 0) {
+        dout(1) << "getdir on " << a << " returns " << r << dendl;
+      }
+    } else if (strcmp(op, "opendir") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int();
+      dir_result_t *dirp;
+      client->opendir(a, &dirp, perms);
+      if (dirp) open_dirs[b] = dirp;
+    } else if (strcmp(op, "closedir") == 0) {
+      int64_t a = t.get_int();
+      client->closedir(open_dirs[a]);
+      open_dirs.erase(a);
+    } else if (strcmp(op, "open") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int(); 
+      int64_t c = t.get_int(); 
+      int64_t d = t.get_int();
+      int64_t fd = client->open(a, b, perms, c);
+      if (fd > 0) open_files[d] = fd;
+    } else if (strcmp(op, "oldopen") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t b = t.get_int(); 
+      int64_t d = t.get_int();
+      int64_t fd = client->open(a, b, perms, 0755);
+      if (fd > 0) open_files[d] = fd;
+    } else if (strcmp(op, "close") == 0) {
+      int64_t id = t.get_int();
+      int64_t fh = open_files[id];
+      if (fh > 0) client->close(fh);
+      open_files.erase(id);
+    } else if (strcmp(op, "lseek") == 0) {
+      int64_t f = t.get_int();
+      int fd = open_files[f];
+      int64_t off = t.get_int();
+      int64_t whence = t.get_int();
+      client->lseek(fd, off, whence);
+    } else if (strcmp(op, "read") == 0) {
+      int64_t f = t.get_int();
+      int64_t size = t.get_int();
+      int64_t off = t.get_int();
+      int64_t fd = open_files[f];
+      if (!metadata_only) {
+	char *b = new char[size];
+	client->read(fd, b, size, off);
+	delete[] b;
+      }
+    } else if (strcmp(op, "write") == 0) {
+      int64_t f = t.get_int();
+      int64_t fd = open_files[f];
+      int64_t size = t.get_int();
+      int64_t off = t.get_int();
+      if (!metadata_only) {
+	char *b = new char[size];
+	memset(b, 1, size);            // let's write 1's!
+	client->write(fd, b, size, off);
+	delete[] b;
+      } else {
+	client->write(fd, NULL, 0, size+off);
+      }
+    } else if (strcmp(op, "truncate") == 0) {
+      const char *a = t.get_string(buf, p);
+      int64_t l = t.get_int();
+      client->truncate(a, l, perms);
+    } else if (strcmp(op, "ftruncate") == 0) {
+      int64_t f = t.get_int();
+      int fd = open_files[f];
+      int64_t l = t.get_int();
+      client->ftruncate(fd, l, perms);
+    } else if (strcmp(op, "fsync") == 0) {
+      int64_t f = t.get_int();
+      int64_t b = t.get_int();
+      int fd = open_files[f];
+      client->fsync(fd, b);
+    } else if (strcmp(op, "chdir") == 0) {
+      const char *a = t.get_string(buf, p);
+      // Client users should remember their path, but since this
+      // is just a synthetic client we ignore it.
+      std::string ignore;
+      client->chdir(a, ignore, perms);
+    } else if (strcmp(op, "statfs") == 0) {
+      struct statvfs stbuf;
+      client->statfs("/", &stbuf, perms);
+    }
+
+    // low level ops ---------------------
+    else if (strcmp(op, "ll_lookup") == 0) {
+      int64_t i = t.get_int();
+      const char *name = t.get_string(buf, p);
+      int64_t r = t.get_int();
+      struct ceph_statx stx;
+      if (ll_inos.count(i)) {
+	  i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	  if (client->ll_lookupx(i1, name, &i2, &stx, CEPH_STATX_INO, 0, perms) == 0)
+	    ll_inos[r] = stx.stx_ino;
+	  client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_forget") == 0) {
+      int64_t i = t.get_int();
+      int64_t n = t.get_int();
+      if (ll_inos.count(i) && 
+	  client->ll_forget(
+	    client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP)), n))
+	ll_inos.erase(i);
+    } else if (strcmp(op, "ll_getattr") == 0) {
+      int64_t i = t.get_int();
+      struct stat attr;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	client->ll_getattr(i1, &attr, perms);
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_setattr") == 0) {
+      int64_t i = t.get_int();
+      struct stat attr;
+      memset(&attr, 0, sizeof(attr));
+      attr.st_mode = t.get_int();
+      attr.st_uid = t.get_int();
+      attr.st_gid = t.get_int();
+      attr.st_size = t.get_int();
+      attr.st_mtime = t.get_int();
+      attr.st_atime = t.get_int();
+      int mask = t.get_int();
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	client->ll_setattr(i1, &attr, mask, perms);
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_readlink") == 0) {
+      int64_t i = t.get_int();
+      if (ll_inos.count(i)) {
+        char buf[PATH_MAX];
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	client->ll_readlink(i1, buf, sizeof(buf), perms);
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_mknod") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      int m = t.get_int();
+      int r = t.get_int();
+      int64_t ri = t.get_int();
+      struct stat attr;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_mknod(i1, n, m, r, &attr, &i2, perms) == 0)
+	  ll_inos[ri] = attr.st_ino;
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_mkdir") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      int m = t.get_int();
+      int64_t ri = t.get_int();
+      struct stat attr;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_mkdir(i1, n, m, &attr, &i2, perms) == 0)
+	  ll_inos[ri] = attr.st_ino;
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_symlink") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      const char *v = t.get_string(buf2, p);
+      int64_t ri = t.get_int();
+      struct stat attr;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_symlink(i1, n, v, &attr, &i2, perms) == 0)
+	  ll_inos[ri] = attr.st_ino;
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_unlink") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	client->ll_unlink(i1, n, perms);
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_rmdir") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	client->ll_rmdir(i1, n, perms);
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_rename") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      int64_t ni = t.get_int();
+      const char *nn = t.get_string(buf2, p);
+      if (ll_inos.count(i) &&
+	  ll_inos.count(ni)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	i2 = client->ll_get_inode(vinodeno_t(ll_inos[ni],CEPH_NOSNAP));
+	client->ll_rename(i1, n, i2, nn, perms);
+	client->ll_put(i1);
+	client->ll_put(i2);
+      }
+    } else if (strcmp(op, "ll_link") == 0) {
+      int64_t i = t.get_int();
+      int64_t ni = t.get_int();
+      const char *nn = t.get_string(buf, p);
+      if (ll_inos.count(i) &&
+	  ll_inos.count(ni)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	i2 = client->ll_get_inode(vinodeno_t(ll_inos[ni],CEPH_NOSNAP));
+	client->ll_link(i1, i2, nn, perms);
+	client->ll_put(i1);
+	client->ll_put(i2);
+      }
+    } else if (strcmp(op, "ll_opendir") == 0) {
+      int64_t i = t.get_int();
+      int64_t r = t.get_int();
+      dir_result_t *dirp;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_opendir(i1, O_RDONLY, &dirp, perms) == 0)
+	  ll_dirs[r] = dirp;
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_releasedir") == 0) {
+      int64_t f = t.get_int();
+      if (ll_dirs.count(f)) {
+	client->ll_releasedir(ll_dirs[f]);
+	ll_dirs.erase(f);
+      }
+    } else if (strcmp(op, "ll_open") == 0) {
+      int64_t i = t.get_int();
+      int64_t f = t.get_int();
+      int64_t r = t.get_int();
+      Fh *fhp;
+      if (ll_inos.count(i)) {
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_open(i1, f, &fhp, perms) == 0)
+	  ll_files[r] = fhp;
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_create") == 0) {
+      int64_t i = t.get_int();
+      const char *n = t.get_string(buf, p);
+      int64_t m = t.get_int();
+      int64_t f = t.get_int();
+      int64_t r = t.get_int();
+      int64_t ri = t.get_int();
+      struct stat attr;
+      if (ll_inos.count(i)) {
+	Fh *fhp;
+	i1 = client->ll_get_inode(vinodeno_t(ll_inos[i],CEPH_NOSNAP));
+	if (client->ll_create(i1, n, m, f, &attr, NULL, &fhp, perms) == 0) {
+	  ll_inos[ri] = attr.st_ino;
+	  ll_files[r] = fhp;
+	}
+	client->ll_put(i1);
+      }
+    } else if (strcmp(op, "ll_read") == 0) {
+      int64_t f = t.get_int();
+      int64_t off = t.get_int();
+      int64_t size = t.get_int();
+      if (ll_files.count(f) &&
+	  !metadata_only) {
+	bufferlist bl;
+	client->ll_read(ll_files[f], off, size, &bl);
+      }
+    } else if (strcmp(op, "ll_write") == 0) {
+      int64_t f = t.get_int();
+      int64_t off = t.get_int();
+      int64_t size = t.get_int();
+      if (ll_files.count(f)) {
+	if (!metadata_only) {
+	  bufferlist bl;
+	  bufferptr bp(size);
+	  bl.push_back(bp);
+	  bp.zero();
+	  client->ll_write(ll_files[f], off, size, bl.c_str());
+	} else {
+	  client->ll_write(ll_files[f], off+size, 0, NULL);
+	}
+      }
+    } else if (strcmp(op, "ll_flush") == 0) {
+      int64_t f = t.get_int();
+      if (!metadata_only &&
+	  ll_files.count(f)) 
+	client->ll_flush(ll_files[f]);
+    } else if (strcmp(op, "ll_fsync") == 0) {
+      int64_t f = t.get_int();
+      if (!metadata_only &&
+	  ll_files.count(f)) 
+	client->ll_fsync(ll_files[f], false); // FIXME dataonly param
+    } else if (strcmp(op, "ll_release") == 0) {
+      int64_t f = t.get_int();
+      if (ll_files.count(f)) {
+	client->ll_release(ll_files[f]);
+	ll_files.erase(f);
+      }
+    } else if (strcmp(op, "ll_statfs") == 0) {
+      int64_t i = t.get_int();
+      if (ll_inos.count(i))
+	{} //client->ll_statfs(vinodeno_t(ll_inos[i],CEPH_NOSNAP), perms);
+    }
+
+
+    // object-level traces
+
+    else if (strcmp(op, "o_stat") == 0) {
+      int64_t oh = t.get_int();
+      int64_t ol = t.get_int();
+      object_t oid = file_object_t(oh, ol);
+      std::unique_lock locker{lock};
+      object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+      uint64_t size;
+      ceph::real_time mtime;
+      client->objecter->stat(oid, oloc, CEPH_NOSNAP, &size, &mtime, 0,
+			     new C_SafeCond(lock, cond, &ack));
+      cond.wait(locker, [&ack] { return ack; });
+    }
+    else if (strcmp(op, "o_read") == 0) {
+      int64_t oh = t.get_int();
+      int64_t ol = t.get_int();
+      int64_t off = t.get_int();
+      int64_t len = t.get_int();
+      object_t oid = file_object_t(oh, ol);
+      object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+      std::unique_lock locker{lock};
+      bufferlist bl;
+      client->objecter->read(oid, oloc, off, len, CEPH_NOSNAP, &bl, 0,
+			     new C_SafeCond(lock, cond, &ack));
+      cond.wait(locker, [&ack] { return ack; });
+    }
+    else if (strcmp(op, "o_write") == 0) {
+      int64_t oh = t.get_int();
+      int64_t ol = t.get_int();
+      int64_t off = t.get_int();
+      int64_t len = t.get_int();
+      object_t oid = file_object_t(oh, ol);
+      object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+      std::unique_lock locker{lock};
+      bufferptr bp(len);
+      bufferlist bl;
+      bl.push_back(bp);
+      SnapContext snapc;
+      client->objecter->write(oid, oloc, off, len, snapc, bl,
+			      ceph::real_clock::now(), 0,
+			      new C_SafeCond(lock, cond, &ack));
+      cond.wait(locker, [&ack] { return ack; });
+    }
+    else if (strcmp(op, "o_zero") == 0) {
+      int64_t oh = t.get_int();
+      int64_t ol = t.get_int();
+      int64_t off = t.get_int();
+      int64_t len = t.get_int();
+      object_t oid = file_object_t(oh, ol);
+      object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+      std::unique_lock locker{lock};
+      SnapContext snapc;
+      client->objecter->zero(oid, oloc, off, len, snapc,
+			     ceph::real_clock::now(), 0,
+			     new C_SafeCond(lock, cond, &ack));
+      cond.wait(locker, [&ack] { return ack; });
+    }
+
+
+    else {
+      dout(0) << (t.get_line()-1) << ": *** trace hit unrecognized symbol '" << op << "' " << dendl;
+      ceph_abort();
+    }
+  }
+
+  dout(10) << "trace finished on line " << t.get_line() << dendl;
+
+  // close open files
+  for (ceph::unordered_map<int64_t, int64_t>::iterator fi = open_files.begin();
+       fi != open_files.end();
+       ++fi) {
+    dout(1) << "leftover close " << fi->second << dendl;
+    if (fi->second > 0) client->close(fi->second);
+  }
+  for (ceph::unordered_map<int64_t, dir_result_t*>::iterator fi = open_dirs.begin();
+       fi != open_dirs.end();
+       ++fi) {
+    dout(1) << "leftover closedir " << fi->second << dendl;
+    if (fi->second != 0) client->closedir(fi->second);
+  }
+  for (ceph::unordered_map<int64_t,Fh*>::iterator fi = ll_files.begin();
+       fi != ll_files.end();
+       ++fi) {
+    dout(1) << "leftover ll_release " << fi->second << dendl;
+    if (fi->second) client->ll_release(fi->second);
+  }
+  for (ceph::unordered_map<int64_t,dir_result_t*>::iterator fi = ll_dirs.begin();
+       fi != ll_dirs.end();
+       ++fi) {
+    dout(1) << "leftover ll_releasedir " << fi->second << dendl;
+    if (fi->second) client->ll_releasedir(fi->second);
+  }
+  
+  return 0;
+}
+
+
+
+int SyntheticClient::clean_dir(string& basedir)
+{
+  // read dir
+  list<string> contents;
+  UserPerm perms = client->pick_my_perms();
+  int r = client->getdir(basedir.c_str(), contents, perms);
+  if (r < 0) {
+    dout(1) << "getdir on " << basedir << " returns " << r << dendl;
+    return r;
+  }
+
+  for (list<string>::iterator it = contents.begin();
+       it != contents.end();
+       ++it) {
+    if (*it == ".") continue;
+    if (*it == "..") continue;
+    string file = basedir + "/" + *it;
+
+    if (time_to_stop()) break;
+
+    struct stat st;
+    int r = client->lstat(file.c_str(), &st, perms);
+    if (r < 0) {
+      dout(1) << "stat error on " << file << " r=" << r << dendl;
+      continue;
+    }
+
+    if ((st.st_mode & S_IFMT) == S_IFDIR) {
+      clean_dir(file);
+      client->rmdir(file.c_str(), perms);
+    } else {
+      client->unlink(file.c_str(), perms);
+    }
+  }
+
+  return 0;
+
+}
+
+
+int SyntheticClient::full_walk(string& basedir) 
+{
+  if (time_to_stop()) return -1;
+
+  list<string> dirq;
+  list<frag_info_t> statq;
+  dirq.push_back(basedir);
+  frag_info_t empty;
+  statq.push_back(empty);
+
+  ceph::unordered_map<inodeno_t, int> nlink;
+  ceph::unordered_map<inodeno_t, int> nlink_seen;
+
+  UserPerm perms = client->pick_my_perms();
+  while (!dirq.empty()) {
+    string dir = dirq.front();
+    frag_info_t expect = statq.front();
+    dirq.pop_front();
+    statq.pop_front();
+
+    frag_info_t actual = empty;
+
+    // read dir
+    list<string> contents;
+    int r = client->getdir(dir.c_str(), contents, perms);
+    if (r < 0) {
+      dout(1) << "getdir on " << dir << " returns " << r << dendl;
+      continue;
+    }
+    
+    for (list<string>::iterator it = contents.begin();
+	 it != contents.end();
+	 ++it) {
+      if (*it == "." ||
+	  *it == "..") 
+	continue;
+      string file = dir + "/" + *it;
+      
+      struct stat st;
+      frag_info_t dirstat;
+      int r = client->lstat(file.c_str(), &st, perms, &dirstat);
+      if (r < 0) {
+	dout(1) << "stat error on " << file << " r=" << r << dendl;
+	continue;
+      }
+
+      nlink_seen[st.st_ino]++;
+      nlink[st.st_ino] = st.st_nlink;
+
+      if (S_ISDIR(st.st_mode))
+	actual.nsubdirs++;
+      else
+	actual.nfiles++;
+
+      // print
+      char *tm = ctime(&st.st_mtime);
+      tm[strlen(tm)-1] = 0;
+      printf("%llx %c%c%c%c%c%c%c%c%c%c %2d %5d %5d %8llu %12s %s\n",
+	     (long long)st.st_ino,
+	     S_ISDIR(st.st_mode) ? 'd':'-',
+	     (st.st_mode & 0400) ? 'r':'-',
+	     (st.st_mode & 0200) ? 'w':'-',
+	     (st.st_mode & 0100) ? 'x':'-',
+	     (st.st_mode & 040) ? 'r':'-',
+	     (st.st_mode & 020) ? 'w':'-',
+	     (st.st_mode & 010) ? 'x':'-',
+	     (st.st_mode & 04) ? 'r':'-',
+	     (st.st_mode & 02) ? 'w':'-',
+	     (st.st_mode & 01) ? 'x':'-',
+	     (int)st.st_nlink,
+	     (int)st.st_uid, (int)st.st_gid,
+	     (long long unsigned)st.st_size,
+	     tm,
+	     file.c_str());
+
+      
+      if ((st.st_mode & S_IFMT) == S_IFDIR) {
+	dirq.push_back(file);
+	statq.push_back(dirstat);
+      }
+    }
+
+    if (dir != "" &&
+	(actual.nsubdirs != expect.nsubdirs ||
+	 actual.nfiles != expect.nfiles)) {
+      dout(0) << dir << ": expected " << expect << dendl;
+      dout(0) << dir << ":      got " << actual << dendl;
+    }
+  }
+
+  for (ceph::unordered_map<inodeno_t,int>::iterator p = nlink.begin(); p != nlink.end(); ++p) {
+    if (nlink_seen[p->first] != p->second)
+      dout(0) << p->first << " nlink " << p->second << " != " << nlink_seen[p->first] << "seen" << dendl;
+  }
+
+  return 0;
+}
+
+
+
+int SyntheticClient::dump_placement(string& fn) {
+  
+  UserPerm perms = client->pick_my_perms();
+
+  // open file
+  int fd = client->open(fn.c_str(), O_RDONLY, perms);
+  dout(5) << "reading from " << fn << " fd " << fd << dendl;
+  if (fd < 0) return fd;
+
+
+  // How big is it?
+  struct stat stbuf;
+  int lstat_result = client->lstat(fn.c_str(), &stbuf, perms);
+  if (lstat_result < 0) {
+    dout(0) << "lstat error for file " << fn << dendl;
+    client->close(fd);
+    return lstat_result;
+  }
+    
+  off_t filesize = stbuf.st_size;
+ 
+  // grab the placement info
+  vector<ObjectExtent> extents;
+  off_t offset = 0;
+  client->enumerate_layout(fd, extents, filesize, offset);
+  client->close(fd);
+
+
+  // run through all the object extents
+  dout(0) << "file size is " << filesize << dendl;
+  dout(0) << "(osd, start, length) tuples for file " << fn << dendl;
+  for (const auto& x : extents) {
+    int osd = client->objecter->with_osdmap([&](const OSDMap& o) {
+	return o.get_pg_acting_primary(o.object_locator_to_pg(x.oid, x.oloc));
+      });
+
+    // run through all the buffer extents
+    for (const auto& be : x.buffer_extents)
+      dout(0) << "OSD " << osd << ", offset " << be.first
+	      << ", length " << be.second << dendl;
+  }
+  return 0;
+}
+
+
+
+int SyntheticClient::make_dirs(const char *basedir, int dirs, int files, int depth)
+{
+  if (time_to_stop()) return 0;
+
+  UserPerm perms = client->pick_my_perms();
+  // make sure base dir exists
+  int r = client->mkdir(basedir, 0755, perms);
+  if (r != 0) {
+    dout(1) << "can't make base dir? " << basedir << dendl;
+    //return -1;
+  }
+
+  // children
+  char d[500];
+  dout(3) << "make_dirs " << basedir << " dirs " << dirs << " files " << files << " depth " << depth << dendl;
+  for (int i=0; i<files; i++) {
+    snprintf(d, sizeof(d), "%s/file.%d", basedir, i);
+    client->mknod(d, 0644, perms);
+  }
+
+  if (depth == 0) return 0;
+
+  for (int i=0; i<dirs; i++) {
+    snprintf(d, sizeof(d), "%s/dir.%d", basedir, i);
+    make_dirs(d, dirs, files, depth-1);
+  }
+  
+  return 0;
+}
+
+int SyntheticClient::stat_dirs(const char *basedir, int dirs, int files, int depth)
+{
+  if (time_to_stop()) return 0;
+
+  UserPerm perms = client->pick_my_perms();
+
+  // make sure base dir exists
+  struct stat st;
+  int r = client->lstat(basedir, &st, perms);
+  if (r != 0) {
+    dout(1) << "can't make base dir? " << basedir << dendl;
+    return -1;
+  }
+
+  // children
+  char d[500];
+  dout(3) << "stat_dirs " << basedir << " dirs " << dirs << " files " << files << " depth " << depth << dendl;
+  for (int i=0; i<files; i++) {
+    snprintf(d, sizeof(d), "%s/file.%d", basedir, i);
+    client->lstat(d, &st, perms);
+  }
+
+  if (depth == 0) return 0;
+
+  for (int i=0; i<dirs; i++) {
+    snprintf(d, sizeof(d), "%s/dir.%d", basedir, i);
+    stat_dirs(d, dirs, files, depth-1);
+  }
+  
+  return 0;
+}
+int SyntheticClient::read_dirs(const char *basedir, int dirs, int files, int depth)
+{
+  if (time_to_stop()) return 0;
+
+  struct stat st;
+
+  // children
+  char d[500];
+  dout(3) << "read_dirs " << basedir << " dirs " << dirs << " files " << files << " depth " << depth << dendl;
+
+  list<string> contents;
+  UserPerm perms = client->pick_my_perms();
+  utime_t s = ceph_clock_now();
+  int r = client->getdir(basedir, contents, perms);
+  utime_t e = ceph_clock_now();
+  e -= s;
+  if (r < 0) {
+    dout(0) << "getdir couldn't readdir " << basedir << ", stopping" << dendl;
+    return -1;
+  }
+
+  for (int i=0; i<files; i++) {
+    snprintf(d, sizeof(d), "%s/file.%d", basedir, i);
+    utime_t s = ceph_clock_now();
+    if (client->lstat(d, &st, perms) < 0) {
+      dout(2) << "read_dirs failed stat on " << d << ", stopping" << dendl;
+      return -1;
+    }
+    utime_t e = ceph_clock_now();
+    e -= s;
+  }
+
+  if (depth > 0) 
+    for (int i=0; i<dirs; i++) {
+      snprintf(d, sizeof(d), "%s/dir.%d", basedir, i);
+      if (read_dirs(d, dirs, files, depth-1) < 0) return -1;
+    }
+
+  return 0;
+}
+
+
+int SyntheticClient::make_files(int num, int count, int priv, bool more)
+{
+  int whoami = client->get_nodeid().v;
+  char d[255];
+  UserPerm perms = client->pick_my_perms();
+
+  if (priv) {
+    for (int c=0; c<count; c++) {
+      snprintf(d, sizeof(d), "dir.%d.run%d", whoami, c);
+      client->mkdir(d, 0755, perms);
+    }
+  } else {
+    // shared
+    if (true || whoami == 0) {
+      for (int c=0; c<count; c++) {
+        snprintf(d, sizeof(d), "dir.%d.run%d", 0, c);
+        client->mkdir(d, 0755, perms);
+      }
+    } else {
+      sleep(2);
+    }
+  }
+  
+  // files
+  struct stat st;
+  utime_t start = ceph_clock_now();
+  for (int c=0; c<count; c++) {
+    for (int n=0; n<num; n++) {
+      snprintf(d, sizeof(d), "dir.%d.run%d/file.client%d.%d", priv ? whoami:0, c, whoami, n);
+
+      client->mknod(d, 0644, perms);
+
+      if (more) {
+        client->lstat(d, &st, perms);
+        int fd = client->open(d, O_RDONLY, perms);
+        client->unlink(d, perms);
+        client->close(fd);
+      }
+
+      if (time_to_stop()) return 0;
+    }
+  }
+  utime_t end = ceph_clock_now();
+  end -= start;
+  dout(0) << "makefiles time is " << end << " or " << ((double)end / (double)num) <<" per file" << dendl;
+
+  return 0;
+}
+
+int SyntheticClient::link_test()
+{
+  char d[255];
+  char e[255];
+
+  UserPerm perms = client->pick_my_perms();
+
+ // create files
+  int num = 200;
+
+  client->mkdir("orig", 0755, perms);
+  client->mkdir("copy", 0755, perms);
+
+  utime_t start = ceph_clock_now();
+  for (int i=0; i<num; i++) {
+    snprintf(d, sizeof(d), "orig/file.%d", i);
+    client->mknod(d, 0755, perms);
+  }
+  utime_t end = ceph_clock_now();
+  end -= start;
+
+  dout(0) << "orig " << end << dendl;
+
+  // link
+  start = ceph_clock_now();
+  for (int i=0; i<num; i++) {
+    snprintf(d, sizeof(d), "orig/file.%d", i);
+    snprintf(e, sizeof(e), "copy/file.%d", i);
+    client->link(d, e, perms);
+  }
+  end = ceph_clock_now();
+  end -= start;
+  dout(0) << "copy " << end << dendl;
+
+  return 0;
+}
+
+
+int SyntheticClient::create_shared(int num)
+{
+  // files
+  UserPerm perms = client->pick_my_perms();
+  char d[255];
+  client->mkdir("test", 0755, perms);
+  for (int n=0; n<num; n++) {
+    snprintf(d, sizeof(d), "test/file.%d", n);
+    client->mknod(d, 0644, perms);
+  }
+  
+  return 0;
+}
+
+int SyntheticClient::open_shared(int num, int count)
+{
+  // files
+  char d[255];
+  UserPerm perms = client->pick_my_perms();
+  for (int c=0; c<count; c++) {
+    // open
+    list<int> fds;
+    for (int n=0; n<num; n++) {
+      snprintf(d, sizeof(d), "test/file.%d", n);
+      int fd = client->open(d, O_RDONLY, perms);
+      if (fd > 0) fds.push_back(fd);
+    }
+
+    if (false && client->get_nodeid() == 0)
+      for (int n=0; n<num; n++) {
+	snprintf(d, sizeof(d), "test/file.%d", n);
+	client->unlink(d, perms);
+      }
+
+    while (!fds.empty()) {
+      int fd = fds.front();
+      fds.pop_front();
+      client->close(fd);
+    }
+  }
+  
+  return 0;
+}
+
+
+// Hits OSD 0 with writes to various files with OSD 0 as the primary.
+int SyntheticClient::overload_osd_0(int n, int size, int wrsize) {
+  UserPerm perms = client->pick_my_perms();
+  // collect a bunch of files starting on OSD 0
+  int left = n;
+  int tried = 0;
+  while (left < 0) {
+
+
+    // pull open a file
+    dout(0) << "in OSD overload" << dendl;
+    string filename = get_sarg(tried);
+    dout(1) << "OSD Overload workload: trying file " << filename << dendl;
+    int fd = client->open(filename.c_str(), O_RDWR|O_CREAT, perms);
+    ++tried;
+
+    // only use the file if its first primary is OSD 0
+    int primary_osd = check_first_primary(fd);
+    if (primary_osd != 0) {
+      client->close(fd);
+      dout(1) << "OSD Overload workload: SKIPPING file " << filename <<
+	" with OSD " << primary_osd << " as first primary. " << dendl;
+      continue;
+    }
+      dout(1) << "OSD Overload workload: USING file " << filename <<
+	" with OSD 0 as first primary. " << dendl;
+
+
+    --left;
+    // do whatever operation we want to do on the file. How about a write?
+    write_fd(fd, size, wrsize);
+  }
+  return 0;
+}
+
+
+// See what the primary is for the first object in this file.
+int SyntheticClient::check_first_primary(int fh)
+{
+  vector<ObjectExtent> extents;
+  client->enumerate_layout(fh, extents, 1, 0);
+  return client->objecter->with_osdmap([&](const OSDMap& o) {
+      return o.get_pg_acting_primary(
+	o.object_locator_to_pg(extents.begin()->oid, extents.begin()->oloc));
+    });
+}
+
+int SyntheticClient::rm_file(string& fn)
+{
+  UserPerm perms = client->pick_my_perms();
+  return client->unlink(fn.c_str(), perms);
+}
+
+int SyntheticClient::write_file(string& fn, int size, loff_t wrsize)   // size is in MB, wrsize in bytes
+{
+  //uint64_t wrsize = 1024*256;
+  char *buf = new char[wrsize+100];   // 1 MB
+  memset(buf, 7, wrsize);
+  int64_t chunks = (uint64_t)size * (uint64_t)(1024*1024) / (uint64_t)wrsize;
+  UserPerm perms = client->pick_my_perms();
+
+  int fd = client->open(fn.c_str(), O_RDWR|O_CREAT, perms);
+  dout(5) << "writing to " << fn << " fd " << fd << dendl;
+  if (fd < 0) {
+    delete[] buf;
+    return fd;
+  }
+
+  utime_t from = ceph_clock_now();
+  utime_t start = from;
+  uint64_t bytes = 0, total = 0;
+
+
+  for (loff_t i=0; i<chunks; i++) {
+    if (time_to_stop()) {
+      dout(0) << "stopping" << dendl;
+      break;
+    }
+    dout(2) << "writing block " << i << "/" << chunks << dendl;
+    
+    // fill buf with a 16 byte fingerprint
+    // 64 bits : file offset
+    // 64 bits : client id
+    // = 128 bits (16 bytes)
+    uint64_t *p = (uint64_t*)buf;
+    while ((char*)p < buf + wrsize) {
+      *p = (uint64_t)i*(uint64_t)wrsize + (uint64_t)((char*)p - buf);      
+      p++;
+      *p = client->get_nodeid().v;
+      p++;
+    }
+
+    client->write(fd, buf, wrsize, i*wrsize);
+    bytes += wrsize;
+    total += wrsize;
+
+    utime_t now = ceph_clock_now();
+    if (now - from >= 1.0) {
+      double el = now - from;
+      dout(0) << "write " << (bytes / el / 1048576.0) << " MB/sec" << dendl;
+      from = now;
+      bytes = 0;
+    }
+  }
+
+  client->fsync(fd, true);
+
+  utime_t stop = ceph_clock_now();
+  double el = stop - start;
+  dout(0) << "write total " << (total / el / 1048576.0) << " MB/sec ("
+	  << total << " bytes in " << el << " seconds)" << dendl;
+
+  client->close(fd);
+  delete[] buf;
+
+  return 0;
+}
+
+int SyntheticClient::write_fd(int fd, int size, int wrsize)   // size is in MB, wrsize in bytes
+{
+  //uint64_t wrsize = 1024*256;
+  char *buf = new char[wrsize+100];   // 1 MB
+  memset(buf, 7, wrsize);
+  uint64_t chunks = (uint64_t)size * (uint64_t)(1024*1024) / (uint64_t)wrsize;
+
+  //dout(5) << "SyntheticClient::write_fd: writing to fd " << fd << dendl;
+  if (fd < 0) {
+    delete[] buf;
+    return fd;
+  }
+
+  for (unsigned i=0; i<chunks; i++) {
+    if (time_to_stop()) {
+      dout(0) << "stopping" << dendl;
+      break;
+    }
+    dout(2) << "writing block " << i << "/" << chunks << dendl;
+    
+    // fill buf with a 16 byte fingerprint
+    // 64 bits : file offset
+    // 64 bits : client id
+    // = 128 bits (16 bytes)
+    uint64_t *p = (uint64_t*)buf;
+    while ((char*)p < buf + wrsize) {
+      *p = (uint64_t)i*(uint64_t)wrsize + (uint64_t)((char*)p - buf);      
+      p++;
+      *p = client->get_nodeid().v;
+      p++;
+    }
+
+    client->write(fd, buf, wrsize, i*wrsize);
+  }
+  client->close(fd);
+  delete[] buf;
+
+  return 0;
+}
+
+
+int SyntheticClient::write_batch(int nfile, int size, int wrsize)
+{
+  for (int i=0; i<nfile; i++) {
+      string sarg1 = get_sarg(i);
+    dout(0) << "Write file " << sarg1 << dendl;
+    write_file(sarg1, size, wrsize);
+  }
+  return 0;
+}
+
+// size is in MB, wrsize in bytes
+int SyntheticClient::read_file(const std::string& fn, int size,
+			       int rdsize, bool ignoreprint)
+{
+  char *buf = new char[rdsize]; 
+  memset(buf, 1, rdsize);
+  uint64_t chunks = (uint64_t)size * (uint64_t)(1024*1024) / (uint64_t)rdsize;
+  UserPerm perms = client->pick_my_perms();
+
+  int fd = client->open(fn.c_str(), O_RDONLY, perms);
+  dout(5) << "reading from " << fn << " fd " << fd << dendl;
+  if (fd < 0) {
+    delete[] buf;
+    return fd;
+  }
+
+  utime_t from = ceph_clock_now();
+  utime_t start = from;
+  uint64_t bytes = 0, total = 0;
+
+  for (unsigned i=0; i<chunks; i++) {
+    if (time_to_stop()) break;
+    dout(2) << "reading block " << i << "/" << chunks << dendl;
+    int r = client->read(fd, buf, rdsize, i*rdsize);
+    if (r < rdsize) {
+      dout(1) << "read_file got r = " << r << ", probably end of file" << dendl;
+      break;
+    }
+ 
+    bytes += rdsize;
+    total += rdsize;
+
+    utime_t now = ceph_clock_now();
+    if (now - from >= 1.0) {
+      double el = now - from;
+      dout(0) << "read " << (bytes / el / 1048576.0) << " MB/sec" << dendl;
+      from = now;
+      bytes = 0;
+    }
+
+    // verify fingerprint
+    int bad = 0;
+    uint64_t *p = (uint64_t*)buf;
+    while ((char*)p + 32 < buf + rdsize) {
+      uint64_t readoff = *p;
+      uint64_t wantoff = (uint64_t)i*(uint64_t)rdsize + (uint64_t)((char*)p - buf);
+      p++;
+      int64_t readclient = *p;
+      p++;
+      if (readoff != wantoff ||
+	  readclient != client->get_nodeid()) {
+        if (!bad && !ignoreprint)
+          dout(0) << "WARNING: wrong data from OSD, block says fileoffset=" << readoff << " client=" << readclient
+		  << ", should be offset " << wantoff << " client " << client->get_nodeid()
+		  << dendl;
+        bad++;
+      }
+    }
+    if (bad && !ignoreprint) 
+      dout(0) << " + " << (bad-1) << " other bad 16-byte bits in this block" << dendl;
+  }
+
+  utime_t stop = ceph_clock_now();
+  double el = stop - start;
+  dout(0) << "read total " << (total / el / 1048576.0) << " MB/sec ("
+	  << total << " bytes in " << el << " seconds)" << dendl;
+
+  client->close(fd);
+  delete[] buf;
+
+  return 0;
+}
+
+
+
+
+class C_Ref : public Context {
+  ceph::mutex& lock;
+  ceph::condition_variable& cond;
+  int *ref;
+public:
+  C_Ref(ceph::mutex &l, ceph::condition_variable &c, int *r)
+    : lock(l), cond(c), ref(r) {
+    lock_guard locker{lock};
+    (*ref)++;
+  }
+  void finish(int) override {
+    lock_guard locker{lock};
+    (*ref)--;
+    cond.notify_all();
+  }
+};
+
+int SyntheticClient::create_objects(int nobj, int osize, int inflight)
+{
+  // divy up
+  int numc = num_client ? num_client : 1;
+
+  int start, inc, end;
+
+  if (1) {
+    // strided
+    start = client->get_nodeid().v; //nobjs % numc;
+    inc = numc;
+    end = start + nobj;
+  } else {
+    // segments
+    start = nobj * client->get_nodeid().v / numc;
+    inc = 1;
+    end = nobj * (client->get_nodeid().v+1) / numc;
+  }
+
+  dout(5) << "create_objects " << nobj << " size=" << osize 
+	  << " .. doing [" << start << "," << end << ") inc " << inc
+	  << dendl;
+  
+  bufferptr bp(osize);
+  bp.zero();
+  bufferlist bl;
+  bl.push_back(bp);
+
+  ceph::mutex lock = ceph::make_mutex("create_objects lock");
+  ceph::condition_variable cond;
+  
+  int unsafe = 0;
+  
+  list<utime_t> starts;
+
+  for (int i=start; i<end; i += inc) {
+    if (time_to_stop()) break;
+
+    object_t oid = file_object_t(999, i);
+    object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+    SnapContext snapc;
+    
+    if (i % inflight == 0) {
+      dout(6) << "create_objects " << i << "/" << (nobj+1) << dendl;
+    }
+    dout(10) << "writing " << oid << dendl;
+
+    starts.push_back(ceph_clock_now());
+    {
+      std::lock_guard locker{client->client_lock};
+      client->objecter->write(oid, oloc, 0, osize, snapc, bl,
+			      ceph::real_clock::now(), 0,
+			      new C_Ref(lock, cond, &unsafe));
+    }
+    {
+      std::unique_lock locker{lock};
+      cond.wait(locker, [&unsafe, inflight, this] {
+        if (unsafe > inflight) {
+	  dout(20) << "waiting for " << unsafe << " unsafe" << dendl;
+	}
+	return unsafe <= inflight;
+      });
+    }
+    utime_t lat = ceph_clock_now();
+    lat -= starts.front();
+    starts.pop_front();
+  }
+  {
+    std::unique_lock locker{lock};
+    cond.wait(locker, [&unsafe, this] {
+      if (unsafe > 0) {
+	dout(10) << "waiting for " << unsafe << " unsafe" << dendl;
+      }
+      return unsafe <= 0;
+    });
+  }
+  dout(5) << "create_objects done" << dendl;
+  return 0;
+}
+
+int SyntheticClient::object_rw(int nobj, int osize, int wrpc, 
+			       int overlappc,
+			       double rskew, double wskew)
+{
+  dout(5) << "object_rw " << nobj << " size=" << osize << " with "
+	  << wrpc << "% writes" 
+	  << ", " << overlappc << "% overlap"
+	  << ", rskew = " << rskew
+	  << ", wskew = " << wskew
+	  << dendl;
+
+  bufferptr bp(osize);
+  bp.zero();
+  bufferlist bl;
+  bl.push_back(bp);
+
+  // start with odd number > nobj
+  rjhash<uint32_t> h;
+  unsigned prime = nobj + 1;             // this is the minimum!
+  prime += h(nobj) % (3*nobj);  // bump it up some
+  prime |= 1;                               // make it odd
+
+  while (true) {
+    unsigned j;
+    for (j=2; j*j<=prime; j++)
+      if (prime % j == 0) break;
+    if (j*j > prime) {
+      break;
+      //cout << "prime " << prime << endl;
+    }
+    prime += 2;
+  }
+
+  ceph::mutex lock = ceph::make_mutex("lock");
+  ceph::condition_variable cond;
+
+  int unack = 0;
+
+  while (1) {
+    if (time_to_stop()) break;
+    
+    // read or write?
+    bool write = (rand() % 100) < wrpc;
+
+    // choose object
+    double r = drand48(); // [0..1)
+    long o;
+    if (write) {
+      o = (long)trunc(pow(r, wskew) * (double)nobj);  // exponentially skew towards 0
+      int pnoremap = (long)(r * 100.0);
+      if (pnoremap >= overlappc) 
+	o = (o*prime) % nobj;    // remap
+    } else {
+      o = (long)trunc(pow(r, rskew) * (double)nobj);  // exponentially skew towards 0
+    }
+    object_t oid = file_object_t(999, o);
+    object_locator_t oloc(SYNCLIENT_FIRST_POOL);
+    SnapContext snapc;
+    
+    client->client_lock.lock();
+    utime_t start = ceph_clock_now();
+    if (write) {
+      dout(10) << "write to " << oid << dendl;
+
+      ObjectOperation m;
+      OSDOp op;
+      op.op.op = CEPH_OSD_OP_WRITE;
+      op.op.extent.offset = 0;
+      op.op.extent.length = osize;
+      op.indata = bl;
+      m.ops.push_back(op);
+      client->objecter->mutate(oid, oloc, m, snapc,
+			       ceph::real_clock::now(), 0,
+			       new C_Ref(lock, cond, &unack));
+    } else {
+      dout(10) << "read from " << oid << dendl;
+      bufferlist inbl;
+      client->objecter->read(oid, oloc, 0, osize, CEPH_NOSNAP, &inbl, 0,
+			     new C_Ref(lock, cond, &unack));
+    }
+    client->client_lock.unlock();
+
+    {
+      std::unique_lock locker{lock};
+      cond.wait(locker, [&unack, this] {
+	if (unack > 0) {
+	  dout(20) << "waiting for " << unack << " unack" << dendl;
+	}
+	return unack <= 0;
+      });
+    }
+
+    utime_t lat = ceph_clock_now();
+    lat -= start;
+  }
+
+  return 0;
+}
+
+
+
+
+
+int SyntheticClient::read_random(string& fn, int size, int rdsize)   // size is in MB, wrsize in bytes
+{
+  UserPerm perms = client->pick_my_perms();
+  uint64_t chunks = (uint64_t)size * (uint64_t)(1024*1024) / (uint64_t)rdsize;
+  int fd = client->open(fn.c_str(), O_RDWR, perms);
+  dout(5) << "reading from " << fn << " fd " << fd << dendl;
+
+  if (fd < 0) return fd;
+  int offset = 0;
+  char * buf = NULL;
+
+  for (unsigned i=0; i<2000; i++) {
+    if (time_to_stop()) break;
+
+    bool read=false;
+
+    time_t seconds;
+    time( &seconds);
+    srand(seconds);
+
+    // use rand instead ??
+    double x = drand48();
+
+    // cleanup before call 'new'
+    if (buf != NULL) {
+	delete[] buf;
+	buf = NULL;
+    }
+    if (x < 0.5) {
+        buf = new char[rdsize]; 
+        memset(buf, 1, rdsize);
+        read=true;
+    } else {
+        buf = new char[rdsize+100];   // 1 MB
+        memset(buf, 7, rdsize);
+    }
+    
+    if (read) {
+        offset=(rand())%(chunks+1);
+        dout(2) << "reading block " << offset << "/" << chunks << dendl;
+
+        int r = client->read(fd, buf, rdsize, offset*rdsize);
+        if (r < rdsize) {
+	  dout(1) << "read_file got r = " << r << ", probably end of file" << dendl;
+	}
+    } else {
+      dout(2) << "writing block " << offset << "/" << chunks << dendl;
+
+      // fill buf with a 16 byte fingerprint
+      // 64 bits : file offset
+      // 64 bits : client id
+      // = 128 bits (16 bytes)
+
+      offset=(rand())%(chunks+1);
+      uint64_t *p = (uint64_t*)buf;
+      while ((char*)p < buf + rdsize) {
+	*p = offset*rdsize + (char*)p - buf;      
+	p++;
+	*p = client->get_nodeid().v;
+	p++;
+      }
+
+      client->write(fd, buf, rdsize,
+                        offset*rdsize);
+    }
+
+    // verify fingerprint
+    if (read) {
+      int bad = 0;
+      int64_t *p = (int64_t*)buf;
+      while ((char*)p + 32 < buf + rdsize) {
+	int64_t readoff = *p;
+	int64_t wantoff = offset*rdsize + (int64_t)((char*)p - buf);
+	p++;
+	int64_t readclient = *p;
+	p++;
+	if (readoff != wantoff || readclient != client->get_nodeid()) {
+	  if (!bad)
+	    dout(0) << "WARNING: wrong data from OSD, block says fileoffset=" << readoff << " client=" << readclient
+		    << ", should be offset " << wantoff << " client " << client->get_nodeid()
+		    << dendl;
+	  bad++;
+	}
+      }
+      if (bad) 
+	dout(0) << " + " << (bad-1) << " other bad 16-byte bits in this block" << dendl;
+    }
+  }
+  
+  client->close(fd);
+  delete[] buf;
+
+  return 0;
+}
+
+int normdist(int min, int max, int stdev) /* specifies input values */
+{
+  /* min: Minimum value; max: Maximum value; stdev: degree of deviation */
+  
+  //int min, max, stdev; {
+  time_t seconds;
+  time( &seconds);
+  srand(seconds);
+  
+  int range, iterate, result;
+  /* declare range, iterate and result as integers, to avoid the need for
+     floating point math*/
+  
+  result = 0;
+  /* ensure result is initialized to 0 */
+  
+  range = max -min;
+  /* calculate range of possible values between the max and min values */
+  
+  iterate = range / stdev;
+  /* this number of iterations ensures the proper shape of the resulting
+     curve */
+  
+  stdev += 1; /* compensation for integer vs. floating point math */
+  for (int c = iterate; c != 0; c--) /* loop through iterations */
+    {
+      //  result += (uniform (1, 100) * stdev) / 100; /* calculate and
+      result += ( (rand()%100 + 1)  * stdev) / 100;
+      // printf("result=%d\n", result );
+    }
+  printf("\n final result=%d\n", result );
+  return result + min; /* send final result back */
+}
+
+int SyntheticClient::read_random_ex(string& fn, int size, int rdsize)   // size is in MB, wrsize in bytes
+{
+  uint64_t chunks = (uint64_t)size * (uint64_t)(1024*1024) / (uint64_t)rdsize;
+  UserPerm perms = client->pick_my_perms();
+  int fd = client->open(fn.c_str(), O_RDWR, perms);
+  dout(5) << "reading from " << fn << " fd " << fd << dendl;
+  
+  if (fd < 0) return fd;
+  int offset = 0;
+  char * buf = NULL;
+  
+  for (unsigned i=0; i<2000; i++) {
+    if (time_to_stop()) break;
+    
+    bool read=false;
+    
+    time_t seconds;
+    time( &seconds);
+    srand(seconds);
+    
+    // use rand instead ??
+    double x = drand48();
+    
+    // cleanup before call 'new'
+    if (buf != NULL) {
+      delete[] buf;
+      buf = NULL;
+    }
+    if (x < 0.5) {
+      buf = new char[rdsize]; 
+      memset(buf, 1, rdsize);
+      read=true;
+    } else {
+      buf = new char[rdsize+100];   // 1 MB
+      memset(buf, 7, rdsize);
+    }
+    
+    if (read) {
+      dout(2) << "reading block " << offset << "/" << chunks << dendl;
+	
+      int r = client->read(fd, buf, rdsize,
+			     offset*rdsize);
+      if (r < rdsize) {
+	dout(1) << "read_file got r = " << r << ", probably end of file" << dendl;
+      }
+    } else {
+        dout(2) << "writing block " << offset << "/" << chunks << dendl;
+	
+	// fill buf with a 16 byte fingerprint
+	// 64 bits : file offset
+	// 64 bits : client id
+	// = 128 bits (16 bytes)
+	
+	int count = rand()%10;
+	
+	for ( int j=0;j<count; j++ ) {
+	  offset=(rand())%(chunks+1);
+	  uint64_t *p = (uint64_t*)buf;
+	  while ((char*)p < buf + rdsize) {
+	    *p = offset*rdsize + (char*)p - buf;      
+	    p++;
+	    *p = client->get_nodeid().v;
+	    p++;
+	  }
+	    
+	  client->write(fd, buf, rdsize, offset*rdsize);
+	}
+    }
+    
+    // verify fingerprint
+    if (read) {
+      int bad = 0;
+      int64_t *p = (int64_t*)buf;
+      while ((char*)p + 32 < buf + rdsize) {
+	int64_t readoff = *p;
+	int64_t wantoff = offset*rdsize + (int64_t)((char*)p - buf);
+	p++;
+	int64_t readclient = *p;
+	p++;
+	if (readoff != wantoff || readclient != client->get_nodeid()) { 
+	  if (!bad)
+	    dout(0) << "WARNING: wrong data from OSD, block says fileoffset=" << readoff << " client=" << readclient
+		    << ", should be offset " << wantoff << " client " << client->get_nodeid()
+		    << dendl;
+	  bad++;
+	}
+      }
+      if (bad) 
+	dout(0) << " + " << (bad-1) << " other bad 16-byte bits in this block" << dendl;
+    }
+  }
+  
+  client->close(fd);
+  delete[] buf;
+
+  return 0;
+}
+
+
+int SyntheticClient::random_walk(int num_req)
+{
+  int left = num_req;
+
+  //dout(1) << "random_walk() will do " << left << " ops" << dendl;
+
+  init_op_dist();  // set up metadata op distribution
+ 
+  UserPerm perms = client->pick_my_perms();
+  while (left > 0) {
+    left--;
+
+    if (time_to_stop()) break;
+
+    // ascend?
+    if (cwd.depth() && !roll_die(::pow((double).9, (double)cwd.depth()))) {
+      dout(DBL) << "die says up" << dendl;
+      up();
+      continue;
+    }
+
+    // descend?
+    if (roll_die(::pow((double).9,(double)cwd.depth())) && !subdirs.empty()) {
+      string s = get_random_subdir();
+      cwd.push_dentry( s );
+      dout(DBL) << "cd " << s << " -> " << cwd << dendl;
+      clear_dir();
+      continue;
+    }
+
+    int op = 0;
+    filepath path;
+
+    if (contents.empty() && roll_die(.3)) {
+      if (did_readdir) {
+        dout(DBL) << "empty dir, up" << dendl;
+        up();
+      } else
+        op = CEPH_MDS_OP_READDIR;
+    } else {
+      op = op_dist.sample();
+    }
+    //dout(DBL) << "op is " << op << dendl;
+
+    int r = 0;
+
+    // do op
+    if (op == CEPH_MDS_OP_UNLINK) {
+      if (contents.empty())
+        op = CEPH_MDS_OP_READDIR;
+      else 
+        r = client->unlink(get_random_sub(), perms);   // will fail on dirs
+    }
+     
+    if (op == CEPH_MDS_OP_RENAME) {
+      if (contents.empty())
+        op = CEPH_MDS_OP_READDIR;
+      else {
+        r = client->rename(get_random_sub(), make_sub("ren"), perms);
+      }
+    }
+    
+    if (op == CEPH_MDS_OP_MKDIR) {
+      r = client->mkdir(make_sub("mkdir"), 0755, perms);
+    }
+    
+    if (op == CEPH_MDS_OP_RMDIR) {
+      if (!subdirs.empty())
+        r = client->rmdir(get_random_subdir(), perms);
+      else
+        r = client->rmdir(cwd.c_str(), perms);     // will pbly fail
+    }
+    
+    if (op == CEPH_MDS_OP_SYMLINK) {
+    }
+    /*
+    if (op == CEPH_MDS_OP_CHMOD) {
+      if (contents.empty())
+        op = CEPH_MDS_OP_READDIR;
+      else
+        r = client->chmod(get_random_sub(), rand() & 0755, perms);
+    }
+    
+    if (op == CEPH_MDS_OP_CHOWN) {
+      if (contents.empty())         r = client->chown(cwd.c_str(), rand(), rand(), perms);
+      else
+        r = client->chown(get_random_sub(), rand(), rand(), perms);
+    }
+     
+    if (op == CEPH_MDS_OP_UTIME) {
+      struct utimbuf b;
+      memset(&b, 1, sizeof(b));
+      if (contents.empty()) 
+        r = client->utime(cwd.c_str(), &b, perms);
+      else
+        r = client->utime(get_random_sub(), &b, perms);
+    }
+    */
+    if (op == CEPH_MDS_OP_LINK) {
+    }
+    
+    if (op == CEPH_MDS_OP_MKNOD) {
+      r = client->mknod(make_sub("mknod"), 0644, perms);
+    }
+     
+    if (op == CEPH_MDS_OP_OPEN) {
+      if (contents.empty())
+        op = CEPH_MDS_OP_READDIR;
+      else {
+        r = client->open(get_random_sub(), O_RDONLY, perms);
+        if (r > 0) {
+          ceph_assert(open_files.count(r) == 0);
+          open_files.insert(r);
+        }
+      }
+    }
+
+    /*if (op == CEPH_MDS_OP_RELEASE) {   // actually, close
+      if (open_files.empty())
+        op = CEPH_MDS_OP_STAT;
+      else {
+        int fh = get_random_fh();
+        r = client->close( fh );
+        if (r == 0) open_files.erase(fh);
+      }
+    }
+    */
+    
+    if (op == CEPH_MDS_OP_GETATTR) {
+      struct stat st;
+      if (contents.empty()) {
+        if (did_readdir) {
+          if (roll_die(.1)) {
+            dout(DBL) << "stat in empty dir, up" << dendl;
+            up();
+          } else {
+            op = CEPH_MDS_OP_MKNOD;
+          }
+        } else
+          op = CEPH_MDS_OP_READDIR;
+      } else
+        r = client->lstat(get_random_sub(), &st, perms);
+    }
+
+    if (op == CEPH_MDS_OP_READDIR) {
+      clear_dir();
+      
+      list<string> c;
+      r = client->getdir(cwd.c_str(), c, perms);
+      
+      for (list<string>::iterator it = c.begin();
+           it != c.end();
+           ++it) {
+        //dout(DBL) << " got " << *it << dendl;
+	ceph_abort();
+	/*contents[*it] = it->second;
+        if (it->second &&
+	    S_ISDIR(it->second->st_mode)) 
+          subdirs.insert(*it);
+	*/
+      }
+      
+      did_readdir = true;
+    }
+      
+    // errors?
+    if (r < 0) {
+      // reevaluate cwd.
+      //while (cwd.depth()) {
+      //if (client->lookup(cwd)) break;   // it's in the cache
+        
+      //dout(DBL) << "r = " << r << ", client doesn't have " << cwd << ", cd .." << dendl;
+      dout(DBL) << "r = " << r << ", client may not have " << cwd << ", cd .." << dendl;
+      up();
+      //}      
+    }
+  }
+
+  // close files
+  dout(DBL) << "closing files" << dendl;
+  while (!open_files.empty()) {
+    int fh = get_random_fh();
+    int r = client->close( fh );
+    if (r == 0) open_files.erase(fh);
+  }
+
+  dout(DBL) << "done" << dendl;
+  return 0;
+}
+
+
+
+
+void SyntheticClient::make_dir_mess(const char *basedir, int n)
+{
+  UserPerm perms = client->pick_my_perms();
+  vector<string> dirs;
+  
+  dirs.push_back(basedir);
+  dirs.push_back(basedir);
+  
+  client->mkdir(basedir, 0755, perms);
+
+  // motivation:
+  //  P(dir) ~ subdirs_of(dir) + 2
+  // from 5-year metadata workload paper in fast'07
+
+  // create dirs
+  for (int i=0; i<n; i++) {
+    // pick a dir
+    int k = rand() % dirs.size();
+    string parent = dirs[k];
+    
+    // pick a name
+    std::stringstream ss;
+    ss << parent << "/" << i;
+    string dir = ss.str();
+
+    // update dirs
+    dirs.push_back(parent);
+    dirs.push_back(dir);
+    dirs.push_back(dir);
+
+    // do it
+    client->mkdir(dir.c_str(), 0755, perms);
+  }
+    
+  
+}
+
+
+
+void SyntheticClient::foo()
+{
+  UserPerm perms = client->pick_my_perms();
+
+  if (1) {
+    // make 2 parallel dirs, link/unlink between them.
+    char a[100], b[100];
+    client->mkdir("/a", 0755, perms);
+    client->mkdir("/b", 0755, perms);
+    for (int i=0; i<10; i++) {
+      snprintf(a, sizeof(a), "/a/%d", i);
+      client->mknod(a, 0644, perms);
+    }
+    while (1) {
+      for (int i=0; i<10; i++) {
+	snprintf(a, sizeof(a), "/a/%d", i);
+	snprintf(b, sizeof(b), "/b/%d", i);
+	client->link(a, b, perms);
+      }
+      for (int i=0; i<10; i++) {
+	snprintf(b, sizeof(b), "/b/%d", i);
+	client->unlink(b, perms);
+      }
+    }
+    return;
+  }
+  if (1) {
+    // bug1.cpp
+    const char *fn = "blah";
+    char buffer[8192]; 
+    client->unlink(fn, perms);
+    int handle = client->open(fn, O_CREAT|O_RDWR, perms, S_IRWXU);
+    ceph_assert(handle>=0);
+    int r=client->write(handle,buffer,8192);
+    ceph_assert(r>=0);
+    r=client->close(handle);
+    ceph_assert(r>=0);
+         
+    handle = client->open(fn, O_RDWR, perms); // open the same  file, it must have some data already
+    ceph_assert(handle>=0);      
+    r=client->read(handle,buffer,8192);
+    ceph_assert(r==8192); //  THIS ASSERTION FAILS with disabled cache
+    r=client->close(handle);
+    ceph_assert(r>=0);
+
+    return;
+  }
+  if (1) {
+    dout(0) << "first" << dendl;
+    int fd = client->open("tester", O_WRONLY|O_CREAT, perms);
+    client->write(fd, "hi there", 0, 8);
+    client->close(fd);
+    dout(0) << "sleep" << dendl;
+    sleep(10);
+    dout(0) << "again" << dendl;
+    fd = client->open("tester", O_WRONLY|O_CREAT, perms);
+    client->write(fd, "hi there", 0, 8);
+    client->close(fd);
+    return;    
+  }
+  if (1) {
+    // open some files
+    srand(0);
+    for (int i=0; i<20; i++) {
+      int s = 5;
+      int a = rand() % s;
+      int b = rand() % s;
+      int c = rand() % s;
+      char src[80];
+      snprintf(src, sizeof(src), "syn.0.0/dir.%d/dir.%d/file.%d", a, b, c);
+      //int fd = 
+      client->open(src, O_RDONLY, perms);
+    }
+
+    return;
+  }
+
+  if (0) {
+    // rename fun
+    for (int i=0; i<100; i++) {
+      int s = 5;
+      int a = rand() % s;
+      int b = rand() % s;
+      int c = rand() % s;
+      int d = rand() % s;
+      int e = rand() % s;
+      int f = rand() % s;
+      char src[80];
+      char dst[80];
+      snprintf(src, sizeof(src), "syn.0.0/dir.%d/dir.%d/file.%d", a, b, c);
+      snprintf(dst, sizeof(dst), "syn.0.0/dir.%d/dir.%d/file.%d", d, e, f);
+      client->rename(src, dst, perms);
+    }
+    return;
+  }
+
+  if (1) {
+    // link fun
+    srand(0);
+    for (int i=0; i<100; i++) {
+      int s = 5;
+      int a = rand() % s;
+      int b = rand() % s;
+      int c = rand() % s;
+      int d = rand() % s;
+      int e = rand() % s;
+      int f = rand() % s;
+      char src[80];
+      char dst[80];
+      snprintf(src, sizeof(src), "syn.0.0/dir.%d/dir.%d/file.%d", a, b, c);
+      snprintf(dst, sizeof(dst), "syn.0.0/dir.%d/dir.%d/newlink.%d", d, e, f);
+      client->link(src, dst, perms);
+    }
+    srand(0);
+    for (int i=0; i<100; i++) {
+      int s = 5;
+      int a = rand() % s;
+      int b = rand() % s;
+      int c = rand() % s;
+      int d = rand() % s;
+      int e = rand() % s;
+      int f = rand() % s;
+      char src[80];
+      char dst[80];
+      snprintf(src, sizeof(src), "syn.0.0/dir.%d/dir.%d/file.%d", a, b, c);
+      snprintf(dst, sizeof(dst), "syn.0.0/dir.%d/dir.%d/newlink.%d", d, e, f);
+      client->unlink(dst, perms);
+    }
+
+    
+    return;
+  }
+
+  // link fun
+  client->mknod("one", 0755, perms);
+  client->mknod("two", 0755, perms);
+  client->link("one", "three", perms);
+  client->mkdir("dir", 0755, perms);
+  client->link("two", "/dir/twolink", perms);
+  client->link("dir/twolink", "four", perms);
+  
+  // unlink fun
+  client->mknod("a", 0644, perms);
+  client->unlink("a", perms);
+  client->mknod("b", 0644, perms);
+  client->link("b", "c", perms);
+  client->unlink("c", perms);
+  client->mkdir("d", 0755, perms);
+  client->unlink("d", perms);
+  client->rmdir("d", perms);
+
+  // rename fun
+  client->mknod("p1", 0644, perms);
+  client->mknod("p2", 0644, perms);
+  client->rename("p1","p2", perms);
+  client->mknod("p3", 0644, perms);
+  client->rename("p3","p4", perms);
+
+  // check dest dir ambiguity thing
+  client->mkdir("dir1", 0755, perms);
+  client->mkdir("dir2", 0755, perms);
+  client->rename("p2", "dir1/p2", perms);
+  client->rename("dir1/p2", "dir2/p2", perms);
+  client->rename("dir2/p2", "/p2", perms);
+  
+  // check primary+remote link merging
+  client->link("p2","p2.l", perms);
+  client->link("p4","p4.l", perms);
+  client->rename("p2.l", "p2", perms);
+  client->rename("p4", "p4.l", perms);
+
+  // check anchor updates
+  client->mknod("dir1/a", 0644, perms);
+  client->link("dir1/a", "da1", perms);
+  client->link("dir1/a", "da2", perms);
+  client->link("da2","da3", perms);
+  client->rename("dir1/a", "dir2/a", perms);
+  client->rename("dir2/a", "da2", perms);
+  client->rename("da1", "da2", perms);
+  client->rename("da2", "da3", perms);
+
+  // check directory renames
+  client->mkdir("dir3", 0755, perms);
+  client->mknod("dir3/asdf", 0644, perms);
+  client->mkdir("dir4", 0755, perms);
+  client->mkdir("dir5", 0755, perms);
+  client->mknod("dir5/asdf", 0644, perms);
+  client->rename("dir3", "dir4", perms); // ok
+  client->rename("dir4", "dir5", perms); // fail
+}
+
+int SyntheticClient::thrash_links(const char *basedir, int dirs, int files, int depth, int n)
+{
+  dout(1) << "thrash_links " << basedir << " " << dirs << " " << files << " " << depth
+	  << " links " << n
+	  << dendl;
+
+  if (time_to_stop()) return 0;
+
+  UserPerm perms = client->pick_my_perms();
+
+  srand(0);
+  if (1) {
+    bool renames = true; // thrash renames too?
+    for (int k=0; k<n; k++) {
+      
+      if (renames && rand() % 10 == 0) {
+	// rename some directories.  whee!
+	int dep = (rand() % depth) + 1;
+	string src = basedir;
+	{
+	  char t[80];
+	  for (int d=0; d<dep; d++) {
+	    int a = rand() % dirs;
+	    snprintf(t, sizeof(t), "/dir.%d", a);
+	    src += t;
+	  }
+	}
+	string dst = basedir;
+	{
+	  char t[80];
+	  for (int d=0; d<dep; d++) {
+	    int a = rand() % dirs;
+	    snprintf(t, sizeof(t), "/dir.%d", a);
+	    dst += t;
+	  }
+	}
+	
+	if (client->rename(dst.c_str(), "/tmp", perms) == 0) {
+	  client->rename(src.c_str(), dst.c_str(), perms);
+	  client->rename("/tmp", src.c_str(), perms);
+	}
+	continue;
+      } 
+      
+      // pick a dest dir
+      string src = basedir;
+      {
+	char t[80];
+	for (int d=0; d<depth; d++) {
+	  int a = rand() % dirs;
+	  snprintf(t, sizeof(t), "/dir.%d", a);
+	  src += t;
+	}
+	int a = rand() % files;
+	snprintf(t, sizeof(t), "/file.%d", a);
+	src += t;
+      }
+      string dst = basedir;
+      {
+	char t[80];
+	for (int d=0; d<depth; d++) {
+	  int a = rand() % dirs;
+	  snprintf(t, sizeof(t), "/dir.%d", a);
+	  dst += t;
+	}
+	int a = rand() % files;
+	snprintf(t, sizeof(t), "/file.%d", a);
+	dst += t;
+      }
+      
+      int o = rand() % 4;
+      switch (o) {
+      case 0: 
+	client->mknod(src.c_str(), 0755, perms);
+	if (renames) client->rename(src.c_str(), dst.c_str(), perms);
+	break;
+      case 1: 
+	client->mknod(src.c_str(), 0755, perms);
+	client->unlink(dst.c_str(), perms);
+	client->link(src.c_str(), dst.c_str(), perms); 
+	break;
+      case 2: client->unlink(src.c_str(), perms); break;
+      case 3: client->unlink(dst.c_str(), perms); break;
+	//case 4: client->mknod(src.c_str(), 0755, perms); break;
+	//case 5: client->mknod(dst.c_str(), 0755, perms); break;
+      }
+    }
+    return 0;
+  }
+
+  if (1) {
+    // now link shit up
+    for (int i=0; i<n; i++) {
+      if (time_to_stop()) return 0;
+      
+      char f[20];
+      
+      // pick a file
+      string file = basedir;
+      
+      if (depth) {
+	int d = rand() % (depth+1);
+	for (int k=0; k<d; k++) {
+	  snprintf(f, sizeof(f), "/dir.%d", rand() % dirs);
+	  file += f;
+	}
+      }
+      snprintf(f, sizeof(f), "/file.%d", rand() % files);
+      file += f;
+      
+      // pick a dir for our link
+      string ln = basedir;
+      if (depth) {
+	int d = rand() % (depth+1);
+	for (int k=0; k<d; k++) {
+	  snprintf(f, sizeof(f), "/dir.%d", rand() % dirs);
+	  ln += f;
+	}
+      }
+      snprintf(f, sizeof(f), "/ln.%d", i);
+      ln += f;
+      
+      client->link(file.c_str(), ln.c_str(), perms);
+    }
+  }
+  return 0;
+}
+
+
+
+
+void SyntheticClient::import_find(const char *base, const char *find, bool data)
+{
+  dout(1) << "import_find " << base << " from " << find << " data=" << data << dendl;
+
+  /* use this to gather the static trace:
+   *
+   *  find . -exec ls -dilsn --time-style=+%s \{\} \;
+   * or if it's wafl,
+   *  find . -path ./.snapshot -prune -o -exec ls -dilsn --time-style=+%s \{\} \;
+   *
+   */
+
+  UserPerm process_perms = client->pick_my_perms();
+
+  if (base[0] != '-') 
+    client->mkdir(base, 0755, process_perms);
+
+  ifstream f(find);
+  ceph_assert(f.is_open());
+  
+  int dirnum = 0;
+
+  while (!f.eof()) {
+    uint64_t ino;
+    int dunno, nlink;
+    string modestring;
+    int uid, gid;
+    off_t size;
+    time_t mtime;
+    string filename;
+    f >> ino;
+    if (f.eof()) break;
+    f >> dunno;
+    f >> modestring;
+    f >> nlink;
+    f >> uid;
+    f >> gid;
+    f >> size;
+    f >> mtime;
+    f.seekg(1, ios::cur);
+    getline(f, filename);
+    UserPerm perms(uid, gid);
+
+    // ignore "."
+    if (filename == ".") continue;
+
+    // remove leading ./
+    ceph_assert(filename[0] == '.' && filename[1] == '/');
+    filename = filename.substr(2);
+
+    // new leading dir?
+    int sp = filename.find("/");
+    if (sp < 0) dirnum++;
+
+    //dout(0) << "leading dir " << filename << " " << dirnum << dendl;
+    if (dirnum % num_client != client->get_nodeid()) {
+      dout(20) << "skipping leading dir " << dirnum << " " << filename << dendl;
+      continue;
+    }
+
+    // parse the mode
+    ceph_assert(modestring.length() == 10);
+    mode_t mode = 0;
+    switch (modestring[0]) {
+    case 'd': mode |= S_IFDIR; break;
+    case 'l': mode |= S_IFLNK; break;
+    default:
+    case '-': mode |= S_IFREG; break;
+    }
+    if (modestring[1] == 'r') mode |= 0400;
+    if (modestring[2] == 'w') mode |= 0200;
+    if (modestring[3] == 'x') mode |= 0100;
+    if (modestring[4] == 'r') mode |= 040;
+    if (modestring[5] == 'w') mode |= 020;
+    if (modestring[6] == 'x') mode |= 010;
+    if (modestring[7] == 'r') mode |= 04;
+    if (modestring[8] == 'w') mode |= 02;
+    if (modestring[9] == 'x') mode |= 01;
+
+    dout(20) << " mode " << modestring << " to " << oct << mode << dec << dendl;
+
+    if (S_ISLNK(mode)) {
+      // target vs destination
+      int pos = filename.find(" -> ");
+      ceph_assert(pos > 0);
+      string link;
+      if (base[0] != '-') {
+	link = base;
+	link += "/";
+      }
+      link += filename.substr(0, pos);
+      string target;
+      if (filename[pos+4] == '/') {
+	if (base[0] != '-') 
+	  target = base;
+	target += filename.substr(pos + 4);
+      } else {
+	target = filename.substr(pos + 4);
+      }
+      dout(10) << "symlink from '" << link << "' -> '" << target << "'" << dendl;
+      client->symlink(target.c_str(), link.c_str(), perms);
+    } else {
+      string f;
+      if (base[0] != '-') {
+	f = base;
+	f += "/";
+      }
+      f += filename;
+      if (S_ISDIR(mode)) {
+	client->mkdir(f.c_str(), mode, perms);
+      } else {
+	int fd = client->open(f.c_str(), O_WRONLY|O_CREAT, perms, mode & 0777);
+	ceph_assert(fd > 0);	
+	if (data) {
+	  client->write(fd, "", 0, size);
+	} else {
+	  client->truncate(f.c_str(), size, perms);
+	}
+	client->close(fd);
+
+	//client->chmod(f.c_str(), mode & 0777, perms, process_perms);
+	client->chown(f.c_str(), uid, gid, process_perms);
+
+	struct utimbuf ut;
+	ut.modtime = mtime;
+	ut.actime = mtime;
+	client->utime(f.c_str(), &ut, perms);
+      }
+    }
+  }
+  
+
+}
+
+
+int SyntheticClient::lookup_hash(inodeno_t ino, inodeno_t dirino,
+				 const char *name, const UserPerm& perms)
+{
+  int r = client->lookup_hash(ino, dirino, name, perms);
+  dout(0) << "lookup_hash(" << ino << ", #" << dirino << "/" << name << ") = " << r << dendl;
+  return r;
+}
+
+int SyntheticClient::lookup_ino(inodeno_t ino, const UserPerm& perms)
+{
+  int r = client->lookup_ino(ino, perms);
+  dout(0) << "lookup_ino(" << ino << ") = " << r << dendl;
+  return r;
+}
+
+int SyntheticClient::chunk_file(string &filename)
+{
+  UserPerm perms = client->pick_my_perms();
+  int fd = client->open(filename.c_str(), O_RDONLY, perms);
+  if (fd < 0)
+    return fd;
+
+  struct stat st;
+  int ret = client->fstat(fd, &st, perms);
+  if (ret < 0) {
+    client->close(fd);
+    return ret;
+  }
+  uint64_t size = st.st_size;
+  dout(0) << "file " << filename << " size is " << size << dendl;
+
+  inode_t inode{};
+  inode.ino = st.st_ino;
+  ret = client->fdescribe_layout(fd, &inode.layout);
+  ceph_assert(ret == 0); // otherwise fstat did a bad thing
+
+  uint64_t pos = 0;
+  bufferlist from_before;
+  while (pos < size) {
+    int get = std::min<int>(size - pos, 1048576);
+
+    ceph::mutex flock = ceph::make_mutex("synclient chunk_file lock");
+    ceph::condition_variable cond;
+    bool done;
+    bufferlist bl;
+    {
+      std::unique_lock locker{flock};
+      Context *onfinish = new C_SafeCond(flock, cond, &done);
+      client->filer->read(inode.ino, &inode.layout, CEPH_NOSNAP, pos, get, &bl, 0,
+			  onfinish);
+      cond.wait(locker, [&done] { return done; });
+    }
+    dout(0) << "got " << bl.length() << " bytes at " << pos << dendl;
+    
+    if (from_before.length()) {
+      dout(0) << " including bit from previous block" << dendl;
+      pos -= from_before.length();
+      from_before.claim_append(bl);
+      bl.swap(from_before);
+    }      
+
+    // ....
+
+    // keep last 32 bytes around
+    from_before.clear();
+    from_before.substr_of(bl, bl.length()-32, 32);
+
+    pos += bl.length();
+  }
+
+  client->close(fd);
+  return 0;
+}
+
+
+
+void SyntheticClient::mksnap(const char *base, const char *name, const UserPerm& perms)
+{
+  client->mksnap(base, name, perms);
+}
+
+void SyntheticClient::rmsnap(const char *base, const char *name, const UserPerm& perms)
+{
+  client->rmsnap(base, name, perms);
+}
+
+void SyntheticClient::mksnapfile(const char *dir)
+{
+  UserPerm perms = client->pick_my_perms();
+  client->mkdir(dir, 0755, perms);
+
+  string f = dir;
+  f += "/foo";
+  int fd = client->open(f.c_str(), O_WRONLY|O_CREAT|O_TRUNC, perms);
+
+  char buf[1048576*4];
+  client->write(fd, buf, sizeof(buf), 0);
+  client->fsync(fd, true);
+  client->close(fd);
+  
+  string s = dir;
+  s += "/.snap/1";
+  client->mkdir(s.c_str(), 0755, perms);
+
+  fd = client->open(f.c_str(), O_WRONLY, perms);
+  client->write(fd, buf, 1048576*2, 1048576);
+  client->fsync(fd, true);
+  client->close(fd);
+}
diff --git a/src/client/SyntheticClient.h b/src/client/SyntheticClient.h
new file mode 100644
index 000000000..faec7e26d
--- /dev/null
+++ b/src/client/SyntheticClient.h
@@ -0,0 +1,281 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+
+#ifndef CEPH_SYNTHETICCLIENT_H
+#define CEPH_SYNTHETICCLIENT_H
+
+#include <pthread.h>
+
+#include "Client.h"
+#include "include/Distribution.h"
+
+#include "Trace.h"
+
+#define SYNCLIENT_FIRST_POOL	0
+
+#define SYNCLIENT_MODE_RANDOMWALK  1
+#define SYNCLIENT_MODE_FULLWALK    2
+#define SYNCLIENT_MODE_REPEATWALK  3
+
+#define SYNCLIENT_MODE_MAKEDIRMESS  7
+#define SYNCLIENT_MODE_MAKEDIRS     8      // dirs files depth
+#define SYNCLIENT_MODE_STATDIRS     9     // dirs files depth
+#define SYNCLIENT_MODE_READDIRS     10     // dirs files depth
+
+#define SYNCLIENT_MODE_MAKEFILES    11     // num count private
+#define SYNCLIENT_MODE_MAKEFILES2   12     // num count private
+#define SYNCLIENT_MODE_CREATESHARED 13     // num
+#define SYNCLIENT_MODE_OPENSHARED   14     // num count
+
+#define SYNCLIENT_MODE_RMFILE      19
+#define SYNCLIENT_MODE_WRITEFILE   20
+#define SYNCLIENT_MODE_READFILE    21
+#define SYNCLIENT_MODE_WRITEBATCH  22
+#define SYNCLIENT_MODE_WRSHARED    23
+#define SYNCLIENT_MODE_READSHARED    24
+#define SYNCLIENT_MODE_RDWRRANDOM    25
+#define SYNCLIENT_MODE_RDWRRANDOM_EX    26
+
+#define SYNCLIENT_MODE_LINKTEST   27
+
+#define SYNCLIENT_MODE_OVERLOAD_OSD_0 28 // two args
+
+#define SYNCLIENT_MODE_DROPCACHE   29
+
+#define SYNCLIENT_MODE_TRACE       30
+
+#define SYNCLIENT_MODE_CREATEOBJECTS 35
+#define SYNCLIENT_MODE_OBJECTRW 36
+
+#define SYNCLIENT_MODE_OPENTEST     40
+#define SYNCLIENT_MODE_OPTEST       41
+
+#define SYNCLIENT_MODE_ONLY        50
+#define SYNCLIENT_MODE_ONLYRANGE   51
+#define SYNCLIENT_MODE_EXCLUDE     52
+#define SYNCLIENT_MODE_EXCLUDERANGE  53
+
+#define SYNCLIENT_MODE_UNTIL       55
+#define SYNCLIENT_MODE_SLEEPUNTIL  56
+
+#define SYNCLIENT_MODE_RANDOMSLEEP  61
+#define SYNCLIENT_MODE_SLEEP        62
+
+#define SYNCLIENT_MODE_DUMP 63
+
+#define SYNCLIENT_MODE_LOOKUPHASH     70
+#define SYNCLIENT_MODE_LOOKUPINO     71
+
+#define SYNCLIENT_MODE_TRUNCATE     200
+
+#define SYNCLIENT_MODE_FOO        100
+#define SYNCLIENT_MODE_THRASHLINKS  101
+
+#define SYNCLIENT_MODE_IMPORTFIND 300
+
+#define SYNCLIENT_MODE_CHUNK    400
+
+#define SYNCLIENT_MODE_MKSNAP 1000
+#define SYNCLIENT_MODE_RMSNAP 1001
+
+#define SYNCLIENT_MODE_MKSNAPFILE 1002
+
+
+
+void parse_syn_options(vector<const char*>& args);
+extern int num_client;
+
+class SyntheticClient {
+  StandaloneClient *client;
+  int whoami;
+
+  pthread_t thread_id;
+
+  Distribution op_dist;
+
+  void init_op_dist();
+  int get_op();
+
+  
+  filepath             cwd;
+  map<string, struct stat*> contents;
+  set<string>          subdirs;
+  bool                 did_readdir;
+  set<int>             open_files;
+
+  void up();
+
+  void clear_dir() {
+    contents.clear();
+    subdirs.clear();
+    did_readdir = false;
+  }
+
+  int get_random_fh() {
+    int r = rand() % open_files.size();
+    set<int>::iterator it = open_files.begin();
+    while (r--) ++it;
+    return *it;
+  }
+
+
+  filepath n1;
+  const char *get_random_subdir() {
+    ceph_assert(!subdirs.empty());
+    int r = ((rand() % subdirs.size()) + (rand() % subdirs.size())) / 2;  // non-uniform distn
+    set<string>::iterator it = subdirs.begin();
+    while (r--) ++it;
+
+    n1 = cwd;
+    n1.push_dentry( *it );
+    return n1.get_path().c_str();
+  }
+  filepath n2;
+  const char *get_random_sub() {
+    ceph_assert(!contents.empty());
+    int r = ((rand() % contents.size()) + (rand() % contents.size())) / 2;  // non-uniform distn
+    if (cwd.depth() && cwd.last_dentry().length()) 
+      r += cwd.last_dentry().c_str()[0];                                         // slightly permuted
+    r %= contents.size();
+
+    map<string,struct stat*>::iterator it = contents.begin();
+    while (r--) ++it;
+
+    n2 = cwd;
+    n2.push_dentry( it->first );
+    return n2.get_path().c_str();
+  }
+  
+  filepath sub;
+  char sub_s[50];
+  const char *make_sub(const char *base) {
+    snprintf(sub_s, sizeof(sub_s), "%s.%d", base, rand() % 100);
+    string f = sub_s;
+    sub = cwd;
+    sub.push_dentry(f);
+    return sub.c_str();
+  }
+
+ public:
+  SyntheticClient(StandaloneClient *client, int w = -1);
+
+  int start_thread();
+  int join_thread();
+
+  int run();
+
+  bool run_me() {
+    if (run_only >= 0) {
+      if (run_only == client->get_nodeid())
+        return true;
+      else
+	return false;
+    }
+    return true;
+  }
+  void did_run_me() {
+    run_only = -1;
+    run_until = utime_t();
+  }
+
+  // run() will do one of these things:
+  list<int> modes;
+  list<string> sargs;
+  list<int> iargs;
+  utime_t run_start;
+  utime_t run_until;
+
+  client_t run_only;
+  client_t exclude;
+
+  string get_sarg(int seq);
+  int get_iarg() {
+    int i = iargs.front();
+    iargs.pop_front();
+    return i;
+  }
+
+  bool time_to_stop() {
+    utime_t now = ceph_clock_now();
+    if (0) cout << "time_to_stop .. now " << now
+		<< " until " << run_until
+		<< " start " << run_start
+		<< std::endl;
+    if (run_until.sec() && now > run_until)
+      return true;
+    else
+      return false;
+  }
+
+  string compose_path(string& prefix, char *rest) {
+    return prefix + rest;
+  }
+
+  int full_walk(string& fromdir);
+  int random_walk(int n);
+
+  int dump_placement(string& fn);
+
+
+  int make_dirs(const char *basedir, int dirs, int files, int depth);
+  int stat_dirs(const char *basedir, int dirs, int files, int depth);
+  int read_dirs(const char *basedir, int dirs, int files, int depth);
+  int make_files(int num, int count, int priv, bool more);
+  int link_test();
+
+  int create_shared(int num);
+  int open_shared(int num, int count);
+
+  int rm_file(string& fn);
+  int write_file(string& fn, int mb, loff_t chunk);
+  int write_fd(int fd, int size, int wrsize);
+
+  int write_batch(int nfile, int mb, int chunk);
+  int read_file(const std::string& fn, int mb, int chunk, bool ignoreprint=false);
+
+  int create_objects(int nobj, int osize, int inflight);
+  int object_rw(int nobj, int osize, int wrpc, int overlap, 
+		double rskew, double wskew);
+
+  int read_random(string& fn, int mb, int chunk);
+  int read_random_ex(string& fn, int mb, int chunk);
+  
+  int overload_osd_0(int n, int sie, int wrsize);
+  int check_first_primary(int fd);
+
+  int clean_dir(string& basedir);
+
+  int play_trace(Trace& t, string& prefix, bool metadata_only=false);
+
+  void make_dir_mess(const char *basedir, int n);
+  void foo();
+
+  int thrash_links(const char *basedir, int dirs, int files, int depth, int n);
+
+  void import_find(const char *basedir, const char *find, bool writedata);
+
+  int lookup_hash(inodeno_t ino, inodeno_t dirino, const char *name,
+		  const UserPerm& perms);
+  int lookup_ino(inodeno_t ino, const UserPerm& perms);
+
+  int chunk_file(string &filename);
+
+  void mksnap(const char *base, const char *name, const UserPerm& perms);
+  void rmsnap(const char *base, const char *name, const UserPerm& perms);
+  void mksnapfile(const char *dir);
+
+};
+
+#endif
diff --git a/src/client/Trace.cc b/src/client/Trace.cc
new file mode 100644
index 000000000..0d07da66c
--- /dev/null
+++ b/src/client/Trace.cc
@@ -0,0 +1,79 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+
+
+#include "Trace.h"
+#include "common/debug.h"
+
+#include <iostream>
+#include <map>
+
+#include "common/config.h"
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+
+
+
+
+
+void Trace::start()
+{
+  //cout << "start" << std::endl;
+  delete fs;
+
+  fs = new ifstream();
+  fs->open(filename);
+  if (!fs->is_open()) {
+    //generic_dout(0) << "** unable to open trace file " << filename << dendl;
+    ceph_abort();
+  }
+  //generic_dout(2) << "opened traced file '" << filename << "'" << dendl;
+  
+  // read first line
+  getline(*fs, line);
+  //cout << "first line is " << line << std::endl;
+
+  _line = 1;
+}
+
+const char *Trace::peek_string(string &buf, const char *prefix)
+{
+  //if (prefix) cout << "prefix '" << prefix << "' line '" << line << "'" << std::endl;
+  if (prefix &&
+      strstr(line.c_str(), "/prefix") == line.c_str()) {
+    buf.clear();
+    buf.append(prefix);
+    buf.append(line.c_str() + strlen("/prefix"));
+  } else {
+    buf = line;
+  }
+  return buf.c_str();
+}
+
+
+const char *Trace::get_string(string &buf, const char *prefix)
+{
+  peek_string(buf, prefix);
+
+  //cout << "buf is " << buf << std::endl;
+  // read next line (and detect eof early)
+  _line++;
+  getline(*fs, line);
+  //cout << "next line is " << line << std::endl;
+
+  return buf.c_str();
+}
diff --git a/src/client/Trace.h b/src/client/Trace.h
new file mode 100644
index 000000000..2f90c4f7c
--- /dev/null
+++ b/src/client/Trace.h
@@ -0,0 +1,67 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+
+#ifndef CEPH_CLIENT_TRACE_H
+#define CEPH_CLIENT_TRACE_H
+
+#include <stdlib.h>
+
+#include <list>
+#include <string>
+#include <fstream>
+using std::list;
+using std::string;
+using std::ifstream;
+
+/*
+
+ this class is more like an iterator over a constant tokenlist (which 
+ is protected by a mutex, see Trace.cc)
+
+ */
+
+class Trace {
+  int _line;
+  const char *filename;
+  ifstream *fs;
+  string line;
+
+ public:
+  explicit Trace(const char* f) : _line(0), filename(f), fs(0) {}
+  ~Trace() { 
+    delete fs; 
+  }
+
+  Trace(const Trace& other);
+  const Trace& operator=(const Trace& other);
+
+  int get_line() { return _line; }
+
+  void start();
+
+  const char *peek_string(string &buf, const char *prefix);
+  const char *get_string(string &buf, const char *prefix);
+
+  int64_t get_int() {
+    string buf;
+    return atoll(get_string(buf, 0));
+  }
+  bool end() {
+    return !fs || fs->eof();
+    //return _cur == _end;
+  }
+};
+
+#endif
diff --git a/src/client/UserPerm.h b/src/client/UserPerm.h
new file mode 100644
index 000000000..bb7c25d3e
--- /dev/null
+++ b/src/client/UserPerm.h
@@ -0,0 +1,93 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2016 Red Hat
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+#ifndef CEPH_CLIENT_USERPERM_H
+#define CEPH_CLIENT_USERPERM_H
+
+struct UserPerm
+{
+private:
+  uid_t m_uid;
+  gid_t m_gid;
+  int gid_count;
+  gid_t *gids;
+  bool alloced_gids;
+  void deep_copy_from(const UserPerm& b) {
+    if (alloced_gids) {
+      delete[] gids;
+      alloced_gids = false;
+    }
+    m_uid = b.m_uid;
+    m_gid = b.m_gid;
+    gid_count = b.gid_count;
+    if (gid_count > 0) {
+      gids = new gid_t[gid_count];
+      alloced_gids = true;
+      for (int i = 0; i < gid_count; ++i) {
+	gids[i] = b.gids[i];
+      }
+    }
+  }
+public:
+  UserPerm() : m_uid(-1), m_gid(-1), gid_count(0),
+	       gids(NULL), alloced_gids(false) {}
+  UserPerm(uid_t uid, gid_t gid, int ngids=0, gid_t *gidlist=NULL) :
+	    m_uid(uid), m_gid(gid), gid_count(ngids),
+	    gids(gidlist), alloced_gids(false) {}
+  UserPerm(const UserPerm& o) : UserPerm() {
+    deep_copy_from(o);
+  }
+  UserPerm(UserPerm && o) {
+    m_uid = o.m_uid;
+    m_gid = o.m_gid;
+    gid_count = o.gid_count;
+    gids = o.gids;
+    alloced_gids = o.alloced_gids;
+    o.gids = NULL;
+    o.gid_count = 0;
+  }
+  ~UserPerm() {
+    if (alloced_gids)
+      delete[] gids;
+  }
+  UserPerm& operator=(const UserPerm& o) {
+    deep_copy_from(o);
+    return *this;
+  }
+
+  uid_t uid() const { return m_uid != (uid_t)-1 ? m_uid : ::geteuid(); }
+  gid_t gid() const { return m_gid != (gid_t)-1 ? m_gid : ::getegid(); }
+  bool gid_in_groups(gid_t id) const {
+    if (id == gid()) return true;
+    for (int i = 0; i < gid_count; ++i) {
+      if (id == gids[i]) return true;
+    }
+    return false;
+  }
+  int get_gids(const gid_t **_gids) const { *_gids = gids; return gid_count; }
+  void init_gids(gid_t* _gids, int count) {
+    gids = _gids;
+    gid_count = count;
+    alloced_gids = true;
+  }
+  void shallow_copy(const UserPerm& o) {
+    m_uid = o.m_uid;
+    m_gid = o.m_gid;
+    gid_count = o.gid_count;
+    gids = o.gids;
+    alloced_gids = false;
+  }
+};
+
+#endif
diff --git a/src/client/barrier.cc b/src/client/barrier.cc
new file mode 100644
index 000000000..7a0c08868
--- /dev/null
+++ b/src/client/barrier.cc
@@ -0,0 +1,196 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ *
+ * Copyright (C) 2012 CohortFS, LLC.
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+#if defined(__FreeBSD__)
+#include <sys/param.h>
+#endif
+
+#include "include/Context.h"
+#include "Client.h"
+#include "barrier.h"
+#include "include/ceph_assert.h"
+
+#undef dout_prefix
+#define dout_prefix *_dout << "client." << whoami << " "
+
+#define dout_subsys ceph_subsys_client
+
+#define cldout(cl, v)  dout_impl((cl)->cct, dout_subsys, v) \
+  *_dout << "client." << cl->whoami << " "
+
+/* C_Block_Sync */
+class C_Block_Sync : public Context {
+private:
+  Client *cl;
+  uint64_t ino;
+  barrier_interval iv;
+  enum CBlockSync_State state;
+  Barrier *barrier;
+  int *rval; /* see Cond.h */
+
+public:
+  boost::intrusive::list_member_hook<> intervals_hook;
+  C_Block_Sync(Client *c, uint64_t i, barrier_interval iv, int *r);
+  void finish(int rval);
+
+  friend class Barrier;
+  friend class BarrierContext;
+};
+
+C_Block_Sync::C_Block_Sync(Client *c, uint64_t i, barrier_interval iv,
+			   int *r=0) :
+  cl(c), ino(i), iv(iv), rval(r)
+{
+  state = CBlockSync_State_None;
+  barrier = NULL;
+
+  cldout(cl, 1) << "C_Block_Sync for " << ino << dendl;
+
+  if (!cl->barriers[ino]) {
+    cl->barriers[ino] = new BarrierContext(cl, ino);
+  }
+  /* XXX current semantics aren't commit-ordered */
+  cl->barriers[ino]->write_nobarrier(*this);
+}
+
+void C_Block_Sync::finish(int r) {
+  cldout(cl, 1) << "C_Block_Sync::finish() for " << ino << " "
+		<< iv << " r==" << r << dendl;
+  if (rval)
+    *rval = r;
+  cl->barriers[ino]->complete(*this);
+}
+
+/* Barrier */
+Barrier::Barrier()
+{ }
+
+Barrier::~Barrier()
+{ }
+
+/* BarrierContext */
+BarrierContext::BarrierContext(Client *c, uint64_t ino) :
+  cl(c), ino(ino)
+{ };
+
+void BarrierContext::write_nobarrier(C_Block_Sync &cbs)
+{
+  std::lock_guard locker(lock);
+  cbs.state = CBlockSync_State_Unclaimed;
+  outstanding_writes.push_back(cbs);
+}
+
+void BarrierContext::write_barrier(C_Block_Sync &cbs)
+{
+  std::unique_lock locker(lock);
+  barrier_interval &iv = cbs.iv;
+
+  { /* find blocking commit--intrusive no help here */
+    BarrierList::iterator iter;
+    bool done = false;
+    for (iter = active_commits.begin();
+	 !done && (iter != active_commits.end());
+	 ++iter) {
+      Barrier &barrier = *iter;
+      while (boost::icl::intersects(barrier.span, iv)) {
+	/*  wait on this */
+	barrier.cond.wait(locker);
+	done = true;
+      }
+    }
+  }
+
+  cbs.state = CBlockSync_State_Unclaimed;
+  outstanding_writes.push_back(cbs);
+
+} /* write_barrier */
+
+void BarrierContext::commit_barrier(barrier_interval &civ)
+{
+    std::unique_lock locker(lock);
+
+    /* we commit outstanding writes--if none exist, we don't care */
+    if (outstanding_writes.size() == 0)
+      return;
+
+    boost::icl::interval_set<uint64_t> cvs;
+    cvs.insert(civ);
+
+    Barrier *barrier = NULL;
+    BlockSyncList::iterator iter, iter2;
+
+    iter = outstanding_writes.begin();
+    while (iter != outstanding_writes.end()) {
+      barrier_interval &iv = iter->iv;
+      if (boost::icl::intersects(cvs, iv)) {
+	C_Block_Sync &a_write = *iter;
+	if (! barrier)
+	  barrier = new Barrier();
+	/* mark the callback */
+	a_write.state = CBlockSync_State_Committing;
+	a_write.barrier = barrier;
+	iter2 = iter++;
+	outstanding_writes.erase(iter2);
+	barrier->write_list.push_back(a_write);
+	barrier->span.insert(iv);
+	/* avoid iter invalidate */
+      } else {
+	++iter;
+      }
+    }
+
+    if (barrier) {
+      active_commits.push_back(*barrier);
+      /* and wait on this */
+      barrier->cond.wait(locker);
+    }
+
+} /* commit_barrier */
+
+void BarrierContext::complete(C_Block_Sync &cbs)
+{
+    std::lock_guard locker(lock);
+    BlockSyncList::iterator iter =
+      BlockSyncList::s_iterator_to(cbs);
+
+    switch (cbs.state) {
+    case CBlockSync_State_Unclaimed:
+      /* cool, no waiting */
+      outstanding_writes.erase(iter);
+      break;
+    case CBlockSync_State_Committing:
+    {
+      Barrier *barrier = iter->barrier;
+      barrier->write_list.erase(iter);
+      /* signal waiters */
+      barrier->cond.notify_all();
+	/* dispose cleared barrier */
+      if (barrier->write_list.size() == 0) {
+	BarrierList::iterator iter2 =
+	  BarrierList::s_iterator_to(*barrier);
+	active_commits.erase(iter2);
+	delete barrier;
+      }
+    }
+    break;
+    default:
+      ceph_abort();
+      break;
+    }
+
+    cbs.state = CBlockSync_State_Completed;
+
+} /* complete */
+
+BarrierContext::~BarrierContext()
+{ }
diff --git a/src/client/barrier.h b/src/client/barrier.h
new file mode 100644
index 000000000..289a3a521
--- /dev/null
+++ b/src/client/barrier.h
@@ -0,0 +1,99 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ *
+ * Copyright (C) 2012 CohortFS, LLC.
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+#ifndef BARRIER_H
+#define BARRIER_H
+
+#include "include/types.h"
+#include <boost/intrusive/list.hpp>
+#define BOOST_ICL_USE_STATIC_BOUNDED_INTERVALS
+#include <boost/icl/interval_set.hpp>
+#include "common/ceph_mutex.h"
+
+class Client;
+
+typedef boost::icl::interval<uint64_t>::type barrier_interval;
+
+
+/*
+ * we keep count of uncommitted writes on the inode, so that
+ * ll_commit_blocks can do the right thing.
+ *
+ * This is just a hacked copy of Ceph's sync callback.
+ */
+
+enum CBlockSync_State
+{
+  CBlockSync_State_None, /* initial state */
+  CBlockSync_State_Unclaimed, /* outstanding write */
+  CBlockSync_State_Committing, /* commit in progress */
+  CBlockSync_State_Completed,
+};
+
+class BarrierContext;
+
+class C_Block_Sync;
+
+typedef boost::intrusive::list< C_Block_Sync,
+				boost::intrusive::member_hook<
+				  C_Block_Sync,
+				  boost::intrusive::list_member_hook<>,
+				  &C_Block_Sync::intervals_hook >
+				> BlockSyncList;
+
+class Barrier
+{
+private:
+  ceph::condition_variable cond;
+  boost::icl::interval_set<uint64_t> span;
+  BlockSyncList write_list;
+
+public:
+  boost::intrusive::list_member_hook<> active_commits_hook;
+
+  Barrier();
+  ~Barrier();
+
+  friend class BarrierContext;
+};
+
+typedef boost::intrusive::list< Barrier,
+				boost::intrusive::member_hook<
+				  Barrier,
+				  boost::intrusive::list_member_hook<>,
+				  &Barrier::active_commits_hook >
+				> BarrierList;
+
+class BarrierContext
+{
+private:
+  Client *cl;
+  uint64_t ino;
+  ceph::mutex lock = ceph::make_mutex("BarrierContext");
+
+  // writes not claimed by a commit
+  BlockSyncList outstanding_writes;
+
+  // commits in progress, with their claimed writes
+  BarrierList active_commits;
+
+public:
+  BarrierContext(Client *c, uint64_t ino);
+  void write_nobarrier(C_Block_Sync &cbs);
+  void write_barrier(C_Block_Sync &cbs);
+  void commit_barrier(barrier_interval &civ);
+  void complete(C_Block_Sync &cbs);
+  ~BarrierContext();
+};
+
+#endif
diff --git a/src/client/fuse_ll.cc b/src/client/fuse_ll.cc
new file mode 100644
index 000000000..fd9d1f55b
--- /dev/null
+++ b/src/client/fuse_ll.cc
@@ -0,0 +1,1680 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software
+ * Foundation.  See file COPYING.
+ *
+ */
+
+#include <sys/file.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <limits.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <unistd.h>
+
+// ceph
+#include "common/errno.h"
+#include "common/safe_io.h"
+#include "include/types.h"
+#include "Client.h"
+#include "Fh.h"
+#include "ioctl.h"
+#include "common/config.h"
+#include "include/ceph_assert.h"
+#include "include/cephfs/ceph_ll_client.h"
+#include "include/ceph_fuse.h"
+
+#include "fuse_ll.h"
+#include <fuse_lowlevel.h>
+
+#define dout_context g_ceph_context
+
+#define FINO_INO(x) ((x) & ((1ull<<48)-1ull))
+#define FINO_STAG(x) ((x) >> 48)
+#define MAKE_FINO(i,s) ((i) | ((int64_t)(s) << 48))
+#define STAG_MASK 0xffff
+#define G_NOSNAP_STAG   0 // for all CEPH_NOSNAP
+#define G_SNAPDIR_STAG  1 // for all CEPH_SNAPDIR
+
+#define MINORBITS	20
+#define MINORMASK	((1U << MINORBITS) - 1)
+
+#define MAJOR(dev)	((unsigned int) ((dev) >> MINORBITS))
+#define MINOR(dev)	((unsigned int) ((dev) & MINORMASK))
+#define MKDEV(ma,mi)	(((ma) << MINORBITS) | (mi))
+
+/*
+ * The dedicated struct for snapid <-> stag map for each ceph
+ * inode, and the stag is a number in range [2, 0xffff], and
+ * the stag number 0 is reserved for CEPH_NOSNAP and 1 is
+ * reserved for CEPH_SNAPDIR.
+ */
+struct ceph_fuse_fake_inode_stag {
+  ceph::unordered_map<uint64_t,int> snap_stag_map;  // <snapid, stagid>
+  ceph::unordered_map<int, uint64_t> stag_snap_map; // <stagid, snapid>
+  int last_stag = 1;
+};
+
+static const ceph::unordered_map<int,int> cephfs_errno_to_system_errno = {
+  {CEPHFS_EBLOCKLISTED,    ESHUTDOWN},
+  {CEPHFS_EPERM,           EPERM},
+  {CEPHFS_ESTALE,          ESTALE},
+  {CEPHFS_ENOSPC,          ENOSPC},
+  {CEPHFS_ETIMEDOUT,       ETIMEDOUT},
+  {CEPHFS_EIO,             EIO},
+  {CEPHFS_ENOTCONN,        ENOTCONN},
+  {CEPHFS_EEXIST,          EEXIST},
+  {CEPHFS_EINTR,           EINTR},
+  {CEPHFS_EINVAL,          EINVAL},
+  {CEPHFS_EBADF,           EBADF},
+  {CEPHFS_EROFS,           EROFS},
+  {CEPHFS_EAGAIN,          EAGAIN},
+  {CEPHFS_EACCES,          EACCES},
+  {CEPHFS_ELOOP,           ELOOP},
+  {CEPHFS_EISDIR,          EISDIR},
+  {CEPHFS_ENOENT,          ENOENT},
+  {CEPHFS_ENOTDIR,         ENOTDIR},
+  {CEPHFS_ENAMETOOLONG,    ENAMETOOLONG},
+  {CEPHFS_EBUSY,           EBUSY},
+  {CEPHFS_EDQUOT,          EDQUOT},
+  {CEPHFS_EFBIG,           EFBIG},
+  {CEPHFS_ERANGE,          ERANGE},
+  {CEPHFS_ENXIO,           ENXIO},
+  {CEPHFS_ECANCELED,       ECANCELED},
+  {CEPHFS_ENODATA,         ENODATA},
+  {CEPHFS_EOPNOTSUPP,      EOPNOTSUPP},
+  {CEPHFS_EXDEV,           EXDEV},
+  {CEPHFS_ENOMEM,          ENOMEM},
+  {CEPHFS_ENOTRECOVERABLE, ENOTRECOVERABLE},
+  {CEPHFS_ENOSYS,          ENOSYS},
+  {CEPHFS_ENOTEMPTY,       ENOTEMPTY},
+  {CEPHFS_EDEADLK,         EDEADLK},
+  {CEPHFS_EDOM,            EDOM},
+  {CEPHFS_EMLINK,          EMLINK},
+  {CEPHFS_ETIME,           ETIME},
+  {CEPHFS_EOLDSNAPC,       EIO} // forcing to EIO for now
+};
+
+/* Requirements:
+ * cephfs_errno >= 0
+ */
+static int get_sys_errno(int cephfs_errno)
+{
+  if (cephfs_errno == 0)
+    return 0;
+
+  auto it = cephfs_errno_to_system_errno.find(cephfs_errno);
+  if (it != cephfs_errno_to_system_errno.end())
+    return it->second;
+  return EIO;
+}
+
+static uint32_t new_encode_dev(dev_t dev)
+{
+	unsigned major = MAJOR(dev);
+	unsigned minor = MINOR(dev);
+	return (minor & 0xff) | (major << 8) | ((minor & ~0xff) << 12);
+}
+
+static dev_t new_decode_dev(uint32_t dev)
+{
+	unsigned major = (dev & 0xfff00) >> 8;
+	unsigned minor = (dev & 0xff) | ((dev >> 12) & 0xfff00);
+	return MKDEV(major, minor);
+}
+
+class CephFuse::Handle {
+public:
+  Handle(Client *c, int fd);
+  ~Handle();
+
+  int init(int argc, const char *argv[]);
+  int start();
+  int loop();
+  void finalize();
+
+  uint64_t fino_snap(uint64_t fino);
+  uint64_t make_fake_ino(inodeno_t ino, snapid_t snapid);
+  Inode * iget(fuse_ino_t fino);
+  void iput(Inode *in);
+
+  int fd_on_success;
+  Client *client;
+
+  struct fuse_session *se = nullptr;
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  struct fuse_cmdline_opts opts;
+  struct fuse_conn_info_opts *conn_opts;
+#else
+  struct fuse_chan *ch = nullptr;
+  char *mountpoint = nullptr;
+#endif
+
+  ceph::mutex stag_lock = ceph::make_mutex("fuse_ll.cc stag_lock");
+
+  // a map of <ceph ino, fino stag/snapid map>
+  ceph::unordered_map<uint64_t, struct ceph_fuse_fake_inode_stag> g_fino_maps;
+
+  pthread_key_t fuse_req_key = 0;
+  void set_fuse_req(fuse_req_t);
+  fuse_req_t get_fuse_req();
+
+  struct fuse_args args;
+};
+
+static int getgroups(fuse_req_t req, gid_t **sgids)
+{
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+  ceph_assert(sgids);
+  int c = fuse_req_getgroups(req, 0, NULL);
+  if (c < 0) {
+    return c;
+  }
+  if (c == 0) {
+    return 0;
+  }
+
+  gid_t *gids = new (std::nothrow) gid_t[c];
+  if (!gids) {
+    return -get_sys_errno(CEPHFS_ENOMEM);
+  }
+  c = fuse_req_getgroups(req, c, gids);
+  if (c < 0) {
+    delete[] gids;
+  } else {
+    *sgids = gids;
+  }
+  return c;
+#endif
+  return -get_sys_errno(CEPHFS_ENOSYS);
+}
+
+static void get_fuse_groups(UserPerm& perms, fuse_req_t req)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)fuse_req_userdata(req);
+  if (cfuse->client->cct->_conf.get_val<bool>("fuse_set_user_groups")) {
+    gid_t *gids = NULL;
+    int count = getgroups(req, &gids);
+
+    if (count > 0) {
+      perms.init_gids(gids, count);
+    } else if (count < 0) {
+      derr << __func__ << ": getgroups failed: " << cpp_strerror(-count)
+	   << dendl;
+    }
+  }
+}
+
+
+static CephFuse::Handle *fuse_ll_req_prepare(fuse_req_t req)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)fuse_req_userdata(req);
+  cfuse->set_fuse_req(req);
+  return cfuse;
+}
+
+static void fuse_ll_lookup(fuse_req_t req, fuse_ino_t parent, const char *name)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct fuse_entry_param fe;
+  Inode *i2, *i1 = cfuse->iget(parent); // see below
+  int r;
+  UserPerm perms(ctx->uid, ctx->gid);
+  get_fuse_groups(perms, req);
+
+  if (!i1)
+  {
+    r = cfuse->client->lookup_ino(parent, perms, &i1);
+    if (r < 0) {
+      fuse_reply_err(req, get_sys_errno(-r));
+      return;
+    }
+  }
+
+  memset(&fe, 0, sizeof(fe));
+  r = cfuse->client->ll_lookup(i1, name, &fe.attr, &i2, perms);
+  if (r >= 0) {
+    fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+    fe.attr.st_rdev = new_encode_dev(fe.attr.st_rdev);
+    fuse_reply_entry(req, &fe);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  // XXX NB, we dont iput(i2) because FUSE will do so in a matching
+  // fuse_ll_forget()
+  cfuse->iput(i1);
+}
+
+static void fuse_ll_forget(fuse_req_t req, fuse_ino_t ino,
+			   long unsigned nlookup)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Inode *in = cfuse->iget(ino);
+  if (in)
+    cfuse->client->ll_forget(in, nlookup+1);
+  fuse_reply_none(req);
+}
+
+static void fuse_ll_getattr(fuse_req_t req, fuse_ino_t ino,
+			    struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct stat stbuf;
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  (void) fi; // XXX
+
+  if (cfuse->client->ll_getattr(in, &stbuf, perms)
+      == 0) {
+    stbuf.st_ino = cfuse->make_fake_ino(stbuf.st_ino, stbuf.st_dev);
+    stbuf.st_rdev = new_encode_dev(stbuf.st_rdev);
+    fuse_reply_attr(req, &stbuf, 0);
+  } else
+    fuse_reply_err(req, ENOENT);
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_setattr(fuse_req_t req, fuse_ino_t ino, struct stat *attr,
+			    int to_set, struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int mask = 0;
+  if (to_set & FUSE_SET_ATTR_MODE) mask |= CEPH_SETATTR_MODE;
+  if (to_set & FUSE_SET_ATTR_UID) mask |= CEPH_SETATTR_UID;
+  if (to_set & FUSE_SET_ATTR_GID) mask |= CEPH_SETATTR_GID;
+  if (to_set & FUSE_SET_ATTR_MTIME) mask |= CEPH_SETATTR_MTIME;
+  if (to_set & FUSE_SET_ATTR_ATIME) mask |= CEPH_SETATTR_ATIME;
+  if (to_set & FUSE_SET_ATTR_SIZE) mask |= CEPH_SETATTR_SIZE;
+#if !defined(__APPLE__)
+  if (to_set & FUSE_SET_ATTR_MTIME_NOW) mask |= CEPH_SETATTR_MTIME_NOW;
+  if (to_set & FUSE_SET_ATTR_ATIME_NOW) mask |= CEPH_SETATTR_ATIME_NOW;
+#endif
+
+  int r = cfuse->client->ll_setattr(in, attr, mask, perms);
+  if (r == 0)
+    fuse_reply_attr(req, attr, 0);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+// XATTRS
+
+static void fuse_ll_setxattr(fuse_req_t req, fuse_ino_t ino, const char *name,
+			     const char *value, size_t size, 
+			     int flags
+#if defined(__APPLE__)
+			     ,uint32_t pos
+#endif
+  )
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_setxattr(in, name, value, size, flags, perms);
+  fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_listxattr(fuse_req_t req, fuse_ino_t ino, size_t size)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  char buf[size];
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_listxattr(in, buf, size, perms);
+  if (size == 0 && r >= 0)
+    fuse_reply_xattr(req, r);
+  else if (r >= 0) 
+    fuse_reply_buf(req, buf, r);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_getxattr(fuse_req_t req, fuse_ino_t ino, const char *name,
+			     size_t size
+#if defined(__APPLE__)
+			     ,uint32_t position
+#endif
+  )
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  char buf[size];
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_getxattr(in, name, buf, size, perms);
+  if (size == 0 && r >= 0)
+    fuse_reply_xattr(req, r);
+  else if (r >= 0)
+    fuse_reply_buf(req, buf, r);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_removexattr(fuse_req_t req, fuse_ino_t ino,
+				const char *name)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_removexattr(in, name, perms);
+  fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_opendir(fuse_req_t req, fuse_ino_t ino,
+			    struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  void *dirp;
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_opendir(in, fi->flags, (dir_result_t **)&dirp,
+				    perms);
+  if (r >= 0) {
+    fi->fh = (uint64_t)dirp;
+    fuse_reply_open(req, fi);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_readlink(fuse_req_t req, fuse_ino_t ino)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  char buf[PATH_MAX + 1];  // leave room for a null terminator
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+  int r = cfuse->client->ll_readlink(in, buf, sizeof(buf) - 1, perms);
+  if (r >= 0) {
+    buf[r] = '\0';
+    fuse_reply_readlink(req, buf);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_mknod(fuse_req_t req, fuse_ino_t parent, const char *name,
+			  mode_t mode, dev_t rdev)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct fuse_entry_param fe;
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *i2, *i1 = cfuse->iget(parent);
+  if (!i1) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  memset(&fe, 0, sizeof(fe));
+
+  int r = cfuse->client->ll_mknod(i1, name, mode, new_decode_dev(rdev),
+				  &fe.attr, &i2, perms);
+  if (r == 0) {
+    fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+    fe.attr.st_rdev = new_encode_dev(fe.attr.st_rdev);
+    fuse_reply_entry(req, &fe);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  // XXX NB, we dont iput(i2) because FUSE will do so in a matching
+  // fuse_ll_forget()
+  cfuse->iput(i1); // iput required
+}
+
+static void fuse_ll_mkdir(fuse_req_t req, fuse_ino_t parent, const char *name,
+			  mode_t mode)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  Inode *i2, *i1;
+  struct fuse_entry_param fe;
+
+  memset(&fe, 0, sizeof(fe));
+  UserPerm perm(ctx->uid, ctx->gid);
+  get_fuse_groups(perm, req);
+#ifdef HAVE_SYS_SYNCFS
+  auto fuse_multithreaded = cfuse->client->cct->_conf.get_val<bool>(
+    "fuse_multithreaded");
+  auto fuse_syncfs_on_mksnap = cfuse->client->cct->_conf.get_val<bool>(
+    "fuse_syncfs_on_mksnap");
+  if (cfuse->fino_snap(parent) == CEPH_SNAPDIR &&
+      fuse_multithreaded && fuse_syncfs_on_mksnap) {
+    int err = 0;
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+    int fd = ::open(cfuse->opts.mountpoint, O_RDONLY | O_DIRECTORY | O_CLOEXEC);
+#else
+    int fd = ::open(cfuse->mountpoint, O_RDONLY | O_DIRECTORY | O_CLOEXEC);
+#endif
+    if (fd < 0) {
+      err = errno;
+    } else {
+      int r = ::syncfs(fd);
+      if (r < 0)
+	err = errno;
+      ::close(fd);
+    }
+    if (err) {
+      fuse_reply_err(req, err);
+      return;
+    }
+  }
+#endif
+
+  i1 = cfuse->iget(parent);
+  if (!i1) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  int r = cfuse->client->ll_mkdir(i1, name, mode, &fe.attr, &i2, perm);
+  if (r == 0) {
+    fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+    fe.attr.st_rdev = new_encode_dev(fe.attr.st_rdev);
+    fuse_reply_entry(req, &fe);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  // XXX NB, we dont iput(i2) because FUSE will do so in a matching
+  // fuse_ll_forget()
+  cfuse->iput(i1); // iput required
+}
+
+static void fuse_ll_unlink(fuse_req_t req, fuse_ino_t parent, const char *name)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perm(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(parent);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perm, req);
+
+  int r = cfuse->client->ll_unlink(in, name, perm);
+  fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_rmdir(fuse_req_t req, fuse_ino_t parent, const char *name)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(parent);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_rmdir(in, name, perms);
+  fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_symlink(fuse_req_t req, const char *existing,
+			    fuse_ino_t parent, const char *name)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct fuse_entry_param fe;
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *i2, *i1 = cfuse->iget(parent);
+  if (!i1) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  memset(&fe, 0, sizeof(fe));
+
+  int r = cfuse->client->ll_symlink(i1, name, existing, &fe.attr, &i2, perms);
+  if (r == 0) {
+    fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+    fe.attr.st_rdev = new_encode_dev(fe.attr.st_rdev);
+    fuse_reply_entry(req, &fe);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  // XXX NB, we dont iput(i2) because FUSE will do so in a matching
+  // fuse_ll_forget()
+  cfuse->iput(i1); // iput required
+}
+
+static void fuse_ll_rename(fuse_req_t req, fuse_ino_t parent, const char *name,
+			   fuse_ino_t newparent, const char *newname
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+                           , unsigned int flags
+#endif
+                           )
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perm(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(parent);
+  Inode *nin = cfuse->iget(newparent);
+  if (!in || !nin) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perm, req);
+
+  int r = cfuse->client->ll_rename(in, name, nin, newname, perm);
+  fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iputs required
+  cfuse->iput(nin);
+}
+
+static void fuse_ll_link(fuse_req_t req, fuse_ino_t ino, fuse_ino_t newparent,
+			 const char *newname)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct fuse_entry_param fe;
+  Inode *in = cfuse->iget(ino);
+  Inode *nin = cfuse->iget(newparent);
+  if (!in || !nin) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  memset(&fe, 0, sizeof(fe));
+  UserPerm perm(ctx->uid, ctx->gid);
+  get_fuse_groups(perm, req);
+
+  /*
+   * Note that we could successfully link, but then fail the subsequent
+   * getattr and return an error. Perhaps we should ignore getattr errors,
+   * but then how do we tell FUSE that the attrs are bogus?
+   */
+  int r = cfuse->client->ll_link(in, nin, newname, perm);
+  if (r == 0) {
+    r = cfuse->client->ll_getattr(in, &fe.attr, perm);
+    if (r == 0) {
+      fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+      fe.attr.st_rdev = new_encode_dev(fe.attr.st_rdev);
+      fuse_reply_entry(req, &fe);
+    }
+  }
+
+  if (r != 0) {
+    /*
+     * Many ll operations in libcephfs return an extra inode reference, but
+     * ll_link currently does not. Still, FUSE needs one for the new dentry,
+     * so we commandeer the reference taken earlier when ll_link is successful.
+     * On error however, we must put that reference.
+     */
+    cfuse->iput(in);
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  cfuse->iput(nin);
+}
+
+static void fuse_ll_open(fuse_req_t req, fuse_ino_t ino,
+			 struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  Fh *fh = NULL;
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_open(in, fi->flags, &fh, perms);
+  if (r == 0) {
+    fi->fh = (uint64_t)fh;
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+    auto fuse_disable_pagecache = cfuse->client->cct->_conf.get_val<bool>(
+      "fuse_disable_pagecache");
+    auto fuse_use_invalidate_cb = cfuse->client->cct->_conf.get_val<bool>(
+      "fuse_use_invalidate_cb");
+    if (fuse_disable_pagecache)
+      fi->direct_io = 1;
+    else if (fuse_use_invalidate_cb)
+      fi->keep_cache = 1;
+#endif
+    fuse_reply_open(req, fi);
+  } else {
+    fuse_reply_err(req, get_sys_errno(-r));
+  }
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_read(fuse_req_t req, fuse_ino_t ino, size_t size, off_t off,
+			 struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+  bufferlist bl;
+  int r = cfuse->client->ll_read(fh, off, size, &bl);
+  if (r >= 0) {
+    vector<iovec> iov;
+    size_t len;
+    struct fuse_bufvec *bufv;
+
+    if (bl.get_num_buffers() > IOV_MAX)
+      bl.rebuild();
+
+    bl.prepare_iov(&iov);
+    len = sizeof(struct fuse_bufvec) + sizeof(struct fuse_buf) * (iov.size() - 1);
+    bufv = (struct fuse_bufvec *)calloc(1, len);
+    if (bufv) {
+      int i = 0;
+      bufv->count = iov.size();
+      for (auto &v: iov) {
+        bufv->buf[i].mem = v.iov_base;
+        bufv->buf[i++].size = v.iov_len;
+      }
+      fuse_reply_data(req, bufv, FUSE_BUF_SPLICE_MOVE);
+      free(bufv);
+      return;
+    }
+    iov.insert(iov.begin(), {0}); // the first one is reserved for fuse_out_header
+    fuse_reply_iov(req, &iov[0], iov.size());
+  } else
+    fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_write(fuse_req_t req, fuse_ino_t ino, const char *buf,
+			   size_t size, off_t off, struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+  int r = cfuse->client->ll_write(fh, off, size, buf);
+  if (r >= 0)
+    fuse_reply_write(req, r);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_flush(fuse_req_t req, fuse_ino_t ino,
+			  struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+  int r = cfuse->client->ll_flush(fh);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+#ifdef FUSE_IOCTL_COMPAT
+static void fuse_ll_ioctl(fuse_req_t req, fuse_ino_t ino,
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 5)
+                          unsigned int cmd,
+#else
+                          int cmd,
+#endif
+                          void *arg, struct fuse_file_info *fi,
+			  unsigned flags, const void *in_buf, size_t in_bufsz, size_t out_bufsz)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+
+  if (flags & FUSE_IOCTL_COMPAT) {
+    fuse_reply_err(req, ENOSYS);
+    return;
+  }
+
+  switch (static_cast<unsigned>(cmd)) {
+    case CEPH_IOC_GET_LAYOUT: {
+      file_layout_t layout;
+      struct ceph_ioctl_layout l;
+      Fh *fh = (Fh*)fi->fh;
+      cfuse->client->ll_file_layout(fh, &layout);
+      l.stripe_unit = layout.stripe_unit;
+      l.stripe_count = layout.stripe_count;
+      l.object_size = layout.object_size;
+      l.data_pool = layout.pool_id;
+      fuse_reply_ioctl(req, 0, &l, sizeof(struct ceph_ioctl_layout));
+    }
+    break;
+    default:
+      fuse_reply_err(req, EINVAL);
+  }
+}
+#endif
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 9)
+
+static void fuse_ll_fallocate(fuse_req_t req, fuse_ino_t ino, int mode,
+                              off_t offset, off_t length,
+                              struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = (Fh*)fi->fh;
+  int r = cfuse->client->ll_fallocate(fh, mode, offset, length);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+#endif
+
+static void fuse_ll_release(fuse_req_t req, fuse_ino_t ino,
+			    struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+  int r = cfuse->client->ll_release(fh);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_fsync(fuse_req_t req, fuse_ino_t ino, int datasync,
+			  struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+  int r = cfuse->client->ll_fsync(fh, datasync);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+struct readdir_context {
+  fuse_req_t req;
+  char *buf;
+  size_t size;
+  size_t pos; /* in buf */
+  uint64_t snap;
+};
+
+/*
+ * return 0 on success, -1 if out of space
+ */
+static int fuse_ll_add_dirent(void *p, struct dirent *de,
+			      struct ceph_statx *stx, off_t next_off,
+			      Inode *in)
+{
+  struct readdir_context *c = (struct readdir_context *)p;
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)fuse_req_userdata(c->req);
+
+  struct stat st;
+  st.st_ino = cfuse->make_fake_ino(stx->stx_ino, c->snap);
+  st.st_mode = stx->stx_mode;
+  st.st_rdev = new_encode_dev(stx->stx_rdev);
+
+  size_t room = c->size - c->pos;
+  size_t entrysize = fuse_add_direntry(c->req, c->buf + c->pos, room,
+				       de->d_name, &st, next_off);
+  if (entrysize > room)
+    return -ENOSPC;
+
+  /* success */
+  c->pos += entrysize;
+  return 0;
+}
+
+static void fuse_ll_readdir(fuse_req_t req, fuse_ino_t ino, size_t size,
+			    off_t off, struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+
+  dir_result_t *dirp = reinterpret_cast<dir_result_t*>(fi->fh);
+  cfuse->client->seekdir(dirp, off);
+
+  struct readdir_context rc;
+  rc.req = req;
+  rc.snap = cfuse->fino_snap(ino);
+  if (rc.snap == CEPH_MAXSNAP) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+  rc.buf = new char[size];
+  rc.size = size;
+  rc.pos = 0;
+
+  int r = cfuse->client->readdir_r_cb(dirp, fuse_ll_add_dirent, &rc);
+  if (r == 0 || r == -CEPHFS_ENOSPC)  /* ignore ENOSPC from our callback */
+    fuse_reply_buf(req, rc.buf, rc.pos);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+  delete[] rc.buf;
+}
+
+static void fuse_ll_releasedir(fuse_req_t req, fuse_ino_t ino,
+			       struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  dir_result_t *dirp = reinterpret_cast<dir_result_t*>(fi->fh);
+  cfuse->client->ll_releasedir(dirp);
+  fuse_reply_err(req, 0);
+}
+
+static void fuse_ll_fsyncdir(fuse_req_t req, fuse_ino_t ino, int datasync,
+			     struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  dir_result_t *dirp = reinterpret_cast<dir_result_t*>(fi->fh);
+  int r = cfuse->client->ll_fsyncdir(dirp);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_access(fuse_req_t req, fuse_ino_t ino, int mask)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->inode_permission(in, perms, mask);
+  fuse_reply_err(req, get_sys_errno(-r));
+  cfuse->iput(in);
+}
+
+static void fuse_ll_create(fuse_req_t req, fuse_ino_t parent, const char *name,
+			   mode_t mode, struct fuse_file_info *fi)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  struct fuse_entry_param fe;
+  Fh *fh = NULL;
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *i1 = cfuse->iget(parent), *i2;
+  if (!i1) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  memset(&fe, 0, sizeof(fe));
+
+  // pass &i2 for the created inode so that ll_create takes an initial ll_ref
+  int r = cfuse->client->ll_create(i1, name, mode, fi->flags, &fe.attr, &i2,
+				   &fh, perms);
+  if (r == 0) {
+    fi->fh = (uint64_t)fh;
+    fe.ino = cfuse->make_fake_ino(fe.attr.st_ino, fe.attr.st_dev);
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+    auto fuse_disable_pagecache = cfuse->client->cct->_conf.get_val<bool>(
+      "fuse_disable_pagecache");
+    auto fuse_use_invalidate_cb = cfuse->client->cct->_conf.get_val<bool>(
+      "fuse_use_invalidate_cb");
+    if (fuse_disable_pagecache)
+      fi->direct_io = 1;
+    else if (fuse_use_invalidate_cb)
+      fi->keep_cache = 1;
+#endif
+    fuse_reply_create(req, &fe, fi);
+  } else
+    fuse_reply_err(req, get_sys_errno(-r));
+  // XXX NB, we dont iput(i2) because FUSE will do so in a matching
+  // fuse_ll_forget()
+  cfuse->iput(i1); // iput required
+}
+
+static void fuse_ll_statfs(fuse_req_t req, fuse_ino_t ino)
+{
+  struct statvfs stbuf;
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  UserPerm perms(ctx->uid, ctx->gid);
+  Inode *in = cfuse->iget(ino);
+  if (!in) {
+    fuse_reply_err(req, get_sys_errno(CEPHFS_EINVAL));
+    return;
+  }
+
+  get_fuse_groups(perms, req);
+
+  int r = cfuse->client->ll_statfs(in, &stbuf, perms);
+  if (r == 0)
+    fuse_reply_statfs(req, &stbuf);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+
+  cfuse->iput(in); // iput required
+}
+
+static void fuse_ll_getlk(fuse_req_t req, fuse_ino_t ino,
+			  struct fuse_file_info *fi, struct flock *lock)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+
+  int r = cfuse->client->ll_getlk(fh, lock, fi->lock_owner);
+  if (r == 0)
+    fuse_reply_lock(req, lock);
+  else
+    fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_setlk(fuse_req_t req, fuse_ino_t ino,
+		          struct fuse_file_info *fi, struct flock *lock, int sleep)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = reinterpret_cast<Fh*>(fi->fh);
+
+  // must use multithread if operation may block
+  auto fuse_multithreaded = cfuse->client->cct->_conf.get_val<bool>(
+    "fuse_multithreaded");
+  if (!fuse_multithreaded && sleep && lock->l_type != F_UNLCK) {
+    fuse_reply_err(req, EDEADLK);
+    return;
+  }
+
+  int r = cfuse->client->ll_setlk(fh, lock, fi->lock_owner, sleep);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+
+static void fuse_ll_interrupt(fuse_req_t req, void* data)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  cfuse->client->ll_interrupt(data);
+}
+
+static void switch_interrupt_cb(void *handle, void* data)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)handle;
+  fuse_req_t req = cfuse->get_fuse_req();
+
+  if (data)
+    fuse_req_interrupt_func(req, fuse_ll_interrupt, data);
+  else
+    fuse_req_interrupt_func(req, NULL, NULL);
+}
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 9)
+static void fuse_ll_flock(fuse_req_t req, fuse_ino_t ino,
+		          struct fuse_file_info *fi, int cmd)
+{
+  CephFuse::Handle *cfuse = fuse_ll_req_prepare(req);
+  Fh *fh = (Fh*)fi->fh;
+
+  // must use multithread if operation may block
+  auto fuse_multithreaded = cfuse->client->cct->_conf.get_val<bool>(
+    "fuse_multithreaded");
+  if (!fuse_multithreaded && !(cmd & (LOCK_NB | LOCK_UN))) {
+    fuse_reply_err(req, EDEADLK);
+    return;
+  }
+
+  int r = cfuse->client->ll_flock(fh, cmd, fi->lock_owner);
+  fuse_reply_err(req, get_sys_errno(-r));
+}
+#endif
+
+#if !defined(__APPLE__)
+static mode_t umask_cb(void *handle)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)handle;
+  fuse_req_t req = cfuse->get_fuse_req();
+  const struct fuse_ctx *ctx = fuse_req_ctx(req);
+  return ctx->umask;
+}
+#endif
+
+static void ino_invalidate_cb(void *handle, vinodeno_t vino, int64_t off,
+			      int64_t len)
+{
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)handle;
+  fuse_ino_t fino = cfuse->make_fake_ino(vino.ino, vino.snapid);
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  fuse_lowlevel_notify_inval_inode(cfuse->se, fino, off, len);
+#else
+  fuse_lowlevel_notify_inval_inode(cfuse->ch, fino, off, len);
+#endif
+#endif
+}
+
+static void dentry_invalidate_cb(void *handle, vinodeno_t dirino,
+				 vinodeno_t ino, const char *name, size_t len)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)handle;
+  fuse_ino_t fdirino = cfuse->make_fake_ino(dirino.ino, dirino.snapid);
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 9)
+  fuse_ino_t fino = 0;
+  if (ino.ino != inodeno_t())
+    fino = cfuse->make_fake_ino(ino.ino, ino.snapid);
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  fuse_lowlevel_notify_delete(cfuse->se, fdirino, fino, name, len);
+#else
+  fuse_lowlevel_notify_delete(cfuse->ch, fdirino, fino, name, len);
+#endif
+#elif FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+  fuse_lowlevel_notify_inval_entry(cfuse->ch, fdirino, name, len);
+#endif
+}
+
+static int remount_cb(void *handle)
+{
+  // used for trimming kernel dcache. when remounting a file system, linux kernel
+  // trims all unused dentries in the file system
+  char cmd[128+PATH_MAX];
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)handle;
+  snprintf(cmd, sizeof(cmd), "LIBMOUNT_FSTAB=/dev/null mount -i -o remount %s",
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+                  cfuse->opts.mountpoint);
+#else
+                  cfuse->mountpoint);
+#endif
+  int r = system(cmd);
+  if (r != 0 && r != -1) {
+    r = WEXITSTATUS(r);
+  }
+
+  return r;
+}
+
+static void do_init(void *data, fuse_conn_info *conn)
+{
+  CephFuse::Handle *cfuse = (CephFuse::Handle *)data;
+  Client *client = cfuse->client;
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  fuse_apply_conn_info_opts(cfuse->conn_opts, conn);
+#endif
+
+  if(conn->capable & FUSE_CAP_SPLICE_MOVE)
+    conn->want |= FUSE_CAP_SPLICE_MOVE;
+
+#if !defined(__APPLE__)
+  if (!client->fuse_default_permissions && client->ll_handle_umask()) {
+    // apply umask in userspace if posix acl is enabled
+    if(conn->capable & FUSE_CAP_DONT_MASK)
+      conn->want |= FUSE_CAP_DONT_MASK;
+  }
+  if(conn->capable & FUSE_CAP_EXPORT_SUPPORT)
+    conn->want |= FUSE_CAP_EXPORT_SUPPORT;
+#endif
+
+  if (cfuse->fd_on_success) {
+    //cout << "fuse init signaling on fd " << fd_on_success << std::endl;
+    // see Preforker::daemonize(), ceph-fuse's parent process expects a `-1`
+    // from a daemonized child process.
+    uint32_t r = -1;
+    int err = safe_write(cfuse->fd_on_success, &r, sizeof(r));
+    if (err) {
+      derr << "fuse_ll: do_init: safe_write failed with error "
+	   << cpp_strerror(err) << dendl;
+      ceph_abort();
+    }
+    //cout << "fuse init done signaling on fd " << fd_on_success << std::endl;
+
+    // close stdout, etc.
+    ::close(0);
+    ::close(1);
+    ::close(2);
+  }
+}
+
+const static struct fuse_lowlevel_ops fuse_ll_oper = {
+ init: do_init,
+ destroy: 0,
+ lookup: fuse_ll_lookup,
+ forget: fuse_ll_forget,
+ getattr: fuse_ll_getattr,
+ setattr: fuse_ll_setattr,
+ readlink: fuse_ll_readlink,
+ mknod: fuse_ll_mknod,
+ mkdir: fuse_ll_mkdir,
+ unlink: fuse_ll_unlink,
+ rmdir: fuse_ll_rmdir,
+ symlink: fuse_ll_symlink,
+ rename: fuse_ll_rename,
+ link: fuse_ll_link,
+ open: fuse_ll_open,
+ read: fuse_ll_read,
+ write: fuse_ll_write,
+ flush: fuse_ll_flush,
+ release: fuse_ll_release,
+ fsync: fuse_ll_fsync,
+ opendir: fuse_ll_opendir,
+ readdir: fuse_ll_readdir,
+ releasedir: fuse_ll_releasedir,
+ fsyncdir: fuse_ll_fsyncdir,
+ statfs: fuse_ll_statfs,
+ setxattr: fuse_ll_setxattr,
+ getxattr: fuse_ll_getxattr,
+ listxattr: fuse_ll_listxattr,
+ removexattr: fuse_ll_removexattr,
+ access: fuse_ll_access,
+ create: fuse_ll_create,
+ getlk: fuse_ll_getlk,
+ setlk: fuse_ll_setlk,
+ bmap: 0,
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 8)
+#ifdef FUSE_IOCTL_COMPAT
+ ioctl: fuse_ll_ioctl,
+#else
+ ioctl: 0,
+#endif
+ poll: 0,
+#endif
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 9)
+ write_buf: 0,
+ retrieve_reply: 0,
+ forget_multi: 0,
+ flock: fuse_ll_flock,
+#endif
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(2, 9)
+ fallocate: fuse_ll_fallocate
+#endif
+};
+
+
+CephFuse::Handle::Handle(Client *c, int fd) :
+  fd_on_success(fd),
+  client(c)
+{
+  memset(&args, 0, sizeof(args));
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  memset(&opts, 0, sizeof(opts));
+#endif
+}
+
+CephFuse::Handle::~Handle()
+{
+  fuse_opt_free_args(&args);
+}
+
+void CephFuse::Handle::finalize()
+{
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  if (se) {
+    fuse_remove_signal_handlers(se);
+    fuse_session_unmount(se);
+    fuse_session_destroy(se);
+  }
+  if (conn_opts)
+    free(conn_opts);
+  if (opts.mountpoint)
+    free(opts.mountpoint);
+#else
+  if (se)
+    fuse_remove_signal_handlers(se);
+  if (ch)
+    fuse_session_remove_chan(ch);
+  if (se)
+    fuse_session_destroy(se);
+  if (ch)
+    fuse_unmount(mountpoint, ch);
+#endif
+
+  pthread_key_delete(fuse_req_key);
+}
+
+int CephFuse::Handle::init(int argc, const char *argv[])
+{
+
+  int r = pthread_key_create(&fuse_req_key, NULL);
+  if (r) {
+    derr << "pthread_key_create failed." << dendl;
+    return r;
+  }
+
+  // set up fuse argc/argv
+  int newargc = 0;
+  const char **newargv = (const char **) malloc((argc + 17) * sizeof(char *));
+  if(!newargv)
+    return ENOMEM;
+
+  newargv[newargc++] = argv[0];
+  newargv[newargc++] = "-f";  // stay in foreground
+
+  auto fuse_allow_other = client->cct->_conf.get_val<bool>(
+    "fuse_allow_other");
+  auto fuse_default_permissions = client->cct->_conf.get_val<bool>(
+    "fuse_default_permissions");
+#if FUSE_VERSION < FUSE_MAKE_VERSION(3, 0)
+  auto fuse_big_writes = client->cct->_conf.get_val<bool>(
+    "fuse_big_writes");
+#endif
+  auto fuse_max_write = client->cct->_conf.get_val<Option::size_t>(
+    "fuse_max_write");
+  auto fuse_atomic_o_trunc = client->cct->_conf.get_val<bool>(
+    "fuse_atomic_o_trunc");
+  auto fuse_splice_read = client->cct->_conf.get_val<bool>(
+    "fuse_splice_read");
+  auto fuse_splice_write = client->cct->_conf.get_val<bool>(
+    "fuse_splice_write");
+  auto fuse_splice_move = client->cct->_conf.get_val<bool>(
+    "fuse_splice_move");
+  auto fuse_debug = client->cct->_conf.get_val<bool>(
+    "fuse_debug");
+
+  if (fuse_allow_other) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "allow_other";
+  }
+  if (fuse_default_permissions) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "default_permissions";
+  }
+#if defined(__linux__)
+#if FUSE_VERSION < FUSE_MAKE_VERSION(3, 0)
+  if (fuse_big_writes) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "big_writes";
+  }
+#endif
+  if (fuse_max_write > 0) {
+    char strsplice[65];
+    newargv[newargc++] = "-o";
+    sprintf(strsplice, "max_write=%zu", (size_t)fuse_max_write);
+    newargv[newargc++] = strsplice;
+  }
+  if (fuse_atomic_o_trunc) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "atomic_o_trunc";
+  }
+  if (fuse_splice_read) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "splice_read";
+  }
+  if (fuse_splice_write) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "splice_write";
+  }
+  if (fuse_splice_move) {
+    newargv[newargc++] = "-o";
+    newargv[newargc++] = "splice_move";
+  }
+#endif
+  if (fuse_debug)
+    newargv[newargc++] = "-d";
+
+  for (int argctr = 1; argctr < argc; argctr++)
+    newargv[newargc++] = argv[argctr];
+
+  derr << "init, newargv = " << newargv << " newargc=" << newargc << dendl;
+  struct fuse_args a = FUSE_ARGS_INIT(newargc, (char**)newargv);
+  args = a;  // Roundabout construction b/c FUSE_ARGS_INIT is for initialization not assignment
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  if (fuse_parse_cmdline(&args, &opts) == -1) {
+#else
+  if (fuse_parse_cmdline(&args, &mountpoint, NULL, NULL) == -1) {
+#endif
+    derr << "fuse_parse_cmdline failed." << dendl;
+    fuse_opt_free_args(&args);
+    free(newargv);
+    return EINVAL;
+  }
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  derr << "init, args.argv = " << args.argv << " args.argc=" << args.argc << dendl;
+  conn_opts = fuse_parse_conn_info_opts(&args);
+  if (!conn_opts) {
+    derr << "fuse_parse_conn_info_opts failed" << dendl;
+    fuse_opt_free_args(&args);
+    free(newargv);
+    return EINVAL;
+  }
+#endif
+
+  ceph_assert(args.allocated);  // Checking fuse has realloc'd args so we can free newargv
+  free(newargv);
+  return 0;
+}
+
+int CephFuse::Handle::start()
+{
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  se = fuse_session_new(&args, &fuse_ll_oper, sizeof(fuse_ll_oper), this);
+  if (!se) {
+    derr << "fuse_session_new failed" << dendl;
+    return EDOM;
+  }
+#else
+  ch = fuse_mount(mountpoint, &args);
+  if (!ch) {
+    derr << "fuse_mount(mountpoint=" << mountpoint << ") failed." << dendl;
+    return EIO;
+  }
+
+  se = fuse_lowlevel_new(&args, &fuse_ll_oper, sizeof(fuse_ll_oper), this);
+  if (!se) {
+    derr << "fuse_lowlevel_new failed" << dendl;
+    return EDOM;
+  }
+#endif
+
+  signal(SIGTERM, SIG_DFL);
+  signal(SIGINT, SIG_DFL);
+  if (fuse_set_signal_handlers(se) == -1) {
+    derr << "fuse_set_signal_handlers failed" << dendl;
+    return ENOSYS;
+  }
+
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  if (fuse_session_mount(se, opts.mountpoint) != 0) {
+    derr << "fuse_session_mount failed" << dendl;
+    return ENOSYS;
+  }
+#else
+  fuse_session_add_chan(se, ch);
+#endif
+
+
+  struct ceph_client_callback_args args = {
+    handle: this,
+    ino_cb: client->cct->_conf.get_val<bool>("fuse_use_invalidate_cb") ?
+      ino_invalidate_cb : NULL,
+    dentry_cb: dentry_invalidate_cb,
+    switch_intr_cb: switch_interrupt_cb,
+#if defined(__linux__)
+    remount_cb: remount_cb,
+#endif
+#if !defined(__APPLE__)
+    umask_cb: umask_cb,
+#endif
+  };
+  client->ll_register_callbacks(&args);
+
+  return 0;
+}
+
+int CephFuse::Handle::loop()
+{
+  auto fuse_multithreaded = client->cct->_conf.get_val<bool>(
+    "fuse_multithreaded");
+  if (fuse_multithreaded) {
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 1)
+    {
+      struct fuse_loop_config conf = { 0 };
+
+      conf.clone_fd = opts.clone_fd;
+      return fuse_session_loop_mt(se, &conf);
+    }
+#elif FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+    return fuse_session_loop_mt(se, opts.clone_fd);
+#else
+    return fuse_session_loop_mt(se);
+#endif
+  } else {
+    return fuse_session_loop(se);
+  }
+}
+
+uint64_t CephFuse::Handle::fino_snap(uint64_t fino)
+{
+  if (fino == FUSE_ROOT_ID)
+    return CEPH_NOSNAP;
+
+  if (client->use_faked_inos()) {
+    vinodeno_t vino  = client->map_faked_ino(fino);
+    return vino.snapid;
+  } else {
+    std::lock_guard l(stag_lock);
+    uint64_t stag = FINO_STAG(fino);
+    if (stag == 0)
+      return CEPH_NOSNAP;
+    else if (stag == 1)
+      return CEPH_SNAPDIR;
+
+    inodeno_t ino = FINO_INO(fino);
+
+    // does the fino_maps for the ino exist ?
+    if (!g_fino_maps.count(ino))
+      return CEPH_MAXSNAP;
+
+    auto &fino_maps = g_fino_maps[ino];
+
+    // does the stagid <--> snapid map exist ?
+    if (!fino_maps.stag_snap_map.count(stag))
+      return CEPH_MAXSNAP;
+
+    // get the snapid
+    return fino_maps.stag_snap_map[stag];
+  }
+}
+
+Inode * CephFuse::Handle::iget(fuse_ino_t fino)
+{
+  if (fino == FUSE_ROOT_ID)
+    return client->get_root();
+
+  if (client->use_faked_inos()) {
+    return client->ll_get_inode((ino_t)fino);
+  } else {
+    uint64_t snap = fino_snap(fino);
+    if (snap == CEPH_MAXSNAP)
+      return NULL;
+    vinodeno_t vino(FINO_INO(fino), snap);
+    return client->ll_get_inode(vino);
+  }
+}
+
+void CephFuse::Handle::iput(Inode *in)
+{
+  client->ll_put(in);
+}
+
+uint64_t CephFuse::Handle::make_fake_ino(inodeno_t ino, snapid_t snapid)
+{
+  if (client->use_faked_inos()) {
+    // already faked by libcephfs
+    if (ino == client->get_root_ino())
+      return FUSE_ROOT_ID;
+
+    return ino;
+  } else {
+    if (snapid == CEPH_NOSNAP && ino == client->get_root_ino())
+      return FUSE_ROOT_ID;
+
+    int stag;
+    if (snapid == CEPH_NOSNAP) {
+      stag = G_NOSNAP_STAG;
+    } else if (snapid == CEPH_SNAPDIR) {
+      stag = G_SNAPDIR_STAG;
+    } else {
+      std::lock_guard l(stag_lock);
+      auto &fino_maps = g_fino_maps[ino]; // will insert it anyway if not exists
+
+      // already exist ?
+      if (fino_maps.snap_stag_map.count(snapid)) {
+        inodeno_t fino = MAKE_FINO(ino, fino_maps.snap_stag_map[snapid]);
+        return fino;
+      }
+
+      // create a new snapid <--> stagid map
+      int first = fino_maps.last_stag & STAG_MASK;
+      stag =  (++fino_maps.last_stag) & STAG_MASK;
+      for (; stag != first; stag = (++fino_maps.last_stag) & STAG_MASK) {
+        // stag 0 is reserved for CEPH_NOSNAP and 1 for CEPH_SNAPDIR
+        if (stag == 0 || stag == 1)
+          continue;
+
+        // the new stag is not used ?
+        if (!fino_maps.stag_snap_map.count(stag)) {
+          fino_maps.snap_stag_map[snapid] = stag;
+          fino_maps.stag_snap_map[stag] = snapid;
+          break;
+        }
+
+        // the stag is already used by a snpaid,
+        // try to free it
+        auto _snapid = fino_maps.stag_snap_map[stag];
+        if (!client->ll_get_snap_ref(_snapid)) {
+          fino_maps.snap_stag_map.erase(_snapid);
+          fino_maps.snap_stag_map[snapid] = stag;
+          fino_maps.stag_snap_map[stag] = snapid;
+          break;
+        }
+      }
+      if (stag == first) {
+        /*
+         * It shouldn't be here because the max snapshots for each
+         * directory is 4_K, and here we have around 64_K, which is
+         * from 0xffff - 2, stags could be used for each directory.
+         *
+         * More detail please see mds 'mds_max_snaps_per_dir' option.
+         */
+        ceph_abort_msg("run out of stag");
+      }
+    }
+
+    inodeno_t fino = MAKE_FINO(ino, stag);
+    //cout << "make_fake_ino " << ino << "." << snapid << " -> " << fino << std::endl;
+    return fino;
+  }
+}
+
+void CephFuse::Handle::set_fuse_req(fuse_req_t req)
+{
+  pthread_setspecific(fuse_req_key, (void*)req);
+}
+
+fuse_req_t CephFuse::Handle::get_fuse_req()
+{
+  return (fuse_req_t) pthread_getspecific(fuse_req_key);
+}
+
+
+CephFuse::CephFuse(Client *c, int fd) : _handle(new CephFuse::Handle(c, fd))
+{
+}
+
+CephFuse::~CephFuse()
+{
+  delete _handle;
+}
+
+int CephFuse::init(int argc, const char *argv[])
+{
+  return _handle->init(argc, argv);
+}
+
+int CephFuse::start()
+{
+  return _handle->start();
+}
+
+int CephFuse::loop()
+{
+  return _handle->loop();
+}
+
+void CephFuse::finalize()
+{
+  return _handle->finalize();
+}
+
+std::string CephFuse::get_mount_point() const
+{
+#if FUSE_VERSION >= FUSE_MAKE_VERSION(3, 0)
+  if (_handle->opts.mountpoint) {
+    return _handle->opts.mountpoint;
+#else
+  if (_handle->mountpoint) {
+    return _handle->mountpoint;
+#endif
+  } else {
+    return "";
+  }
+}
diff --git a/src/client/fuse_ll.h b/src/client/fuse_ll.h
new file mode 100644
index 000000000..85e587baf
--- /dev/null
+++ b/src/client/fuse_ll.h
@@ -0,0 +1,28 @@
+// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- 
+// vim: ts=8 sw=2 smarttab
+/*
+ * Ceph - scalable distributed file system
+ *
+ * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
+ *
+ * This is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License version 2.1, as published by the Free Software 
+ * Foundation.  See file COPYING.
+ * 
+ */
+
+class CephFuse {
+public:
+  CephFuse(Client *c, int fd);
+  ~CephFuse();
+  int init(int argc, const char *argv[]);
+  int start();
+  int mount();
+  int loop();
+  void finalize();
+  class Handle;
+  std::string get_mount_point() const;
+private:
+  CephFuse::Handle *_handle;
+};
diff --git a/src/client/hypertable/CephBroker.cc b/src/client/hypertable/CephBroker.cc
new file mode 100644
index 000000000..596e72287
--- /dev/null
+++ b/src/client/hypertable/CephBroker.cc
@@ -0,0 +1,526 @@
+/** -*- C++ -*-
+ * Copyright (C) 2009-2011 New Dream Network
+ *
+ * This file is part of Hypertable.
+ *
+ * Hypertable is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or any later version.
+ *
+ * Hypertable is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with Hypertable. If not, see <http://www.gnu.org/licenses/>
+ *
+ * Authors:
+ * Gregory Farnum <gfarnum@gmail.com>
+ * Colin McCabe <cmccabe@alumni.cmu.edu>
+ */
+
+#include "Common/Compat.h"
+
+#include "CephBroker.h"
+#include "Common/Error.h"
+#include "Common/FileUtils.h"
+#include "Common/Filesystem.h"
+#include "Common/System.h"
+
+#include <cephfs/libcephfs.h>
+#include <dirent.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <poll.h>
+#include <string>
+#include <sys/types.h>
+#include <sys/uio.h>
+#include <unistd.h>
+
+using namespace Hypertable;
+
+std::atomic<int> CephBroker::ms_next_fd{0};
+
+/* A thread-safe version of strerror */
+static std::string cpp_strerror(int err)
+{
+  char buf[128];
+  if (err < 0)
+    err = -err;
+  std::ostringstream oss;
+  oss << strerror_r(err, buf, sizeof(buf));
+  return oss.str();
+}
+
+OpenFileDataCeph::OpenFileDataCeph(struct ceph_mount_info *cmount_, const String& fname,
+				   int _fd, int _flags) 
+  : cmount(cmount_), fd(_fd), flags(_flags), filename(fname)
+{
+}
+
+OpenFileDataCeph::~OpenFileDataCeph() {
+  ceph_close(cmount, fd);
+}
+
+CephBroker::CephBroker(PropertiesPtr& cfg)
+  : cmount(NULL)
+{
+  int ret;
+  String id(cfg->get_str("CephBroker.Id"));
+  m_verbose = cfg->get_bool("Hypertable.Verbose");
+  m_root_dir = cfg->get_str("CephBroker.RootDir");
+  String mon_addr(cfg->get_str("CephBroker.MonAddr"));
+
+  HT_INFO("Calling ceph_create");
+  ret = ceph_create(&cmount, id.empty() ? NULL : id.c_str());
+  if (ret) {
+    throw Hypertable::Exception(ret, "ceph_create failed");
+  }
+  ret = ceph_conf_set(cmount, "mon_host", mon_addr.c_str());
+  if (ret) {
+    ceph_shutdown(cmount);
+    throw Hypertable::Exception(ret, "ceph_conf_set(mon_addr) failed");
+  }
+
+  // For Ceph debugging, uncomment these lines
+  //ceph_conf_set(cmount, "debug_client", "1");
+  //ceph_conf_set(cmount, "debug_ms", "1");
+
+  HT_INFO("Calling ceph_mount");
+  ret = ceph_mount(cmount, m_root_dir.empty() ? NULL : m_root_dir.c_str());
+  if (ret) {
+    ceph_shutdown(cmount);
+    throw Hypertable::Exception(ret, "ceph_mount failed");
+  }
+  HT_INFO("Mounted Ceph filesystem.");
+}
+
+CephBroker::~CephBroker()
+{
+  ceph_shutdown(cmount);
+  cmount = NULL;
+}
+
+void CephBroker::open(ResponseCallbackOpen *cb, const char *fname,
+		      uint32_t flags, uint32_t bufsz) {
+  int fd, ceph_fd;
+  String abspath;
+  HT_DEBUGF("open file='%s' bufsz=%d", fname, bufsz);
+
+  make_abs_path(fname, abspath);
+
+  fd = atomic_inc_return(&ms_next_fd);
+
+  if ((ceph_fd = ceph_open(cmount, abspath.c_str(), O_RDONLY, 0)) < 0) {
+    report_error(cb, -ceph_fd);
+    return;
+  }
+  HT_INFOF("open (%s) fd=%" PRIu32 " ceph_fd=%d", fname, fd, ceph_fd);
+
+  {
+    struct sockaddr_in addr;
+    OpenFileDataCephPtr fdata(new OpenFileDataCeph(cmount, abspath, ceph_fd, O_RDONLY));
+
+    cb->get_address(addr);
+
+    m_open_file_map.create(fd, addr, fdata);
+
+    cb->response(fd);
+  }
+}
+
+void CephBroker::create(ResponseCallbackOpen *cb, const char *fname, uint32_t flags,
+			int32_t bufsz, int16_t replication, int64_t blksz){
+  int fd, ceph_fd;
+  int oflags;
+  String abspath;
+
+  make_abs_path(fname, abspath);
+  HT_DEBUGF("create file='%s' flags=%u bufsz=%d replication=%d blksz=%lld",
+            fname, flags, bufsz, (int)replication, (Lld)blksz);
+
+  fd = atomic_inc_return(&ms_next_fd);
+
+  if (flags & Filesystem::OPEN_FLAG_OVERWRITE)
+    oflags = O_WRONLY | O_CREAT | O_TRUNC;
+  else
+    oflags = O_WRONLY | O_CREAT | O_APPEND;
+
+  //make sure the directories in the path exist
+  String directory = abspath.substr(0, abspath.rfind('/'));
+  int r;
+  HT_INFOF("Calling mkdirs on %s", directory.c_str());
+  if((r=ceph_mkdirs(cmount, directory.c_str(), 0644)) < 0 && r!=-CEPHFS_EEXIST) {
+    HT_ERRORF("create failed on mkdirs: dname='%s' - %d", directory.c_str(), -r);
+    report_error(cb, -r);
+    return;
+  }
+
+  //create file
+  if ((ceph_fd = ceph_open(cmount, abspath.c_str(), oflags, 0644)) < 0) {
+    std::string errs(cpp_strerror(-ceph_fd));
+    HT_ERRORF("open failed: file=%s - %s",  abspath.c_str(), errs.c_str());
+    report_error(cb, ceph_fd);
+    return;
+  }
+
+  HT_INFOF("create %s  = %d", fname, ceph_fd);
+
+  {
+    struct sockaddr_in addr;
+    OpenFileDataCephPtr fdata (new OpenFileDataCeph(cmount, fname, ceph_fd, O_WRONLY));
+
+    cb->get_address(addr);
+
+    m_open_file_map.create(fd, addr, fdata);
+
+    cb->response(fd);
+  }
+}
+
+void CephBroker::close(ResponseCallback *cb, uint32_t fd) {
+  if (m_verbose) {
+    HT_INFOF("close fd=%" PRIu32, fd);
+  }
+  OpenFileDataCephPtr fdata;
+  m_open_file_map.get(fd, fdata);
+  m_open_file_map.remove(fd);
+  cb->response_ok();
+}
+
+void CephBroker::read(ResponseCallbackRead *cb, uint32_t fd, uint32_t amount) {
+  OpenFileDataCephPtr fdata;
+  ssize_t nread;
+  int64_t offset;
+  StaticBuffer buf(new uint8_t [amount], amount);
+
+  HT_DEBUGF("read fd=%" PRIu32 " amount = %d", fd, amount);
+
+  if (!m_open_file_map.get(fd, fdata)) {
+    char errbuf[32];
+    sprintf(errbuf, "%" PRIu32, fd);
+    cb->error(Error::DFSBROKER_BAD_FILE_HANDLE, errbuf);
+    HT_ERRORF("bad file handle: %" PRIu32, fd);
+    return;
+  }
+
+  if ((offset = ceph_lseek(cmount, fdata->fd, 0, SEEK_CUR)) < 0) {
+    std::string errs(cpp_strerror(offset));
+    HT_ERRORF("lseek failed: fd=%" PRIu32 " ceph_fd=%d offset=0 SEEK_CUR - %s",
+	      fd, fdata->fd, errs.c_str());
+    report_error(cb, offset);
+    return;
+  }
+
+  if ((nread = ceph_read(cmount, fdata->fd, (char *)buf.base, amount, 0)) < 0 ) {
+    HT_ERRORF("read failed: fd=%" PRIu32 " ceph_fd=%d amount=%d", fd, fdata->fd, amount);
+    report_error(cb, -nread);
+    return;
+  }
+
+  buf.size = nread;
+  cb->response((uint64_t)offset, buf);
+}
+
+void CephBroker::append(ResponseCallbackAppend *cb, uint32_t fd,
+			uint32_t amount, const void *data, bool sync)
+{
+  OpenFileDataCephPtr fdata;
+  ssize_t nwritten;
+  int64_t offset;
+
+  HT_DEBUG_OUT << "append fd="<< fd <<" amount="<< amount <<" data='"
+	       << format_bytes(20, data, amount) <<" sync="<< sync << HT_END;
+
+  if (!m_open_file_map.get(fd, fdata)) {
+    char errbuf[32];
+    sprintf(errbuf, "%" PRIu32, fd);
+    cb->error(Error::DFSBROKER_BAD_FILE_HANDLE, errbuf);
+    return;
+  }
+
+  if ((offset = ceph_lseek(cmount, fdata->fd, 0, SEEK_CUR)) < 0) {
+    std::string errs(cpp_strerror(offset));
+    HT_ERRORF("lseek failed: fd=%" PRIu32 " ceph_fd=%d offset=0 SEEK_CUR - %s", fd, fdata->fd,
+              errs.c_str());
+    report_error(cb, offset);
+    return;
+  }
+
+  if ((nwritten = ceph_write(cmount, fdata->fd, (const char *)data, amount, 0)) < 0) {
+    std::string errs(cpp_strerror(nwritten));
+    HT_ERRORF("write failed: fd=%" PRIu32 " ceph_fd=%d amount=%d - %s",
+	      fd, fdata->fd, amount, errs.c_str());
+    report_error(cb, -nwritten);
+    return;
+  }
+
+  int r;
+  if (sync && ((r = ceph_fsync(cmount, fdata->fd, true)) != 0)) {
+    std::string errs(cpp_strerror(errno));
+    HT_ERRORF("flush failed: fd=%" PRIu32 " ceph_fd=%d - %s", fd, fdata->fd, errs.c_str());
+    report_error(cb, r);
+    return;
+  }
+
+  cb->response((uint64_t)offset, nwritten);
+}
+
+void CephBroker::seek(ResponseCallback *cb, uint32_t fd, uint64_t offset) {
+  OpenFileDataCephPtr fdata;
+
+  HT_DEBUGF("seek fd=%" PRIu32 " offset=%llu", fd, (Llu)offset);
+
+  if (!m_open_file_map.get(fd, fdata)) {
+    char errbuf[32];
+    sprintf(errbuf, "%" PRIu32, fd);
+    cb->error(Error::DFSBROKER_BAD_FILE_HANDLE, errbuf);
+    return;
+  }
+  loff_t res = ceph_lseek(cmount, fdata->fd, offset, SEEK_SET);
+  if (res < 0) {
+    std::string errs(cpp_strerror((int)res));
+    HT_ERRORF("lseek failed: fd=%" PRIu32 " ceph_fd=%d offset=%llu - %s",
+	      fd, fdata->fd, (Llu)offset, errs.c_str());
+    report_error(cb, offset);
+    return;
+  }
+
+  cb->response_ok();
+}
+
+void CephBroker::remove(ResponseCallback *cb, const char *fname) {
+  String abspath;
+  
+  HT_DEBUGF("remove file='%s'", fname);
+  
+  make_abs_path(fname, abspath);
+  
+  int r;
+  if ((r = ceph_unlink(cmount, abspath.c_str())) < 0) {
+    std::string errs(cpp_strerror(r));
+    HT_ERRORF("unlink failed: file='%s' - %s", abspath.c_str(), errs.c_str());
+    report_error(cb, r);
+    return;
+  }
+  cb->response_ok();
+}
+
+void CephBroker::length(ResponseCallbackLength *cb, const char *fname, bool) {
+  int r;
+  struct ceph_statx stx;
+
+  HT_DEBUGF("length file='%s'", fname);
+
+  if ((r = ceph_statx(cmount, fname, &stx, CEPH_STATX_SIZE, AT_SYMLINK_NOFOLLOW)) < 0) {
+    String abspath;
+    make_abs_path(fname, abspath);
+    std::string errs(cpp_strerror(r));
+    HT_ERRORF("length (stat) failed: file='%s' - %s", abspath.c_str(), errs.c_str());
+    report_error(cb,- r);
+    return;
+  }
+  cb->response(stx.stx_size);
+}
+
+void CephBroker::pread(ResponseCallbackRead *cb, uint32_t fd, uint64_t offset,
+		       uint32_t amount, bool) {
+  OpenFileDataCephPtr fdata;
+  ssize_t nread;
+  StaticBuffer buf(new uint8_t [amount], amount);
+
+  HT_DEBUGF("pread fd=%" PRIu32 " offset=%llu amount=%d", fd, (Llu)offset, amount);
+
+  if (!m_open_file_map.get(fd, fdata)) {
+    char errbuf[32];
+    sprintf(errbuf, "%" PRIu32, fd);
+    cb->error(Error::DFSBROKER_BAD_FILE_HANDLE, errbuf);
+    return;
+  }
+
+  if ((nread = ceph_read(cmount, fdata->fd, (char *)buf.base, amount, offset)) < 0) {
+    std::string errs(cpp_strerror(nread));
+    HT_ERRORF("pread failed: fd=%" PRIu32 " ceph_fd=%d amount=%d offset=%llu - %s",
+	      fd, fdata->fd, amount, (Llu)offset, errs.c_str());
+    report_error(cb, nread);
+    return;
+  }
+
+  buf.size = nread;
+
+  cb->response(offset, buf);
+}
+
+void CephBroker::mkdirs(ResponseCallback *cb, const char *dname) {
+  String absdir;
+
+  HT_DEBUGF("mkdirs dir='%s'", dname);
+
+  make_abs_path(dname, absdir);
+  int r;
+  if((r=ceph_mkdirs(cmount, absdir.c_str(), 0644)) < 0 && r!=-CEPHFS_EEXIST) {
+    HT_ERRORF("mkdirs failed: dname='%s' - %d", absdir.c_str(), -r);
+    report_error(cb, -r);
+    return;
+  }
+  cb->response_ok();
+}
+
+void CephBroker::rmdir(ResponseCallback *cb, const char *dname) {
+  String absdir;
+  int r;
+
+  make_abs_path(dname, absdir);
+  if((r = rmdir_recursive(absdir.c_str())) < 0) {
+      HT_ERRORF("failed to remove dir %s, got error %d", absdir.c_str(), r);
+      report_error(cb, -r);
+      return;
+  }
+  cb->response_ok();
+}
+
+int CephBroker::rmdir_recursive(const char *directory) {
+  struct ceph_dir_result *dirp;
+  struct dirent de;
+  struct ceph_statx stx;
+  int r;
+  if ((r = ceph_opendir(cmount, directory, &dirp)) < 0)
+    return r; //failed to open
+  while ((r = ceph_readdirplus_r(cmount, dirp, &de, &stx, CEPH_STATX_INO, AT_STATX_DONT_SYNC, NULL)) > 0) {
+    String new_dir = de.d_name;
+    if(!(new_dir.compare(".")==0 || new_dir.compare("..")==0)) {
+      new_dir = directory;
+      new_dir += '/';
+      new_dir += de.d_name;
+      if (S_ISDIR(stx.stx_mode)) { //it's a dir, clear it out...
+	if((r=rmdir_recursive(new_dir.c_str())) < 0) return r;
+      } else { //delete this file
+	if((r=ceph_unlink(cmount, new_dir.c_str())) < 0) return r;
+      }
+    }
+  }
+  if (r < 0) return r; //we got an error
+  if ((r = ceph_closedir(cmount, dirp)) < 0) return r;
+  return ceph_rmdir(cmount, directory);
+}
+
+void CephBroker::flush(ResponseCallback *cb, uint32_t fd) {
+  OpenFileDataCephPtr fdata;
+
+  HT_DEBUGF("flush fd=%" PRIu32, fd);
+
+  if (!m_open_file_map.get(fd, fdata)) {
+    char errbuf[32];
+    sprintf(errbuf, "%" PRIu32, fd);
+    cb->error(Error::DFSBROKER_BAD_FILE_HANDLE, errbuf);
+    return;
+  }
+
+  int r;
+  if ((r = ceph_fsync(cmount, fdata->fd, true)) != 0) {
+    std::string errs(cpp_strerror(r));
+    HT_ERRORF("flush failed: fd=%" PRIu32 " ceph_fd=%d - %s", fd, fdata->fd, errs.c_str());
+    report_error(cb, -r);
+    return;
+  }
+
+  cb->response_ok();
+}
+
+void CephBroker::status(ResponseCallback *cb) {
+  cb->response_ok();
+  /*perhaps a total cheat, but both the local and Kosmos brokers
+    included in Hypertable also do this. */
+}
+
+void CephBroker::shutdown(ResponseCallback *cb) {
+  m_open_file_map.remove_all();
+  cb->response_ok();
+  poll(0, 0, 2000);
+}
+
+void CephBroker::readdir(ResponseCallbackReaddir *cb, const char *dname) {
+  std::vector<String> listing;
+  String absdir;
+
+  HT_DEBUGF("Readdir dir='%s'", dname);
+
+  //get from ceph in a buffer
+  make_abs_path(dname, absdir);
+
+  struct ceph_dir_result *dirp;
+  ceph_opendir(cmount, absdir.c_str(), &dirp);
+  int r;
+  int buflen = 100; //good default?
+  char *buf = new char[buflen];
+  String *ent;
+  int bufpos;
+  while (1) {
+    r = ceph_getdnames(cmount, dirp, buf, buflen);
+    if (r==-CEPHFS_ERANGE) { //expand the buffer
+      delete [] buf;
+      buflen *= 2;
+      buf = new char[buflen];
+      continue;
+    }
+    if (r<=0) break;
+
+    //if we make it here, we got at least one name, maybe more
+    bufpos = 0;
+    while (bufpos<r) {//make new strings and add them to listing
+      ent = new String(buf+bufpos);
+      if (ent->compare(".") && ent->compare(".."))
+	listing.push_back(*ent);
+      bufpos+=ent->size()+1;
+      delete ent;
+    }
+  }
+  delete [] buf;
+  ceph_closedir(cmount, dirp);
+
+  if (r < 0) report_error(cb, -r); //Ceph shouldn't return r<0 on getdnames
+  //(except for ERANGE) so if it happens this is bad
+  cb->response(listing);
+}
+
+void CephBroker::exists(ResponseCallbackExists *cb, const char *fname) {
+  String abspath;
+  struct ceph_statx stx;
+  
+  HT_DEBUGF("exists file='%s'", fname);
+  make_abs_path(fname, abspath);
+  cb->response(ceph_statx(cmount, abspath.c_str(), &stx, 0, AT_SYMLINK_NOFOLLOW) == 0);
+}
+
+void CephBroker::rename(ResponseCallback *cb, const char *src, const char *dst) {
+  String src_abs;
+  String dest_abs;
+  int r;
+
+  make_abs_path(src, src_abs);
+  make_abs_path(dst, dest_abs);
+  if ((r = ceph_rename(cmount, src_abs.c_str(), dest_abs.c_str())) <0 ) {
+    report_error(cb, r);
+    return;
+  }
+  cb->response_ok();
+}
+
+void CephBroker::debug(ResponseCallback *cb, int32_t command,
+		       StaticBuffer &serialized_parameters) {
+  HT_ERROR("debug commands not implemented!");
+  cb->error(Error::NOT_IMPLEMENTED, format("Debug commands not supported"));
+}
+
+void CephBroker::report_error(ResponseCallback *cb, int error) {
+  char errbuf[128];
+  errbuf[0] = 0;
+
+  strerror_r(error, errbuf, 128);
+
+  cb->error(Error::DFSBROKER_IO_ERROR, errbuf);
+}
+
+
diff --git a/src/client/hypertable/CephBroker.h b/src/client/hypertable/CephBroker.h
new file mode 100644
index 000000000..d2df38909
--- /dev/null
+++ b/src/client/hypertable/CephBroker.h
@@ -0,0 +1,117 @@
+/** -*- C++ -*-
+ * Copyright (C) 2009-2011 New Dream Network
+ *
+ * This file is part of Hypertable.
+ *
+ * Hypertable is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version 2
+ * of the License, or any later version.
+ *
+ * Hypertable is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with Hypertable. If not, see <http://www.gnu.org/licenses/>
+ *
+ * Authors:
+ * Gregory Farnum <gfarnum@gmail.com>
+ * Colin McCabe <cmccabe@alumni.cmu.edu>
+ */
+
+#ifndef HYPERTABLE_CEPHBROKER_H
+#define HYPERTABLE_CEPHBROKER_H
+
+extern "C" {
+#include <unistd.h>
+}
+#include <atomic>
+#include "Common/String.h"
+#include "Common/Properties.h"
+
+#include "DfsBroker/Lib/Broker.h"
+
+#include <cephfs/libcephfs.h>
+
+namespace Hypertable {
+  using namespace DfsBroker;
+  /**
+   *
+   */
+  class OpenFileDataCeph : public OpenFileData {
+  public:
+    OpenFileDataCeph(struct ceph_mount_info *cmount_, const String& fname,
+		     int _fd, int _flags);
+    virtual ~OpenFileDataCeph();
+    struct ceph_mount_info *cmount;
+    int fd;
+    int flags;
+    String filename;
+  };
+
+  /**
+   *
+   */
+  class OpenFileDataCephPtr : public OpenFileDataPtr {
+  public:
+    OpenFileDataCephPtr() : OpenFileDataPtr() { }
+    explicit OpenFileDataCephPtr(OpenFileDataCeph *ofdl) : OpenFileDataPtr(ofdl, true) { }
+    OpenFileDataCeph *operator->() const { return static_cast<OpenFileDataCeph *>(get()); }
+  };
+
+  /**
+   *
+   */
+  class CephBroker : public DfsBroker::Broker {
+  public:
+    explicit CephBroker(PropertiesPtr& cfg);
+    virtual ~CephBroker();
+
+    virtual void open(ResponseCallbackOpen *cb, const char *fname,
+                      uint32_t flags, uint32_t bufsz);
+    virtual void
+    create(ResponseCallbackOpen *cb, const char *fname, uint32_t flags,
+           int32_t bufsz, int16_t replication, int64_t blksz);
+    virtual void close(ResponseCallback *cb, uint32_t fd);
+    virtual void read(ResponseCallbackRead *cb, uint32_t fd, uint32_t amount);
+    virtual void append(ResponseCallbackAppend *cb, uint32_t fd,
+                        uint32_t amount, const void *data, bool sync);
+    virtual void seek(ResponseCallback *cb, uint32_t fd, uint64_t offset);
+    virtual void remove(ResponseCallback *cb, const char *fname);
+    virtual void length(ResponseCallbackLength *cb, const char *fname, bool);
+    virtual void pread(ResponseCallbackRead *cb, uint32_t fd, uint64_t offset,
+                       uint32_t amount, bool);
+    virtual void mkdirs(ResponseCallback *cb, const char *dname);
+    virtual void rmdir(ResponseCallback *cb, const char *dname);
+    virtual void flush(ResponseCallback *cb, uint32_t fd);
+    virtual void status(ResponseCallback *cb);
+    virtual void shutdown(ResponseCallback *cb);
+    virtual void readdir(ResponseCallbackReaddir *cb, const char *dname);
+    virtual void exists(ResponseCallbackExists *cb, const char *fname);
+    virtual void rename(ResponseCallback *cb, const char *src, const char *dst);
+    virtual void debug(ResponseCallback *, int32_t command,
+                       StaticBuffer &serialized_parameters);
+
+  private:
+    struct ceph_mount_info *cmount;
+    static std::atomic<int> ms_next_fd;
+
+    virtual void report_error(ResponseCallback *cb, int error);
+
+    void make_abs_path(const char *fname, String& abs) {
+      if (fname[0] == '/')
+	abs = fname;
+      else
+	abs = m_root_dir + "/" + fname;
+    }
+
+    int rmdir_recursive(const char *directory);
+
+    bool m_verbose;
+    String m_root_dir;
+  };
+}
+
+#endif //HYPERTABLE_CEPH_BROKER_H
diff --git a/src/client/ioctl.h b/src/client/ioctl.h
new file mode 100644
index 000000000..047543946
--- /dev/null
+++ b/src/client/ioctl.h
@@ -0,0 +1,51 @@
+#ifndef FS_CEPH_IOCTL_H
+#define FS_CEPH_IOCTL_H
+
+#include "include/int_types.h"
+
+#if defined(__linux__)
+#include <linux/ioctl.h>
+#include <linux/types.h>
+#elif defined(__APPLE__) || defined(__FreeBSD__)
+#include <sys/ioctl.h>
+#include <sys/types.h>
+#endif
+
+#define CEPH_IOCTL_MAGIC 0x97
+
+/* just use u64 to align sanely on all archs */
+struct ceph_ioctl_layout {
+	__u64 stripe_unit, stripe_count, object_size;
+	__u64 data_pool;
+	__s64 unused;
+};
+
+#define CEPH_IOC_GET_LAYOUT _IOR(CEPH_IOCTL_MAGIC, 1,		\
+				   struct ceph_ioctl_layout)
+#define CEPH_IOC_SET_LAYOUT _IOW(CEPH_IOCTL_MAGIC, 2,		\
+				   struct ceph_ioctl_layout)
+#define CEPH_IOC_SET_LAYOUT_POLICY _IOW(CEPH_IOCTL_MAGIC, 5,	\
+				   struct ceph_ioctl_layout)
+
+/*
+ * Extract identity, address of the OSD and object storing a given
+ * file offset.
+ */
+struct ceph_ioctl_dataloc {
+	__u64 file_offset;           /* in+out: file offset */
+	__u64 object_offset;         /* out: offset in object */
+	__u64 object_no;             /* out: object # */
+	__u64 object_size;           /* out: object size */
+	char object_name[64];        /* out: object name */
+	__u64 block_offset;          /* out: offset in block */
+	__u64 block_size;            /* out: block length */
+	__s64 osd;                   /* out: osd # */
+	struct sockaddr_storage osd_addr; /* out: osd address */
+};
+
+#define CEPH_IOC_GET_DATALOC _IOWR(CEPH_IOCTL_MAGIC, 3,	\
+				   struct ceph_ioctl_dataloc)
+
+#define CEPH_IOC_LAZYIO _IO(CEPH_IOCTL_MAGIC, 4)
+
+#endif
diff --git a/src/client/posix_acl.cc b/src/client/posix_acl.cc
new file mode 100644
index 000000000..909bb0c2b
--- /dev/null
+++ b/src/client/posix_acl.cc
@@ -0,0 +1,288 @@
+#include "include/compat.h"
+#include "include/types.h"
+#include "include/fs_types.h"
+#include <sys/stat.h>
+#include "posix_acl.h"
+#include "UserPerm.h"
+
+int posix_acl_check(const void *xattr, size_t size)
+{
+  const acl_ea_header *header;
+  if (size < sizeof(*header))
+    return -1;
+  header = reinterpret_cast<const acl_ea_header*>(xattr);
+  ceph_le32 expected_version;
+  expected_version = ACL_EA_VERSION;
+  if (header->a_version != expected_version)
+    return -1;
+
+  const acl_ea_entry *entry = header->a_entries;
+  size -= sizeof(*header);
+  if (size % sizeof(*entry))
+    return -1;
+
+  int count = size / sizeof(*entry);
+  if (count == 0)
+    return 0;
+
+  int state = ACL_USER_OBJ;
+  int needs_mask = 0;
+  for (int i = 0; i < count; ++i) {
+    __u16 tag = entry->e_tag;
+    switch(tag) {
+    case ACL_USER_OBJ:
+      if (state == ACL_USER_OBJ) {
+        state = ACL_USER;
+        break;
+      }
+      return -1;
+    case ACL_USER:
+      if (state != ACL_USER)
+        return -1;
+      needs_mask = 1;
+      break;
+    case ACL_GROUP_OBJ:
+      if (state == ACL_USER) {
+        state = ACL_GROUP;
+        break;
+      }
+      return -1;
+    case ACL_GROUP:
+      if (state != ACL_GROUP)
+        return -1;
+      needs_mask = 1;
+      break;
+    case ACL_MASK:
+      if (state != ACL_GROUP)
+        return -1;
+      state = ACL_OTHER;
+      break;
+    case ACL_OTHER:
+      if (state == ACL_OTHER ||
+          (state == ACL_GROUP && !needs_mask)) {
+        state = 0;
+        break;
+      }
+      // fall-thru
+    default:
+      return -1;
+    }
+    ++entry;
+  }
+
+  return state == 0 ? count : -1;
+}
+
+int posix_acl_equiv_mode(const void *xattr, size_t size, mode_t *mode_p)
+{
+  if (posix_acl_check(xattr, size) < 0)
+    return -CEPHFS_EINVAL;
+
+  int not_equiv = 0;
+  mode_t mode = 0;
+
+  const acl_ea_header *header = reinterpret_cast<const acl_ea_header*>(xattr);
+  const acl_ea_entry *entry = header->a_entries;
+  int count = (size - sizeof(*header)) / sizeof(*entry);
+  for (int i = 0; i < count; ++i) {
+    __u16 tag = entry->e_tag;
+    __u16 perm = entry->e_perm;
+    switch(tag) {
+      case ACL_USER_OBJ:
+	mode |= (perm & S_IRWXO) << 6;
+	break;
+      case ACL_GROUP_OBJ:
+	mode |= (perm & S_IRWXO) << 3;
+	break;
+      case ACL_OTHER:
+	mode |= perm & S_IRWXO;
+	break;
+      case ACL_MASK:
+	mode = (mode & ~S_IRWXG) | ((perm & S_IRWXO) << 3);
+	/* fall through */
+      case ACL_USER:
+      case ACL_GROUP:
+	not_equiv = 1;
+	break;
+      default:
+	return -CEPHFS_EINVAL;
+    }
+    ++entry;
+  }
+  if (mode_p)
+    *mode_p = (*mode_p & ~ACCESSPERMS) | mode;
+  return not_equiv;
+}
+
+int posix_acl_inherit_mode(bufferptr& acl, mode_t *mode_p)
+{
+  if (posix_acl_check(acl.c_str(), acl.length()) <= 0)
+    return -CEPHFS_EIO;
+
+  acl_ea_entry *group_entry = NULL, *mask_entry = NULL;
+  mode_t mode = *mode_p;
+  int not_equiv = 0;
+
+  acl_ea_header *header = reinterpret_cast<acl_ea_header*>(acl.c_str());
+  acl_ea_entry *entry = header->a_entries;
+  int count = (acl.length() - sizeof(*header)) / sizeof(*entry);
+  for (int i = 0; i < count; ++i) {
+    __u16 tag = entry->e_tag;
+    __u16 perm = entry->e_perm;
+    switch(tag) {
+      case ACL_USER_OBJ:
+	perm &= (mode >> 6) | ~S_IRWXO;
+	mode &= (perm << 6) | ~S_IRWXU;
+	entry->e_perm = perm;
+	break;
+      case ACL_USER:
+      case ACL_GROUP:
+	not_equiv = 1;
+	break;
+      case ACL_GROUP_OBJ:
+	group_entry = entry;
+	break;
+      case ACL_OTHER:
+	perm &= mode | ~S_IRWXO;
+	mode &= perm | ~S_IRWXO;
+	entry->e_perm = perm;
+	break;
+      case ACL_MASK:
+	mask_entry = entry;
+	not_equiv = 1;
+	break;
+      default:
+	return -CEPHFS_EIO;
+
+    }
+    ++entry;
+  }
+
+  if (mask_entry) {
+    __u16 perm = mask_entry->e_perm;
+    perm &= (mode >> 3) | ~S_IRWXO;
+    mode &= (perm << 3) | ~S_IRWXG;
+    mask_entry->e_perm = perm;
+  } else {
+    if (!group_entry)
+      return -CEPHFS_EIO;
+    __u16 perm = group_entry->e_perm;
+    perm &= (mode >> 3) | ~S_IRWXO;
+    mode &= (perm << 3) | ~S_IRWXG;
+    group_entry->e_perm = perm;
+  }
+
+  *mode_p = (*mode_p & ~ACCESSPERMS) | mode;
+  return not_equiv;
+}
+
+int posix_acl_access_chmod(bufferptr& acl, mode_t mode)
+{
+  if (posix_acl_check(acl.c_str(), acl.length()) <= 0)
+    return -CEPHFS_EIO;
+
+  acl_ea_entry *group_entry = NULL, *mask_entry = NULL;
+
+  acl_ea_header *header = reinterpret_cast<acl_ea_header*>(acl.c_str());
+  acl_ea_entry *entry = header->a_entries;
+  int count = (acl.length() - sizeof(*header)) / sizeof(*entry);
+  for (int i = 0; i < count; ++i) {
+    __u16 tag = entry->e_tag;
+    switch(tag) {
+      case ACL_USER_OBJ:
+	entry->e_perm = (mode & S_IRWXU) >> 6;
+	break;
+      case ACL_GROUP_OBJ:
+	group_entry = entry;
+	break;
+      case ACL_MASK:
+	mask_entry = entry;
+	break;
+      case ACL_OTHER:
+	entry->e_perm = mode & S_IRWXO;
+	break;
+      default:
+	break;
+    }
+    ++entry;
+  }
+
+  if (mask_entry) {
+    mask_entry->e_perm = (mode & S_IRWXG) >> 3;
+  } else {
+    if (!group_entry)
+      return -CEPHFS_EIO;
+    group_entry->e_perm = (mode & S_IRWXG) >> 3;
+  }
+  return 0;
+}
+
+int posix_acl_permits(const bufferptr& acl, uid_t i_uid, gid_t i_gid,
+			 const UserPerm& perms, unsigned want)
+{
+  if (posix_acl_check(acl.c_str(), acl.length()) < 0)
+    return -CEPHFS_EIO;
+
+  const acl_ea_header *header = reinterpret_cast<const acl_ea_header*>(acl.c_str());
+  const acl_ea_entry *entry = header->a_entries;
+  const acl_ea_entry *next_entry;
+  __u16 perm, tag;
+  __u32 id;
+  int group_found = 0;
+  int idx;
+  int count = (acl.length() - sizeof(*header)) / sizeof(*entry);
+  for (idx = 0; idx < count; ++idx) {
+    tag = entry->e_tag;
+    perm = entry->e_perm;
+    switch(tag) {
+      case ACL_USER_OBJ:
+	if (i_uid == perms.uid())
+	  goto check_perm;
+	break;
+      case ACL_USER:
+	id = entry->e_id;
+	if (id == perms.uid())
+	  goto check_mask;
+	break;
+      case ACL_GROUP_OBJ:
+	/* fall through */
+      case ACL_GROUP:
+	id = (tag == ACL_GROUP_OBJ) ? i_gid : entry->e_id;
+	if (perms.gid_in_groups(id)) {
+	  group_found = 1;
+	  if ((perm & want) == want)
+	    goto check_mask;
+	}
+	break;
+      case ACL_MASK:
+	break;
+      case ACL_OTHER:
+	if (group_found)
+	  return -CEPHFS_EACCES;
+	else
+	  goto check_perm;
+	break;
+      default:
+	return -CEPHFS_EIO;
+    }
+    ++entry;
+  }
+  return -CEPHFS_EIO;
+
+check_mask:
+  next_entry = entry + 1;
+  for (++idx; idx < count; ++idx) {
+    tag = next_entry->e_tag;
+    if (tag == ACL_MASK) {
+      __u16 mask = next_entry->e_perm;
+      if ((perm & mask & want) == want)
+	return 0;
+      return -CEPHFS_EACCES;
+    }
+    ++next_entry;
+  }
+check_perm:
+  if ((perm & want) == want)
+    return 0;
+  return -CEPHFS_EACCES;
+}
diff --git a/src/client/posix_acl.h b/src/client/posix_acl.h
new file mode 100644
index 000000000..4afcc3fe2
--- /dev/null
+++ b/src/client/posix_acl.h
@@ -0,0 +1,35 @@
+#ifndef CEPH_POSIX_ACL
+#define CEPH_POSIX_ACL
+
+#define ACL_EA_VERSION          0x0002
+
+#define ACL_USER_OBJ            0x01
+#define ACL_USER                0x02
+#define ACL_GROUP_OBJ           0x04
+#define ACL_GROUP               0x08
+#define ACL_MASK                0x10
+#define ACL_OTHER               0x20
+
+#define ACL_EA_ACCESS  "system.posix_acl_access"
+#define ACL_EA_DEFAULT "system.posix_acl_default"
+
+typedef struct {
+  ceph_le16       e_tag;
+  ceph_le16       e_perm;
+  ceph_le32       e_id;
+} acl_ea_entry;
+
+typedef struct {
+  ceph_le32       a_version;
+  acl_ea_entry    a_entries[0];
+} acl_ea_header;
+
+class UserPerm;
+
+int posix_acl_check(const void *xattr, size_t size);
+int posix_acl_equiv_mode(const void *xattr, size_t size, mode_t *mode_p);
+int posix_acl_inherit_mode(bufferptr& acl, mode_t *mode_p);
+int posix_acl_access_chmod(bufferptr& acl, mode_t mode);
+int posix_acl_permits(const bufferptr& acl, uid_t i_uid, gid_t i_gid,
+		      const UserPerm& groups, unsigned want);
+#endif
diff --git a/src/client/test_ioctls.c b/src/client/test_ioctls.c
new file mode 100644
index 000000000..23fa835c5
--- /dev/null
+++ b/src/client/test_ioctls.c
@@ -0,0 +1,122 @@
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <limits.h>
+
+#include <sys/ioctl.h>
+#include <netinet/in.h>
+#include <sys/socket.h>
+#include <netdb.h>
+
+#include "ioctl.h"
+
+char new_file_name[PATH_MAX];
+
+int main(int argc, char **argv)
+{
+	char *fn;
+	int fd, err;
+	struct ceph_ioctl_layout l;
+	struct ceph_ioctl_dataloc dl;
+
+	if (argc < 3) {
+		printf("usage: ceph_test_ioctls <filename> <offset>\n");
+		return 1;
+	}
+	fn = argv[1];
+
+	fd = open(fn, O_CREAT|O_RDWR, 0644);
+	if (fd < 0) {
+		perror("couldn't open file");
+		return 1;
+	}
+
+	/* get layout */
+        err = ioctl(fd, CEPH_IOC_GET_LAYOUT, (unsigned long)&l);
+        if (err < 0) {
+                perror("ioctl IOC_GET_LAYOUT error");
+                return 1;
+        }
+        printf("layout:\n stripe_unit %lld\n stripe_count %lld\n object_size %lld\n data_pool %lld\n",
+               (long long)l.stripe_unit, (long long)l.stripe_count, (long long)l.object_size, (long long)l.data_pool);
+
+
+        /* set layout */
+        l.stripe_unit = 1048576;
+        l.stripe_count = 2;
+        err = ioctl(fd, CEPH_IOC_SET_LAYOUT, (unsigned long)&l);
+        if (err < 0) {
+               perror("ioctl IOC_SET_LAYOUT error");
+               return 1;
+        }
+        printf("set layout, writing to file\n");
+
+	printf("file %s\n", fn);
+	/* get layout again */
+	err = ioctl(fd, CEPH_IOC_GET_LAYOUT, (unsigned long)&l);
+	if (err < 0) {
+		perror("ioctl IOC_GET_LAYOUT error");
+		return 1;
+	}
+	printf("layout:\n stripe_unit %lld\n stripe_count %lld\n object_size %lld\n data_pool %lld\n",
+	       (long long)l.stripe_unit, (long long)l.stripe_count, (long long)l.object_size, (long long)l.data_pool);
+
+	/* dataloc */
+	dl.file_offset = atoll(argv[2]);
+	err = ioctl(fd, CEPH_IOC_GET_DATALOC, (unsigned long)&dl);
+	if (err < 0) {
+		perror("ioctl IOC_GET_DATALOC error");
+		return 1;
+	}
+
+	printf("dataloc:\n");
+	printf(" file_offset %lld (of object start)\n", (long long)dl.file_offset);
+	printf(" object '%s'\n object_offset %lld\n object_size %lld object_no %lld\n",
+	       dl.object_name, (long long)dl.object_offset, (long long)dl.object_size, (long long)dl.object_no);
+	printf(" block_offset %lld\n block_size %lld\n",
+	       (long long)dl.block_offset, (long long)dl.block_size);
+
+	char buf[80];
+	getnameinfo((struct sockaddr *)&dl.osd_addr, sizeof(dl.osd_addr), buf, sizeof(buf), 0, 0, NI_NUMERICHOST);
+	printf(" osd%lld %s\n", (long long)dl.osd, buf);
+
+	if (argc < 4)
+	  return 0;
+
+	/* set dir default layout */
+	printf("testing dir policy setting\n");
+	fd = open(argv[3], O_RDONLY);
+        if (fd < 0) {
+                perror("couldn't open dir");
+                return 1;
+        }
+
+        l.object_size = 1048576;
+        l.stripe_count = 1;
+        err = ioctl(fd, CEPH_IOC_SET_LAYOUT_POLICY, (unsigned long)&l);
+        if (err < 0) {
+               perror("ioctl IOC_SET_LAYOUT_POLICY error");
+               return 1;
+        }
+        printf("set layout, creating file\n");
+
+        snprintf(new_file_name, sizeof(new_file_name),
+		 "%s/testfile", argv[3]);
+        fd = open(new_file_name, O_CREAT | O_RDWR, 0644);
+        if (fd < 0) {
+                perror("couldn't open file");
+                return 1;
+        }
+        err = ioctl(fd, CEPH_IOC_GET_LAYOUT, (unsigned long)&l);
+        if (err < 0) {
+                perror("ioctl IOC_GET_LAYOUT error");
+                return 1;
+        }
+        printf("layout:\n stripe_unit %lld\n stripe_count %lld\n object_size %lld\n data_pool %lld\n",
+               (long long)l.stripe_unit, (long long)l.stripe_count, (long long)l.object_size, (long long)l.data_pool);
+        return 0;
+}