diff options
Diffstat (limited to 'src/libradosstriper')
| -rw-r--r-- | src/libradosstriper/CMakeLists.txt | 17 | ||||
| -rw-r--r-- | src/libradosstriper/MultiAioCompletionImpl.cc | 71 | ||||
| -rw-r--r-- | src/libradosstriper/MultiAioCompletionImpl.h | 183 | ||||
| -rw-r--r-- | src/libradosstriper/RadosStriperImpl.cc | 1651 | ||||
| -rw-r--r-- | src/libradosstriper/RadosStriperImpl.h | 274 | ||||
| -rw-r--r-- | src/libradosstriper/libradosstriper.cc | 669 |
6 files changed, 2865 insertions, 0 deletions
diff --git a/src/libradosstriper/CMakeLists.txt b/src/libradosstriper/CMakeLists.txt new file mode 100644 index 00000000..a6919246 --- /dev/null +++ b/src/libradosstriper/CMakeLists.txt @@ -0,0 +1,17 @@ +set(libradosstriper_srcs + libradosstriper.cc + RadosStriperImpl.cc + MultiAioCompletionImpl.cc) +add_library(radosstriper ${CEPH_SHARED} + ${libradosstriper_srcs}) +target_link_libraries(radosstriper + PRIVATE + librados + librados_impl cls_lock_client osdc ceph-common + pthread ${CRYPTO_LIBS} ${EXTRALIBS}) +set_target_properties(radosstriper PROPERTIES + OUPUT_NAME radosstriper + VERSION 1.0.0 + SOVERSION 1) + +install(TARGETS radosstriper DESTINATION ${CMAKE_INSTALL_LIBDIR}) diff --git a/src/libradosstriper/MultiAioCompletionImpl.cc b/src/libradosstriper/MultiAioCompletionImpl.cc new file mode 100644 index 00000000..f13d2c95 --- /dev/null +++ b/src/libradosstriper/MultiAioCompletionImpl.cc @@ -0,0 +1,71 @@ +// -*- 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) 2014 Sebastien Ponce <sebastien.ponce@cern.ch> + * + * 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 "common/dout.h" + +#include "libradosstriper/MultiAioCompletionImpl.h" + +void libradosstriper::MultiAioCompletionImpl::complete_request(ssize_t r) +{ + lock.Lock(); + if (rval >= 0) { + if (r < 0 && r != -EEXIST) + rval = r; + else if (r > 0) + rval += r; + } + ceph_assert(pending_complete); + int count = --pending_complete; + if (!count && !building) { + complete(); + } + put_unlock(); +} + +void libradosstriper::MultiAioCompletionImpl::safe_request(ssize_t r) +{ + lock.Lock(); + if (rval >= 0) { + if (r < 0 && r != -EEXIST) + rval = r; + } + ceph_assert(pending_safe); + int count = --pending_safe; + if (!count && !building) { + safe(); + } + put_unlock(); +} + +void libradosstriper::MultiAioCompletionImpl::finish_adding_requests() +{ + lock.Lock(); + ceph_assert(building); + building = false; + if (!pending_complete) + complete(); + if (!pending_safe) + safe(); + lock.Unlock(); +} + +void intrusive_ptr_add_ref(libradosstriper::MultiAioCompletionImpl* ptr) +{ + ptr->get(); +} + +void intrusive_ptr_release(libradosstriper::MultiAioCompletionImpl* ptr) +{ + ptr->put(); +} diff --git a/src/libradosstriper/MultiAioCompletionImpl.h b/src/libradosstriper/MultiAioCompletionImpl.h new file mode 100644 index 00000000..24188ab6 --- /dev/null +++ b/src/libradosstriper/MultiAioCompletionImpl.h @@ -0,0 +1,183 @@ +// -*- 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) 2014 Sebastien Ponce <sebastien.ponce@cern.ch> + * + * 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_LIBRADOSSTRIPERSTRIPER_MULTIAIOCOMPLETIONIMPL_H +#define CEPH_LIBRADOSSTRIPERSTRIPER_MULTIAIOCOMPLETIONIMPL_H + +#include "common/Cond.h" +#include "common/Mutex.h" + +#include "include/radosstriper/libradosstriper.hpp" + +struct libradosstriper::MultiAioCompletionImpl { + + Mutex lock; + Cond cond; + int ref, rval; + int pending_complete, pending_safe; + rados_callback_t callback_complete, callback_safe; + void *callback_complete_arg, *callback_safe_arg; + bool building; ///< true if we are still building this completion + bufferlist bl; /// only used for read case in C api of rados striper + std::list<bufferlist*> bllist; /// keep temporary buffer lists used for destriping + + MultiAioCompletionImpl() : lock("MultiAioCompletionImpl lock", false, false), + ref(1), rval(0), + pending_complete(0), pending_safe(0), + callback_complete(0), callback_safe(0), + callback_complete_arg(0), callback_safe_arg(0), + building(true) {}; + + ~MultiAioCompletionImpl() { + // deallocate temporary buffer lists + for (std::list<bufferlist*>::iterator it = bllist.begin(); + it != bllist.end(); + it++) { + delete *it; + } + bllist.clear(); + } + + int set_complete_callback(void *cb_arg, rados_callback_t cb) { + lock.Lock(); + callback_complete = cb; + callback_complete_arg = cb_arg; + lock.Unlock(); + return 0; + } + int set_safe_callback(void *cb_arg, rados_callback_t cb) { + lock.Lock(); + callback_safe = cb; + callback_safe_arg = cb_arg; + lock.Unlock(); + return 0; + } + int wait_for_complete() { + lock.Lock(); + while (pending_complete) + cond.Wait(lock); + lock.Unlock(); + return 0; + } + int wait_for_safe() { + lock.Lock(); + while (pending_safe) + cond.Wait(lock); + lock.Unlock(); + return 0; + } + bool is_complete() { + lock.Lock(); + int r = pending_complete; + lock.Unlock(); + return 0 == r; + } + bool is_safe() { + lock.Lock(); + int r = pending_safe; + lock.Unlock(); + return r == 0; + } + void wait_for_complete_and_cb() { + lock.Lock(); + while (pending_complete || callback_complete) + cond.Wait(lock); + lock.Unlock(); + } + void wait_for_safe_and_cb() { + lock.Lock(); + while (pending_safe || callback_safe) + cond.Wait(lock); + lock.Unlock(); + } + bool is_complete_and_cb() { + lock.Lock(); + bool r = ((0 == pending_complete) && !callback_complete); + lock.Unlock(); + return r; + } + bool is_safe_and_cb() { + lock.Lock(); + int r = ((0 == pending_safe) && !callback_safe); + lock.Unlock(); + return r; + } + int get_return_value() { + lock.Lock(); + int r = rval; + lock.Unlock(); + return r; + } + void get() { + lock.Lock(); + _get(); + lock.Unlock(); + } + void _get() { + ceph_assert(lock.is_locked()); + ceph_assert(ref > 0); + ++ref; + } + void put() { + lock.Lock(); + put_unlock(); + } + void put_unlock() { + ceph_assert(ref > 0); + int n = --ref; + lock.Unlock(); + if (!n) + delete this; + } + void add_request() { + lock.Lock(); + pending_complete++; + _get(); + pending_safe++; + _get(); + lock.Unlock(); + } + void add_safe_request() { + lock.Lock(); + pending_complete++; + _get(); + lock.Unlock(); + } + void complete() { + ceph_assert(lock.is_locked()); + if (callback_complete) { + callback_complete(this, callback_complete_arg); + callback_complete = 0; + } + cond.Signal(); + } + void safe() { + ceph_assert(lock.is_locked()); + if (callback_safe) { + callback_safe(this, callback_safe_arg); + callback_safe = 0; + } + cond.Signal(); + }; + + void complete_request(ssize_t r); + void safe_request(ssize_t r); + void finish_adding_requests(); + +}; + +void intrusive_ptr_add_ref(libradosstriper::MultiAioCompletionImpl*); +void intrusive_ptr_release(libradosstriper::MultiAioCompletionImpl*); + +#endif // CEPH_LIBRADOSSTRIPERSTRIPER_MULTIAIOCOMPLETIONIMPL_H diff --git a/src/libradosstriper/RadosStriperImpl.cc b/src/libradosstriper/RadosStriperImpl.cc new file mode 100644 index 00000000..1a3eb99f --- /dev/null +++ b/src/libradosstriper/RadosStriperImpl.cc @@ -0,0 +1,1651 @@ +// -*- 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) 2014 Sebastien Ponce <sebastien.ponce@cern.ch> + * + * 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 <boost/algorithm/string/replace.hpp> + +#include "libradosstriper/RadosStriperImpl.h" + +#include <errno.h> + +#include <sstream> +#include <iomanip> +#include <algorithm> + +#include "include/types.h" +#include "include/uuid.h" +#include "include/ceph_fs.h" +#include "common/dout.h" +#include "common/strtol.h" +#include "osdc/Striper.h" +#include "librados/AioCompletionImpl.h" +#include <cls/lock/cls_lock_client.h> + +/* + * This file contents the actual implementation of the rados striped objects interface. + * + * Striped objects are stored in rados in a set of regular rados objects, after their + * content has been striped using the osdc/Striper interface. + * + * The external attributes of the striped object are mapped to the attributes of the + * first underlying object. This first object has a set of extra external attributes + * storing the layout of the striped object for future read back. These attributes are : + * - striper.layout.object_size : the size of rados objects used. + * Must be a multiple of striper.layout.stripe_unit + * - striper.layout.stripe_unit : the size of a stripe unit + * - striper.layout.stripe_count : the number of stripes used + * - striper.size : total striped object size + * + * In general operations on striped objects are not atomic. + * However, a certain number of safety guards have been put to make the interface closer + * to atomicity : + * - each data operation takes a shared lock on the first rados object for the + * whole time of the operation + * - the remove and trunc operations take an exclusive lock on the first rados object + * for the whole time of the operation + * This makes sure that no removal/truncation of a striped object occurs while + * data operations are happening and vice versa. It thus makes sure that the layout + * of a striped object does not change during data operation, which is essential for + * data consistency. + * + * Still the writing to a striped object is not atomic. This means in particular that + * the size of an object may not be in sync with its content at all times. + * As the size is always guaranteed to be updated first and in an atomic way, and as + * sparse striped objects are supported (see below), what will typically happen is + * that a reader that comes too soon after a write will read 0s instead of the actual + * data. + * + * Note that remove handles the pieces of the striped object in reverse order, + * so that the head object is removed last, making the completion of the deletion atomic. + * + * Striped objects can be sparse, typically in case data was written at the end of the + * striped object only. In such a case, some rados objects constituing the striped object + * may be missing. Other can be partial (only the beginning will have data) + * When dealing with such sparse striped files, missing objects are detected and + * considered as full of 0s. They are however not created until real data is written + * to them. + * + * There are a number of missing features/improvements that could be implemented. + * Here are some ideas : + * - implementation of missing entry points (compared to rados) + * In particular : clone_range, sparse_read, exec, aio_flush_async, tmaps, omaps, ... + * + */ + +#define dout_subsys ceph_subsys_rados +#undef dout_prefix +#define dout_prefix *_dout << "libradosstriper: " + +/// size of xattr buffer +#define XATTR_BUFFER_SIZE 32 + +/// names of the different xattr entries +#define XATTR_LAYOUT_STRIPE_UNIT "striper.layout.stripe_unit" +#define XATTR_LAYOUT_STRIPE_COUNT "striper.layout.stripe_count" +#define XATTR_LAYOUT_OBJECT_SIZE "striper.layout.object_size" +#define XATTR_SIZE "striper.size" +#define LOCK_PREFIX "lock." + +/// name of the lock used on objects to ensure layout stability during IO +#define RADOS_LOCK_NAME "striper.lock" + +/// format of the extension of rados objects created for a given striped object +#define RADOS_OBJECT_EXTENSION_FORMAT ".%016llx" + +/// default object layout +struct ceph_file_layout default_file_layout = { + init_le32(1<<22), // fl_stripe_unit + init_le32(1), // fl_stripe_count + init_le32(1<<22), // fl_object_size + init_le32(0), // fl_cas_hash + init_le32(0), // fl_object_stripe_unit + init_le32(-1), // fl_unused + init_le32(-1), // fl_pg_pool +}; + +using libradosstriper::MultiAioCompletionImplPtr; + +namespace { + +///////////////////////// CompletionData ///////////////////////////// + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous operations + */ +struct CompletionData : RefCountedObject { + /// constructor + CompletionData(libradosstriper::RadosStriperImpl * striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion = 0, + int n = 1); + /// destructor + ~CompletionData() override; + /// complete method + void complete(int r); + /// striper to be used to handle the write completion + libradosstriper::RadosStriperImpl *m_striper; + /// striped object concerned by the write operation + std::string m_soid; + /// shared lock to be released at completion + std::string m_lockCookie; + /// completion handler + librados::IoCtxImpl::C_aio_Complete *m_ack; +}; + +CompletionData::CompletionData +(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + int n) : + RefCountedObject(striper->cct(), n), + m_striper(striper), m_soid(soid), m_lockCookie(lockCookie), m_ack(0) { + m_striper->get(); + if (userCompletion) { + m_ack = new librados::IoCtxImpl::C_aio_Complete(userCompletion); + userCompletion->io = striper->m_ioCtxImpl; + } +} + +CompletionData::~CompletionData() { + if (m_ack) delete m_ack; + m_striper->put(); +} + +void CompletionData::complete(int r) { + if (m_ack) m_ack->finish(r); +} + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous read operations + */ +struct ReadCompletionData : CompletionData { + /// bufferlist containing final result + bufferlist* m_bl; + /// extents that will be read + std::vector<ObjectExtent>* m_extents; + /// intermediate results + std::vector<bufferlist>* m_resultbl; + /// return code of read completion, to be remembered until unlocking happened + int m_readRc; + /// completion object for the unlocking of the striped object at the end of the read + librados::AioCompletion *m_unlockCompletion; + /// constructor + ReadCompletionData(libradosstriper::RadosStriperImpl * striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + bufferlist* bl, + std::vector<ObjectExtent>* extents, + std::vector<bufferlist>* resultbl, + int n); + /// destructor + ~ReadCompletionData() override; + /// complete method for when reading is over + void complete_read(int r); + /// complete method for when object is unlocked + void complete_unlock(int r); +}; + +ReadCompletionData::ReadCompletionData +(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + bufferlist* bl, + std::vector<ObjectExtent>* extents, + std::vector<bufferlist>* resultbl, + int n) : + CompletionData(striper, soid, lockCookie, userCompletion, n), + m_bl(bl), m_extents(extents), m_resultbl(resultbl), m_readRc(0), + m_unlockCompletion(0) {} + +ReadCompletionData::~ReadCompletionData() { + m_unlockCompletion->release(); + delete m_extents; + delete m_resultbl; +} + +void ReadCompletionData::complete_read(int r) { + // gather data into final buffer + Striper::StripedReadResult readResult; + vector<bufferlist>::iterator bit = m_resultbl->begin(); + for (vector<ObjectExtent>::iterator eit = m_extents->begin(); + eit != m_extents->end(); + ++eit, ++bit) { + readResult.add_partial_result(m_striper->cct(), *bit, eit->buffer_extents); + } + m_bl->clear(); + readResult.assemble_result(m_striper->cct(), *m_bl, true); + // Remember return code + m_readRc = r; +} + +void ReadCompletionData::complete_unlock(int r) { + // call parent's completion method + // Note that we ignore the return code of the unlock as we cannot do much about it + CompletionData::complete(m_readRc?m_readRc:m_bl->length()); +} + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous write operations + */ +struct WriteCompletionData : CompletionData { + /// safe completion handler + librados::IoCtxImpl::C_aio_Complete *m_safe; + /// completion object for the unlocking of the striped object at the end of the write + librados::AioCompletion *m_unlockCompletion; + /// return code of write completion, to be remembered until unlocking happened + int m_writeRc; + /// constructor + WriteCompletionData(libradosstriper::RadosStriperImpl * striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + int n); + /// destructor + ~WriteCompletionData() override; + /// complete method for when writing is over + void complete_write(int r); + /// complete method for when object is unlocked + void complete_unlock(int r); + /// safe method + void safe(int r); +}; + +WriteCompletionData::WriteCompletionData +(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + int n) : + CompletionData(striper, soid, lockCookie, userCompletion, n), m_safe(0), + m_unlockCompletion(0), m_writeRc(0) { + if (userCompletion) { + m_safe = new librados::IoCtxImpl::C_aio_Complete(userCompletion); + } +} + +WriteCompletionData::~WriteCompletionData() { + m_unlockCompletion->release(); + if (m_safe) delete m_safe; +} + +void WriteCompletionData::complete_unlock(int r) { + // call parent's completion method + // Note that we ignore the return code of the unlock as we cannot do much about it + CompletionData::complete(m_writeRc); +} + +void WriteCompletionData::complete_write(int r) { + // Remember return code + m_writeRc = r; +} + +void WriteCompletionData::safe(int r) { + if (m_safe) m_safe->finish(r); +} + +struct RemoveCompletionData : CompletionData { + /// removal flags + int flags; + /** + * constructor + * note that the constructed object will take ownership of the lock + */ + RemoveCompletionData(libradosstriper::RadosStriperImpl * striper, + const std::string& soid, + const std::string& lockCookie, + librados::AioCompletionImpl *userCompletion, + int flags = 0) : + CompletionData(striper, soid, lockCookie, userCompletion), flags(flags) {} +}; + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous truncate operations + */ +struct TruncateCompletionData : RefCountedObject { + /// constructor + TruncateCompletionData(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + uint64_t size) : + RefCountedObject(striper->cct()), + m_striper(striper), m_soid(soid), m_size(size) { + m_striper->get(); + } + /// destructor + ~TruncateCompletionData() override { + m_striper->put(); + } + /// striper to be used + libradosstriper::RadosStriperImpl *m_striper; + /// striped object concerned by the truncate operation + std::string m_soid; + /// the final size of the truncated object + uint64_t m_size; +}; + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous read operations of a Rados File + */ +struct RadosReadCompletionData : RefCountedObject { + /// constructor + RadosReadCompletionData(MultiAioCompletionImplPtr multiAioCompl, + uint64_t expectedBytes, + bufferlist *bl, + CephContext *context, + int n = 1) : + RefCountedObject(context, n), + m_multiAioCompl(multiAioCompl), m_expectedBytes(expectedBytes), m_bl(bl) {} + /// the multi asynch io completion object to be used + MultiAioCompletionImplPtr m_multiAioCompl; + /// the expected number of bytes + uint64_t m_expectedBytes; + /// the bufferlist object where data have been written + bufferlist *m_bl; +}; + +/** + * struct handling (most of) the data needed to pass to the call back + * function in asynchronous stat operations. + * Inherited by the actual type for adding time information in different + * versions (time_t or struct timespec) + */ +struct BasicStatCompletionData : CompletionData { + /// constructor + BasicStatCompletionData(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + librados::AioCompletionImpl *userCompletion, + libradosstriper::MultiAioCompletionImpl *multiCompletion, + uint64_t *psize, + int n = 1) : + CompletionData(striper, soid, "", userCompletion, n), + m_multiCompletion(multiCompletion), m_psize(psize), + m_statRC(0), m_getxattrRC(0) {}; + // MultiAioCompletionImpl used to handle the double aysnc + // call in the back (stat + getxattr) + libradosstriper::MultiAioCompletionImpl *m_multiCompletion; + // where to store the size of first objct + // this will be ignored but we need a place to store it when + // async stat is called + uint64_t m_objectSize; + // where to store the file size + uint64_t *m_psize; + /// the bufferlist object used for the getxattr call + bufferlist m_bl; + /// return code of the stat + int m_statRC; + /// return code of the getxattr + int m_getxattrRC; +}; + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous stat operations. + * Simple templated extension of BasicStatCompletionData. + * The template parameter is the type of the time information + * (used with time_t for stat and struct timespec for stat2) + */ +template<class TimeType> +struct StatCompletionData : BasicStatCompletionData { + /// constructor + StatCompletionData(libradosstriper::RadosStriperImpl* striper, + const std::string& soid, + librados::AioCompletionImpl *userCompletion, + libradosstriper::MultiAioCompletionImpl *multiCompletion, + uint64_t *psize, + TimeType *pmtime, + int n = 1) : + BasicStatCompletionData(striper, soid, userCompletion, multiCompletion, psize, n), + m_pmtime(pmtime) {}; + // where to store the file time + TimeType *m_pmtime; +}; + +/** + * struct handling the data needed to pass to the call back + * function in asynchronous remove operations of a Rados File + */ +struct RadosRemoveCompletionData : RefCountedObject { + /// constructor + RadosRemoveCompletionData(MultiAioCompletionImplPtr multiAioCompl, + CephContext *context) : + RefCountedObject(context, 2), + m_multiAioCompl(multiAioCompl) {}; + /// the multi asynch io completion object to be used + MultiAioCompletionImplPtr m_multiAioCompl; +}; + + +} // namespace { + +///////////////////////// constructor ///////////////////////////// + +libradosstriper::RadosStriperImpl::RadosStriperImpl(librados::IoCtx& ioctx, librados::IoCtxImpl *ioctx_impl) : + m_refCnt(0),lock("RadosStriper Refcont", false, false), m_radosCluster(ioctx), m_ioCtx(ioctx), m_ioCtxImpl(ioctx_impl), + m_layout(default_file_layout) {} + +///////////////////////// layout ///////////////////////////// + +int libradosstriper::RadosStriperImpl::setObjectLayoutStripeUnit +(unsigned int stripe_unit) +{ + /* stripe unit must be non-zero, 64k increment */ + if (!stripe_unit || (stripe_unit & (CEPH_MIN_STRIPE_UNIT-1))) + return -EINVAL; + m_layout.fl_stripe_unit = stripe_unit; + return 0; +} + +int libradosstriper::RadosStriperImpl::setObjectLayoutStripeCount +(unsigned int stripe_count) +{ + /* stripe count must be non-zero */ + if (!stripe_count) + return -EINVAL; + m_layout.fl_stripe_count = stripe_count; + return 0; +} + +int libradosstriper::RadosStriperImpl::setObjectLayoutObjectSize +(unsigned int object_size) +{ + /* object size must be non-zero, 64k increment */ + if (!object_size || (object_size & (CEPH_MIN_STRIPE_UNIT-1))) + return -EINVAL; + /* object size must be a multiple of stripe unit */ + if (object_size < m_layout.fl_stripe_unit || + object_size % m_layout.fl_stripe_unit) + return -EINVAL; + m_layout.fl_object_size = object_size; + return 0; +} + +///////////////////////// xattrs ///////////////////////////// + +int libradosstriper::RadosStriperImpl::getxattr(const object_t& soid, + const char *name, + bufferlist& bl) +{ + std::string firstObjOid = getObjectId(soid, 0); + return m_ioCtx.getxattr(firstObjOid, name, bl); +} + +int libradosstriper::RadosStriperImpl::setxattr(const object_t& soid, + const char *name, + bufferlist& bl) +{ + std::string firstObjOid = getObjectId(soid, 0); + return m_ioCtx.setxattr(firstObjOid, name, bl); +} + +int libradosstriper::RadosStriperImpl::getxattrs(const object_t& soid, + map<string, bufferlist>& attrset) +{ + std::string firstObjOid = getObjectId(soid, 0); + int rc = m_ioCtx.getxattrs(firstObjOid, attrset); + if (rc) return rc; + // cleanup internal attributes dedicated to striping and locking + attrset.erase(XATTR_LAYOUT_STRIPE_UNIT); + attrset.erase(XATTR_LAYOUT_STRIPE_COUNT); + attrset.erase(XATTR_LAYOUT_OBJECT_SIZE); + attrset.erase(XATTR_SIZE); + attrset.erase(std::string(LOCK_PREFIX) + RADOS_LOCK_NAME); + return rc; +} + +int libradosstriper::RadosStriperImpl::rmxattr(const object_t& soid, + const char *name) +{ + std::string firstObjOid = getObjectId(soid, 0); + return m_ioCtx.rmxattr(firstObjOid, name); +} + +///////////////////////// io ///////////////////////////// + +int libradosstriper::RadosStriperImpl::write(const std::string& soid, + const bufferlist& bl, + size_t len, + uint64_t off) +{ + // open the object. This will create it if needed, retrieve its layout + // and size and take a shared lock on it + ceph_file_layout layout; + std::string lockCookie; + int rc = createAndOpenStripedObject(soid, &layout, len+off, &lockCookie, true); + if (rc) return rc; + return write_in_open_object(soid, layout, lockCookie, bl, len, off); +} + +int libradosstriper::RadosStriperImpl::append(const std::string& soid, + const bufferlist& bl, + size_t len) +{ + // open the object. This will create it if needed, retrieve its layout + // and size and take a shared lock on it + ceph_file_layout layout; + uint64_t size = len; + std::string lockCookie; + int rc = openStripedObjectForWrite(soid, &layout, &size, &lockCookie, false); + if (rc) return rc; + return write_in_open_object(soid, layout, lockCookie, bl, len, size); +} + +int libradosstriper::RadosStriperImpl::write_full(const std::string& soid, + const bufferlist& bl) +{ + int rc = trunc(soid, 0); + if (rc && rc != -ENOENT) return rc; // ENOENT is obviously ok + return write(soid, bl, bl.length(), 0); +} + +int libradosstriper::RadosStriperImpl::read(const std::string& soid, + bufferlist* bl, + size_t len, + uint64_t off) +{ + // create a completion object + librados::AioCompletionImpl c; + // call asynchronous method + int rc = aio_read(soid, &c, bl, len, off); + // and wait for completion + if (!rc) { + // wait for completion + c.wait_for_complete_and_cb(); + // return result + rc = c.get_return_value(); + } + return rc; +} + +///////////////////////// asynchronous io ///////////////////////////// + +int libradosstriper::RadosStriperImpl::aio_write(const std::string& soid, + librados::AioCompletionImpl *c, + const bufferlist& bl, + size_t len, + uint64_t off) +{ + ceph_file_layout layout; + std::string lockCookie; + int rc = createAndOpenStripedObject(soid, &layout, len+off, &lockCookie, true); + if (rc) return rc; + return aio_write_in_open_object(soid, c, layout, lockCookie, bl, len, off); +} + +int libradosstriper::RadosStriperImpl::aio_append(const std::string& soid, + librados::AioCompletionImpl *c, + const bufferlist& bl, + size_t len) +{ + ceph_file_layout layout; + uint64_t size = len; + std::string lockCookie; + int rc = openStripedObjectForWrite(soid, &layout, &size, &lockCookie, false); + if (rc) return rc; + // create a completion object + return aio_write_in_open_object(soid, c, layout, lockCookie, bl, len, size); +} + +int libradosstriper::RadosStriperImpl::aio_write_full(const std::string& soid, + librados::AioCompletionImpl *c, + const bufferlist& bl) +{ + int rc = trunc(soid, 0); + if (rc) return rc; + return aio_write(soid, c, bl, bl.length(), 0); +} + +static void rados_read_aio_unlock_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<ReadCompletionData*>(arg); + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + cdata->complete_unlock(comp->rval); + cdata->put(); +} + +static void striper_read_aio_req_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<ReadCompletionData*>(arg); + // launch the async unlocking of the object + cdata->m_striper->aio_unlockObject(cdata->m_soid, cdata->m_lockCookie, cdata->m_unlockCompletion); + // complete the read part in parallel + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + cdata->complete_read(comp->rval); +} + +static void rados_req_read_safe(rados_completion_t c, void *arg) +{ + auto data = reinterpret_cast<RadosReadCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + // ENOENT means that we are dealing with a sparse file. This is fine, + // data (0s) will be created on the fly by the rados_req_read_complete method + if (rc == -ENOENT) rc = 0; + auto multiAioComp = data->m_multiAioCompl; + multiAioComp->safe_request(rc); + data->put(); +} + +static void rados_req_read_complete(rados_completion_t c, void *arg) +{ + auto data = reinterpret_cast<RadosReadCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + // We need to handle the case of sparse files here + if (rc == -ENOENT) { + // the object did not exist at all. This can happen for sparse files. + // we consider we've read 0 bytes and it will fall into next case + rc = 0; + } + if (rc >= 0 && (((uint64_t)rc) < data->m_expectedBytes)) { + // only partial data were present in the object (or the object did not + // even exist if we've gone through previous case). + // This is typical of sparse file and we need to complete with 0s. + unsigned int lenOfZeros = data->m_expectedBytes-rc; + unsigned int existingDataToZero = min(data->m_bl->length()-rc, lenOfZeros); + if (existingDataToZero > 0) { + data->m_bl->zero(rc, existingDataToZero); + } + if (lenOfZeros > existingDataToZero) { + ceph::bufferptr zeros(ceph::buffer::create(lenOfZeros-existingDataToZero)); + zeros.zero(); + data->m_bl->push_back(zeros); + } + rc = data->m_expectedBytes; + } + auto multiAioComp = data->m_multiAioCompl; + multiAioComp->complete_request(rc); + data->put(); +} + +int libradosstriper::RadosStriperImpl::aio_read(const std::string& soid, + librados::AioCompletionImpl *c, + bufferlist* bl, + size_t len, + uint64_t off) +{ + // open the object. This will retrieve its layout and size + // and take a shared lock on it + ceph_file_layout layout; + uint64_t size; + std::string lockCookie; + int rc = openStripedObjectForRead(soid, &layout, &size, &lockCookie); + if (rc) return rc; + // find out the actual number of bytes we can read + uint64_t read_len; + if (off >= size) { + // nothing to read ! We are done. + read_len = 0; + } else { + read_len = min(len, (size_t)(size-off)); + } + // get list of extents to be read from + vector<ObjectExtent> *extents = new vector<ObjectExtent>(); + if (read_len > 0) { + std::string format = soid; + boost::replace_all(format, "%", "%%"); + format += RADOS_OBJECT_EXTENSION_FORMAT; + file_layout_t l; + l.from_legacy(layout); + Striper::file_to_extents(cct(), format.c_str(), &l, off, read_len, + 0, *extents); + } + + // create a completion object and transfer ownership of extents and resultbl + vector<bufferlist> *resultbl = new vector<bufferlist>(extents->size()); + ReadCompletionData *cdata = new ReadCompletionData(this, soid, lockCookie, c, + bl, extents, resultbl, 1); + c->is_read = true; + c->io = m_ioCtxImpl; + // create a completion for the unlocking of the striped object at the end of the read + librados::AioCompletion *unlock_completion = + librados::Rados::aio_create_completion(cdata, rados_read_aio_unlock_complete, 0); + cdata->m_unlockCompletion = unlock_completion; + // create the multiCompletion object handling the reads + MultiAioCompletionImplPtr nc{new libradosstriper::MultiAioCompletionImpl, + false}; + nc->set_complete_callback(cdata, striper_read_aio_req_complete); + // go through the extents + int r = 0, i = 0; + for (vector<ObjectExtent>::iterator p = extents->begin(); p != extents->end(); ++p) { + // create a buffer list describing where to place data read from current extend + bufferlist *oid_bl = &((*resultbl)[i++]); + for (vector<pair<uint64_t,uint64_t> >::iterator q = p->buffer_extents.begin(); + q != p->buffer_extents.end(); + ++q) { + bufferlist buffer_bl; + buffer_bl.substr_of(*bl, q->first, q->second); + oid_bl->append(buffer_bl); + } + // read all extends of a given object in one go + nc->add_request(); + // we need 2 references on data as both rados_req_read_safe and rados_req_read_complete + // will release one + RadosReadCompletionData *data = new RadosReadCompletionData(nc, p->length, oid_bl, cct(), 2); + librados::AioCompletion *rados_completion = + librados::Rados::aio_create_completion(data, rados_req_read_complete, rados_req_read_safe); + r = m_ioCtx.aio_read(p->oid.name, rados_completion, oid_bl, p->length, p->offset); + rados_completion->release(); + if (r < 0) + break; + } + nc->finish_adding_requests(); + return r; +} + +int libradosstriper::RadosStriperImpl::aio_read(const std::string& soid, + librados::AioCompletionImpl *c, + char* buf, + size_t len, + uint64_t off) +{ + // create a buffer list and store it inside the completion object + c->bl.clear(); + c->bl.push_back(buffer::create_static(len, buf)); + // call the bufferlist version of this method + return aio_read(soid, c, &c->bl, len, off); +} + +int libradosstriper::RadosStriperImpl::aio_flush() +{ + int ret; + // pass to the rados level + ret = m_ioCtx.aio_flush(); + if (ret < 0) + return ret; + //wait all CompletionData are released + lock.Lock(); + while (m_refCnt > 1) + cond.Wait(lock); + lock.Unlock(); + return ret; +} + +///////////////////////// stat and deletion ///////////////////////////// + +int libradosstriper::RadosStriperImpl::stat(const std::string& soid, uint64_t *psize, time_t *pmtime) +{ + // create a completion object + librados::AioCompletionImpl c; + // call asynchronous version of stat + int rc = aio_stat(soid, &c, psize, pmtime); + if (rc == 0) { + // wait for completion of the remove + c.wait_for_complete(); + // get result + rc = c.get_return_value(); + } + return rc; +} + +static void striper_stat_aio_stat_complete(rados_completion_t c, void *arg) { + auto data = reinterpret_cast<BasicStatCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + if (rc == -ENOENT) { + // remember this has failed + data->m_statRC = rc; + } + data->m_multiCompletion->complete_request(rc); + data->put(); +} + +static void striper_stat_aio_getxattr_complete(rados_completion_t c, void *arg) { + auto data = reinterpret_cast<BasicStatCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + // We need to handle the case of sparse files here + if (rc < 0) { + // remember this has failed + data->m_getxattrRC = rc; + } else { + // this intermediate string allows to add a null terminator before calling strtol + std::string err; + std::string strsize(data->m_bl.c_str(), data->m_bl.length()); + *data->m_psize = strict_strtoll(strsize.c_str(), 10, &err); + if (!err.empty()) { + lderr(data->m_striper->cct()) << XATTR_SIZE << " : " << err << dendl; + data->m_getxattrRC = -EINVAL; + } + rc = 0; + } + data->m_multiCompletion->complete_request(rc); + data->put(); +} + +static void striper_stat_aio_req_complete(rados_striper_multi_completion_t c, + void *arg) { + auto data = reinterpret_cast<BasicStatCompletionData*>(arg); + if (data->m_statRC) { + data->complete(data->m_statRC); + } else { + if (data->m_getxattrRC < 0) { + data->complete(data->m_getxattrRC); + } else { + data->complete(0); + } + } + data->put(); +} + +template<class TimeType> +int libradosstriper::RadosStriperImpl::aio_generic_stat +(const std::string& soid, + librados::AioCompletionImpl *c, + uint64_t *psize, + TimeType *pmtime, + typename libradosstriper::RadosStriperImpl::StatFunction<TimeType>::Type statFunction) +{ + // use a MultiAioCompletion object for dealing with the fact + // that we'll do 2 asynchronous calls in parallel + MultiAioCompletionImplPtr multi_completion{ + new libradosstriper::MultiAioCompletionImpl, false}; + // Data object used for passing context to asynchronous calls + std::string firstObjOid = getObjectId(soid, 0); + StatCompletionData<TimeType> *cdata = + new StatCompletionData<TimeType>(this, firstObjOid, c, + multi_completion.get(), psize, pmtime, 4); + multi_completion->set_complete_callback(cdata, striper_stat_aio_req_complete); + // use a regular AioCompletion for the stat async call + librados::AioCompletion *stat_completion = + librados::Rados::aio_create_completion(cdata, striper_stat_aio_stat_complete, 0); + multi_completion->add_safe_request(); + object_t obj(firstObjOid); + int rc = (m_ioCtxImpl->*statFunction)(obj, stat_completion->pc, + &cdata->m_objectSize, cdata->m_pmtime); + stat_completion->release(); + if (rc < 0) { + // nothing is really started so cancel everything + delete cdata; + return rc; + } + // use a regular AioCompletion for the getxattr async call + librados::AioCompletion *getxattr_completion = + librados::Rados::aio_create_completion(cdata, striper_stat_aio_getxattr_complete, 0); + multi_completion->add_safe_request(); + // in parallel, get the pmsize from the first object asynchronously + rc = m_ioCtxImpl->aio_getxattr(obj, getxattr_completion->pc, + XATTR_SIZE, cdata->m_bl); + getxattr_completion->release(); + multi_completion->finish_adding_requests(); + if (rc < 0) { + // the async stat is ongoing, so we need to go on + // we mark the getxattr as failed in the data object + cdata->m_getxattrRC = rc; + multi_completion->complete_request(rc); + return rc; + } + cdata->put(); + return 0; +} + +int libradosstriper::RadosStriperImpl::aio_stat(const std::string& soid, + librados::AioCompletionImpl *c, + uint64_t *psize, + time_t *pmtime) +{ + return aio_generic_stat<time_t>(soid, c, psize, pmtime, &librados::IoCtxImpl::aio_stat); +} + +int libradosstriper::RadosStriperImpl::stat2(const std::string& soid, uint64_t *psize, struct timespec *pts) +{ + // create a completion object + librados::AioCompletionImpl c; + // call asynchronous version of stat + int rc = aio_stat2(soid, &c, psize, pts); + if (rc == 0) { + // wait for completion of the remove + c.wait_for_complete_and_cb(); + // get result + rc = c.get_return_value(); + } + return rc; +} + +int libradosstriper::RadosStriperImpl::aio_stat2(const std::string& soid, + librados::AioCompletionImpl *c, + uint64_t *psize, + struct timespec *pts) +{ + return aio_generic_stat<struct timespec>(soid, c, psize, pts, &librados::IoCtxImpl::aio_stat2); +} + +static void rados_req_remove_complete(rados_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<RadosRemoveCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + // in case the object did not exist, it means we had a sparse file, all is fine + if (rc == -ENOENT) { + rc = 0; + } + cdata->m_multiAioCompl->complete_request(rc); + cdata->put(); +} + +static void rados_req_remove_safe(rados_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<RadosRemoveCompletionData*>(arg); + int rc = rados_aio_get_return_value(c); + // in case the object did not exist, it means we had a sparse file, all is fine + if (rc == -ENOENT) { + rc = 0; + } + cdata->m_multiAioCompl->safe_request(rc); + cdata->put(); +} + +static void striper_remove_aio_req_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<RemoveCompletionData*>(arg); + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + ldout(cdata->m_striper->cct(), 10) + << "RadosStriperImpl : striper_remove_aio_req_complete called for " + << cdata->m_soid << dendl; + int rc = comp->rval; + if (rc == 0) { + // All went fine, synchronously remove first object + rc = cdata->m_striper->m_ioCtx.remove(cdata->m_striper->getObjectId(cdata->m_soid, 0), + cdata->flags); + } else { + lderr(cdata->m_striper->cct()) + << "RadosStriperImpl : deletion/truncation incomplete for " << cdata->m_soid + << ", as errors were encountered. The file is left present but it's content " + << " has been partially removed" + << dendl; + } + cdata->complete(rc); + cdata->put(); +} + +int libradosstriper::RadosStriperImpl::remove(const std::string& soid, int flags) +{ + // create a completion object + librados::AioCompletionImpl c; + // call asynchronous version of remove + int rc = aio_remove(soid, &c, flags); + if (rc == 0) { + // wait for completion of the remove + c.wait_for_complete_and_cb(); + // get result + rc = c.get_return_value(); + } + return rc; +} + +int libradosstriper::RadosStriperImpl::aio_remove(const std::string& soid, + librados::AioCompletionImpl *c, + int flags) +{ + // the RemoveCompletionData object will lock the given soid for the duration + // of the removal + std::string lockCookie = getUUID(); + int rc = m_ioCtx.lock_exclusive(getObjectId(soid, 0), RADOS_LOCK_NAME, lockCookie, "", 0, 0); + if (rc) return rc; + // create CompletionData for the async remove call + RemoveCompletionData *cdata = new RemoveCompletionData(this, soid, lockCookie, c, flags); + MultiAioCompletionImplPtr multi_completion{ + new libradosstriper::MultiAioCompletionImpl, false}; + multi_completion->set_complete_callback(cdata, striper_remove_aio_req_complete); + // call asynchronous internal version of remove + ldout(cct(), 10) + << "RadosStriperImpl : Aio_remove starting for " + << soid << dendl; + rc = internal_aio_remove(soid, multi_completion); + return rc; +} + +int libradosstriper::RadosStriperImpl::internal_aio_remove( + const std::string& soid, + MultiAioCompletionImplPtr multi_completion, + int flags) +{ + std::string firstObjOid = getObjectId(soid, 0); + try { + // check size and get number of rados objects to delete + uint64_t nb_objects = 0; + bufferlist bl2; + int rc = getxattr(soid, XATTR_SIZE, bl2); + if (rc < 0) { + // no object size (or not able to get it) + // try to find the number of object "by hand" + uint64_t psize; + time_t pmtime; + while (!m_ioCtx.stat(getObjectId(soid, nb_objects), &psize, &pmtime)) { + nb_objects++; + } + } else { + // count total number of rados objects in the striped object + std::string err; + // this intermediate string allows to add a null terminator before calling strtol + std::string strsize(bl2.c_str(), bl2.length()); + uint64_t size = strict_strtoll(strsize.c_str(), 10, &err); + if (!err.empty()) { + lderr(cct()) << XATTR_SIZE << " : " << err << dendl; + + return -EINVAL; + } + uint64_t object_size = m_layout.fl_object_size; + uint64_t su = m_layout.fl_stripe_unit; + uint64_t stripe_count = m_layout.fl_stripe_count; + uint64_t nb_complete_sets = size / (object_size*stripe_count); + uint64_t remaining_data = size % (object_size*stripe_count); + uint64_t remaining_stripe_units = (remaining_data + su -1) / su; + uint64_t remaining_objects = std::min(remaining_stripe_units, stripe_count); + nb_objects = nb_complete_sets * stripe_count + remaining_objects; + } + // delete rados objects in reverse order + // Note that we do not drop the first object. This one will only be dropped + // if all other removals have been successful, and this is done in the + // callback of the multi_completion object + int rcr = 0; + for (int i = nb_objects-1; i >= 1; i--) { + multi_completion->add_request(); + RadosRemoveCompletionData *data = + new RadosRemoveCompletionData(multi_completion, cct()); + librados::AioCompletion *rados_completion = + librados::Rados::aio_create_completion(data, + rados_req_remove_complete, + rados_req_remove_safe); + if (flags == 0) { + rcr = m_ioCtx.aio_remove(getObjectId(soid, i), rados_completion); + } else { + rcr = m_ioCtx.aio_remove(getObjectId(soid, i), rados_completion, flags); + } + rados_completion->release(); + if (rcr < 0 and -ENOENT != rcr) { + lderr(cct()) << "RadosStriperImpl::remove : deletion incomplete for " << soid + << ", as " << getObjectId(soid, i) << " could not be deleted (rc=" << rc << ")" + << dendl; + break; + } + } + // we are over adding requests to the multi_completion object + multi_completion->finish_adding_requests(); + // return + return rcr; + } catch (ErrorCode &e) { + // error caught when trying to take the exclusive lock + return e.m_code; + } + +} + +int libradosstriper::RadosStriperImpl::trunc(const std::string& soid, uint64_t size) +{ + // lock the object in exclusive mode + std::string firstObjOid = getObjectId(soid, 0); + librados::ObjectWriteOperation op; + op.assert_exists(); + std::string lockCookie = RadosStriperImpl::getUUID(); + utime_t dur = utime_t(); + rados::cls::lock::lock(&op, RADOS_LOCK_NAME, LOCK_EXCLUSIVE, lockCookie, "", "", dur, 0); + int rc = m_ioCtx.operate(firstObjOid, &op); + if (rc) return rc; + // load layout and size + ceph_file_layout layout; + uint64_t original_size; + rc = internal_get_layout_and_size(firstObjOid, &layout, &original_size); + if (!rc) { + if (size < original_size) { + rc = truncate(soid, original_size, size, layout); + } else if (size > original_size) { + rc = grow(soid, original_size, size, layout); + } + } + // unlock object, ignore return code as we cannot do much + m_ioCtx.unlock(firstObjOid, RADOS_LOCK_NAME, lockCookie); + // final return + return rc; +} + + +///////////////////////// private helpers ///////////////////////////// + +std::string libradosstriper::RadosStriperImpl::getObjectId(const object_t& soid, + long long unsigned objectno) +{ + std::ostringstream s; + s << soid << '.' << std::setfill ('0') << std::setw(16) << std::hex << objectno; + return s.str(); +} + +void libradosstriper::RadosStriperImpl::unlockObject(const std::string& soid, + const std::string& lockCookie) +{ + // unlock the shared lock on the first rados object + std::string firstObjOid = getObjectId(soid, 0); + m_ioCtx.unlock(firstObjOid, RADOS_LOCK_NAME, lockCookie); +} + +void libradosstriper::RadosStriperImpl::aio_unlockObject(const std::string& soid, + const std::string& lockCookie, + librados::AioCompletion *c) +{ + // unlock the shared lock on the first rados object + std::string firstObjOid = getObjectId(soid, 0); + m_ioCtx.aio_unlock(firstObjOid, RADOS_LOCK_NAME, lockCookie, c); +} + +static void rados_write_aio_unlock_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<WriteCompletionData*>(arg); + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + cdata->complete_unlock(comp->rval); + cdata->put(); +} + +static void striper_write_aio_req_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<WriteCompletionData*>(arg); + // launch the async unlocking of the object + cdata->m_striper->aio_unlockObject(cdata->m_soid, cdata->m_lockCookie, cdata->m_unlockCompletion); + // complete the write part in parallel + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + cdata->complete_write(comp->rval); + cdata->put(); +} + +static void striper_write_aio_req_safe(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<WriteCompletionData*>(arg); + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + cdata->safe(comp->rval); + cdata->put(); +} + +int libradosstriper::RadosStriperImpl::write_in_open_object(const std::string& soid, + const ceph_file_layout& layout, + const std::string& lockCookie, + const bufferlist& bl, + size_t len, + uint64_t off) { + // create a completion object to be passed to the callbacks of the multicompletion + // we need 3 references as striper_write_aio_req_complete will release two and + // striper_write_aio_req_safe will release one + WriteCompletionData *cdata = new WriteCompletionData(this, soid, lockCookie, 0, 3); + cdata->get(); // local ref + // create a completion object for the unlocking of the striped object at the end of the write + librados::AioCompletion *unlock_completion = + librados::Rados::aio_create_completion(cdata, rados_write_aio_unlock_complete, 0); + cdata->m_unlockCompletion = unlock_completion; + // create the multicompletion that will handle the write completion + MultiAioCompletionImplPtr c{new libradosstriper::MultiAioCompletionImpl, + false}; + c->set_complete_callback(cdata, striper_write_aio_req_complete); + c->set_safe_callback(cdata, striper_write_aio_req_safe); + // call the asynchronous API + int rc = internal_aio_write(soid, c, bl, len, off, layout); + if (!rc) { + // wait for completion and safety of data + c->wait_for_complete_and_cb(); + c->wait_for_safe_and_cb(); + // wait for the unlocking + unlock_completion->wait_for_complete(); + // return result + rc = c->get_return_value(); + } + cdata->put(); + return rc; +} + +int libradosstriper::RadosStriperImpl::aio_write_in_open_object(const std::string& soid, + librados::AioCompletionImpl *c, + const ceph_file_layout& layout, + const std::string& lockCookie, + const bufferlist& bl, + size_t len, + uint64_t off) { + // create a completion object to be passed to the callbacks of the multicompletion + // we need 3 references as striper_write_aio_req_complete will release two and + // striper_write_aio_req_safe will release one + WriteCompletionData *cdata = new WriteCompletionData(this, soid, lockCookie, c, 3); + cdata->get(); // local ref + m_ioCtxImpl->get(); + c->io = m_ioCtxImpl; + // create a completion object for the unlocking of the striped object at the end of the write + librados::AioCompletion *unlock_completion = + librados::Rados::aio_create_completion(cdata, rados_write_aio_unlock_complete, 0); + cdata->m_unlockCompletion = unlock_completion; + // create the multicompletion that will handle the write completion + libradosstriper::MultiAioCompletionImplPtr nc{ + new libradosstriper::MultiAioCompletionImpl, false}; + nc->set_complete_callback(cdata, striper_write_aio_req_complete); + nc->set_safe_callback(cdata, striper_write_aio_req_safe); + // internal asynchronous API + int rc = internal_aio_write(soid, nc, bl, len, off, layout); + cdata->put(); + return rc; +} + +static void rados_req_write_safe(rados_completion_t c, void *arg) +{ + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(arg); + comp->safe_request(rados_aio_get_return_value(c)); +} + +static void rados_req_write_complete(rados_completion_t c, void *arg) +{ + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(arg); + comp->complete_request(rados_aio_get_return_value(c)); +} + +int +libradosstriper::RadosStriperImpl::internal_aio_write(const std::string& soid, + libradosstriper::MultiAioCompletionImplPtr c, + const bufferlist& bl, + size_t len, + uint64_t off, + const ceph_file_layout& layout) +{ + int r = 0; + // Do not try anything if we are called with empty buffer, + // file_to_extents would raise an exception + if (len > 0) { + // get list of extents to be written to + vector<ObjectExtent> extents; + std::string format = soid; + boost::replace_all(format, "%", "%%"); + format += RADOS_OBJECT_EXTENSION_FORMAT; + file_layout_t l; + l.from_legacy(layout); + Striper::file_to_extents(cct(), format.c_str(), &l, off, len, 0, extents); + // go through the extents + for (vector<ObjectExtent>::iterator p = extents.begin(); p != extents.end(); ++p) { + // assemble pieces of a given object into a single buffer list + bufferlist oid_bl; + for (vector<pair<uint64_t,uint64_t> >::iterator q = p->buffer_extents.begin(); + q != p->buffer_extents.end(); + ++q) { + bufferlist buffer_bl; + buffer_bl.substr_of(bl, q->first, q->second); + oid_bl.append(buffer_bl); + } + // and write the object + c->add_request(); + librados::AioCompletion *rados_completion = + librados::Rados::aio_create_completion(c.get(), + rados_req_write_complete, + rados_req_write_safe); + r = m_ioCtx.aio_write(p->oid.name, rados_completion, oid_bl, + p->length, p->offset); + rados_completion->release(); + if (r < 0) + break; + } + } + c->finish_adding_requests(); + return r; +} + +int libradosstriper::RadosStriperImpl::extract_uint32_attr +(std::map<std::string, bufferlist> &attrs, + const std::string& key, + ceph_le32 *value) +{ + std::map<std::string, bufferlist>::iterator attrsIt = attrs.find(key); + if (attrsIt != attrs.end()) { + // this intermediate string allows to add a null terminator before calling strtol + std::string strvalue(attrsIt->second.c_str(), attrsIt->second.length()); + std::string err; + *value = strict_strtol(strvalue.c_str(), 10, &err); + if (!err.empty()) { + lderr(cct()) << key << " : " << err << dendl; + return -EINVAL; + } + } else { + return -ENOENT; + } + return 0; +} + +int libradosstriper::RadosStriperImpl::extract_sizet_attr +(std::map<std::string, bufferlist> &attrs, + const std::string& key, + size_t *value) +{ + std::map<std::string, bufferlist>::iterator attrsIt = attrs.find(key); + if (attrsIt != attrs.end()) { + // this intermediate string allows to add a null terminator before calling strtol + std::string strvalue(attrsIt->second.c_str(), attrsIt->second.length()); + std::string err; + *value = strict_strtoll(strvalue.c_str(), 10, &err); + if (!err.empty()) { + lderr(cct()) << key << " : " << err << dendl; + return -EINVAL; + } + } else { + return -ENOENT; + } + return 0; +} + +int libradosstriper::RadosStriperImpl::internal_get_layout_and_size( + const std::string& oid, + ceph_file_layout *layout, + uint64_t *size) +{ + // get external attributes of the first rados object + std::map<std::string, bufferlist> attrs; + int rc = m_ioCtx.getxattrs(oid, attrs); + if (rc) return rc; + // deal with stripe_unit + rc = extract_uint32_attr(attrs, XATTR_LAYOUT_STRIPE_UNIT, &layout->fl_stripe_unit); + if (rc) return rc; + // deal with stripe_count + rc = extract_uint32_attr(attrs, XATTR_LAYOUT_STRIPE_COUNT, &layout->fl_stripe_count); + if (rc) return rc; + // deal with object_size + rc = extract_uint32_attr(attrs, XATTR_LAYOUT_OBJECT_SIZE, &layout->fl_object_size); + if (rc) return rc; + // deal with size + size_t ssize; + rc = extract_sizet_attr(attrs, XATTR_SIZE, &ssize); + if (rc) { + return rc; + } + *size = ssize; + // make valgrind happy by setting unused fl_pg_pool + layout->fl_pg_pool = 0; + return 0; +} + +int libradosstriper::RadosStriperImpl::openStripedObjectForRead( + const std::string& soid, + ceph_file_layout *layout, + uint64_t *size, + std::string *lockCookie) +{ + // take a lock the first rados object, if it exists and gets its size + // check, lock and size reading must be atomic and are thus done within a single operation + librados::ObjectWriteOperation op; + op.assert_exists(); + *lockCookie = getUUID(); + utime_t dur = utime_t(); + rados::cls::lock::lock(&op, RADOS_LOCK_NAME, LOCK_SHARED, *lockCookie, "Tag", "", dur, 0); + std::string firstObjOid = getObjectId(soid, 0); + int rc = m_ioCtx.operate(firstObjOid, &op); + if (rc) { + // error case (including -ENOENT) + return rc; + } + rc = internal_get_layout_and_size(firstObjOid, layout, size); + if (rc) { + unlockObject(soid, *lockCookie); + lderr(cct()) << "RadosStriperImpl::openStripedObjectForRead : " + << "could not load layout and size for " + << soid << " : rc = " << rc << dendl; + } + return rc; +} + +int libradosstriper::RadosStriperImpl::openStripedObjectForWrite(const std::string& soid, + ceph_file_layout *layout, + uint64_t *size, + std::string *lockCookie, + bool isFileSizeAbsolute) +{ + // take a lock the first rados object, if it exists + // check and lock must be atomic and are thus done within a single operation + librados::ObjectWriteOperation op; + op.assert_exists(); + *lockCookie = getUUID(); + utime_t dur = utime_t(); + rados::cls::lock::lock(&op, RADOS_LOCK_NAME, LOCK_SHARED, *lockCookie, "Tag", "", dur, 0); + std::string firstObjOid = getObjectId(soid, 0); + int rc = m_ioCtx.operate(firstObjOid, &op); + if (rc) { + if (rc == -ENOENT) { + // object does not exist, delegate to createEmptyStripedObject + int rc = createAndOpenStripedObject(soid, layout, *size, lockCookie, isFileSizeAbsolute); + // return original size + *size = 0; + return rc; + } else { + return rc; + } + } + // all fine + uint64_t curSize; + rc = internal_get_layout_and_size(firstObjOid, layout, &curSize); + if (rc) { + unlockObject(soid, *lockCookie); + lderr(cct()) << "RadosStriperImpl::openStripedObjectForWrite : " + << "could not load layout and size for " + << soid << " : rc = " << rc << dendl; + return rc; + } + // atomically update object size, only if smaller than current one + if (!isFileSizeAbsolute) + *size += curSize; + librados::ObjectWriteOperation writeOp; + writeOp.cmpxattr(XATTR_SIZE, LIBRADOS_CMPXATTR_OP_GT, *size); + std::ostringstream oss; + oss << *size; + bufferlist bl; + bl.append(oss.str()); + writeOp.setxattr(XATTR_SIZE, bl); + rc = m_ioCtx.operate(firstObjOid, &writeOp); + // return current size + *size = curSize; + // handle case where objectsize is already bigger than size + if (-ECANCELED == rc) + rc = 0; + if (rc) { + unlockObject(soid, *lockCookie); + lderr(cct()) << "RadosStriperImpl::openStripedObjectForWrite : " + << "could not set new size for " + << soid << " : rc = " << rc << dendl; + } + return rc; +} + +int libradosstriper::RadosStriperImpl::createAndOpenStripedObject(const std::string& soid, + ceph_file_layout *layout, + uint64_t size, + std::string *lockCookie, + bool isFileSizeAbsolute) +{ + // build atomic write operation + librados::ObjectWriteOperation writeOp; + writeOp.create(true); + // object_size + std::ostringstream oss_object_size; + oss_object_size << m_layout.fl_object_size; + bufferlist bl_object_size; + bl_object_size.append(oss_object_size.str()); + writeOp.setxattr(XATTR_LAYOUT_OBJECT_SIZE, bl_object_size); + // stripe unit + std::ostringstream oss_stripe_unit; + oss_stripe_unit << m_layout.fl_stripe_unit; + bufferlist bl_stripe_unit; + bl_stripe_unit.append(oss_stripe_unit.str()); + writeOp.setxattr(XATTR_LAYOUT_STRIPE_UNIT, bl_stripe_unit); + // stripe count + std::ostringstream oss_stripe_count; + oss_stripe_count << m_layout.fl_stripe_count; + bufferlist bl_stripe_count; + bl_stripe_count.append(oss_stripe_count.str()); + writeOp.setxattr(XATTR_LAYOUT_STRIPE_COUNT, bl_stripe_count); + // size + std::ostringstream oss_size; + oss_size << (isFileSizeAbsolute?size:0); + bufferlist bl_size; + bl_size.append(oss_size.str()); + writeOp.setxattr(XATTR_SIZE, bl_size); + // effectively change attributes + std::string firstObjOid = getObjectId(soid, 0); + int rc = m_ioCtx.operate(firstObjOid, &writeOp); + // in case of error (but no EEXIST which would mean the object existed), return + if (rc && -EEXIST != rc) return rc; + // Otherwise open the object + uint64_t fileSize = size; + return openStripedObjectForWrite(soid, layout, &fileSize, lockCookie, isFileSizeAbsolute); +} + +static void striper_truncate_aio_req_complete(rados_striper_multi_completion_t c, void *arg) +{ + auto cdata = reinterpret_cast<TruncateCompletionData*>(arg); + libradosstriper::MultiAioCompletionImpl *comp = + reinterpret_cast<libradosstriper::MultiAioCompletionImpl*>(c); + if (0 == comp->rval) { + // all went fine, change size in the external attributes + std::ostringstream oss; + oss << cdata->m_size; + bufferlist bl; + bl.append(oss.str()); + cdata->m_striper->setxattr(cdata->m_soid, XATTR_SIZE, bl); + } + cdata->put(); +} + +int libradosstriper::RadosStriperImpl::truncate(const std::string& soid, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout) +{ + TruncateCompletionData *cdata = new TruncateCompletionData(this, soid, size); + libradosstriper::MultiAioCompletionImplPtr multi_completion{ + new libradosstriper::MultiAioCompletionImpl, false}; + multi_completion->set_complete_callback(cdata, striper_truncate_aio_req_complete); + // call asynchrous version of truncate + int rc = aio_truncate(soid, multi_completion, original_size, size, layout); + // wait for completion of the truncation + multi_completion->finish_adding_requests(); + multi_completion->wait_for_complete_and_cb(); + // return result + if (rc == 0) { + rc = multi_completion->get_return_value(); + } + return rc; +} + +int libradosstriper::RadosStriperImpl::aio_truncate +(const std::string& soid, + libradosstriper::MultiAioCompletionImplPtr multi_completion, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout) +{ + // handle the underlying rados objects. 3 cases here : + // -- the objects belonging to object sets entirely located + // before the truncation are unchanged + // -- the objects belonging to the object set where the + // truncation took place are truncated or removed + // -- the objects belonging to object sets entirely located + // after the truncation are removed + // Note that we do it backward and that we change the size in + // the external attributes only at the end. This make sure that + // no rados object stays behind if we remove the striped object + // after a truncation has failed + uint64_t trunc_objectsetno = size / layout.fl_object_size / layout.fl_stripe_count; + uint64_t last_objectsetno = original_size / layout.fl_object_size / layout.fl_stripe_count; + bool exists = false; + for (int64_t objectno = (last_objectsetno+1) * layout.fl_stripe_count-1; + objectno >= (int64_t)((trunc_objectsetno + 1) * layout.fl_stripe_count); + objectno--) { + // if no object existed so far, check object existence + if (!exists) { + uint64_t nb_full_object_set = objectno / layout.fl_stripe_count; + uint64_t object_index_in_set = objectno % layout.fl_stripe_count; + uint64_t set_start_off = nb_full_object_set * layout.fl_object_size * layout.fl_stripe_count; + uint64_t object_start_off = set_start_off + object_index_in_set * layout.fl_stripe_unit; + exists = (original_size > object_start_off); + } + if (exists) { + // remove asynchronously + multi_completion->add_request(); + RadosRemoveCompletionData *data = + new RadosRemoveCompletionData(multi_completion, cct()); + librados::AioCompletion *rados_completion = + librados::Rados::aio_create_completion(data, + rados_req_remove_complete, + rados_req_remove_safe); + int rc = m_ioCtx.aio_remove(getObjectId(soid, objectno), rados_completion); + rados_completion->release(); + // in case the object did not exist, it means we had a sparse file, all is fine + if (rc && rc != -ENOENT) return rc; + } + } + for (int64_t objectno = ((trunc_objectsetno + 1) * layout.fl_stripe_count) -1; + objectno >= (int64_t)(trunc_objectsetno * layout.fl_stripe_count); + objectno--) { + // if no object existed so far, check object existence + if (!exists) { + uint64_t object_start_off = ((objectno / layout.fl_stripe_count) * layout.fl_object_size) + + ((objectno % layout.fl_stripe_count) * layout.fl_stripe_unit); + exists = (original_size > object_start_off); + } + if (exists) { + // truncate + file_layout_t l; + l.from_legacy(layout); + uint64_t new_object_size = Striper::object_truncate_size(cct(), &l, objectno, size); + int rc; + if (new_object_size > 0 or 0 == objectno) { + // trunc is synchronous as there is no async version + // but note that only a single object will be truncated + // reducing the overload to a fixed amount + rc = m_ioCtx.trunc(getObjectId(soid, objectno), new_object_size); + } else { + // removes are asynchronous in order to speed up truncations of big files + multi_completion->add_request(); + RadosRemoveCompletionData *data = + new RadosRemoveCompletionData(multi_completion, cct()); + librados::AioCompletion *rados_completion = + librados::Rados::aio_create_completion(data, + rados_req_remove_complete, + rados_req_remove_safe); + rc = m_ioCtx.aio_remove(getObjectId(soid, objectno), rados_completion); + rados_completion->release(); + } + // in case the object did not exist, it means we had a sparse file, all is fine + if (rc && rc != -ENOENT) return rc; + } + } + return 0; +} + +int libradosstriper::RadosStriperImpl::grow(const std::string& soid, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout) +{ + // handle the underlying rados objects. As we support sparse objects, + // we only have to change the size in the external attributes + std::ostringstream oss; + oss << size; + bufferlist bl; + bl.append(oss.str()); + int rc = m_ioCtx.setxattr(getObjectId(soid, 0), XATTR_SIZE, bl); + return rc; +} + +std::string libradosstriper::RadosStriperImpl::getUUID() +{ + struct uuid_d uuid; + uuid.generate_random(); + char suuid[37]; + uuid.print(suuid); + return std::string(suuid); +} diff --git a/src/libradosstriper/RadosStriperImpl.h b/src/libradosstriper/RadosStriperImpl.h new file mode 100644 index 00000000..6622471e --- /dev/null +++ b/src/libradosstriper/RadosStriperImpl.h @@ -0,0 +1,274 @@ +// -*- 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) 2014 Sebastien Ponce <sebastien.ponce@cern.ch> + * + * 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_LIBRADOSSTRIPER_RADOSSTRIPERIMPL_H +#define CEPH_LIBRADOSSTRIPER_RADOSSTRIPERIMPL_H + +#include <string> + +#include "include/rados/librados.h" +#include "include/rados/librados.hpp" +#include "include/radosstriper/libradosstriper.h" +#include "include/radosstriper/libradosstriper.hpp" +#include "MultiAioCompletionImpl.h" + +#include "librados/IoCtxImpl.h" +#include "librados/AioCompletionImpl.h" +#include "common/RefCountedObj.h" + +namespace libradosstriper { + +using MultiAioCompletionImplPtr = + boost::intrusive_ptr<MultiAioCompletionImpl>; + +struct RadosStriperImpl { + + /** + * exception wrapper around an error code + */ + struct ErrorCode { + ErrorCode(int error) : m_code(error) {}; + int m_code; + }; + + /* + * Constructor + * @param cluster_name name of the cluster, can be NULL + * @param client_name has 2 meanings depending on cluster_name + * - if cluster_name is null : this is the client id + * - else : this is the full client name in format type.id + */ + RadosStriperImpl(librados::IoCtx& ioctx, librados::IoCtxImpl *ioctx_impl); + /// Destructor + ~RadosStriperImpl() {}; + + // configuration + int setObjectLayoutStripeUnit(unsigned int stripe_unit); + int setObjectLayoutStripeCount(unsigned int stripe_count); + int setObjectLayoutObjectSize(unsigned int object_size); + + // xattrs + int getxattr(const object_t& soid, const char *name, bufferlist& bl); + int setxattr(const object_t& soid, const char *name, bufferlist& bl); + int getxattrs(const object_t& soid, map<string, bufferlist>& attrset); + int rmxattr(const object_t& soid, const char *name); + + // io + int write(const std::string& soid, const bufferlist& bl, size_t len, uint64_t off); + int append(const std::string& soid, const bufferlist& bl, size_t len); + int write_full(const std::string& soid, const bufferlist& bl); + int read(const std::string& soid, bufferlist* pbl, size_t len, uint64_t off); + + // asynchronous io + int aio_write(const std::string& soid, librados::AioCompletionImpl *c, + const bufferlist& bl, size_t len, uint64_t off); + int aio_append(const std::string& soid, librados::AioCompletionImpl *c, + const bufferlist& bl, size_t len); + int aio_write_full(const std::string& soid, librados::AioCompletionImpl *c, + const bufferlist& bl); + int aio_read(const std::string& soid, librados::AioCompletionImpl *c, + bufferlist* pbl, size_t len, uint64_t off); + int aio_read(const std::string& soid, librados::AioCompletionImpl *c, + char* buf, size_t len, uint64_t off); + int aio_flush(); + + // stat, deletion and truncation + int stat(const std::string& soid, uint64_t *psize, time_t *pmtime); + int stat2(const std::string& soid, uint64_t *psize, struct timespec *pts); + template<class TimeType> + struct StatFunction { + typedef int (librados::IoCtxImpl::*Type) (const object_t& oid, + librados::AioCompletionImpl *c, + uint64_t *psize, TimeType *pmtime); + }; + template<class TimeType> + int aio_generic_stat(const std::string& soid, librados::AioCompletionImpl *c, + uint64_t *psize, TimeType *pmtime, + typename StatFunction<TimeType>::Type statFunction); + int aio_stat(const std::string& soid, librados::AioCompletionImpl *c, + uint64_t *psize, time_t *pmtime); + int aio_stat2(const std::string& soid, librados::AioCompletionImpl *c, + uint64_t *psize, struct timespec *pts); + int remove(const std::string& soid, int flags=0); + int trunc(const std::string& soid, uint64_t size); + + // asynchronous remove. Note that the removal is not 100% parallelized : + // the removal of the first rados object of the striped object will be + // done via a syncrhonous call after the completion of all other removals. + // These are done asynchrounously and in parallel + int aio_remove(const std::string& soid, librados::AioCompletionImpl *c, int flags=0); + + // reference counting + void get() { + lock.Lock(); + m_refCnt ++ ; + lock.Unlock(); + } + void put() { + bool deleteme = false; + lock.Lock(); + m_refCnt --; + if (m_refCnt == 0) + deleteme = true; + cond.Signal(); + lock.Unlock(); + if (deleteme) + delete this; + } + + // objectid manipulation + std::string getObjectId(const object_t& soid, long long unsigned objectno); + + // opening and closing of striped objects + void unlockObject(const std::string& soid, + const std::string& lockCookie); + void aio_unlockObject(const std::string& soid, + const std::string& lockCookie, + librados::AioCompletion *c); + + // internal versions of IO method + int write_in_open_object(const std::string& soid, + const ceph_file_layout& layout, + const std::string& lockCookie, + const bufferlist& bl, + size_t len, + uint64_t off); + int aio_write_in_open_object(const std::string& soid, + librados::AioCompletionImpl *c, + const ceph_file_layout& layout, + const std::string& lockCookie, + const bufferlist& bl, + size_t len, + uint64_t off); + int internal_aio_write(const std::string& soid, + MultiAioCompletionImplPtr c, + const bufferlist& bl, + size_t len, + uint64_t off, + const ceph_file_layout& layout); + + int extract_uint32_attr(std::map<std::string, bufferlist> &attrs, + const std::string& key, + ceph_le32 *value); + + int extract_sizet_attr(std::map<std::string, bufferlist> &attrs, + const std::string& key, + size_t *value); + + int internal_get_layout_and_size(const std::string& oid, + ceph_file_layout *layout, + uint64_t *size); + + int internal_aio_remove(const std::string& soid, + MultiAioCompletionImplPtr multi_completion, + int flags=0); + + /** + * opens an existing striped object and takes a shared lock on it + * @return 0 if everything is ok and the lock was taken. -errcode otherwise + * In particulae, if the striped object does not exists, -ENOENT is returned + * In case the return code in not 0, no lock is taken + */ + int openStripedObjectForRead(const std::string& soid, + ceph_file_layout *layout, + uint64_t *size, + std::string *lockCookie); + + /** + * opens an existing striped object, takes a shared lock on it + * and sets its size to the size it will have after the write. + * In case the striped object does not exists, it will create it by + * calling createOrOpenStripedObject. + * @param layout this is filled with the layout of the file + * @param size new size of the file (together with isFileSizeAbsolute) + * In case of success, this is filled with the size of the file before the opening + * @param isFileSizeAbsolute if false, this means that the given size should + * be added to the current file size (append mode) + * @return 0 if everything is ok and the lock was taken. -errcode otherwise + * In case the return code in not 0, no lock is taken + */ + int openStripedObjectForWrite(const std::string& soid, + ceph_file_layout *layout, + uint64_t *size, + std::string *lockCookie, + bool isFileSizeAbsolute); + /** + * creates an empty striped object with the given size and opens it calling + * openStripedObjectForWrite, which implies taking a shared lock on it + * Also deals with the cases where the object was created in the mean time + * @param isFileSizeAbsolute if false, this means that the given size should + * be added to the current file size (append mode). This of course only makes + * sense in case the striped object already exists + * @return 0 if everything is ok and the lock was taken. -errcode otherwise + * In case the return code in not 0, no lock is taken + */ + int createAndOpenStripedObject(const std::string& soid, + ceph_file_layout *layout, + uint64_t size, + std::string *lockCookie, + bool isFileSizeAbsolute); + + /** + * truncates an object synchronously. Should only be called with size < original_size + */ + int truncate(const std::string& soid, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout); + + /** + * truncates an object asynchronously. Should only be called with size < original_size + * note that the method is not 100% asynchronous, only the removal of rados objects + * is, the (potential) truncation of the rados object residing just at the truncation + * point is synchronous for lack of asynchronous truncation in the rados layer + */ + int aio_truncate(const std::string& soid, + MultiAioCompletionImplPtr c, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout); + + /** + * grows an object (adding 0s). Should only be called with size > original_size + */ + int grow(const std::string& soid, + uint64_t original_size, + uint64_t size, + ceph_file_layout &layout); + + /** + * creates a unique identifier + */ + static std::string getUUID(); + + CephContext *cct() { + return (CephContext*)m_radosCluster.cct(); + } + + // reference counting + Cond cond; + int m_refCnt; + Mutex lock; + + + // Context + librados::Rados m_radosCluster; + librados::IoCtx m_ioCtx; + librados::IoCtxImpl *m_ioCtxImpl; + + // Default layout + ceph_file_layout m_layout; +}; +} +#endif diff --git a/src/libradosstriper/libradosstriper.cc b/src/libradosstriper/libradosstriper.cc new file mode 100644 index 00000000..9aa764b9 --- /dev/null +++ b/src/libradosstriper/libradosstriper.cc @@ -0,0 +1,669 @@ +// -*- 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) 2014 Sebastien Ponce <sebastien.ponce@cern.ch> + * + * 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 <errno.h> + +#include "libradosstriper/RadosStriperImpl.h" +#include "libradosstriper/MultiAioCompletionImpl.h" + +#include "include/types.h" + +#include "include/radosstriper/libradosstriper.h" +#include "include/radosstriper/libradosstriper.hpp" +#include "librados/RadosXattrIter.h" + +/* + * This file implements the rados striper API. + * There are 2 flavours of it : + * - the C API, found in include/rados/libradosstriper.h + * - the C++ API, found in include/rados/libradosstriper.hpp + */ + +///////////////////////////// C++ API ////////////////////////////// + +libradosstriper::MultiAioCompletion::~MultiAioCompletion() +{ + ceph_assert(pc->ref == 1); + pc->put(); +} + +int libradosstriper::MultiAioCompletion::set_complete_callback +(void *cb_arg, rados_callback_t cb) +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->set_complete_callback(cb_arg, cb); +} + +int libradosstriper::MultiAioCompletion::set_safe_callback +(void *cb_arg, rados_callback_t cb) +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->set_safe_callback(cb_arg, cb); +} + +void libradosstriper::MultiAioCompletion::wait_for_complete() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + c->wait_for_complete(); +} + +void libradosstriper::MultiAioCompletion::wait_for_safe() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + c->wait_for_safe(); +} + +bool libradosstriper::MultiAioCompletion::is_complete() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->is_complete(); +} + +bool libradosstriper::MultiAioCompletion::is_safe() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->is_safe(); +} + +void libradosstriper::MultiAioCompletion::wait_for_complete_and_cb() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + c->wait_for_complete_and_cb(); +} + +void libradosstriper::MultiAioCompletion::MultiAioCompletion::wait_for_safe_and_cb() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + c->wait_for_safe_and_cb(); +} + +bool libradosstriper::MultiAioCompletion::is_complete_and_cb() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->is_complete_and_cb(); +} + +bool libradosstriper::MultiAioCompletion::is_safe_and_cb() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->is_safe_and_cb(); +} + +int libradosstriper::MultiAioCompletion::get_return_value() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + return c->get_return_value(); +} + +void libradosstriper::MultiAioCompletion::release() +{ + MultiAioCompletionImpl *c = (MultiAioCompletionImpl *)pc; + c->put(); + delete this; +} + +libradosstriper::RadosStriper::RadosStriper() : + rados_striper_impl(0) +{ +} + +void libradosstriper::RadosStriper::to_rados_striper_t(RadosStriper &striper, rados_striper_t *s) +{ + *s = (rados_striper_t)striper.rados_striper_impl; + striper.rados_striper_impl->get(); +} + +libradosstriper::RadosStriper::RadosStriper(const RadosStriper& rs) +{ + rados_striper_impl = rs.rados_striper_impl; + if (rados_striper_impl) { + rados_striper_impl->get(); + } +} + +libradosstriper::RadosStriper& libradosstriper::RadosStriper::operator=(const RadosStriper& rs) +{ + if (rados_striper_impl) + rados_striper_impl->put(); + rados_striper_impl = rs.rados_striper_impl; + rados_striper_impl->get(); + return *this; +} + +libradosstriper::RadosStriper::~RadosStriper() +{ + if (rados_striper_impl) + rados_striper_impl->put(); + rados_striper_impl = 0; +} + +int libradosstriper::RadosStriper::striper_create(librados::IoCtx& ioctx, + RadosStriper *striper) +{ + try { + striper->rados_striper_impl = new libradosstriper::RadosStriperImpl(ioctx, ioctx.io_ctx_impl); + striper->rados_striper_impl->get(); + } catch (int rc) { + return rc; + } + return 0; +} + +int libradosstriper::RadosStriper::set_object_layout_stripe_unit +(unsigned int stripe_unit) +{ + return rados_striper_impl->setObjectLayoutStripeUnit(stripe_unit); +} + +int libradosstriper::RadosStriper::set_object_layout_stripe_count +(unsigned int stripe_count) +{ + return rados_striper_impl->setObjectLayoutStripeCount(stripe_count); +} + +int libradosstriper::RadosStriper::set_object_layout_object_size +(unsigned int object_size) +{ + return rados_striper_impl->setObjectLayoutObjectSize(object_size); +} + +int libradosstriper::RadosStriper::getxattr(const std::string& oid, const char *name, bufferlist& bl) +{ + return rados_striper_impl->getxattr(oid, name, bl); +} + +int libradosstriper::RadosStriper::setxattr(const std::string& oid, const char *name, bufferlist& bl) +{ + return rados_striper_impl->setxattr(oid, name, bl); +} + +int libradosstriper::RadosStriper::rmxattr(const std::string& oid, const char *name) +{ + return rados_striper_impl->rmxattr(oid, name); +} + +int libradosstriper::RadosStriper::getxattrs(const std::string& oid, + std::map<std::string, bufferlist>& attrset) +{ + return rados_striper_impl->getxattrs(oid, attrset); +} + +int libradosstriper::RadosStriper::write(const std::string& soid, + const bufferlist& bl, + size_t len, + uint64_t off) +{ + return rados_striper_impl->write(soid, bl, len, off); +} + +int libradosstriper::RadosStriper::write_full(const std::string& soid, + const bufferlist& bl) +{ + return rados_striper_impl->write_full(soid, bl); +} + +int libradosstriper::RadosStriper::append(const std::string& soid, + const bufferlist& bl, + size_t len) +{ + return rados_striper_impl->append(soid, bl, len); +} + +int libradosstriper::RadosStriper::aio_write(const std::string& soid, + librados::AioCompletion *c, + const bufferlist& bl, + size_t len, + uint64_t off) +{ + return rados_striper_impl->aio_write(soid, c->pc, bl, len, off); +} + +int libradosstriper::RadosStriper::aio_write_full(const std::string& soid, + librados::AioCompletion *c, + const bufferlist& bl) +{ + return rados_striper_impl->aio_write_full(soid, c->pc, bl); +} + +int libradosstriper::RadosStriper::aio_append(const std::string& soid, + librados::AioCompletion *c, + const bufferlist& bl, + size_t len) +{ + return rados_striper_impl->aio_append(soid, c->pc, bl, len); +} + +int libradosstriper::RadosStriper::read(const std::string& soid, + bufferlist* bl, + size_t len, + uint64_t off) +{ + bl->clear(); + bl->push_back(buffer::create(len)); + return rados_striper_impl->read(soid, bl, len, off); +} + +int libradosstriper::RadosStriper::aio_read(const std::string& soid, + librados::AioCompletion *c, + bufferlist* bl, + size_t len, + uint64_t off) +{ + bl->clear(); + bl->push_back(buffer::create(len)); + return rados_striper_impl->aio_read(soid, c->pc, bl, len, off); +} + +int libradosstriper::RadosStriper::stat(const std::string& soid, uint64_t *psize, time_t *pmtime) +{ + return rados_striper_impl->stat(soid, psize, pmtime); +} + +int libradosstriper::RadosStriper::aio_stat(const std::string& soid, + librados::AioCompletion *c, + uint64_t *psize, + time_t *pmtime) +{ + return rados_striper_impl->aio_stat(soid, c->pc, psize, pmtime); +} + +int libradosstriper::RadosStriper::stat2(const std::string& soid, uint64_t *psize, struct timespec *pts) +{ + return rados_striper_impl->stat2(soid, psize, pts); +} + +int libradosstriper::RadosStriper::aio_stat2(const std::string& soid, + librados::AioCompletion *c, + uint64_t *psize, + struct timespec *pts) +{ + return rados_striper_impl->aio_stat2(soid, c->pc, psize, pts); +} + +int libradosstriper::RadosStriper::remove(const std::string& soid) +{ + return rados_striper_impl->remove(soid); +} + +int libradosstriper::RadosStriper::aio_remove(const std::string& soid, + librados::AioCompletion *c) +{ + return rados_striper_impl->aio_remove(soid, c->pc); +} + +int libradosstriper::RadosStriper::remove(const std::string& soid, int flags) +{ + return rados_striper_impl->remove(soid, flags); +} + +int libradosstriper::RadosStriper::aio_remove(const std::string& soid, + librados::AioCompletion *c, + int flags) +{ + return rados_striper_impl->aio_remove(soid, c->pc, flags); +} + +int libradosstriper::RadosStriper::trunc(const std::string& soid, uint64_t size) +{ + return rados_striper_impl->trunc(soid, size); +} + +int libradosstriper::RadosStriper::aio_flush() +{ + return rados_striper_impl->aio_flush(); +} + +libradosstriper::MultiAioCompletion* libradosstriper::RadosStriper::multi_aio_create_completion() +{ + MultiAioCompletionImpl *c = new MultiAioCompletionImpl; + return new MultiAioCompletion(c); +} + +libradosstriper::MultiAioCompletion* +libradosstriper::RadosStriper::multi_aio_create_completion(void *cb_arg, + librados::callback_t cb_complete, + librados::callback_t cb_safe) +{ + MultiAioCompletionImpl *c; + int r = rados_striper_multi_aio_create_completion(cb_arg, cb_complete, cb_safe, (void**)&c); + ceph_assert(r == 0); + return new MultiAioCompletion(c); +} + +///////////////////////////// C API ////////////////////////////// + +extern "C" int rados_striper_create(rados_ioctx_t ioctx, + rados_striper_t *striper) +{ + librados::IoCtx ctx; + librados::IoCtx::from_rados_ioctx_t(ioctx, ctx); + libradosstriper::RadosStriper striperp; + int rc = libradosstriper::RadosStriper::striper_create(ctx, &striperp); + if (0 == rc) + libradosstriper::RadosStriper::to_rados_striper_t(striperp, striper); + return rc; +} + +extern "C" void rados_striper_destroy(rados_striper_t striper) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + impl->put(); +} + +extern "C" int rados_striper_set_object_layout_stripe_unit(rados_striper_t striper, + unsigned int stripe_unit) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->setObjectLayoutStripeUnit(stripe_unit); +} + +extern "C" int rados_striper_set_object_layout_stripe_count(rados_striper_t striper, + unsigned int stripe_count) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->setObjectLayoutStripeCount(stripe_count); +} + +extern "C" int rados_striper_set_object_layout_object_size(rados_striper_t striper, + unsigned int object_size) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->setObjectLayoutObjectSize(object_size); +} + +extern "C" int rados_striper_write(rados_striper_t striper, + const char *soid, + const char *buf, + size_t len, + uint64_t off) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->write(soid, bl, len, off); +} + +extern "C" int rados_striper_write_full(rados_striper_t striper, + const char *soid, + const char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->write_full(soid, bl); +} + + +extern "C" int rados_striper_append(rados_striper_t striper, + const char *soid, + const char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->append(soid, bl, len); +} + +extern "C" int rados_striper_read(rados_striper_t striper, + const char *soid, + char *buf, + size_t len, + uint64_t off) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bufferptr bp = buffer::create_static(len, buf); + bl.push_back(bp); + int ret = impl->read(soid, &bl, len, off); + if (ret >= 0) { + if (bl.length() > len) + return -ERANGE; + if (!bl.is_provided_buffer(buf)) + bl.copy(0, bl.length(), buf); + ret = bl.length(); // hrm :/ + } + return ret; +} + +extern "C" int rados_striper_remove(rados_striper_t striper, const char* soid) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->remove(soid); +} + +extern "C" int rados_striper_trunc(rados_striper_t striper, const char* soid, uint64_t size) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->trunc(soid, size); +} + +extern "C" int rados_striper_getxattr(rados_striper_t striper, + const char *oid, + const char *name, + char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + object_t obj(oid); + bufferlist bl; + int ret = impl->getxattr(oid, name, bl); + if (ret >= 0) { + if (bl.length() > len) + return -ERANGE; + bl.copy(0, bl.length(), buf); + ret = bl.length(); + } + return ret; +} + +extern "C" int rados_striper_setxattr(rados_striper_t striper, + const char *oid, + const char *name, + const char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + object_t obj(oid); + bufferlist bl; + bl.append(buf, len); + return impl->setxattr(obj, name, bl); +} + +extern "C" int rados_striper_rmxattr(rados_striper_t striper, + const char *oid, + const char *name) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + object_t obj(oid); + return impl->rmxattr(obj, name); +} + +extern "C" int rados_striper_getxattrs(rados_striper_t striper, + const char *oid, + rados_xattrs_iter_t *iter) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + object_t obj(oid); + librados::RadosXattrsIter *it = new librados::RadosXattrsIter(); + if (!it) + return -ENOMEM; + int ret = impl->getxattrs(obj, it->attrset); + if (ret) { + delete it; + return ret; + } + it->i = it->attrset.begin(); + *iter = it; + return 0; +} + +extern "C" int rados_striper_getxattrs_next(rados_xattrs_iter_t iter, + const char **name, + const char **val, + size_t *len) +{ + return rados_getxattrs_next(iter, name, val, len); +} + +extern "C" void rados_striper_getxattrs_end(rados_xattrs_iter_t iter) +{ + return rados_getxattrs_end(iter); +} + +extern "C" int rados_striper_stat(rados_striper_t striper, + const char* soid, + uint64_t *psize, + time_t *pmtime) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->stat(soid, psize, pmtime); +} + +extern "C" int rados_striper_multi_aio_create_completion(void *cb_arg, + rados_callback_t cb_complete, + rados_callback_t cb_safe, + rados_striper_multi_completion_t *pc) +{ + libradosstriper::MultiAioCompletionImpl *c = new libradosstriper::MultiAioCompletionImpl; + if (cb_complete) + c->set_complete_callback(cb_arg, cb_complete); + if (cb_safe) + c->set_safe_callback(cb_arg, cb_safe); + *pc = c; + return 0; +} + +extern "C" void rados_striper_multi_aio_wait_for_complete(rados_striper_multi_completion_t c) +{ + ((libradosstriper::MultiAioCompletionImpl*)c)->wait_for_complete(); +} + +extern "C" void rados_striper_multi_aio_wait_for_safe(rados_striper_multi_completion_t c) +{ + ((libradosstriper::MultiAioCompletionImpl*)c)->wait_for_safe(); +} + +extern "C" int rados_striper_multi_aio_is_complete(rados_striper_multi_completion_t c) +{ + return ((libradosstriper::MultiAioCompletionImpl*)c)->is_complete(); +} + +extern "C" int rados_striper_multi_aio_is_safe(rados_striper_multi_completion_t c) +{ + return ((libradosstriper::MultiAioCompletionImpl*)c)->is_safe(); +} + +extern "C" void rados_striper_multi_aio_wait_for_complete_and_cb(rados_striper_multi_completion_t c) +{ + ((libradosstriper::MultiAioCompletionImpl*)c)->wait_for_complete_and_cb(); +} + +extern "C" void rados_striper_multi_aio_wait_for_safe_and_cb(rados_striper_multi_completion_t c) +{ + ((libradosstriper::MultiAioCompletionImpl*)c)->wait_for_safe_and_cb(); +} + +extern "C" int rados_striper_multi_aio_is_complete_and_cb(rados_striper_multi_completion_t c) +{ + return ((libradosstriper::MultiAioCompletionImpl*)c)->is_complete_and_cb(); +} + +extern "C" int rados_striper_multi_aio_is_safe_and_cb(rados_striper_multi_completion_t c) +{ + return ((libradosstriper::MultiAioCompletionImpl*)c)->is_safe_and_cb(); +} + +extern "C" int rados_striper_multi_aio_get_return_value(rados_striper_multi_completion_t c) +{ + return ((libradosstriper::MultiAioCompletionImpl*)c)->get_return_value(); +} + +extern "C" void rados_striper_multi_aio_release(rados_striper_multi_completion_t c) +{ + ((libradosstriper::MultiAioCompletionImpl*)c)->put(); +} + +extern "C" int rados_striper_aio_write(rados_striper_t striper, + const char* soid, + rados_completion_t completion, + const char *buf, + size_t len, + uint64_t off) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->aio_write(soid, (librados::AioCompletionImpl*)completion, bl, len, off); +} + +extern "C" int rados_striper_aio_append(rados_striper_t striper, + const char* soid, + rados_completion_t completion, + const char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->aio_append(soid, (librados::AioCompletionImpl*)completion, bl, len); +} + +extern "C" int rados_striper_aio_write_full(rados_striper_t striper, + const char* soid, + rados_completion_t completion, + const char *buf, + size_t len) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + bufferlist bl; + bl.append(buf, len); + return impl->aio_write_full(soid, (librados::AioCompletionImpl*)completion, bl); +} + +extern "C" int rados_striper_aio_read(rados_striper_t striper, + const char *soid, + rados_completion_t completion, + char *buf, + size_t len, + uint64_t off) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->aio_read(soid, (librados::AioCompletionImpl*)completion, buf, len, off); +} + +extern "C" int rados_striper_aio_remove(rados_striper_t striper, + const char* soid, + rados_completion_t completion) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->aio_remove(soid, (librados::AioCompletionImpl*)completion); +} + +extern "C" void rados_striper_aio_flush(rados_striper_t striper) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + impl->aio_flush(); +} + +extern "C" int rados_striper_aio_stat(rados_striper_t striper, + const char* soid, + rados_completion_t completion, + uint64_t *psize, + time_t *pmtime) +{ + libradosstriper::RadosStriperImpl *impl = (libradosstriper::RadosStriperImpl *)striper; + return impl->aio_stat(soid, (librados::AioCompletionImpl*)completion, psize, pmtime); +} |