diff options
Diffstat (limited to 'fs/ocfs2/dlmfs')
-rw-r--r-- | fs/ocfs2/dlmfs/Makefile | 4 | ||||
-rw-r--r-- | fs/ocfs2/dlmfs/dlmfs.c | 630 | ||||
-rw-r--r-- | fs/ocfs2/dlmfs/userdlm.c | 682 | ||||
-rw-r--r-- | fs/ocfs2/dlmfs/userdlm.h | 95 |
4 files changed, 1411 insertions, 0 deletions
diff --git a/fs/ocfs2/dlmfs/Makefile b/fs/ocfs2/dlmfs/Makefile new file mode 100644 index 000000000..c7895f65b --- /dev/null +++ b/fs/ocfs2/dlmfs/Makefile @@ -0,0 +1,4 @@ +# SPDX-License-Identifier: GPL-2.0-only +obj-$(CONFIG_OCFS2_FS) += ocfs2_dlmfs.o + +ocfs2_dlmfs-objs := userdlm.o dlmfs.o diff --git a/fs/ocfs2/dlmfs/dlmfs.c b/fs/ocfs2/dlmfs/dlmfs.c new file mode 100644 index 000000000..8b2020f92 --- /dev/null +++ b/fs/ocfs2/dlmfs/dlmfs.c @@ -0,0 +1,630 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * dlmfs.c + * + * Code which implements the kernel side of a minimal userspace + * interface to our DLM. This file handles the virtual file system + * used for communication with userspace. Credit should go to ramfs, + * which was a template for the fs side of this module. + * + * Copyright (C) 2003, 2004 Oracle. All rights reserved. + */ + +/* Simple VFS hooks based on: */ +/* + * Resizable simple ram filesystem for Linux. + * + * Copyright (C) 2000 Linus Torvalds. + * 2000 Transmeta Corp. + */ + +#include <linux/module.h> +#include <linux/fs.h> +#include <linux/pagemap.h> +#include <linux/types.h> +#include <linux/slab.h> +#include <linux/highmem.h> +#include <linux/init.h> +#include <linux/string.h> +#include <linux/backing-dev.h> +#include <linux/poll.h> + +#include <linux/uaccess.h> + +#include "../stackglue.h" +#include "userdlm.h" + +#define MLOG_MASK_PREFIX ML_DLMFS +#include "../cluster/masklog.h" + + +static const struct super_operations dlmfs_ops; +static const struct file_operations dlmfs_file_operations; +static const struct inode_operations dlmfs_dir_inode_operations; +static const struct inode_operations dlmfs_root_inode_operations; +static const struct inode_operations dlmfs_file_inode_operations; +static struct kmem_cache *dlmfs_inode_cache; + +struct workqueue_struct *user_dlm_worker; + + + +/* + * These are the ABI capabilities of dlmfs. + * + * Over time, dlmfs has added some features that were not part of the + * initial ABI. Unfortunately, some of these features are not detectable + * via standard usage. For example, Linux's default poll always returns + * EPOLLIN, so there is no way for a caller of poll(2) to know when dlmfs + * added poll support. Instead, we provide this list of new capabilities. + * + * Capabilities is a read-only attribute. We do it as a module parameter + * so we can discover it whether dlmfs is built in, loaded, or even not + * loaded. + * + * The ABI features are local to this machine's dlmfs mount. This is + * distinct from the locking protocol, which is concerned with inter-node + * interaction. + * + * Capabilities: + * - bast : EPOLLIN against the file descriptor of a held lock + * signifies a bast fired on the lock. + */ +#define DLMFS_CAPABILITIES "bast stackglue" +static int param_set_dlmfs_capabilities(const char *val, + const struct kernel_param *kp) +{ + printk(KERN_ERR "%s: readonly parameter\n", kp->name); + return -EINVAL; +} +static int param_get_dlmfs_capabilities(char *buffer, + const struct kernel_param *kp) +{ + return strlcpy(buffer, DLMFS_CAPABILITIES, + strlen(DLMFS_CAPABILITIES) + 1); +} +module_param_call(capabilities, param_set_dlmfs_capabilities, + param_get_dlmfs_capabilities, NULL, 0444); +MODULE_PARM_DESC(capabilities, DLMFS_CAPABILITIES); + + +/* + * decodes a set of open flags into a valid lock level and a set of flags. + * returns < 0 if we have invalid flags + * flags which mean something to us: + * O_RDONLY -> PRMODE level + * O_WRONLY -> EXMODE level + * + * O_NONBLOCK -> NOQUEUE + */ +static int dlmfs_decode_open_flags(int open_flags, + int *level, + int *flags) +{ + if (open_flags & (O_WRONLY|O_RDWR)) + *level = DLM_LOCK_EX; + else + *level = DLM_LOCK_PR; + + *flags = 0; + if (open_flags & O_NONBLOCK) + *flags |= DLM_LKF_NOQUEUE; + + return 0; +} + +static int dlmfs_file_open(struct inode *inode, + struct file *file) +{ + int status, level, flags; + struct dlmfs_filp_private *fp = NULL; + struct dlmfs_inode_private *ip; + + if (S_ISDIR(inode->i_mode)) + BUG(); + + mlog(0, "open called on inode %lu, flags 0x%x\n", inode->i_ino, + file->f_flags); + + status = dlmfs_decode_open_flags(file->f_flags, &level, &flags); + if (status < 0) + goto bail; + + /* We don't want to honor O_APPEND at read/write time as it + * doesn't make sense for LVB writes. */ + file->f_flags &= ~O_APPEND; + + fp = kmalloc(sizeof(*fp), GFP_NOFS); + if (!fp) { + status = -ENOMEM; + goto bail; + } + fp->fp_lock_level = level; + + ip = DLMFS_I(inode); + + status = user_dlm_cluster_lock(&ip->ip_lockres, level, flags); + if (status < 0) { + /* this is a strange error to return here but I want + * to be able userspace to be able to distinguish a + * valid lock request from one that simply couldn't be + * granted. */ + if (flags & DLM_LKF_NOQUEUE && status == -EAGAIN) + status = -ETXTBSY; + kfree(fp); + goto bail; + } + + file->private_data = fp; +bail: + return status; +} + +static int dlmfs_file_release(struct inode *inode, + struct file *file) +{ + int level; + struct dlmfs_inode_private *ip = DLMFS_I(inode); + struct dlmfs_filp_private *fp = file->private_data; + + if (S_ISDIR(inode->i_mode)) + BUG(); + + mlog(0, "close called on inode %lu\n", inode->i_ino); + + if (fp) { + level = fp->fp_lock_level; + if (level != DLM_LOCK_IV) + user_dlm_cluster_unlock(&ip->ip_lockres, level); + + kfree(fp); + file->private_data = NULL; + } + + return 0; +} + +/* + * We do ->setattr() just to override size changes. Our size is the size + * of the LVB and nothing else. + */ +static int dlmfs_file_setattr(struct user_namespace *mnt_userns, + struct dentry *dentry, struct iattr *attr) +{ + int error; + struct inode *inode = d_inode(dentry); + + attr->ia_valid &= ~ATTR_SIZE; + error = setattr_prepare(&init_user_ns, dentry, attr); + if (error) + return error; + + setattr_copy(&init_user_ns, inode, attr); + mark_inode_dirty(inode); + return 0; +} + +static __poll_t dlmfs_file_poll(struct file *file, poll_table *wait) +{ + __poll_t event = 0; + struct inode *inode = file_inode(file); + struct dlmfs_inode_private *ip = DLMFS_I(inode); + + poll_wait(file, &ip->ip_lockres.l_event, wait); + + spin_lock(&ip->ip_lockres.l_lock); + if (ip->ip_lockres.l_flags & USER_LOCK_BLOCKED) + event = EPOLLIN | EPOLLRDNORM; + spin_unlock(&ip->ip_lockres.l_lock); + + return event; +} + +static ssize_t dlmfs_file_read(struct file *file, + char __user *buf, + size_t count, + loff_t *ppos) +{ + char lvb[DLM_LVB_LEN]; + + if (!user_dlm_read_lvb(file_inode(file), lvb)) + return 0; + + return simple_read_from_buffer(buf, count, ppos, lvb, sizeof(lvb)); +} + +static ssize_t dlmfs_file_write(struct file *filp, + const char __user *buf, + size_t count, + loff_t *ppos) +{ + char lvb_buf[DLM_LVB_LEN]; + int bytes_left; + struct inode *inode = file_inode(filp); + + mlog(0, "inode %lu, count = %zu, *ppos = %llu\n", + inode->i_ino, count, *ppos); + + if (*ppos >= DLM_LVB_LEN) + return -ENOSPC; + + /* don't write past the lvb */ + if (count > DLM_LVB_LEN - *ppos) + count = DLM_LVB_LEN - *ppos; + + if (!count) + return 0; + + bytes_left = copy_from_user(lvb_buf, buf, count); + count -= bytes_left; + if (count) + user_dlm_write_lvb(inode, lvb_buf, count); + + *ppos = *ppos + count; + mlog(0, "wrote %zu bytes\n", count); + return count; +} + +static void dlmfs_init_once(void *foo) +{ + struct dlmfs_inode_private *ip = + (struct dlmfs_inode_private *) foo; + + ip->ip_conn = NULL; + ip->ip_parent = NULL; + + inode_init_once(&ip->ip_vfs_inode); +} + +static struct inode *dlmfs_alloc_inode(struct super_block *sb) +{ + struct dlmfs_inode_private *ip; + + ip = alloc_inode_sb(sb, dlmfs_inode_cache, GFP_NOFS); + if (!ip) + return NULL; + + return &ip->ip_vfs_inode; +} + +static void dlmfs_free_inode(struct inode *inode) +{ + kmem_cache_free(dlmfs_inode_cache, DLMFS_I(inode)); +} + +static void dlmfs_evict_inode(struct inode *inode) +{ + int status; + struct dlmfs_inode_private *ip; + struct user_lock_res *lockres; + int teardown; + + clear_inode(inode); + + mlog(0, "inode %lu\n", inode->i_ino); + + ip = DLMFS_I(inode); + lockres = &ip->ip_lockres; + + if (S_ISREG(inode->i_mode)) { + spin_lock(&lockres->l_lock); + teardown = !!(lockres->l_flags & USER_LOCK_IN_TEARDOWN); + spin_unlock(&lockres->l_lock); + if (!teardown) { + status = user_dlm_destroy_lock(lockres); + if (status < 0) + mlog_errno(status); + } + iput(ip->ip_parent); + goto clear_fields; + } + + mlog(0, "we're a directory, ip->ip_conn = 0x%p\n", ip->ip_conn); + /* we must be a directory. If required, lets unregister the + * dlm context now. */ + if (ip->ip_conn) + user_dlm_unregister(ip->ip_conn); +clear_fields: + ip->ip_parent = NULL; + ip->ip_conn = NULL; +} + +static struct inode *dlmfs_get_root_inode(struct super_block *sb) +{ + struct inode *inode = new_inode(sb); + umode_t mode = S_IFDIR | 0755; + + if (inode) { + inode->i_ino = get_next_ino(); + inode_init_owner(&init_user_ns, inode, NULL, mode); + inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode); + inc_nlink(inode); + + inode->i_fop = &simple_dir_operations; + inode->i_op = &dlmfs_root_inode_operations; + } + + return inode; +} + +static struct inode *dlmfs_get_inode(struct inode *parent, + struct dentry *dentry, + umode_t mode) +{ + struct super_block *sb = parent->i_sb; + struct inode * inode = new_inode(sb); + struct dlmfs_inode_private *ip; + + if (!inode) + return NULL; + + inode->i_ino = get_next_ino(); + inode_init_owner(&init_user_ns, inode, parent, mode); + inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode); + + ip = DLMFS_I(inode); + ip->ip_conn = DLMFS_I(parent)->ip_conn; + + switch (mode & S_IFMT) { + default: + /* for now we don't support anything other than + * directories and regular files. */ + BUG(); + break; + case S_IFREG: + inode->i_op = &dlmfs_file_inode_operations; + inode->i_fop = &dlmfs_file_operations; + + i_size_write(inode, DLM_LVB_LEN); + + user_dlm_lock_res_init(&ip->ip_lockres, dentry); + + /* released at clear_inode time, this insures that we + * get to drop the dlm reference on each lock *before* + * we call the unregister code for releasing parent + * directories. */ + ip->ip_parent = igrab(parent); + BUG_ON(!ip->ip_parent); + break; + case S_IFDIR: + inode->i_op = &dlmfs_dir_inode_operations; + inode->i_fop = &simple_dir_operations; + + /* directory inodes start off with i_nlink == + * 2 (for "." entry) */ + inc_nlink(inode); + break; + } + return inode; +} + +/* + * File creation. Allocate an inode, and we're done.. + */ +/* SMP-safe */ +static int dlmfs_mkdir(struct user_namespace * mnt_userns, + struct inode * dir, + struct dentry * dentry, + umode_t mode) +{ + int status; + struct inode *inode = NULL; + const struct qstr *domain = &dentry->d_name; + struct dlmfs_inode_private *ip; + struct ocfs2_cluster_connection *conn; + + mlog(0, "mkdir %.*s\n", domain->len, domain->name); + + /* verify that we have a proper domain */ + if (domain->len >= GROUP_NAME_MAX) { + status = -EINVAL; + mlog(ML_ERROR, "invalid domain name for directory.\n"); + goto bail; + } + + inode = dlmfs_get_inode(dir, dentry, mode | S_IFDIR); + if (!inode) { + status = -ENOMEM; + mlog_errno(status); + goto bail; + } + + ip = DLMFS_I(inode); + + conn = user_dlm_register(domain); + if (IS_ERR(conn)) { + status = PTR_ERR(conn); + mlog(ML_ERROR, "Error %d could not register domain \"%.*s\"\n", + status, domain->len, domain->name); + goto bail; + } + ip->ip_conn = conn; + + inc_nlink(dir); + d_instantiate(dentry, inode); + dget(dentry); /* Extra count - pin the dentry in core */ + + status = 0; +bail: + if (status < 0) + iput(inode); + return status; +} + +static int dlmfs_create(struct user_namespace *mnt_userns, + struct inode *dir, + struct dentry *dentry, + umode_t mode, + bool excl) +{ + int status = 0; + struct inode *inode; + const struct qstr *name = &dentry->d_name; + + mlog(0, "create %.*s\n", name->len, name->name); + + /* verify name is valid and doesn't contain any dlm reserved + * characters */ + if (name->len >= USER_DLM_LOCK_ID_MAX_LEN || + name->name[0] == '$') { + status = -EINVAL; + mlog(ML_ERROR, "invalid lock name, %.*s\n", name->len, + name->name); + goto bail; + } + + inode = dlmfs_get_inode(dir, dentry, mode | S_IFREG); + if (!inode) { + status = -ENOMEM; + mlog_errno(status); + goto bail; + } + + d_instantiate(dentry, inode); + dget(dentry); /* Extra count - pin the dentry in core */ +bail: + return status; +} + +static int dlmfs_unlink(struct inode *dir, + struct dentry *dentry) +{ + int status; + struct inode *inode = d_inode(dentry); + + mlog(0, "unlink inode %lu\n", inode->i_ino); + + /* if there are no current holders, or none that are waiting + * to acquire a lock, this basically destroys our lockres. */ + status = user_dlm_destroy_lock(&DLMFS_I(inode)->ip_lockres); + if (status < 0) { + mlog(ML_ERROR, "unlink %pd, error %d from destroy\n", + dentry, status); + goto bail; + } + status = simple_unlink(dir, dentry); +bail: + return status; +} + +static int dlmfs_fill_super(struct super_block * sb, + void * data, + int silent) +{ + sb->s_maxbytes = MAX_LFS_FILESIZE; + sb->s_blocksize = PAGE_SIZE; + sb->s_blocksize_bits = PAGE_SHIFT; + sb->s_magic = DLMFS_MAGIC; + sb->s_op = &dlmfs_ops; + sb->s_root = d_make_root(dlmfs_get_root_inode(sb)); + if (!sb->s_root) + return -ENOMEM; + return 0; +} + +static const struct file_operations dlmfs_file_operations = { + .open = dlmfs_file_open, + .release = dlmfs_file_release, + .poll = dlmfs_file_poll, + .read = dlmfs_file_read, + .write = dlmfs_file_write, + .llseek = default_llseek, +}; + +static const struct inode_operations dlmfs_dir_inode_operations = { + .create = dlmfs_create, + .lookup = simple_lookup, + .unlink = dlmfs_unlink, +}; + +/* this way we can restrict mkdir to only the toplevel of the fs. */ +static const struct inode_operations dlmfs_root_inode_operations = { + .lookup = simple_lookup, + .mkdir = dlmfs_mkdir, + .rmdir = simple_rmdir, +}; + +static const struct super_operations dlmfs_ops = { + .statfs = simple_statfs, + .alloc_inode = dlmfs_alloc_inode, + .free_inode = dlmfs_free_inode, + .evict_inode = dlmfs_evict_inode, + .drop_inode = generic_delete_inode, +}; + +static const struct inode_operations dlmfs_file_inode_operations = { + .getattr = simple_getattr, + .setattr = dlmfs_file_setattr, +}; + +static struct dentry *dlmfs_mount(struct file_system_type *fs_type, + int flags, const char *dev_name, void *data) +{ + return mount_nodev(fs_type, flags, data, dlmfs_fill_super); +} + +static struct file_system_type dlmfs_fs_type = { + .owner = THIS_MODULE, + .name = "ocfs2_dlmfs", + .mount = dlmfs_mount, + .kill_sb = kill_litter_super, +}; +MODULE_ALIAS_FS("ocfs2_dlmfs"); + +static int __init init_dlmfs_fs(void) +{ + int status; + int cleanup_inode = 0, cleanup_worker = 0; + + dlmfs_inode_cache = kmem_cache_create("dlmfs_inode_cache", + sizeof(struct dlmfs_inode_private), + 0, (SLAB_HWCACHE_ALIGN|SLAB_RECLAIM_ACCOUNT| + SLAB_MEM_SPREAD|SLAB_ACCOUNT), + dlmfs_init_once); + if (!dlmfs_inode_cache) { + status = -ENOMEM; + goto bail; + } + cleanup_inode = 1; + + user_dlm_worker = alloc_workqueue("user_dlm", WQ_MEM_RECLAIM, 0); + if (!user_dlm_worker) { + status = -ENOMEM; + goto bail; + } + cleanup_worker = 1; + + user_dlm_set_locking_protocol(); + status = register_filesystem(&dlmfs_fs_type); +bail: + if (status) { + if (cleanup_inode) + kmem_cache_destroy(dlmfs_inode_cache); + if (cleanup_worker) + destroy_workqueue(user_dlm_worker); + } else + printk("OCFS2 User DLM kernel interface loaded\n"); + return status; +} + +static void __exit exit_dlmfs_fs(void) +{ + unregister_filesystem(&dlmfs_fs_type); + + destroy_workqueue(user_dlm_worker); + + /* + * Make sure all delayed rcu free inodes are flushed before we + * destroy cache. + */ + rcu_barrier(); + kmem_cache_destroy(dlmfs_inode_cache); + +} + +MODULE_AUTHOR("Oracle"); +MODULE_LICENSE("GPL"); +MODULE_DESCRIPTION("OCFS2 DLM-Filesystem"); + +module_init(init_dlmfs_fs) +module_exit(exit_dlmfs_fs) diff --git a/fs/ocfs2/dlmfs/userdlm.c b/fs/ocfs2/dlmfs/userdlm.c new file mode 100644 index 000000000..617c92e7b --- /dev/null +++ b/fs/ocfs2/dlmfs/userdlm.c @@ -0,0 +1,682 @@ +// SPDX-License-Identifier: GPL-2.0-or-later +/* + * userdlm.c + * + * Code which implements the kernel side of a minimal userspace + * interface to our DLM. + * + * Many of the functions here are pared down versions of dlmglue.c + * functions. + * + * Copyright (C) 2003, 2004 Oracle. All rights reserved. + */ + +#include <linux/signal.h> +#include <linux/sched/signal.h> + +#include <linux/module.h> +#include <linux/fs.h> +#include <linux/types.h> +#include <linux/crc32.h> + +#include "../ocfs2_lockingver.h" +#include "../stackglue.h" +#include "userdlm.h" + +#define MLOG_MASK_PREFIX ML_DLMFS +#include "../cluster/masklog.h" + + +static inline struct user_lock_res *user_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb) +{ + return container_of(lksb, struct user_lock_res, l_lksb); +} + +static inline int user_check_wait_flag(struct user_lock_res *lockres, + int flag) +{ + int ret; + + spin_lock(&lockres->l_lock); + ret = lockres->l_flags & flag; + spin_unlock(&lockres->l_lock); + + return ret; +} + +static inline void user_wait_on_busy_lock(struct user_lock_res *lockres) + +{ + wait_event(lockres->l_event, + !user_check_wait_flag(lockres, USER_LOCK_BUSY)); +} + +static inline void user_wait_on_blocked_lock(struct user_lock_res *lockres) + +{ + wait_event(lockres->l_event, + !user_check_wait_flag(lockres, USER_LOCK_BLOCKED)); +} + +/* I heart container_of... */ +static inline struct ocfs2_cluster_connection * +cluster_connection_from_user_lockres(struct user_lock_res *lockres) +{ + struct dlmfs_inode_private *ip; + + ip = container_of(lockres, + struct dlmfs_inode_private, + ip_lockres); + return ip->ip_conn; +} + +static struct inode * +user_dlm_inode_from_user_lockres(struct user_lock_res *lockres) +{ + struct dlmfs_inode_private *ip; + + ip = container_of(lockres, + struct dlmfs_inode_private, + ip_lockres); + return &ip->ip_vfs_inode; +} + +static inline void user_recover_from_dlm_error(struct user_lock_res *lockres) +{ + spin_lock(&lockres->l_lock); + lockres->l_flags &= ~USER_LOCK_BUSY; + spin_unlock(&lockres->l_lock); +} + +#define user_log_dlm_error(_func, _stat, _lockres) do { \ + mlog(ML_ERROR, "Dlm error %d while calling %s on " \ + "resource %.*s\n", _stat, _func, \ + _lockres->l_namelen, _lockres->l_name); \ +} while (0) + +/* WARNING: This function lives in a world where the only three lock + * levels are EX, PR, and NL. It *will* have to be adjusted when more + * lock types are added. */ +static inline int user_highest_compat_lock_level(int level) +{ + int new_level = DLM_LOCK_EX; + + if (level == DLM_LOCK_EX) + new_level = DLM_LOCK_NL; + else if (level == DLM_LOCK_PR) + new_level = DLM_LOCK_PR; + return new_level; +} + +static void user_ast(struct ocfs2_dlm_lksb *lksb) +{ + struct user_lock_res *lockres = user_lksb_to_lock_res(lksb); + int status; + + mlog(ML_BASTS, "AST fired for lockres %.*s, level %d => %d\n", + lockres->l_namelen, lockres->l_name, lockres->l_level, + lockres->l_requested); + + spin_lock(&lockres->l_lock); + + status = ocfs2_dlm_lock_status(&lockres->l_lksb); + if (status) { + mlog(ML_ERROR, "lksb status value of %u on lockres %.*s\n", + status, lockres->l_namelen, lockres->l_name); + spin_unlock(&lockres->l_lock); + return; + } + + mlog_bug_on_msg(lockres->l_requested == DLM_LOCK_IV, + "Lockres %.*s, requested ivmode. flags 0x%x\n", + lockres->l_namelen, lockres->l_name, lockres->l_flags); + + /* we're downconverting. */ + if (lockres->l_requested < lockres->l_level) { + if (lockres->l_requested <= + user_highest_compat_lock_level(lockres->l_blocking)) { + lockres->l_blocking = DLM_LOCK_NL; + lockres->l_flags &= ~USER_LOCK_BLOCKED; + } + } + + lockres->l_level = lockres->l_requested; + lockres->l_requested = DLM_LOCK_IV; + lockres->l_flags |= USER_LOCK_ATTACHED; + lockres->l_flags &= ~USER_LOCK_BUSY; + + spin_unlock(&lockres->l_lock); + + wake_up(&lockres->l_event); +} + +static inline void user_dlm_grab_inode_ref(struct user_lock_res *lockres) +{ + struct inode *inode; + inode = user_dlm_inode_from_user_lockres(lockres); + if (!igrab(inode)) + BUG(); +} + +static void user_dlm_unblock_lock(struct work_struct *work); + +static void __user_dlm_queue_lockres(struct user_lock_res *lockres) +{ + if (!(lockres->l_flags & USER_LOCK_QUEUED)) { + user_dlm_grab_inode_ref(lockres); + + INIT_WORK(&lockres->l_work, user_dlm_unblock_lock); + + queue_work(user_dlm_worker, &lockres->l_work); + lockres->l_flags |= USER_LOCK_QUEUED; + } +} + +static void __user_dlm_cond_queue_lockres(struct user_lock_res *lockres) +{ + int queue = 0; + + if (!(lockres->l_flags & USER_LOCK_BLOCKED)) + return; + + switch (lockres->l_blocking) { + case DLM_LOCK_EX: + if (!lockres->l_ex_holders && !lockres->l_ro_holders) + queue = 1; + break; + case DLM_LOCK_PR: + if (!lockres->l_ex_holders) + queue = 1; + break; + default: + BUG(); + } + + if (queue) + __user_dlm_queue_lockres(lockres); +} + +static void user_bast(struct ocfs2_dlm_lksb *lksb, int level) +{ + struct user_lock_res *lockres = user_lksb_to_lock_res(lksb); + + mlog(ML_BASTS, "BAST fired for lockres %.*s, blocking %d, level %d\n", + lockres->l_namelen, lockres->l_name, level, lockres->l_level); + + spin_lock(&lockres->l_lock); + lockres->l_flags |= USER_LOCK_BLOCKED; + if (level > lockres->l_blocking) + lockres->l_blocking = level; + + __user_dlm_queue_lockres(lockres); + spin_unlock(&lockres->l_lock); + + wake_up(&lockres->l_event); +} + +static void user_unlock_ast(struct ocfs2_dlm_lksb *lksb, int status) +{ + struct user_lock_res *lockres = user_lksb_to_lock_res(lksb); + + mlog(ML_BASTS, "UNLOCK AST fired for lockres %.*s, flags 0x%x\n", + lockres->l_namelen, lockres->l_name, lockres->l_flags); + + if (status) + mlog(ML_ERROR, "dlm returns status %d\n", status); + + spin_lock(&lockres->l_lock); + /* The teardown flag gets set early during the unlock process, + * so test the cancel flag to make sure that this ast isn't + * for a concurrent cancel. */ + if (lockres->l_flags & USER_LOCK_IN_TEARDOWN + && !(lockres->l_flags & USER_LOCK_IN_CANCEL)) { + lockres->l_level = DLM_LOCK_IV; + } else if (status == DLM_CANCELGRANT) { + /* We tried to cancel a convert request, but it was + * already granted. Don't clear the busy flag - the + * ast should've done this already. */ + BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL)); + lockres->l_flags &= ~USER_LOCK_IN_CANCEL; + goto out_noclear; + } else { + BUG_ON(!(lockres->l_flags & USER_LOCK_IN_CANCEL)); + /* Cancel succeeded, we want to re-queue */ + lockres->l_requested = DLM_LOCK_IV; /* cancel an + * upconvert + * request. */ + lockres->l_flags &= ~USER_LOCK_IN_CANCEL; + /* we want the unblock thread to look at it again + * now. */ + if (lockres->l_flags & USER_LOCK_BLOCKED) + __user_dlm_queue_lockres(lockres); + } + + lockres->l_flags &= ~USER_LOCK_BUSY; +out_noclear: + spin_unlock(&lockres->l_lock); + + wake_up(&lockres->l_event); +} + +/* + * This is the userdlmfs locking protocol version. + * + * See fs/ocfs2/dlmglue.c for more details on locking versions. + */ +static struct ocfs2_locking_protocol user_dlm_lproto = { + .lp_max_version = { + .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR, + .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR, + }, + .lp_lock_ast = user_ast, + .lp_blocking_ast = user_bast, + .lp_unlock_ast = user_unlock_ast, +}; + +static inline void user_dlm_drop_inode_ref(struct user_lock_res *lockres) +{ + struct inode *inode; + inode = user_dlm_inode_from_user_lockres(lockres); + iput(inode); +} + +static void user_dlm_unblock_lock(struct work_struct *work) +{ + int new_level, status; + struct user_lock_res *lockres = + container_of(work, struct user_lock_res, l_work); + struct ocfs2_cluster_connection *conn = + cluster_connection_from_user_lockres(lockres); + + mlog(0, "lockres %.*s\n", lockres->l_namelen, lockres->l_name); + + spin_lock(&lockres->l_lock); + + mlog_bug_on_msg(!(lockres->l_flags & USER_LOCK_QUEUED), + "Lockres %.*s, flags 0x%x\n", + lockres->l_namelen, lockres->l_name, lockres->l_flags); + + /* notice that we don't clear USER_LOCK_BLOCKED here. If it's + * set, we want user_ast clear it. */ + lockres->l_flags &= ~USER_LOCK_QUEUED; + + /* It's valid to get here and no longer be blocked - if we get + * several basts in a row, we might be queued by the first + * one, the unblock thread might run and clear the queued + * flag, and finally we might get another bast which re-queues + * us before our ast for the downconvert is called. */ + if (!(lockres->l_flags & USER_LOCK_BLOCKED)) { + mlog(ML_BASTS, "lockres %.*s USER_LOCK_BLOCKED\n", + lockres->l_namelen, lockres->l_name); + spin_unlock(&lockres->l_lock); + goto drop_ref; + } + + if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) { + mlog(ML_BASTS, "lockres %.*s USER_LOCK_IN_TEARDOWN\n", + lockres->l_namelen, lockres->l_name); + spin_unlock(&lockres->l_lock); + goto drop_ref; + } + + if (lockres->l_flags & USER_LOCK_BUSY) { + if (lockres->l_flags & USER_LOCK_IN_CANCEL) { + mlog(ML_BASTS, "lockres %.*s USER_LOCK_IN_CANCEL\n", + lockres->l_namelen, lockres->l_name); + spin_unlock(&lockres->l_lock); + goto drop_ref; + } + + lockres->l_flags |= USER_LOCK_IN_CANCEL; + spin_unlock(&lockres->l_lock); + + status = ocfs2_dlm_unlock(conn, &lockres->l_lksb, + DLM_LKF_CANCEL); + if (status) + user_log_dlm_error("ocfs2_dlm_unlock", status, lockres); + goto drop_ref; + } + + /* If there are still incompat holders, we can exit safely + * without worrying about re-queueing this lock as that will + * happen on the last call to user_cluster_unlock. */ + if ((lockres->l_blocking == DLM_LOCK_EX) + && (lockres->l_ex_holders || lockres->l_ro_holders)) { + spin_unlock(&lockres->l_lock); + mlog(ML_BASTS, "lockres %.*s, EX/PR Holders %u,%u\n", + lockres->l_namelen, lockres->l_name, + lockres->l_ex_holders, lockres->l_ro_holders); + goto drop_ref; + } + + if ((lockres->l_blocking == DLM_LOCK_PR) + && lockres->l_ex_holders) { + spin_unlock(&lockres->l_lock); + mlog(ML_BASTS, "lockres %.*s, EX Holders %u\n", + lockres->l_namelen, lockres->l_name, + lockres->l_ex_holders); + goto drop_ref; + } + + /* yay, we can downconvert now. */ + new_level = user_highest_compat_lock_level(lockres->l_blocking); + lockres->l_requested = new_level; + lockres->l_flags |= USER_LOCK_BUSY; + mlog(ML_BASTS, "lockres %.*s, downconvert %d => %d\n", + lockres->l_namelen, lockres->l_name, lockres->l_level, new_level); + spin_unlock(&lockres->l_lock); + + /* need lock downconvert request now... */ + status = ocfs2_dlm_lock(conn, new_level, &lockres->l_lksb, + DLM_LKF_CONVERT|DLM_LKF_VALBLK, + lockres->l_name, + lockres->l_namelen); + if (status) { + user_log_dlm_error("ocfs2_dlm_lock", status, lockres); + user_recover_from_dlm_error(lockres); + } + +drop_ref: + user_dlm_drop_inode_ref(lockres); +} + +static inline void user_dlm_inc_holders(struct user_lock_res *lockres, + int level) +{ + switch(level) { + case DLM_LOCK_EX: + lockres->l_ex_holders++; + break; + case DLM_LOCK_PR: + lockres->l_ro_holders++; + break; + default: + BUG(); + } +} + +/* predict what lock level we'll be dropping down to on behalf + * of another node, and return true if the currently wanted + * level will be compatible with it. */ +static inline int +user_may_continue_on_blocked_lock(struct user_lock_res *lockres, + int wanted) +{ + BUG_ON(!(lockres->l_flags & USER_LOCK_BLOCKED)); + + return wanted <= user_highest_compat_lock_level(lockres->l_blocking); +} + +int user_dlm_cluster_lock(struct user_lock_res *lockres, + int level, + int lkm_flags) +{ + int status, local_flags; + struct ocfs2_cluster_connection *conn = + cluster_connection_from_user_lockres(lockres); + + if (level != DLM_LOCK_EX && + level != DLM_LOCK_PR) { + mlog(ML_ERROR, "lockres %.*s: invalid request!\n", + lockres->l_namelen, lockres->l_name); + status = -EINVAL; + goto bail; + } + + mlog(ML_BASTS, "lockres %.*s, level %d, flags = 0x%x\n", + lockres->l_namelen, lockres->l_name, level, lkm_flags); + +again: + if (signal_pending(current)) { + status = -ERESTARTSYS; + goto bail; + } + + spin_lock(&lockres->l_lock); + if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) { + spin_unlock(&lockres->l_lock); + status = -EAGAIN; + goto bail; + } + + /* We only compare against the currently granted level + * here. If the lock is blocked waiting on a downconvert, + * we'll get caught below. */ + if ((lockres->l_flags & USER_LOCK_BUSY) && + (level > lockres->l_level)) { + /* is someone sitting in dlm_lock? If so, wait on + * them. */ + spin_unlock(&lockres->l_lock); + + user_wait_on_busy_lock(lockres); + goto again; + } + + if ((lockres->l_flags & USER_LOCK_BLOCKED) && + (!user_may_continue_on_blocked_lock(lockres, level))) { + /* is the lock is currently blocked on behalf of + * another node */ + spin_unlock(&lockres->l_lock); + + user_wait_on_blocked_lock(lockres); + goto again; + } + + if (level > lockres->l_level) { + local_flags = lkm_flags | DLM_LKF_VALBLK; + if (lockres->l_level != DLM_LOCK_IV) + local_flags |= DLM_LKF_CONVERT; + + lockres->l_requested = level; + lockres->l_flags |= USER_LOCK_BUSY; + spin_unlock(&lockres->l_lock); + + BUG_ON(level == DLM_LOCK_IV); + BUG_ON(level == DLM_LOCK_NL); + + /* call dlm_lock to upgrade lock now */ + status = ocfs2_dlm_lock(conn, level, &lockres->l_lksb, + local_flags, lockres->l_name, + lockres->l_namelen); + if (status) { + if ((lkm_flags & DLM_LKF_NOQUEUE) && + (status != -EAGAIN)) + user_log_dlm_error("ocfs2_dlm_lock", + status, lockres); + user_recover_from_dlm_error(lockres); + goto bail; + } + + user_wait_on_busy_lock(lockres); + goto again; + } + + user_dlm_inc_holders(lockres, level); + spin_unlock(&lockres->l_lock); + + status = 0; +bail: + return status; +} + +static inline void user_dlm_dec_holders(struct user_lock_res *lockres, + int level) +{ + switch(level) { + case DLM_LOCK_EX: + BUG_ON(!lockres->l_ex_holders); + lockres->l_ex_holders--; + break; + case DLM_LOCK_PR: + BUG_ON(!lockres->l_ro_holders); + lockres->l_ro_holders--; + break; + default: + BUG(); + } +} + +void user_dlm_cluster_unlock(struct user_lock_res *lockres, + int level) +{ + if (level != DLM_LOCK_EX && + level != DLM_LOCK_PR) { + mlog(ML_ERROR, "lockres %.*s: invalid request!\n", + lockres->l_namelen, lockres->l_name); + return; + } + + spin_lock(&lockres->l_lock); + user_dlm_dec_holders(lockres, level); + __user_dlm_cond_queue_lockres(lockres); + spin_unlock(&lockres->l_lock); +} + +void user_dlm_write_lvb(struct inode *inode, + const char *val, + unsigned int len) +{ + struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres; + char *lvb; + + BUG_ON(len > DLM_LVB_LEN); + + spin_lock(&lockres->l_lock); + + BUG_ON(lockres->l_level < DLM_LOCK_EX); + lvb = ocfs2_dlm_lvb(&lockres->l_lksb); + memcpy(lvb, val, len); + + spin_unlock(&lockres->l_lock); +} + +bool user_dlm_read_lvb(struct inode *inode, char *val) +{ + struct user_lock_res *lockres = &DLMFS_I(inode)->ip_lockres; + char *lvb; + bool ret = true; + + spin_lock(&lockres->l_lock); + + BUG_ON(lockres->l_level < DLM_LOCK_PR); + if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)) { + lvb = ocfs2_dlm_lvb(&lockres->l_lksb); + memcpy(val, lvb, DLM_LVB_LEN); + } else + ret = false; + + spin_unlock(&lockres->l_lock); + return ret; +} + +void user_dlm_lock_res_init(struct user_lock_res *lockres, + struct dentry *dentry) +{ + memset(lockres, 0, sizeof(*lockres)); + + spin_lock_init(&lockres->l_lock); + init_waitqueue_head(&lockres->l_event); + lockres->l_level = DLM_LOCK_IV; + lockres->l_requested = DLM_LOCK_IV; + lockres->l_blocking = DLM_LOCK_IV; + + /* should have been checked before getting here. */ + BUG_ON(dentry->d_name.len >= USER_DLM_LOCK_ID_MAX_LEN); + + memcpy(lockres->l_name, + dentry->d_name.name, + dentry->d_name.len); + lockres->l_namelen = dentry->d_name.len; +} + +int user_dlm_destroy_lock(struct user_lock_res *lockres) +{ + int status = -EBUSY; + struct ocfs2_cluster_connection *conn = + cluster_connection_from_user_lockres(lockres); + + mlog(ML_BASTS, "lockres %.*s\n", lockres->l_namelen, lockres->l_name); + + spin_lock(&lockres->l_lock); + if (lockres->l_flags & USER_LOCK_IN_TEARDOWN) { + spin_unlock(&lockres->l_lock); + goto bail; + } + + lockres->l_flags |= USER_LOCK_IN_TEARDOWN; + + while (lockres->l_flags & USER_LOCK_BUSY) { + spin_unlock(&lockres->l_lock); + + user_wait_on_busy_lock(lockres); + + spin_lock(&lockres->l_lock); + } + + if (lockres->l_ro_holders || lockres->l_ex_holders) { + lockres->l_flags &= ~USER_LOCK_IN_TEARDOWN; + spin_unlock(&lockres->l_lock); + goto bail; + } + + status = 0; + if (!(lockres->l_flags & USER_LOCK_ATTACHED)) { + /* + * lock is never requested, leave USER_LOCK_IN_TEARDOWN set + * to avoid new lock request coming in. + */ + spin_unlock(&lockres->l_lock); + goto bail; + } + + lockres->l_flags |= USER_LOCK_BUSY; + spin_unlock(&lockres->l_lock); + + status = ocfs2_dlm_unlock(conn, &lockres->l_lksb, DLM_LKF_VALBLK); + if (status) { + spin_lock(&lockres->l_lock); + lockres->l_flags &= ~USER_LOCK_IN_TEARDOWN; + lockres->l_flags &= ~USER_LOCK_BUSY; + spin_unlock(&lockres->l_lock); + user_log_dlm_error("ocfs2_dlm_unlock", status, lockres); + goto bail; + } + + user_wait_on_busy_lock(lockres); + + status = 0; +bail: + return status; +} + +static void user_dlm_recovery_handler_noop(int node_num, + void *recovery_data) +{ + /* We ignore recovery events */ + return; +} + +void user_dlm_set_locking_protocol(void) +{ + ocfs2_stack_glue_set_max_proto_version(&user_dlm_lproto.lp_max_version); +} + +struct ocfs2_cluster_connection *user_dlm_register(const struct qstr *name) +{ + int rc; + struct ocfs2_cluster_connection *conn; + + rc = ocfs2_cluster_connect_agnostic(name->name, name->len, + &user_dlm_lproto, + user_dlm_recovery_handler_noop, + NULL, &conn); + if (rc) + mlog_errno(rc); + + return rc ? ERR_PTR(rc) : conn; +} + +void user_dlm_unregister(struct ocfs2_cluster_connection *conn) +{ + ocfs2_cluster_disconnect(conn, 0); +} diff --git a/fs/ocfs2/dlmfs/userdlm.h b/fs/ocfs2/dlmfs/userdlm.h new file mode 100644 index 000000000..47ba18eac --- /dev/null +++ b/fs/ocfs2/dlmfs/userdlm.h @@ -0,0 +1,95 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ +/* + * userdlm.h + * + * Userspace dlm defines + * + * Copyright (C) 2002, 2004 Oracle. All rights reserved. + */ + + +#ifndef USERDLM_H +#define USERDLM_H + +#include <linux/module.h> +#include <linux/fs.h> +#include <linux/types.h> +#include <linux/workqueue.h> + +/* user_lock_res->l_flags flags. */ +#define USER_LOCK_ATTACHED (0x00000001) /* we have initialized + * the lvb */ +#define USER_LOCK_BUSY (0x00000002) /* we are currently in + * dlm_lock */ +#define USER_LOCK_BLOCKED (0x00000004) /* blocked waiting to + * downconvert*/ +#define USER_LOCK_IN_TEARDOWN (0x00000008) /* we're currently + * destroying this + * lock. */ +#define USER_LOCK_QUEUED (0x00000010) /* lock is on the + * workqueue */ +#define USER_LOCK_IN_CANCEL (0x00000020) + +struct user_lock_res { + spinlock_t l_lock; + + int l_flags; + +#define USER_DLM_LOCK_ID_MAX_LEN 32 + char l_name[USER_DLM_LOCK_ID_MAX_LEN]; + int l_namelen; + int l_level; + unsigned int l_ro_holders; + unsigned int l_ex_holders; + struct ocfs2_dlm_lksb l_lksb; + + int l_requested; + int l_blocking; + + wait_queue_head_t l_event; + + struct work_struct l_work; +}; + +extern struct workqueue_struct *user_dlm_worker; + +void user_dlm_lock_res_init(struct user_lock_res *lockres, + struct dentry *dentry); +int user_dlm_destroy_lock(struct user_lock_res *lockres); +int user_dlm_cluster_lock(struct user_lock_res *lockres, + int level, + int lkm_flags); +void user_dlm_cluster_unlock(struct user_lock_res *lockres, + int level); +void user_dlm_write_lvb(struct inode *inode, + const char *val, + unsigned int len); +bool user_dlm_read_lvb(struct inode *inode, char *val); +struct ocfs2_cluster_connection *user_dlm_register(const struct qstr *name); +void user_dlm_unregister(struct ocfs2_cluster_connection *conn); +void user_dlm_set_locking_protocol(void); + +struct dlmfs_inode_private { + struct ocfs2_cluster_connection *ip_conn; + + struct user_lock_res ip_lockres; /* unused for directories. */ + struct inode *ip_parent; + + struct inode ip_vfs_inode; +}; + +static inline struct dlmfs_inode_private * +DLMFS_I(struct inode *inode) +{ + return container_of(inode, + struct dlmfs_inode_private, + ip_vfs_inode); +} + +struct dlmfs_filp_private { + int fp_lock_level; +}; + +#define DLMFS_MAGIC 0x76a9f425 + +#endif /* USERDLM_H */ |