1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Code which implements an OCFS2 specific interface to our DLM.
8 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
26 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
30 #include <linux/kthread.h>
31 #include <linux/pagemap.h>
32 #include <linux/debugfs.h>
33 #include <linux/seq_file.h>
34 #include <linux/time.h>
35 #include <linux/quotaops.h>
37 #define MLOG_MASK_PREFIX ML_DLM_GLUE
38 #include <cluster/masklog.h>
41 #include "ocfs2_lockingver.h"
46 #include "extent_map.h"
48 #include "heartbeat.h"
51 #include "stackglue.h"
56 #include "refcounttree.h"
58 #include "buffer_head_io.h"
60 struct ocfs2_mask_waiter {
61 struct list_head mw_item;
63 struct completion mw_complete;
64 unsigned long mw_mask;
65 unsigned long mw_goal;
66 #ifdef CONFIG_OCFS2_FS_STATS
67 unsigned long long mw_lock_start;
71 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres);
72 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres);
73 static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres);
74 static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres);
77 * Return value from ->downconvert_worker functions.
79 * These control the precise actions of ocfs2_unblock_lock()
80 * and ocfs2_process_blocked_lock()
83 enum ocfs2_unblock_action {
84 UNBLOCK_CONTINUE = 0, /* Continue downconvert */
85 UNBLOCK_CONTINUE_POST = 1, /* Continue downconvert, fire
86 * ->post_unlock callback */
87 UNBLOCK_STOP_POST = 2, /* Do not downconvert, fire
88 * ->post_unlock() callback. */
91 struct ocfs2_unblock_ctl {
93 enum ocfs2_unblock_action unblock_action;
96 /* Lockdep class keys */
97 struct lock_class_key lockdep_keys[OCFS2_NUM_LOCK_TYPES];
99 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
101 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
103 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
106 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
109 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
110 struct ocfs2_lock_res *lockres);
112 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres);
114 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
116 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
119 #define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
121 /* This aids in debugging situations where a bad LVB might be involved. */
122 static void ocfs2_dump_meta_lvb_info(u64 level,
123 const char *function,
125 struct ocfs2_lock_res *lockres)
127 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
129 mlog(level, "LVB information for %s (called from %s:%u):\n",
130 lockres->l_name, function, line);
131 mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
132 lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
133 be32_to_cpu(lvb->lvb_igeneration));
134 mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
135 (unsigned long long)be64_to_cpu(lvb->lvb_isize),
136 be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
137 be16_to_cpu(lvb->lvb_imode));
138 mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
139 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
140 (long long)be64_to_cpu(lvb->lvb_iatime_packed),
141 (long long)be64_to_cpu(lvb->lvb_ictime_packed),
142 (long long)be64_to_cpu(lvb->lvb_imtime_packed),
143 be32_to_cpu(lvb->lvb_iattr));
148 * OCFS2 Lock Resource Operations
150 * These fine tune the behavior of the generic dlmglue locking infrastructure.
152 * The most basic of lock types can point ->l_priv to their respective
153 * struct ocfs2_super and allow the default actions to manage things.
155 * Right now, each lock type also needs to implement an init function,
156 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
157 * should be called when the lock is no longer needed (i.e., object
160 struct ocfs2_lock_res_ops {
162 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
163 * this callback if ->l_priv is not an ocfs2_super pointer
165 struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
168 * Optionally called in the downconvert thread after a
169 * successful downconvert. The lockres will not be referenced
170 * after this callback is called, so it is safe to free
173 * The exact semantics of when this is called are controlled
174 * by ->downconvert_worker()
176 void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
179 * Allow a lock type to add checks to determine whether it is
180 * safe to downconvert a lock. Return 0 to re-queue the
181 * downconvert at a later time, nonzero to continue.
183 * For most locks, the default checks that there are no
184 * incompatible holders are sufficient.
186 * Called with the lockres spinlock held.
188 int (*check_downconvert)(struct ocfs2_lock_res *, int);
191 * Allows a lock type to populate the lock value block. This
192 * is called on downconvert, and when we drop a lock.
194 * Locks that want to use this should set LOCK_TYPE_USES_LVB
195 * in the flags field.
197 * Called with the lockres spinlock held.
199 void (*set_lvb)(struct ocfs2_lock_res *);
202 * Called from the downconvert thread when it is determined
203 * that a lock will be downconverted. This is called without
204 * any locks held so the function can do work that might
205 * schedule (syncing out data, etc).
207 * This should return any one of the ocfs2_unblock_action
208 * values, depending on what it wants the thread to do.
210 int (*downconvert_worker)(struct ocfs2_lock_res *, int);
213 * LOCK_TYPE_* flags which describe the specific requirements
214 * of a lock type. Descriptions of each individual flag follow.
220 * Some locks want to "refresh" potentially stale data when a
221 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
222 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
223 * individual lockres l_flags member from the ast function. It is
224 * expected that the locking wrapper will clear the
225 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
227 #define LOCK_TYPE_REQUIRES_REFRESH 0x1
230 * Indicate that a lock type makes use of the lock value block. The
231 * ->set_lvb lock type callback must be defined.
233 #define LOCK_TYPE_USES_LVB 0x2
235 static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
236 .get_osb = ocfs2_get_inode_osb,
240 static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops = {
241 .get_osb = ocfs2_get_inode_osb,
242 .check_downconvert = ocfs2_check_meta_downconvert,
243 .set_lvb = ocfs2_set_meta_lvb,
244 .downconvert_worker = ocfs2_data_convert_worker,
245 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
248 static struct ocfs2_lock_res_ops ocfs2_super_lops = {
249 .flags = LOCK_TYPE_REQUIRES_REFRESH,
252 static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
256 static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops = {
260 static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops = {
261 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
264 static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
265 .get_osb = ocfs2_get_dentry_osb,
266 .post_unlock = ocfs2_dentry_post_unlock,
267 .downconvert_worker = ocfs2_dentry_convert_worker,
271 static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
272 .get_osb = ocfs2_get_inode_osb,
276 static struct ocfs2_lock_res_ops ocfs2_flock_lops = {
277 .get_osb = ocfs2_get_file_osb,
281 static struct ocfs2_lock_res_ops ocfs2_qinfo_lops = {
282 .set_lvb = ocfs2_set_qinfo_lvb,
283 .get_osb = ocfs2_get_qinfo_osb,
284 .flags = LOCK_TYPE_REQUIRES_REFRESH | LOCK_TYPE_USES_LVB,
287 static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops = {
288 .check_downconvert = ocfs2_check_refcount_downconvert,
289 .downconvert_worker = ocfs2_refcount_convert_worker,
293 static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
295 return lockres->l_type == OCFS2_LOCK_TYPE_META ||
296 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
297 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
300 static inline struct ocfs2_lock_res *ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb *lksb)
302 return container_of(lksb, struct ocfs2_lock_res, l_lksb);
305 static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
307 BUG_ON(!ocfs2_is_inode_lock(lockres));
309 return (struct inode *) lockres->l_priv;
312 static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
314 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
316 return (struct ocfs2_dentry_lock *)lockres->l_priv;
319 static inline struct ocfs2_mem_dqinfo *ocfs2_lock_res_qinfo(struct ocfs2_lock_res *lockres)
321 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_QINFO);
323 return (struct ocfs2_mem_dqinfo *)lockres->l_priv;
326 static inline struct ocfs2_refcount_tree *
327 ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res *res)
329 return container_of(res, struct ocfs2_refcount_tree, rf_lockres);
332 static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
334 if (lockres->l_ops->get_osb)
335 return lockres->l_ops->get_osb(lockres);
337 return (struct ocfs2_super *)lockres->l_priv;
340 static int ocfs2_lock_create(struct ocfs2_super *osb,
341 struct ocfs2_lock_res *lockres,
344 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
346 static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
347 struct ocfs2_lock_res *lockres,
348 int level, unsigned long caller_ip);
349 static inline void ocfs2_cluster_unlock(struct ocfs2_super *osb,
350 struct ocfs2_lock_res *lockres,
353 __ocfs2_cluster_unlock(osb, lockres, level, _RET_IP_);
356 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
357 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
358 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
359 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
360 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
361 struct ocfs2_lock_res *lockres);
362 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
364 #define ocfs2_log_dlm_error(_func, _err, _lockres) do { \
365 if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY) \
366 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \
367 _err, _func, _lockres->l_name); \
369 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n", \
370 _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name, \
371 (unsigned int)ocfs2_get_dentry_lock_ino(_lockres)); \
373 static int ocfs2_downconvert_thread(void *arg);
374 static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
375 struct ocfs2_lock_res *lockres);
376 static int ocfs2_inode_lock_update(struct inode *inode,
377 struct buffer_head **bh);
378 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
379 static inline int ocfs2_highest_compat_lock_level(int level);
380 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
382 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
383 struct ocfs2_lock_res *lockres,
386 unsigned int generation);
387 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
388 struct ocfs2_lock_res *lockres);
389 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
390 struct ocfs2_lock_res *lockres);
393 static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
402 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
404 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
405 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
406 (long long)blkno, generation);
408 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
410 mlog(0, "built lock resource with name: %s\n", name);
415 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
417 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
418 struct ocfs2_dlm_debug *dlm_debug)
420 mlog(0, "Add tracking for lockres %s\n", res->l_name);
422 spin_lock(&ocfs2_dlm_tracking_lock);
423 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
424 spin_unlock(&ocfs2_dlm_tracking_lock);
427 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
429 spin_lock(&ocfs2_dlm_tracking_lock);
430 if (!list_empty(&res->l_debug_list))
431 list_del_init(&res->l_debug_list);
432 spin_unlock(&ocfs2_dlm_tracking_lock);
435 #ifdef CONFIG_OCFS2_FS_STATS
436 static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
438 res->l_lock_num_prmode = 0;
439 res->l_lock_num_prmode_failed = 0;
440 res->l_lock_total_prmode = 0;
441 res->l_lock_max_prmode = 0;
442 res->l_lock_num_exmode = 0;
443 res->l_lock_num_exmode_failed = 0;
444 res->l_lock_total_exmode = 0;
445 res->l_lock_max_exmode = 0;
446 res->l_lock_refresh = 0;
449 static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
450 struct ocfs2_mask_waiter *mw, int ret)
452 unsigned long long *num, *sum;
453 unsigned int *max, *failed;
454 struct timespec ts = current_kernel_time();
455 unsigned long long time = timespec_to_ns(&ts) - mw->mw_lock_start;
457 if (level == LKM_PRMODE) {
458 num = &res->l_lock_num_prmode;
459 sum = &res->l_lock_total_prmode;
460 max = &res->l_lock_max_prmode;
461 failed = &res->l_lock_num_prmode_failed;
462 } else if (level == LKM_EXMODE) {
463 num = &res->l_lock_num_exmode;
464 sum = &res->l_lock_total_exmode;
465 max = &res->l_lock_max_exmode;
466 failed = &res->l_lock_num_exmode_failed;
478 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
480 lockres->l_lock_refresh++;
483 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
485 struct timespec ts = current_kernel_time();
486 mw->mw_lock_start = timespec_to_ns(&ts);
489 static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
492 static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
493 int level, struct ocfs2_mask_waiter *mw, int ret)
496 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
499 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
504 static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
505 struct ocfs2_lock_res *res,
506 enum ocfs2_lock_type type,
507 struct ocfs2_lock_res_ops *ops,
514 res->l_level = DLM_LOCK_IV;
515 res->l_requested = DLM_LOCK_IV;
516 res->l_blocking = DLM_LOCK_IV;
517 res->l_action = OCFS2_AST_INVALID;
518 res->l_unlock_action = OCFS2_UNLOCK_INVALID;
520 res->l_flags = OCFS2_LOCK_INITIALIZED;
522 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
524 ocfs2_init_lock_stats(res);
525 #ifdef CONFIG_DEBUG_LOCK_ALLOC
526 if (type != OCFS2_LOCK_TYPE_OPEN)
527 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
528 &lockdep_keys[type], 0);
530 res->l_lockdep_map.key = NULL;
534 void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
536 /* This also clears out the lock status block */
537 memset(res, 0, sizeof(struct ocfs2_lock_res));
538 spin_lock_init(&res->l_lock);
539 init_waitqueue_head(&res->l_event);
540 INIT_LIST_HEAD(&res->l_blocked_list);
541 INIT_LIST_HEAD(&res->l_mask_waiters);
544 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
545 enum ocfs2_lock_type type,
546 unsigned int generation,
549 struct ocfs2_lock_res_ops *ops;
552 case OCFS2_LOCK_TYPE_RW:
553 ops = &ocfs2_inode_rw_lops;
555 case OCFS2_LOCK_TYPE_META:
556 ops = &ocfs2_inode_inode_lops;
558 case OCFS2_LOCK_TYPE_OPEN:
559 ops = &ocfs2_inode_open_lops;
562 mlog_bug_on_msg(1, "type: %d\n", type);
563 ops = NULL; /* thanks, gcc */
567 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
568 generation, res->l_name);
569 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
572 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
574 struct inode *inode = ocfs2_lock_res_inode(lockres);
576 return OCFS2_SB(inode->i_sb);
579 static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
581 struct ocfs2_mem_dqinfo *info = lockres->l_priv;
583 return OCFS2_SB(info->dqi_gi.dqi_sb);
586 static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
588 struct ocfs2_file_private *fp = lockres->l_priv;
590 return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
593 static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
595 __be64 inode_blkno_be;
597 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
600 return be64_to_cpu(inode_blkno_be);
603 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
605 struct ocfs2_dentry_lock *dl = lockres->l_priv;
607 return OCFS2_SB(dl->dl_inode->i_sb);
610 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
611 u64 parent, struct inode *inode)
614 u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
615 __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
616 struct ocfs2_lock_res *lockres = &dl->dl_lockres;
618 ocfs2_lock_res_init_once(lockres);
621 * Unfortunately, the standard lock naming scheme won't work
622 * here because we have two 16 byte values to use. Instead,
623 * we'll stuff the inode number as a binary value. We still
624 * want error prints to show something without garbling the
625 * display, so drop a null byte in there before the inode
626 * number. A future version of OCFS2 will likely use all
627 * binary lock names. The stringified names have been a
628 * tremendous aid in debugging, but now that the debugfs
629 * interface exists, we can mangle things there if need be.
631 * NOTE: We also drop the standard "pad" value (the total lock
632 * name size stays the same though - the last part is all
633 * zeros due to the memset in ocfs2_lock_res_init_once()
635 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
637 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
640 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
642 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
645 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
646 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
650 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
651 struct ocfs2_super *osb)
653 /* Superblock lockres doesn't come from a slab so we call init
654 * once on it manually. */
655 ocfs2_lock_res_init_once(res);
656 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
658 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
659 &ocfs2_super_lops, osb);
662 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
663 struct ocfs2_super *osb)
665 /* Rename lockres doesn't come from a slab so we call init
666 * once on it manually. */
667 ocfs2_lock_res_init_once(res);
668 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
669 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
670 &ocfs2_rename_lops, osb);
673 static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
674 struct ocfs2_super *osb)
676 /* nfs_sync lockres doesn't come from a slab so we call init
677 * once on it manually. */
678 ocfs2_lock_res_init_once(res);
679 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
680 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
681 &ocfs2_nfs_sync_lops, osb);
684 static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
685 struct ocfs2_super *osb)
687 ocfs2_lock_res_init_once(res);
688 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
689 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
690 &ocfs2_orphan_scan_lops, osb);
693 void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
694 struct ocfs2_file_private *fp)
696 struct inode *inode = fp->fp_file->f_mapping->host;
697 struct ocfs2_inode_info *oi = OCFS2_I(inode);
699 ocfs2_lock_res_init_once(lockres);
700 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
701 inode->i_generation, lockres->l_name);
702 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
703 OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
705 lockres->l_flags |= OCFS2_LOCK_NOCACHE;
708 void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
709 struct ocfs2_mem_dqinfo *info)
711 ocfs2_lock_res_init_once(lockres);
712 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
714 ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
715 OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
719 void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
720 struct ocfs2_super *osb, u64 ref_blkno,
721 unsigned int generation)
723 ocfs2_lock_res_init_once(lockres);
724 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
725 generation, lockres->l_name);
726 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
727 &ocfs2_refcount_block_lops, osb);
730 void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
734 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
737 ocfs2_remove_lockres_tracking(res);
739 mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
740 "Lockres %s is on the blocked list\n",
742 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
743 "Lockres %s has mask waiters pending\n",
745 mlog_bug_on_msg(spin_is_locked(&res->l_lock),
746 "Lockres %s is locked\n",
748 mlog_bug_on_msg(res->l_ro_holders,
749 "Lockres %s has %u ro holders\n",
750 res->l_name, res->l_ro_holders);
751 mlog_bug_on_msg(res->l_ex_holders,
752 "Lockres %s has %u ex holders\n",
753 res->l_name, res->l_ex_holders);
755 /* Need to clear out the lock status block for the dlm */
756 memset(&res->l_lksb, 0, sizeof(res->l_lksb));
762 static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
771 lockres->l_ex_holders++;
774 lockres->l_ro_holders++;
783 static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
792 BUG_ON(!lockres->l_ex_holders);
793 lockres->l_ex_holders--;
796 BUG_ON(!lockres->l_ro_holders);
797 lockres->l_ro_holders--;
805 /* WARNING: This function lives in a world where the only three lock
806 * levels are EX, PR, and NL. It *will* have to be adjusted when more
807 * lock types are added. */
808 static inline int ocfs2_highest_compat_lock_level(int level)
810 int new_level = DLM_LOCK_EX;
812 if (level == DLM_LOCK_EX)
813 new_level = DLM_LOCK_NL;
814 else if (level == DLM_LOCK_PR)
815 new_level = DLM_LOCK_PR;
819 static void lockres_set_flags(struct ocfs2_lock_res *lockres,
820 unsigned long newflags)
822 struct ocfs2_mask_waiter *mw, *tmp;
824 assert_spin_locked(&lockres->l_lock);
826 lockres->l_flags = newflags;
828 list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
829 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
832 list_del_init(&mw->mw_item);
834 complete(&mw->mw_complete);
837 static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
839 lockres_set_flags(lockres, lockres->l_flags | or);
841 static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
844 lockres_set_flags(lockres, lockres->l_flags & ~clear);
847 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
851 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
852 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
853 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
854 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
856 lockres->l_level = lockres->l_requested;
857 if (lockres->l_level <=
858 ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
859 lockres->l_blocking = DLM_LOCK_NL;
860 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
862 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
867 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
871 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
872 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
874 /* Convert from RO to EX doesn't really need anything as our
875 * information is already up to data. Convert from NL to
876 * *anything* however should mark ourselves as needing an
878 if (lockres->l_level == DLM_LOCK_NL &&
879 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
880 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
882 lockres->l_level = lockres->l_requested;
885 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
886 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
887 * downconverting the lock before the upconvert has fully completed.
889 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
891 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
896 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
900 BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
901 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
903 if (lockres->l_requested > DLM_LOCK_NL &&
904 !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
905 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
906 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
908 lockres->l_level = lockres->l_requested;
909 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
910 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
915 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
918 int needs_downconvert = 0;
921 assert_spin_locked(&lockres->l_lock);
923 if (level > lockres->l_blocking) {
924 /* only schedule a downconvert if we haven't already scheduled
925 * one that goes low enough to satisfy the level we're
926 * blocking. this also catches the case where we get
928 if (ocfs2_highest_compat_lock_level(level) <
929 ocfs2_highest_compat_lock_level(lockres->l_blocking))
930 needs_downconvert = 1;
932 lockres->l_blocking = level;
935 if (needs_downconvert)
936 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
938 mlog_exit(needs_downconvert);
939 return needs_downconvert;
943 * OCFS2_LOCK_PENDING and l_pending_gen.
945 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
946 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
947 * for more details on the race.
949 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
950 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
951 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
952 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
953 * the caller is going to try to clear PENDING again. If nothing else is
954 * happening, __lockres_clear_pending() sees PENDING is unset and does
957 * But what if another path (eg downconvert thread) has just started a
958 * new locking action? The other path has re-set PENDING. Our path
959 * cannot clear PENDING, because that will re-open the original race
965 * ocfs2_cluster_lock()
970 * ocfs2_locking_ast() ocfs2_downconvert_thread()
971 * clear PENDING ocfs2_unblock_lock()
974 * ocfs2_prepare_downconvert()
984 * So as you can see, we now have a window where l_lock is not held,
985 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
987 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
988 * set by ocfs2_prepare_downconvert(). That wasn't nice.
990 * To solve this we introduce l_pending_gen. A call to
991 * lockres_clear_pending() will only do so when it is passed a generation
992 * number that matches the lockres. lockres_set_pending() will return the
993 * current generation number. When ocfs2_cluster_lock() goes to clear
994 * PENDING, it passes the generation it got from set_pending(). In our
995 * example above, the generation numbers will *not* match. Thus,
996 * ocfs2_cluster_lock() will not clear the PENDING set by
997 * ocfs2_prepare_downconvert().
1000 /* Unlocked version for ocfs2_locking_ast() */
1001 static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
1002 unsigned int generation,
1003 struct ocfs2_super *osb)
1005 assert_spin_locked(&lockres->l_lock);
1008 * The ast and locking functions can race us here. The winner
1009 * will clear pending, the loser will not.
1011 if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
1012 (lockres->l_pending_gen != generation))
1015 lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
1016 lockres->l_pending_gen++;
1019 * The downconvert thread may have skipped us because we
1020 * were PENDING. Wake it up.
1022 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1023 ocfs2_wake_downconvert_thread(osb);
1026 /* Locked version for callers of ocfs2_dlm_lock() */
1027 static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1028 unsigned int generation,
1029 struct ocfs2_super *osb)
1031 unsigned long flags;
1033 spin_lock_irqsave(&lockres->l_lock, flags);
1034 __lockres_clear_pending(lockres, generation, osb);
1035 spin_unlock_irqrestore(&lockres->l_lock, flags);
1038 static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1040 assert_spin_locked(&lockres->l_lock);
1041 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1043 lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1045 return lockres->l_pending_gen;
1049 static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1051 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1052 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1053 int needs_downconvert;
1054 unsigned long flags;
1056 BUG_ON(level <= DLM_LOCK_NL);
1058 mlog(0, "BAST fired for lockres %s, blocking %d, level %d type %s\n",
1059 lockres->l_name, level, lockres->l_level,
1060 ocfs2_lock_type_string(lockres->l_type));
1063 * We can skip the bast for locks which don't enable caching -
1064 * they'll be dropped at the earliest possible time anyway.
1066 if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1069 spin_lock_irqsave(&lockres->l_lock, flags);
1070 needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1071 if (needs_downconvert)
1072 ocfs2_schedule_blocked_lock(osb, lockres);
1073 spin_unlock_irqrestore(&lockres->l_lock, flags);
1075 wake_up(&lockres->l_event);
1077 ocfs2_wake_downconvert_thread(osb);
1080 static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1082 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1083 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1084 unsigned long flags;
1087 spin_lock_irqsave(&lockres->l_lock, flags);
1089 status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1091 if (status == -EAGAIN) {
1092 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1097 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1098 lockres->l_name, status);
1099 spin_unlock_irqrestore(&lockres->l_lock, flags);
1103 switch(lockres->l_action) {
1104 case OCFS2_AST_ATTACH:
1105 ocfs2_generic_handle_attach_action(lockres);
1106 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1108 case OCFS2_AST_CONVERT:
1109 ocfs2_generic_handle_convert_action(lockres);
1111 case OCFS2_AST_DOWNCONVERT:
1112 ocfs2_generic_handle_downconvert_action(lockres);
1115 mlog(ML_ERROR, "lockres %s: ast fired with invalid action: %u "
1116 "lockres flags = 0x%lx, unlock action: %u\n",
1117 lockres->l_name, lockres->l_action, lockres->l_flags,
1118 lockres->l_unlock_action);
1122 /* set it to something invalid so if we get called again we
1124 lockres->l_action = OCFS2_AST_INVALID;
1126 /* Did we try to cancel this lock? Clear that state */
1127 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1128 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1131 * We may have beaten the locking functions here. We certainly
1132 * know that dlm_lock() has been called :-)
1133 * Because we can't have two lock calls in flight at once, we
1134 * can use lockres->l_pending_gen.
1136 __lockres_clear_pending(lockres, lockres->l_pending_gen, osb);
1138 wake_up(&lockres->l_event);
1139 spin_unlock_irqrestore(&lockres->l_lock, flags);
1142 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1145 unsigned long flags;
1148 spin_lock_irqsave(&lockres->l_lock, flags);
1149 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1150 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1152 lockres->l_action = OCFS2_AST_INVALID;
1154 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1155 spin_unlock_irqrestore(&lockres->l_lock, flags);
1157 wake_up(&lockres->l_event);
1161 /* Note: If we detect another process working on the lock (i.e.,
1162 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1163 * to do the right thing in that case.
1165 static int ocfs2_lock_create(struct ocfs2_super *osb,
1166 struct ocfs2_lock_res *lockres,
1171 unsigned long flags;
1176 mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1179 spin_lock_irqsave(&lockres->l_lock, flags);
1180 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1181 (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1182 spin_unlock_irqrestore(&lockres->l_lock, flags);
1186 lockres->l_action = OCFS2_AST_ATTACH;
1187 lockres->l_requested = level;
1188 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1189 gen = lockres_set_pending(lockres);
1190 spin_unlock_irqrestore(&lockres->l_lock, flags);
1192 ret = ocfs2_dlm_lock(osb->cconn,
1197 OCFS2_LOCK_ID_MAX_LEN - 1);
1198 lockres_clear_pending(lockres, gen, osb);
1200 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1201 ocfs2_recover_from_dlm_error(lockres, 1);
1204 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1211 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1214 unsigned long flags;
1217 spin_lock_irqsave(&lockres->l_lock, flags);
1218 ret = lockres->l_flags & flag;
1219 spin_unlock_irqrestore(&lockres->l_lock, flags);
1224 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1227 wait_event(lockres->l_event,
1228 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1231 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1234 wait_event(lockres->l_event,
1235 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1238 /* predict what lock level we'll be dropping down to on behalf
1239 * of another node, and return true if the currently wanted
1240 * level will be compatible with it. */
1241 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1244 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1246 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1249 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1251 INIT_LIST_HEAD(&mw->mw_item);
1252 init_completion(&mw->mw_complete);
1253 ocfs2_init_start_time(mw);
1256 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1258 wait_for_completion(&mw->mw_complete);
1259 /* Re-arm the completion in case we want to wait on it again */
1260 INIT_COMPLETION(mw->mw_complete);
1261 return mw->mw_status;
1264 static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1265 struct ocfs2_mask_waiter *mw,
1269 BUG_ON(!list_empty(&mw->mw_item));
1271 assert_spin_locked(&lockres->l_lock);
1273 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1278 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
1279 * if the mask still hadn't reached its goal */
1280 static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1281 struct ocfs2_mask_waiter *mw)
1283 unsigned long flags;
1286 spin_lock_irqsave(&lockres->l_lock, flags);
1287 if (!list_empty(&mw->mw_item)) {
1288 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1291 list_del_init(&mw->mw_item);
1292 init_completion(&mw->mw_complete);
1294 spin_unlock_irqrestore(&lockres->l_lock, flags);
1300 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1301 struct ocfs2_lock_res *lockres)
1305 ret = wait_for_completion_interruptible(&mw->mw_complete);
1307 lockres_remove_mask_waiter(lockres, mw);
1309 ret = mw->mw_status;
1310 /* Re-arm the completion in case we want to wait on it again */
1311 INIT_COMPLETION(mw->mw_complete);
1315 static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1316 struct ocfs2_lock_res *lockres,
1321 unsigned long caller_ip)
1323 struct ocfs2_mask_waiter mw;
1324 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1325 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1326 unsigned long flags;
1328 int noqueue_attempted = 0;
1332 ocfs2_init_mask_waiter(&mw);
1334 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1335 lkm_flags |= DLM_LKF_VALBLK;
1340 spin_lock_irqsave(&lockres->l_lock, flags);
1342 if (catch_signals && signal_pending(current)) {
1347 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1348 "Cluster lock called on freeing lockres %s! flags "
1349 "0x%lx\n", lockres->l_name, lockres->l_flags);
1351 /* We only compare against the currently granted level
1352 * here. If the lock is blocked waiting on a downconvert,
1353 * we'll get caught below. */
1354 if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1355 level > lockres->l_level) {
1356 /* is someone sitting in dlm_lock? If so, wait on
1358 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1363 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1365 * We've upconverted. If the lock now has a level we can
1366 * work with, we take it. If, however, the lock is not at the
1367 * required level, we go thru the full cycle. One way this could
1368 * happen is if a process requesting an upconvert to PR is
1369 * closely followed by another requesting upconvert to an EX.
1370 * If the process requesting EX lands here, we want it to
1371 * continue attempting to upconvert and let the process
1372 * requesting PR take the lock.
1373 * If multiple processes request upconvert to PR, the first one
1374 * here will take the lock. The others will have to go thru the
1375 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1376 * downconvert request.
1378 if (level <= lockres->l_level)
1379 goto update_holders;
1382 if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1383 !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1384 /* is the lock is currently blocked on behalf of
1386 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1391 if (level > lockres->l_level) {
1392 if (noqueue_attempted > 0) {
1396 if (lkm_flags & DLM_LKF_NOQUEUE)
1397 noqueue_attempted = 1;
1399 if (lockres->l_action != OCFS2_AST_INVALID)
1400 mlog(ML_ERROR, "lockres %s has action %u pending\n",
1401 lockres->l_name, lockres->l_action);
1403 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1404 lockres->l_action = OCFS2_AST_ATTACH;
1405 lkm_flags &= ~DLM_LKF_CONVERT;
1407 lockres->l_action = OCFS2_AST_CONVERT;
1408 lkm_flags |= DLM_LKF_CONVERT;
1411 lockres->l_requested = level;
1412 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1413 gen = lockres_set_pending(lockres);
1414 spin_unlock_irqrestore(&lockres->l_lock, flags);
1416 BUG_ON(level == DLM_LOCK_IV);
1417 BUG_ON(level == DLM_LOCK_NL);
1419 mlog(0, "lock %s, convert from %d to level = %d\n",
1420 lockres->l_name, lockres->l_level, level);
1422 /* call dlm_lock to upgrade lock now */
1423 ret = ocfs2_dlm_lock(osb->cconn,
1428 OCFS2_LOCK_ID_MAX_LEN - 1);
1429 lockres_clear_pending(lockres, gen, osb);
1431 if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1433 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1436 ocfs2_recover_from_dlm_error(lockres, 1);
1440 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1443 /* At this point we've gone inside the dlm and need to
1444 * complete our work regardless. */
1447 /* wait for busy to clear and carry on */
1452 /* Ok, if we get here then we're good to go. */
1453 ocfs2_inc_holders(lockres, level);
1457 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1459 spin_unlock_irqrestore(&lockres->l_lock, flags);
1462 * This is helping work around a lock inversion between the page lock
1463 * and dlm locks. One path holds the page lock while calling aops
1464 * which block acquiring dlm locks. The voting thread holds dlm
1465 * locks while acquiring page locks while down converting data locks.
1466 * This block is helping an aop path notice the inversion and back
1467 * off to unlock its page lock before trying the dlm lock again.
1469 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1470 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1472 if (lockres_remove_mask_waiter(lockres, &mw))
1478 ret = ocfs2_wait_for_mask(&mw);
1483 ocfs2_update_lock_stats(lockres, level, &mw, ret);
1485 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1486 if (!ret && lockres->l_lockdep_map.key != NULL) {
1487 if (level == DLM_LOCK_PR)
1488 rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1489 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1492 rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1493 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1501 static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1502 struct ocfs2_lock_res *lockres,
1507 return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1512 static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1513 struct ocfs2_lock_res *lockres,
1515 unsigned long caller_ip)
1517 unsigned long flags;
1520 spin_lock_irqsave(&lockres->l_lock, flags);
1521 ocfs2_dec_holders(lockres, level);
1522 ocfs2_downconvert_on_unlock(osb, lockres);
1523 spin_unlock_irqrestore(&lockres->l_lock, flags);
1524 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1525 if (lockres->l_lockdep_map.key != NULL)
1526 rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1531 static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1532 struct ocfs2_lock_res *lockres,
1536 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1537 unsigned long flags;
1538 u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1540 spin_lock_irqsave(&lockres->l_lock, flags);
1541 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1542 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1543 spin_unlock_irqrestore(&lockres->l_lock, flags);
1545 return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1548 /* Grants us an EX lock on the data and metadata resources, skipping
1549 * the normal cluster directory lookup. Use this ONLY on newly created
1550 * inodes which other nodes can't possibly see, and which haven't been
1551 * hashed in the inode hash yet. This can give us a good performance
1552 * increase as it'll skip the network broadcast normally associated
1553 * with creating a new lock resource. */
1554 int ocfs2_create_new_inode_locks(struct inode *inode)
1557 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1560 BUG_ON(!ocfs2_inode_is_new(inode));
1564 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1566 /* NOTE: That we don't increment any of the holder counts, nor
1567 * do we add anything to a journal handle. Since this is
1568 * supposed to be a new inode which the cluster doesn't know
1569 * about yet, there is no need to. As far as the LVB handling
1570 * is concerned, this is basically like acquiring an EX lock
1571 * on a resource which has an invalid one -- we'll set it
1572 * valid when we release the EX. */
1574 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1581 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1582 * don't use a generation in their lock names.
1584 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1590 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1601 int ocfs2_rw_lock(struct inode *inode, int write)
1604 struct ocfs2_lock_res *lockres;
1605 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1611 mlog(0, "inode %llu take %s RW lock\n",
1612 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1613 write ? "EXMODE" : "PRMODE");
1615 if (ocfs2_mount_local(osb)) {
1620 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1622 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1624 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1633 void ocfs2_rw_unlock(struct inode *inode, int write)
1635 int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1636 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1637 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1641 mlog(0, "inode %llu drop %s RW lock\n",
1642 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1643 write ? "EXMODE" : "PRMODE");
1645 if (!ocfs2_mount_local(osb))
1646 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1652 * ocfs2_open_lock always get PR mode lock.
1654 int ocfs2_open_lock(struct inode *inode)
1657 struct ocfs2_lock_res *lockres;
1658 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1664 mlog(0, "inode %llu take PRMODE open lock\n",
1665 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1667 if (ocfs2_mount_local(osb))
1670 lockres = &OCFS2_I(inode)->ip_open_lockres;
1672 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1682 int ocfs2_try_open_lock(struct inode *inode, int write)
1684 int status = 0, level;
1685 struct ocfs2_lock_res *lockres;
1686 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1692 mlog(0, "inode %llu try to take %s open lock\n",
1693 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1694 write ? "EXMODE" : "PRMODE");
1696 if (ocfs2_mount_local(osb))
1699 lockres = &OCFS2_I(inode)->ip_open_lockres;
1701 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1704 * The file system may already holding a PRMODE/EXMODE open lock.
1705 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1706 * other nodes and the -EAGAIN will indicate to the caller that
1707 * this inode is still in use.
1709 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1710 level, DLM_LKF_NOQUEUE, 0);
1718 * ocfs2_open_unlock unlock PR and EX mode open locks.
1720 void ocfs2_open_unlock(struct inode *inode)
1722 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1723 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1727 mlog(0, "inode %llu drop open lock\n",
1728 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1730 if (ocfs2_mount_local(osb))
1733 if(lockres->l_ro_holders)
1734 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1736 if(lockres->l_ex_holders)
1737 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1744 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1748 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1749 unsigned long flags;
1750 struct ocfs2_mask_waiter mw;
1752 ocfs2_init_mask_waiter(&mw);
1755 spin_lock_irqsave(&lockres->l_lock, flags);
1756 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1757 ret = ocfs2_prepare_cancel_convert(osb, lockres);
1759 spin_unlock_irqrestore(&lockres->l_lock, flags);
1760 ret = ocfs2_cancel_convert(osb, lockres);
1767 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1768 spin_unlock_irqrestore(&lockres->l_lock, flags);
1770 ocfs2_wait_for_mask(&mw);
1776 * We may still have gotten the lock, in which case there's no
1777 * point to restarting the syscall.
1779 if (lockres->l_level == level)
1782 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1783 lockres->l_flags, lockres->l_level, lockres->l_action);
1785 spin_unlock_irqrestore(&lockres->l_lock, flags);
1792 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1793 * flock() calls. The locking approach this requires is sufficiently
1794 * different from all other cluster lock types that we implement a
1795 * seperate path to the "low-level" dlm calls. In particular:
1797 * - No optimization of lock levels is done - we take at exactly
1798 * what's been requested.
1800 * - No lock caching is employed. We immediately downconvert to
1801 * no-lock at unlock time. This also means flock locks never go on
1802 * the blocking list).
1804 * - Since userspace can trivially deadlock itself with flock, we make
1805 * sure to allow cancellation of a misbehaving applications flock()
1808 * - Access to any flock lockres doesn't require concurrency, so we
1809 * can simplify the code by requiring the caller to guarantee
1810 * serialization of dlmglue flock calls.
1812 int ocfs2_file_lock(struct file *file, int ex, int trylock)
1814 int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1815 unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1816 unsigned long flags;
1817 struct ocfs2_file_private *fp = file->private_data;
1818 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1819 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1820 struct ocfs2_mask_waiter mw;
1822 ocfs2_init_mask_waiter(&mw);
1824 if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1825 (lockres->l_level > DLM_LOCK_NL)) {
1827 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1828 "level: %u\n", lockres->l_name, lockres->l_flags,
1833 spin_lock_irqsave(&lockres->l_lock, flags);
1834 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1835 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1836 spin_unlock_irqrestore(&lockres->l_lock, flags);
1839 * Get the lock at NLMODE to start - that way we
1840 * can cancel the upconvert request if need be.
1842 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1848 ret = ocfs2_wait_for_mask(&mw);
1853 spin_lock_irqsave(&lockres->l_lock, flags);
1856 lockres->l_action = OCFS2_AST_CONVERT;
1857 lkm_flags |= DLM_LKF_CONVERT;
1858 lockres->l_requested = level;
1859 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1861 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1862 spin_unlock_irqrestore(&lockres->l_lock, flags);
1864 ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
1865 lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
1867 if (!trylock || (ret != -EAGAIN)) {
1868 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1872 ocfs2_recover_from_dlm_error(lockres, 1);
1873 lockres_remove_mask_waiter(lockres, &mw);
1877 ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
1878 if (ret == -ERESTARTSYS) {
1880 * Userspace can cause deadlock itself with
1881 * flock(). Current behavior locally is to allow the
1882 * deadlock, but abort the system call if a signal is
1883 * received. We follow this example, otherwise a
1884 * poorly written program could sit in kernel until
1887 * Handling this is a bit more complicated for Ocfs2
1888 * though. We can't exit this function with an
1889 * outstanding lock request, so a cancel convert is
1890 * required. We intentionally overwrite 'ret' - if the
1891 * cancel fails and the lock was granted, it's easier
1892 * to just bubble success back up to the user.
1894 ret = ocfs2_flock_handle_signal(lockres, level);
1895 } else if (!ret && (level > lockres->l_level)) {
1896 /* Trylock failed asynchronously */
1903 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1904 lockres->l_name, ex, trylock, ret);
1908 void ocfs2_file_unlock(struct file *file)
1912 unsigned long flags;
1913 struct ocfs2_file_private *fp = file->private_data;
1914 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1915 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1916 struct ocfs2_mask_waiter mw;
1918 ocfs2_init_mask_waiter(&mw);
1920 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
1923 if (lockres->l_level == DLM_LOCK_NL)
1926 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1927 lockres->l_name, lockres->l_flags, lockres->l_level,
1930 spin_lock_irqsave(&lockres->l_lock, flags);
1932 * Fake a blocking ast for the downconvert code.
1934 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
1935 lockres->l_blocking = DLM_LOCK_EX;
1937 gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
1938 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1939 spin_unlock_irqrestore(&lockres->l_lock, flags);
1941 ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
1947 ret = ocfs2_wait_for_mask(&mw);
1952 static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
1953 struct ocfs2_lock_res *lockres)
1959 /* If we know that another node is waiting on our lock, kick
1960 * the downconvert thread * pre-emptively when we reach a release
1962 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
1963 switch(lockres->l_blocking) {
1965 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
1969 if (!lockres->l_ex_holders)
1978 ocfs2_wake_downconvert_thread(osb);
1983 #define OCFS2_SEC_BITS 34
1984 #define OCFS2_SEC_SHIFT (64 - 34)
1985 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
1987 /* LVB only has room for 64 bits of time here so we pack it for
1989 static u64 ocfs2_pack_timespec(struct timespec *spec)
1992 u64 sec = spec->tv_sec;
1993 u32 nsec = spec->tv_nsec;
1995 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2000 /* Call this with the lockres locked. I am reasonably sure we don't
2001 * need ip_lock in this function as anyone who would be changing those
2002 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2003 static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2005 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2006 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2007 struct ocfs2_meta_lvb *lvb;
2011 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2014 * Invalidate the LVB of a deleted inode - this way other
2015 * nodes are forced to go to disk and discover the new inode
2018 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2019 lvb->lvb_version = 0;
2023 lvb->lvb_version = OCFS2_LVB_VERSION;
2024 lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
2025 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2026 lvb->lvb_iuid = cpu_to_be32(inode->i_uid);
2027 lvb->lvb_igid = cpu_to_be32(inode->i_gid);
2028 lvb->lvb_imode = cpu_to_be16(inode->i_mode);
2029 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink);
2030 lvb->lvb_iatime_packed =
2031 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2032 lvb->lvb_ictime_packed =
2033 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2034 lvb->lvb_imtime_packed =
2035 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2036 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr);
2037 lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2038 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2041 mlog_meta_lvb(0, lockres);
2046 static void ocfs2_unpack_timespec(struct timespec *spec,
2049 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2050 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2053 static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2055 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2056 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2057 struct ocfs2_meta_lvb *lvb;
2061 mlog_meta_lvb(0, lockres);
2063 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2065 /* We're safe here without the lockres lock... */
2066 spin_lock(&oi->ip_lock);
2067 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2068 i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2070 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2071 oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2072 ocfs2_set_inode_flags(inode);
2074 /* fast-symlinks are a special case */
2075 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2076 inode->i_blocks = 0;
2078 inode->i_blocks = ocfs2_inode_sector_count(inode);
2080 inode->i_uid = be32_to_cpu(lvb->lvb_iuid);
2081 inode->i_gid = be32_to_cpu(lvb->lvb_igid);
2082 inode->i_mode = be16_to_cpu(lvb->lvb_imode);
2083 inode->i_nlink = be16_to_cpu(lvb->lvb_inlink);
2084 ocfs2_unpack_timespec(&inode->i_atime,
2085 be64_to_cpu(lvb->lvb_iatime_packed));
2086 ocfs2_unpack_timespec(&inode->i_mtime,
2087 be64_to_cpu(lvb->lvb_imtime_packed));
2088 ocfs2_unpack_timespec(&inode->i_ctime,
2089 be64_to_cpu(lvb->lvb_ictime_packed));
2090 spin_unlock(&oi->ip_lock);
2095 static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2096 struct ocfs2_lock_res *lockres)
2098 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2100 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2101 && lvb->lvb_version == OCFS2_LVB_VERSION
2102 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2107 /* Determine whether a lock resource needs to be refreshed, and
2108 * arbitrate who gets to refresh it.
2110 * 0 means no refresh needed.
2112 * > 0 means you need to refresh this and you MUST call
2113 * ocfs2_complete_lock_res_refresh afterwards. */
2114 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2116 unsigned long flags;
2122 spin_lock_irqsave(&lockres->l_lock, flags);
2123 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2124 spin_unlock_irqrestore(&lockres->l_lock, flags);
2128 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2129 spin_unlock_irqrestore(&lockres->l_lock, flags);
2131 ocfs2_wait_on_refreshing_lock(lockres);
2135 /* Ok, I'll be the one to refresh this lock. */
2136 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2137 spin_unlock_irqrestore(&lockres->l_lock, flags);
2145 /* If status is non zero, I'll mark it as not being in refresh
2146 * anymroe, but i won't clear the needs refresh flag. */
2147 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2150 unsigned long flags;
2153 spin_lock_irqsave(&lockres->l_lock, flags);
2154 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2156 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2157 spin_unlock_irqrestore(&lockres->l_lock, flags);
2159 wake_up(&lockres->l_event);
2164 /* may or may not return a bh if it went to disk. */
2165 static int ocfs2_inode_lock_update(struct inode *inode,
2166 struct buffer_head **bh)
2169 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2170 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2171 struct ocfs2_dinode *fe;
2172 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2176 if (ocfs2_mount_local(osb))
2179 spin_lock(&oi->ip_lock);
2180 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2181 mlog(0, "Orphaned inode %llu was deleted while we "
2182 "were waiting on a lock. ip_flags = 0x%x\n",
2183 (unsigned long long)oi->ip_blkno, oi->ip_flags);
2184 spin_unlock(&oi->ip_lock);
2188 spin_unlock(&oi->ip_lock);
2190 if (!ocfs2_should_refresh_lock_res(lockres))
2193 /* This will discard any caching information we might have had
2194 * for the inode metadata. */
2195 ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2197 ocfs2_extent_map_trunc(inode, 0);
2199 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2200 mlog(0, "Trusting LVB on inode %llu\n",
2201 (unsigned long long)oi->ip_blkno);
2202 ocfs2_refresh_inode_from_lvb(inode);
2204 /* Boo, we have to go to disk. */
2205 /* read bh, cast, ocfs2_refresh_inode */
2206 status = ocfs2_read_inode_block(inode, bh);
2211 fe = (struct ocfs2_dinode *) (*bh)->b_data;
2213 /* This is a good chance to make sure we're not
2214 * locking an invalid object. ocfs2_read_inode_block()
2215 * already checked that the inode block is sane.
2217 * We bug on a stale inode here because we checked
2218 * above whether it was wiped from disk. The wiping
2219 * node provides a guarantee that we receive that
2220 * message and can mark the inode before dropping any
2221 * locks associated with it. */
2222 mlog_bug_on_msg(inode->i_generation !=
2223 le32_to_cpu(fe->i_generation),
2224 "Invalid dinode %llu disk generation: %u "
2225 "inode->i_generation: %u\n",
2226 (unsigned long long)oi->ip_blkno,
2227 le32_to_cpu(fe->i_generation),
2228 inode->i_generation);
2229 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2230 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2231 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2232 (unsigned long long)oi->ip_blkno,
2233 (unsigned long long)le64_to_cpu(fe->i_dtime),
2234 le32_to_cpu(fe->i_flags));
2236 ocfs2_refresh_inode(inode, fe);
2237 ocfs2_track_lock_refresh(lockres);
2242 ocfs2_complete_lock_res_refresh(lockres, status);
2248 static int ocfs2_assign_bh(struct inode *inode,
2249 struct buffer_head **ret_bh,
2250 struct buffer_head *passed_bh)
2255 /* Ok, the update went to disk for us, use the
2257 *ret_bh = passed_bh;
2263 status = ocfs2_read_inode_block(inode, ret_bh);
2271 * returns < 0 error if the callback will never be called, otherwise
2272 * the result of the lock will be communicated via the callback.
2274 int ocfs2_inode_lock_full_nested(struct inode *inode,
2275 struct buffer_head **ret_bh,
2280 int status, level, acquired;
2282 struct ocfs2_lock_res *lockres = NULL;
2283 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2284 struct buffer_head *local_bh = NULL;
2290 mlog(0, "inode %llu, take %s META lock\n",
2291 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2292 ex ? "EXMODE" : "PRMODE");
2296 /* We'll allow faking a readonly metadata lock for
2298 if (ocfs2_is_hard_readonly(osb)) {
2304 if (ocfs2_mount_local(osb))
2307 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2308 ocfs2_wait_for_recovery(osb);
2310 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2311 level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2313 if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2314 dlm_flags |= DLM_LKF_NOQUEUE;
2316 status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2317 arg_flags, subclass, _RET_IP_);
2319 if (status != -EAGAIN && status != -EIOCBRETRY)
2324 /* Notify the error cleanup path to drop the cluster lock. */
2327 /* We wait twice because a node may have died while we were in
2328 * the lower dlm layers. The second time though, we've
2329 * committed to owning this lock so we don't allow signals to
2330 * abort the operation. */
2331 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2332 ocfs2_wait_for_recovery(osb);
2336 * We only see this flag if we're being called from
2337 * ocfs2_read_locked_inode(). It means we're locking an inode
2338 * which hasn't been populated yet, so clear the refresh flag
2339 * and let the caller handle it.
2341 if (inode->i_state & I_NEW) {
2344 ocfs2_complete_lock_res_refresh(lockres, 0);
2348 /* This is fun. The caller may want a bh back, or it may
2349 * not. ocfs2_inode_lock_update definitely wants one in, but
2350 * may or may not read one, depending on what's in the
2351 * LVB. The result of all of this is that we've *only* gone to
2352 * disk if we have to, so the complexity is worthwhile. */
2353 status = ocfs2_inode_lock_update(inode, &local_bh);
2355 if (status != -ENOENT)
2361 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2370 if (ret_bh && (*ret_bh)) {
2375 ocfs2_inode_unlock(inode, ex);
2386 * This is working around a lock inversion between tasks acquiring DLM
2387 * locks while holding a page lock and the downconvert thread which
2388 * blocks dlm lock acquiry while acquiring page locks.
2390 * ** These _with_page variantes are only intended to be called from aop
2391 * methods that hold page locks and return a very specific *positive* error
2392 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2394 * The DLM is called such that it returns -EAGAIN if it would have
2395 * blocked waiting for the downconvert thread. In that case we unlock
2396 * our page so the downconvert thread can make progress. Once we've
2397 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2398 * that called us can bubble that back up into the VFS who will then
2399 * immediately retry the aop call.
2401 * We do a blocking lock and immediate unlock before returning, though, so that
2402 * the lock has a great chance of being cached on this node by the time the VFS
2403 * calls back to retry the aop. This has a potential to livelock as nodes
2404 * ping locks back and forth, but that's a risk we're willing to take to avoid
2405 * the lock inversion simply.
2407 int ocfs2_inode_lock_with_page(struct inode *inode,
2408 struct buffer_head **ret_bh,
2414 ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2415 if (ret == -EAGAIN) {
2417 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2418 ocfs2_inode_unlock(inode, ex);
2419 ret = AOP_TRUNCATED_PAGE;
2425 int ocfs2_inode_lock_atime(struct inode *inode,
2426 struct vfsmount *vfsmnt,
2432 ret = ocfs2_inode_lock(inode, NULL, 0);
2439 * If we should update atime, we will get EX lock,
2440 * otherwise we just get PR lock.
2442 if (ocfs2_should_update_atime(inode, vfsmnt)) {
2443 struct buffer_head *bh = NULL;
2445 ocfs2_inode_unlock(inode, 0);
2446 ret = ocfs2_inode_lock(inode, &bh, 1);
2452 if (ocfs2_should_update_atime(inode, vfsmnt))
2453 ocfs2_update_inode_atime(inode, bh);
2463 void ocfs2_inode_unlock(struct inode *inode,
2466 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2467 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2468 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2472 mlog(0, "inode %llu drop %s META lock\n",
2473 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2474 ex ? "EXMODE" : "PRMODE");
2476 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
2477 !ocfs2_mount_local(osb))
2478 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
2483 int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2485 struct ocfs2_lock_res *lockres;
2486 struct ocfs2_orphan_scan_lvb *lvb;
2489 if (ocfs2_is_hard_readonly(osb))
2492 if (ocfs2_mount_local(osb))
2495 lockres = &osb->osb_orphan_scan.os_lockres;
2496 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2500 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2501 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2502 lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2503 *seqno = be32_to_cpu(lvb->lvb_os_seqno);
2505 *seqno = osb->osb_orphan_scan.os_seqno + 1;
2510 void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2512 struct ocfs2_lock_res *lockres;
2513 struct ocfs2_orphan_scan_lvb *lvb;
2515 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2516 lockres = &osb->osb_orphan_scan.os_lockres;
2517 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2518 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2519 lvb->lvb_os_seqno = cpu_to_be32(seqno);
2520 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2524 int ocfs2_super_lock(struct ocfs2_super *osb,
2528 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2529 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2533 if (ocfs2_is_hard_readonly(osb))
2536 if (ocfs2_mount_local(osb))
2539 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2545 /* The super block lock path is really in the best position to
2546 * know when resources covered by the lock need to be
2547 * refreshed, so we do it here. Of course, making sense of
2548 * everything is up to the caller :) */
2549 status = ocfs2_should_refresh_lock_res(lockres);
2555 status = ocfs2_refresh_slot_info(osb);
2557 ocfs2_complete_lock_res_refresh(lockres, status);
2561 ocfs2_track_lock_refresh(lockres);
2568 void ocfs2_super_unlock(struct ocfs2_super *osb,
2571 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2572 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2574 if (!ocfs2_mount_local(osb))
2575 ocfs2_cluster_unlock(osb, lockres, level);
2578 int ocfs2_rename_lock(struct ocfs2_super *osb)
2581 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2583 if (ocfs2_is_hard_readonly(osb))
2586 if (ocfs2_mount_local(osb))
2589 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2596 void ocfs2_rename_unlock(struct ocfs2_super *osb)
2598 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2600 if (!ocfs2_mount_local(osb))
2601 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2604 int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2607 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2609 if (ocfs2_is_hard_readonly(osb))
2612 if (ocfs2_mount_local(osb))
2615 status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2618 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2623 void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2625 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2627 if (!ocfs2_mount_local(osb))
2628 ocfs2_cluster_unlock(osb, lockres,
2629 ex ? LKM_EXMODE : LKM_PRMODE);
2632 int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2635 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2636 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2637 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2641 if (ocfs2_is_hard_readonly(osb))
2644 if (ocfs2_mount_local(osb))
2647 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2654 void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2656 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2657 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2658 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2660 if (!ocfs2_mount_local(osb))
2661 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2664 /* Reference counting of the dlm debug structure. We want this because
2665 * open references on the debug inodes can live on after a mount, so
2666 * we can't rely on the ocfs2_super to always exist. */
2667 static void ocfs2_dlm_debug_free(struct kref *kref)
2669 struct ocfs2_dlm_debug *dlm_debug;
2671 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2676 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2679 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2682 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2684 kref_get(&debug->d_refcnt);
2687 struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2689 struct ocfs2_dlm_debug *dlm_debug;
2691 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2693 mlog_errno(-ENOMEM);
2697 kref_init(&dlm_debug->d_refcnt);
2698 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2699 dlm_debug->d_locking_state = NULL;
2704 /* Access to this is arbitrated for us via seq_file->sem. */
2705 struct ocfs2_dlm_seq_priv {
2706 struct ocfs2_dlm_debug *p_dlm_debug;
2707 struct ocfs2_lock_res p_iter_res;
2708 struct ocfs2_lock_res p_tmp_res;
2711 static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2712 struct ocfs2_dlm_seq_priv *priv)
2714 struct ocfs2_lock_res *iter, *ret = NULL;
2715 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2717 assert_spin_locked(&ocfs2_dlm_tracking_lock);
2719 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2720 /* discover the head of the list */
2721 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2722 mlog(0, "End of list found, %p\n", ret);
2726 /* We track our "dummy" iteration lockres' by a NULL
2728 if (iter->l_ops != NULL) {
2737 static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2739 struct ocfs2_dlm_seq_priv *priv = m->private;
2740 struct ocfs2_lock_res *iter;
2742 spin_lock(&ocfs2_dlm_tracking_lock);
2743 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2745 /* Since lockres' have the lifetime of their container
2746 * (which can be inodes, ocfs2_supers, etc) we want to
2747 * copy this out to a temporary lockres while still
2748 * under the spinlock. Obviously after this we can't
2749 * trust any pointers on the copy returned, but that's
2750 * ok as the information we want isn't typically held
2752 priv->p_tmp_res = *iter;
2753 iter = &priv->p_tmp_res;
2755 spin_unlock(&ocfs2_dlm_tracking_lock);
2760 static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2764 static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2766 struct ocfs2_dlm_seq_priv *priv = m->private;
2767 struct ocfs2_lock_res *iter = v;
2768 struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2770 spin_lock(&ocfs2_dlm_tracking_lock);
2771 iter = ocfs2_dlm_next_res(iter, priv);
2772 list_del_init(&dummy->l_debug_list);
2774 list_add(&dummy->l_debug_list, &iter->l_debug_list);
2775 priv->p_tmp_res = *iter;
2776 iter = &priv->p_tmp_res;
2778 spin_unlock(&ocfs2_dlm_tracking_lock);
2783 /* So that debugfs.ocfs2 can determine which format is being used */
2784 #define OCFS2_DLM_DEBUG_STR_VERSION 2
2785 static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2789 struct ocfs2_lock_res *lockres = v;
2794 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2796 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2797 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2799 (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2801 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2803 seq_printf(m, "%d\t"
2814 lockres->l_unlock_action,
2815 lockres->l_ro_holders,
2816 lockres->l_ex_holders,
2817 lockres->l_requested,
2818 lockres->l_blocking);
2820 /* Dump the raw LVB */
2821 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2822 for(i = 0; i < DLM_LVB_LEN; i++)
2823 seq_printf(m, "0x%x\t", lvb[i]);
2825 #ifdef CONFIG_OCFS2_FS_STATS
2826 # define lock_num_prmode(_l) (_l)->l_lock_num_prmode
2827 # define lock_num_exmode(_l) (_l)->l_lock_num_exmode
2828 # define lock_num_prmode_failed(_l) (_l)->l_lock_num_prmode_failed
2829 # define lock_num_exmode_failed(_l) (_l)->l_lock_num_exmode_failed
2830 # define lock_total_prmode(_l) (_l)->l_lock_total_prmode
2831 # define lock_total_exmode(_l) (_l)->l_lock_total_exmode
2832 # define lock_max_prmode(_l) (_l)->l_lock_max_prmode
2833 # define lock_max_exmode(_l) (_l)->l_lock_max_exmode
2834 # define lock_refresh(_l) (_l)->l_lock_refresh
2836 # define lock_num_prmode(_l) (0ULL)
2837 # define lock_num_exmode(_l) (0ULL)
2838 # define lock_num_prmode_failed(_l) (0)
2839 # define lock_num_exmode_failed(_l) (0)
2840 # define lock_total_prmode(_l) (0ULL)
2841 # define lock_total_exmode(_l) (0ULL)
2842 # define lock_max_prmode(_l) (0)
2843 # define lock_max_exmode(_l) (0)
2844 # define lock_refresh(_l) (0)
2846 /* The following seq_print was added in version 2 of this output */
2847 seq_printf(m, "%llu\t"
2856 lock_num_prmode(lockres),
2857 lock_num_exmode(lockres),
2858 lock_num_prmode_failed(lockres),
2859 lock_num_exmode_failed(lockres),
2860 lock_total_prmode(lockres),
2861 lock_total_exmode(lockres),
2862 lock_max_prmode(lockres),
2863 lock_max_exmode(lockres),
2864 lock_refresh(lockres));
2867 seq_printf(m, "\n");
2871 static const struct seq_operations ocfs2_dlm_seq_ops = {
2872 .start = ocfs2_dlm_seq_start,
2873 .stop = ocfs2_dlm_seq_stop,
2874 .next = ocfs2_dlm_seq_next,
2875 .show = ocfs2_dlm_seq_show,
2878 static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
2880 struct seq_file *seq = (struct seq_file *) file->private_data;
2881 struct ocfs2_dlm_seq_priv *priv = seq->private;
2882 struct ocfs2_lock_res *res = &priv->p_iter_res;
2884 ocfs2_remove_lockres_tracking(res);
2885 ocfs2_put_dlm_debug(priv->p_dlm_debug);
2886 return seq_release_private(inode, file);
2889 static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
2892 struct ocfs2_dlm_seq_priv *priv;
2893 struct seq_file *seq;
2894 struct ocfs2_super *osb;
2896 priv = kzalloc(sizeof(struct ocfs2_dlm_seq_priv), GFP_KERNEL);
2902 osb = inode->i_private;
2903 ocfs2_get_dlm_debug(osb->osb_dlm_debug);
2904 priv->p_dlm_debug = osb->osb_dlm_debug;
2905 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
2907 ret = seq_open(file, &ocfs2_dlm_seq_ops);
2914 seq = (struct seq_file *) file->private_data;
2915 seq->private = priv;
2917 ocfs2_add_lockres_tracking(&priv->p_iter_res,
2924 static const struct file_operations ocfs2_dlm_debug_fops = {
2925 .open = ocfs2_dlm_debug_open,
2926 .release = ocfs2_dlm_debug_release,
2928 .llseek = seq_lseek,
2931 static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
2934 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2936 dlm_debug->d_locking_state = debugfs_create_file("locking_state",
2938 osb->osb_debug_root,
2940 &ocfs2_dlm_debug_fops);
2941 if (!dlm_debug->d_locking_state) {
2944 "Unable to create locking state debugfs file.\n");
2948 ocfs2_get_dlm_debug(dlm_debug);
2953 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
2955 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2958 debugfs_remove(dlm_debug->d_locking_state);
2959 ocfs2_put_dlm_debug(dlm_debug);
2963 int ocfs2_dlm_init(struct ocfs2_super *osb)
2966 struct ocfs2_cluster_connection *conn = NULL;
2970 if (ocfs2_mount_local(osb)) {
2975 status = ocfs2_dlm_init_debug(osb);
2981 /* launch downconvert thread */
2982 osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc");
2983 if (IS_ERR(osb->dc_task)) {
2984 status = PTR_ERR(osb->dc_task);
2985 osb->dc_task = NULL;
2990 /* for now, uuid == domain */
2991 status = ocfs2_cluster_connect(osb->osb_cluster_stack,
2993 strlen(osb->uuid_str),
2994 ocfs2_do_node_down, osb,
3001 status = ocfs2_cluster_this_node(&osb->node_num);
3005 "could not find this host's node number\n");
3006 ocfs2_cluster_disconnect(conn, 0);
3011 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3012 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3013 ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3014 ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3021 ocfs2_dlm_shutdown_debug(osb);
3023 kthread_stop(osb->dc_task);
3030 void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3035 ocfs2_drop_osb_locks(osb);
3038 * Now that we have dropped all locks and ocfs2_dismount_volume()
3039 * has disabled recovery, the DLM won't be talking to us. It's
3040 * safe to tear things down before disconnecting the cluster.
3044 kthread_stop(osb->dc_task);
3045 osb->dc_task = NULL;
3048 ocfs2_lock_res_free(&osb->osb_super_lockres);
3049 ocfs2_lock_res_free(&osb->osb_rename_lockres);
3050 ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3051 ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3053 ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3056 ocfs2_dlm_shutdown_debug(osb);
3061 static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
3063 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
3064 unsigned long flags;
3068 mlog(0, "UNLOCK AST called on lock %s, action = %d\n", lockres->l_name,
3069 lockres->l_unlock_action);
3071 spin_lock_irqsave(&lockres->l_lock, flags);
3073 mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
3074 "unlock_action %d\n", error, lockres->l_name,
3075 lockres->l_unlock_action);
3076 spin_unlock_irqrestore(&lockres->l_lock, flags);
3081 switch(lockres->l_unlock_action) {
3082 case OCFS2_UNLOCK_CANCEL_CONVERT:
3083 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
3084 lockres->l_action = OCFS2_AST_INVALID;
3085 /* Downconvert thread may have requeued this lock, we
3086 * need to wake it. */
3087 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3088 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
3090 case OCFS2_UNLOCK_DROP_LOCK:
3091 lockres->l_level = DLM_LOCK_IV;
3097 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
3098 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
3099 wake_up(&lockres->l_event);
3100 spin_unlock_irqrestore(&lockres->l_lock, flags);
3105 static int ocfs2_drop_lock(struct ocfs2_super *osb,
3106 struct ocfs2_lock_res *lockres)
3109 unsigned long flags;
3112 /* We didn't get anywhere near actually using this lockres. */
3113 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3116 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3117 lkm_flags |= DLM_LKF_VALBLK;
3119 spin_lock_irqsave(&lockres->l_lock, flags);
3121 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3122 "lockres %s, flags 0x%lx\n",
3123 lockres->l_name, lockres->l_flags);
3125 while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3126 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3127 "%u, unlock_action = %u\n",
3128 lockres->l_name, lockres->l_flags, lockres->l_action,
3129 lockres->l_unlock_action);
3131 spin_unlock_irqrestore(&lockres->l_lock, flags);
3133 /* XXX: Today we just wait on any busy
3134 * locks... Perhaps we need to cancel converts in the
3136 ocfs2_wait_on_busy_lock(lockres);
3138 spin_lock_irqsave(&lockres->l_lock, flags);
3141 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3142 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3143 lockres->l_level == DLM_LOCK_EX &&
3144 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3145 lockres->l_ops->set_lvb(lockres);
3148 if (lockres->l_flags & OCFS2_LOCK_BUSY)
3149 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3151 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3152 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3154 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3155 spin_unlock_irqrestore(&lockres->l_lock, flags);
3159 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3161 /* make sure we never get here while waiting for an ast to
3163 BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3165 /* is this necessary? */
3166 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3167 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3168 spin_unlock_irqrestore(&lockres->l_lock, flags);
3170 mlog(0, "lock %s\n", lockres->l_name);
3172 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3174 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3175 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3176 ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3179 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3182 ocfs2_wait_on_busy_lock(lockres);
3188 /* Mark the lockres as being dropped. It will no longer be
3189 * queued if blocking, but we still may have to wait on it
3190 * being dequeued from the downconvert thread before we can consider
3193 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3194 void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res *lockres)
3197 struct ocfs2_mask_waiter mw;
3198 unsigned long flags;
3200 ocfs2_init_mask_waiter(&mw);
3202 spin_lock_irqsave(&lockres->l_lock, flags);
3203 lockres->l_flags |= OCFS2_LOCK_FREEING;
3204 while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3205 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3206 spin_unlock_irqrestore(&lockres->l_lock, flags);
3208 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3210 status = ocfs2_wait_for_mask(&mw);
3214 spin_lock_irqsave(&lockres->l_lock, flags);
3216 spin_unlock_irqrestore(&lockres->l_lock, flags);
3219 void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3220 struct ocfs2_lock_res *lockres)
3224 ocfs2_mark_lockres_freeing(lockres);
3225 ret = ocfs2_drop_lock(osb, lockres);
3230 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3232 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3233 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3234 ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3235 ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3238 int ocfs2_drop_inode_locks(struct inode *inode)
3244 /* No need to call ocfs2_mark_lockres_freeing here -
3245 * ocfs2_clear_inode has done it for us. */
3247 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3248 &OCFS2_I(inode)->ip_open_lockres);
3254 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3255 &OCFS2_I(inode)->ip_inode_lockres);
3258 if (err < 0 && !status)
3261 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3262 &OCFS2_I(inode)->ip_rw_lockres);
3265 if (err < 0 && !status)
3272 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3275 assert_spin_locked(&lockres->l_lock);
3277 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3279 if (lockres->l_level <= new_level) {
3280 mlog(ML_ERROR, "lockres->l_level (%d) <= new_level (%d)\n",
3281 lockres->l_level, new_level);
3285 mlog(0, "lock %s, new_level = %d, l_blocking = %d\n",
3286 lockres->l_name, new_level, lockres->l_blocking);
3288 lockres->l_action = OCFS2_AST_DOWNCONVERT;
3289 lockres->l_requested = new_level;
3290 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3291 return lockres_set_pending(lockres);
3294 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3295 struct ocfs2_lock_res *lockres,
3298 unsigned int generation)
3301 u32 dlm_flags = DLM_LKF_CONVERT;
3306 dlm_flags |= DLM_LKF_VALBLK;
3308 ret = ocfs2_dlm_lock(osb->cconn,
3313 OCFS2_LOCK_ID_MAX_LEN - 1);
3314 lockres_clear_pending(lockres, generation, osb);
3316 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3317 ocfs2_recover_from_dlm_error(lockres, 1);
3327 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3328 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3329 struct ocfs2_lock_res *lockres)
3331 assert_spin_locked(&lockres->l_lock);
3334 mlog(0, "lock %s\n", lockres->l_name);
3336 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3337 /* If we're already trying to cancel a lock conversion
3338 * then just drop the spinlock and allow the caller to
3339 * requeue this lock. */
3341 mlog(0, "Lockres %s, skip convert\n", lockres->l_name);
3345 /* were we in a convert when we got the bast fire? */
3346 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3347 lockres->l_action != OCFS2_AST_DOWNCONVERT);
3348 /* set things up for the unlockast to know to just
3349 * clear out the ast_action and unset busy, etc. */
3350 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3352 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3353 "lock %s, invalid flags: 0x%lx\n",
3354 lockres->l_name, lockres->l_flags);
3359 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3360 struct ocfs2_lock_res *lockres)
3365 mlog(0, "lock %s\n", lockres->l_name);
3367 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3370 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3371 ocfs2_recover_from_dlm_error(lockres, 0);
3374 mlog(0, "lock %s return from ocfs2_dlm_unlock\n", lockres->l_name);
3380 static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3381 struct ocfs2_lock_res *lockres,
3382 struct ocfs2_unblock_ctl *ctl)
3384 unsigned long flags;
3394 spin_lock_irqsave(&lockres->l_lock, flags);
3398 * Is it still blocking? If not, we have no more work to do.
3400 if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3401 BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3402 spin_unlock_irqrestore(&lockres->l_lock, flags);
3407 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3409 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3410 * exists entirely for one reason - another thread has set
3411 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3413 * If we do ocfs2_cancel_convert() before the other thread
3414 * calls dlm_lock(), our cancel will do nothing. We will
3415 * get no ast, and we will have no way of knowing the
3416 * cancel failed. Meanwhile, the other thread will call
3417 * into dlm_lock() and wait...forever.
3419 * Why forever? Because another node has asked for the
3420 * lock first; that's why we're here in unblock_lock().
3422 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3423 * set, we just requeue the unblock. Only when the other
3424 * thread has called dlm_lock() and cleared PENDING will
3425 * we then cancel their request.
3427 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3428 * at the same time they set OCFS2_DLM_BUSY. They must
3429 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3431 if (lockres->l_flags & OCFS2_LOCK_PENDING)
3435 ret = ocfs2_prepare_cancel_convert(osb, lockres);
3436 spin_unlock_irqrestore(&lockres->l_lock, flags);
3438 ret = ocfs2_cancel_convert(osb, lockres);
3446 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3447 * set when the ast is received for an upconvert just before the
3448 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3449 * on the heels of the ast, we want to delay the downconvert just
3450 * enough to allow the up requestor to do its task. Because this
3451 * lock is in the blocked queue, the lock will be downconverted
3452 * as soon as the requestor is done with the lock.
3454 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3458 * How can we block and yet be at NL? We were trying to upconvert
3459 * from NL and got canceled. The code comes back here, and now
3460 * we notice and clear BLOCKING.
3462 if (lockres->l_level == DLM_LOCK_NL) {
3463 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3464 lockres->l_blocking = DLM_LOCK_NL;
3465 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3466 spin_unlock_irqrestore(&lockres->l_lock, flags);
3470 /* if we're blocking an exclusive and we have *any* holders,
3472 if ((lockres->l_blocking == DLM_LOCK_EX)
3473 && (lockres->l_ex_holders || lockres->l_ro_holders))
3476 /* If it's a PR we're blocking, then only
3477 * requeue if we've got any EX holders */
3478 if (lockres->l_blocking == DLM_LOCK_PR &&
3479 lockres->l_ex_holders)
3483 * Can we get a lock in this state if the holder counts are
3484 * zero? The meta data unblock code used to check this.
3486 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3487 && (lockres->l_flags & OCFS2_LOCK_REFRESHING))
3490 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3492 if (lockres->l_ops->check_downconvert
3493 && !lockres->l_ops->check_downconvert(lockres, new_level))
3496 /* If we get here, then we know that there are no more
3497 * incompatible holders (and anyone asking for an incompatible
3498 * lock is blocked). We can now downconvert the lock */
3499 if (!lockres->l_ops->downconvert_worker)
3502 /* Some lockres types want to do a bit of work before
3503 * downconverting a lock. Allow that here. The worker function
3504 * may sleep, so we save off a copy of what we're blocking as
3505 * it may change while we're not holding the spin lock. */
3506 blocking = lockres->l_blocking;
3507 level = lockres->l_level;
3508 spin_unlock_irqrestore(&lockres->l_lock, flags);
3510 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3512 if (ctl->unblock_action == UNBLOCK_STOP_POST)
3515 spin_lock_irqsave(&lockres->l_lock, flags);
3516 if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3517 /* If this changed underneath us, then we can't drop
3525 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3526 if (lockres->l_level == DLM_LOCK_EX)
3530 * We only set the lvb if the lock has been fully
3531 * refreshed - otherwise we risk setting stale
3532 * data. Otherwise, there's no need to actually clear
3533 * out the lvb here as it's value is still valid.
3535 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3536 lockres->l_ops->set_lvb(lockres);
3539 gen = ocfs2_prepare_downconvert(lockres, new_level);
3540 spin_unlock_irqrestore(&lockres->l_lock, flags);
3541 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3549 spin_unlock_irqrestore(&lockres->l_lock, flags);
3556 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3559 struct inode *inode;
3560 struct address_space *mapping;
3562 inode = ocfs2_lock_res_inode(lockres);
3563 mapping = inode->i_mapping;
3565 if (!S_ISREG(inode->i_mode))
3569 * We need this before the filemap_fdatawrite() so that it can
3570 * transfer the dirty bit from the PTE to the
3571 * page. Unfortunately this means that even for EX->PR
3572 * downconverts, we'll lose our mappings and have to build
3575 unmap_mapping_range(mapping, 0, 0, 0);
3577 if (filemap_fdatawrite(mapping)) {
3578 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3579 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3581 sync_mapping_buffers(mapping);
3582 if (blocking == DLM_LOCK_EX) {
3583 truncate_inode_pages(mapping, 0);
3585 /* We only need to wait on the I/O if we're not also
3586 * truncating pages because truncate_inode_pages waits
3587 * for us above. We don't truncate pages if we're
3588 * blocking anything < EXMODE because we want to keep
3589 * them around in that case. */
3590 filemap_fdatawait(mapping);
3594 return UNBLOCK_CONTINUE;
3597 static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3598 struct ocfs2_lock_res *lockres,
3601 int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3603 BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3604 BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3609 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3613 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3616 struct inode *inode = ocfs2_lock_res_inode(lockres);
3618 return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3621 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3623 struct inode *inode = ocfs2_lock_res_inode(lockres);
3625 __ocfs2_stuff_meta_lvb(inode);
3629 * Does the final reference drop on our dentry lock. Right now this
3630 * happens in the downconvert thread, but we could choose to simplify the
3631 * dlmglue API and push these off to the ocfs2_wq in the future.
3633 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3634 struct ocfs2_lock_res *lockres)
3636 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3637 ocfs2_dentry_lock_put(osb, dl);
3641 * d_delete() matching dentries before the lock downconvert.
3643 * At this point, any process waiting to destroy the
3644 * dentry_lock due to last ref count is stopped by the
3645 * OCFS2_LOCK_QUEUED flag.
3647 * We have two potential problems
3649 * 1) If we do the last reference drop on our dentry_lock (via dput)
3650 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3651 * the downconvert to finish. Instead we take an elevated
3652 * reference and push the drop until after we've completed our
3653 * unblock processing.
3655 * 2) There might be another process with a final reference,
3656 * waiting on us to finish processing. If this is the case, we
3657 * detect it and exit out - there's no more dentries anyway.
3659 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3662 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3663 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3664 struct dentry *dentry;
3665 unsigned long flags;
3669 * This node is blocking another node from getting a read
3670 * lock. This happens when we've renamed within a
3671 * directory. We've forced the other nodes to d_delete(), but
3672 * we never actually dropped our lock because it's still
3673 * valid. The downconvert code will retain a PR for this node,
3674 * so there's no further work to do.
3676 if (blocking == DLM_LOCK_PR)
3677 return UNBLOCK_CONTINUE;
3680 * Mark this inode as potentially orphaned. The code in
3681 * ocfs2_delete_inode() will figure out whether it actually
3682 * needs to be freed or not.
3684 spin_lock(&oi->ip_lock);
3685 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3686 spin_unlock(&oi->ip_lock);
3689 * Yuck. We need to make sure however that the check of
3690 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3691 * respect to a reference decrement or the setting of that
3694 spin_lock_irqsave(&lockres->l_lock, flags);
3695 spin_lock(&dentry_attach_lock);
3696 if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3701 spin_unlock(&dentry_attach_lock);
3702 spin_unlock_irqrestore(&lockres->l_lock, flags);
3704 mlog(0, "extra_ref = %d\n", extra_ref);
3707 * We have a process waiting on us in ocfs2_dentry_iput(),
3708 * which means we can't have any more outstanding
3709 * aliases. There's no need to do any more work.
3712 return UNBLOCK_CONTINUE;
3714 spin_lock(&dentry_attach_lock);
3716 dentry = ocfs2_find_local_alias(dl->dl_inode,
3717 dl->dl_parent_blkno, 1);
3720 spin_unlock(&dentry_attach_lock);
3722 mlog(0, "d_delete(%.*s);\n", dentry->d_name.len,
3723 dentry->d_name.name);
3726 * The following dcache calls may do an
3727 * iput(). Normally we don't want that from the
3728 * downconverting thread, but in this case it's ok
3729 * because the requesting node already has an
3730 * exclusive lock on the inode, so it can't be queued
3731 * for a downconvert.
3736 spin_lock(&dentry_attach_lock);
3738 spin_unlock(&dentry_attach_lock);
3741 * If we are the last holder of this dentry lock, there is no
3742 * reason to downconvert so skip straight to the unlock.
3744 if (dl->dl_count == 1)
3745 return UNBLOCK_STOP_POST;
3747 return UNBLOCK_CONTINUE_POST;
3750 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
3753 struct ocfs2_refcount_tree *tree =
3754 ocfs2_lock_res_refcount_tree(lockres);
3756 return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
3759 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
3762 struct ocfs2_refcount_tree *tree =
3763 ocfs2_lock_res_refcount_tree(lockres);
3765 ocfs2_metadata_cache_purge(&tree->rf_ci);
3767 return UNBLOCK_CONTINUE;
3770 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
3772 struct ocfs2_qinfo_lvb *lvb;
3773 struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
3774 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3775 oinfo->dqi_gi.dqi_type);
3779 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3780 lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
3781 lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
3782 lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
3783 lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
3784 lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
3785 lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
3786 lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
3791 void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3793 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3794 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3795 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3798 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3799 ocfs2_cluster_unlock(osb, lockres, level);
3803 static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
3805 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3806 oinfo->dqi_gi.dqi_type);
3807 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3808 struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3809 struct buffer_head *bh = NULL;
3810 struct ocfs2_global_disk_dqinfo *gdinfo;
3813 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
3814 lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
3815 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
3816 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
3817 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
3818 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
3819 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
3820 oinfo->dqi_gi.dqi_free_entry =
3821 be32_to_cpu(lvb->lvb_free_entry);
3823 status = ocfs2_read_quota_block(oinfo->dqi_gqinode, 0, &bh);
3828 gdinfo = (struct ocfs2_global_disk_dqinfo *)
3829 (bh->b_data + OCFS2_GLOBAL_INFO_OFF);
3830 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
3831 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
3832 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
3833 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
3834 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
3835 oinfo->dqi_gi.dqi_free_entry =
3836 le32_to_cpu(gdinfo->dqi_free_entry);
3838 ocfs2_track_lock_refresh(lockres);
3845 /* Lock quota info, this function expects at least shared lock on the quota file
3846 * so that we can safely refresh quota info from disk. */
3847 int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3849 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3850 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3851 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3856 /* On RO devices, locking really isn't needed... */
3857 if (ocfs2_is_hard_readonly(osb)) {
3862 if (ocfs2_mount_local(osb))
3865 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3870 if (!ocfs2_should_refresh_lock_res(lockres))
3872 /* OK, we have the lock but we need to refresh the quota info */
3873 status = ocfs2_refresh_qinfo(oinfo);
3875 ocfs2_qinfo_unlock(oinfo, ex);
3876 ocfs2_complete_lock_res_refresh(lockres, status);
3882 int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
3885 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3886 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3887 struct ocfs2_super *osb = lockres->l_priv;
3890 if (ocfs2_is_hard_readonly(osb))
3893 if (ocfs2_mount_local(osb))
3896 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3903 void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
3905 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3906 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3907 struct ocfs2_super *osb = lockres->l_priv;
3909 if (!ocfs2_mount_local(osb))
3910 ocfs2_cluster_unlock(osb, lockres, level);
3914 * This is the filesystem locking protocol. It provides the lock handling
3915 * hooks for the underlying DLM. It has a maximum version number.
3916 * The version number allows interoperability with systems running at
3917 * the same major number and an equal or smaller minor number.
3919 * Whenever the filesystem does new things with locks (adds or removes a
3920 * lock, orders them differently, does different things underneath a lock),
3921 * the version must be changed. The protocol is negotiated when joining
3922 * the dlm domain. A node may join the domain if its major version is
3923 * identical to all other nodes and its minor version is greater than
3924 * or equal to all other nodes. When its minor version is greater than
3925 * the other nodes, it will run at the minor version specified by the
3928 * If a locking change is made that will not be compatible with older
3929 * versions, the major number must be increased and the minor version set
3930 * to zero. If a change merely adds a behavior that can be disabled when
3931 * speaking to older versions, the minor version must be increased. If a
3932 * change adds a fully backwards compatible change (eg, LVB changes that
3933 * are just ignored by older versions), the version does not need to be
3936 static struct ocfs2_locking_protocol lproto = {
3938 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
3939 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
3941 .lp_lock_ast = ocfs2_locking_ast,
3942 .lp_blocking_ast = ocfs2_blocking_ast,
3943 .lp_unlock_ast = ocfs2_unlock_ast,
3946 void ocfs2_set_locking_protocol(void)
3948 ocfs2_stack_glue_set_locking_protocol(&lproto);
3952 static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3953 struct ocfs2_lock_res *lockres)
3956 struct ocfs2_unblock_ctl ctl = {0, 0,};
3957 unsigned long flags;
3959 /* Our reference to the lockres in this function can be
3960 * considered valid until we remove the OCFS2_LOCK_QUEUED
3966 BUG_ON(!lockres->l_ops);
3968 mlog(0, "lockres %s blocked.\n", lockres->l_name);
3970 /* Detect whether a lock has been marked as going away while
3971 * the downconvert thread was processing other things. A lock can
3972 * still be marked with OCFS2_LOCK_FREEING after this check,
3973 * but short circuiting here will still save us some
3975 spin_lock_irqsave(&lockres->l_lock, flags);
3976 if (lockres->l_flags & OCFS2_LOCK_FREEING)
3978 spin_unlock_irqrestore(&lockres->l_lock, flags);
3980 status = ocfs2_unblock_lock(osb, lockres, &ctl);
3984 spin_lock_irqsave(&lockres->l_lock, flags);
3986 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
3987 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
3989 ocfs2_schedule_blocked_lock(osb, lockres);
3991 mlog(0, "lockres %s, requeue = %s.\n", lockres->l_name,
3992 ctl.requeue ? "yes" : "no");
3993 spin_unlock_irqrestore(&lockres->l_lock, flags);
3995 if (ctl.unblock_action != UNBLOCK_CONTINUE
3996 && lockres->l_ops->post_unlock)
3997 lockres->l_ops->post_unlock(osb, lockres);
4002 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
4003 struct ocfs2_lock_res *lockres)
4007 assert_spin_locked(&lockres->l_lock);
4009 if (lockres->l_flags & OCFS2_LOCK_FREEING) {
4010 /* Do not schedule a lock for downconvert when it's on
4011 * the way to destruction - any nodes wanting access
4012 * to the resource will get it soon. */
4013 mlog(0, "Lockres %s won't be scheduled: flags 0x%lx\n",
4014 lockres->l_name, lockres->l_flags);
4018 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4020 spin_lock(&osb->dc_task_lock);
4021 if (list_empty(&lockres->l_blocked_list)) {
4022 list_add_tail(&lockres->l_blocked_list,
4023 &osb->blocked_lock_list);
4024 osb->blocked_lock_count++;
4026 spin_unlock(&osb->dc_task_lock);
4031 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4033 unsigned long processed;
4034 struct ocfs2_lock_res *lockres;
4038 spin_lock(&osb->dc_task_lock);
4039 /* grab this early so we know to try again if a state change and
4040 * wake happens part-way through our work */
4041 osb->dc_work_sequence = osb->dc_wake_sequence;
4043 processed = osb->blocked_lock_count;
4045 BUG_ON(list_empty(&osb->blocked_lock_list));
4047 lockres = list_entry(osb->blocked_lock_list.next,
4048 struct ocfs2_lock_res, l_blocked_list);
4049 list_del_init(&lockres->l_blocked_list);
4050 osb->blocked_lock_count--;
4051 spin_unlock(&osb->dc_task_lock);
4056 ocfs2_process_blocked_lock(osb, lockres);
4058 spin_lock(&osb->dc_task_lock);
4060 spin_unlock(&osb->dc_task_lock);
4065 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4069 spin_lock(&osb->dc_task_lock);
4070 if (list_empty(&osb->blocked_lock_list))
4073 spin_unlock(&osb->dc_task_lock);
4077 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4079 int should_wake = 0;
4081 spin_lock(&osb->dc_task_lock);
4082 if (osb->dc_work_sequence != osb->dc_wake_sequence)
4084 spin_unlock(&osb->dc_task_lock);
4089 static int ocfs2_downconvert_thread(void *arg)
4092 struct ocfs2_super *osb = arg;
4094 /* only quit once we've been asked to stop and there is no more
4096 while (!(kthread_should_stop() &&
4097 ocfs2_downconvert_thread_lists_empty(osb))) {
4099 wait_event_interruptible(osb->dc_event,
4100 ocfs2_downconvert_thread_should_wake(osb) ||
4101 kthread_should_stop());
4103 mlog(0, "downconvert_thread: awoken\n");
4105 ocfs2_downconvert_thread_do_work(osb);
4108 osb->dc_task = NULL;
4112 void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4114 spin_lock(&osb->dc_task_lock);
4115 /* make sure the voting thread gets a swipe at whatever changes
4116 * the caller may have made to the voting state */
4117 osb->dc_wake_sequence++;
4118 spin_unlock(&osb->dc_task_lock);
4119 wake_up(&osb->dc_event);