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 ktime_t 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,
400 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
402 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
403 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
404 (long long)blkno, generation);
406 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
408 mlog(0, "built lock resource with name: %s\n", name);
411 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
413 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
414 struct ocfs2_dlm_debug *dlm_debug)
416 mlog(0, "Add tracking for lockres %s\n", res->l_name);
418 spin_lock(&ocfs2_dlm_tracking_lock);
419 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
420 spin_unlock(&ocfs2_dlm_tracking_lock);
423 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
425 spin_lock(&ocfs2_dlm_tracking_lock);
426 if (!list_empty(&res->l_debug_list))
427 list_del_init(&res->l_debug_list);
428 spin_unlock(&ocfs2_dlm_tracking_lock);
431 #ifdef CONFIG_OCFS2_FS_STATS
432 static void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
434 res->l_lock_refresh = 0;
435 memset(&res->l_lock_prmode, 0, sizeof(struct ocfs2_lock_stats));
436 memset(&res->l_lock_exmode, 0, sizeof(struct ocfs2_lock_stats));
439 static void ocfs2_update_lock_stats(struct ocfs2_lock_res *res, int level,
440 struct ocfs2_mask_waiter *mw, int ret)
444 struct ocfs2_lock_stats *stats;
446 if (level == LKM_PRMODE)
447 stats = &res->l_lock_prmode;
448 else if (level == LKM_EXMODE)
449 stats = &res->l_lock_exmode;
453 kt = ktime_sub(ktime_get(), mw->mw_lock_start);
454 usec = ktime_to_us(kt);
457 stats->ls_total += ktime_to_ns(kt);
459 if (unlikely(stats->ls_gets == 0)) {
461 stats->ls_total = ktime_to_ns(kt);
464 if (stats->ls_max < usec)
465 stats->ls_max = usec;
471 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
473 lockres->l_lock_refresh++;
476 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
478 mw->mw_lock_start = ktime_get();
481 static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res *res)
484 static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res *res,
485 int level, struct ocfs2_mask_waiter *mw, int ret)
488 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res *lockres)
491 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter *mw)
496 static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
497 struct ocfs2_lock_res *res,
498 enum ocfs2_lock_type type,
499 struct ocfs2_lock_res_ops *ops,
506 res->l_level = DLM_LOCK_IV;
507 res->l_requested = DLM_LOCK_IV;
508 res->l_blocking = DLM_LOCK_IV;
509 res->l_action = OCFS2_AST_INVALID;
510 res->l_unlock_action = OCFS2_UNLOCK_INVALID;
512 res->l_flags = OCFS2_LOCK_INITIALIZED;
514 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
516 ocfs2_init_lock_stats(res);
517 #ifdef CONFIG_DEBUG_LOCK_ALLOC
518 if (type != OCFS2_LOCK_TYPE_OPEN)
519 lockdep_init_map(&res->l_lockdep_map, ocfs2_lock_type_strings[type],
520 &lockdep_keys[type], 0);
522 res->l_lockdep_map.key = NULL;
526 void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
528 /* This also clears out the lock status block */
529 memset(res, 0, sizeof(struct ocfs2_lock_res));
530 spin_lock_init(&res->l_lock);
531 init_waitqueue_head(&res->l_event);
532 INIT_LIST_HEAD(&res->l_blocked_list);
533 INIT_LIST_HEAD(&res->l_mask_waiters);
536 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
537 enum ocfs2_lock_type type,
538 unsigned int generation,
541 struct ocfs2_lock_res_ops *ops;
544 case OCFS2_LOCK_TYPE_RW:
545 ops = &ocfs2_inode_rw_lops;
547 case OCFS2_LOCK_TYPE_META:
548 ops = &ocfs2_inode_inode_lops;
550 case OCFS2_LOCK_TYPE_OPEN:
551 ops = &ocfs2_inode_open_lops;
554 mlog_bug_on_msg(1, "type: %d\n", type);
555 ops = NULL; /* thanks, gcc */
559 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
560 generation, res->l_name);
561 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
564 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
566 struct inode *inode = ocfs2_lock_res_inode(lockres);
568 return OCFS2_SB(inode->i_sb);
571 static struct ocfs2_super *ocfs2_get_qinfo_osb(struct ocfs2_lock_res *lockres)
573 struct ocfs2_mem_dqinfo *info = lockres->l_priv;
575 return OCFS2_SB(info->dqi_gi.dqi_sb);
578 static struct ocfs2_super *ocfs2_get_file_osb(struct ocfs2_lock_res *lockres)
580 struct ocfs2_file_private *fp = lockres->l_priv;
582 return OCFS2_SB(fp->fp_file->f_mapping->host->i_sb);
585 static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
587 __be64 inode_blkno_be;
589 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
592 return be64_to_cpu(inode_blkno_be);
595 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
597 struct ocfs2_dentry_lock *dl = lockres->l_priv;
599 return OCFS2_SB(dl->dl_inode->i_sb);
602 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
603 u64 parent, struct inode *inode)
606 u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
607 __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
608 struct ocfs2_lock_res *lockres = &dl->dl_lockres;
610 ocfs2_lock_res_init_once(lockres);
613 * Unfortunately, the standard lock naming scheme won't work
614 * here because we have two 16 byte values to use. Instead,
615 * we'll stuff the inode number as a binary value. We still
616 * want error prints to show something without garbling the
617 * display, so drop a null byte in there before the inode
618 * number. A future version of OCFS2 will likely use all
619 * binary lock names. The stringified names have been a
620 * tremendous aid in debugging, but now that the debugfs
621 * interface exists, we can mangle things there if need be.
623 * NOTE: We also drop the standard "pad" value (the total lock
624 * name size stays the same though - the last part is all
625 * zeros due to the memset in ocfs2_lock_res_init_once()
627 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
629 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
632 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
634 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
637 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
638 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
642 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
643 struct ocfs2_super *osb)
645 /* Superblock lockres doesn't come from a slab so we call init
646 * once on it manually. */
647 ocfs2_lock_res_init_once(res);
648 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
650 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
651 &ocfs2_super_lops, osb);
654 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
655 struct ocfs2_super *osb)
657 /* Rename lockres doesn't come from a slab so we call init
658 * once on it manually. */
659 ocfs2_lock_res_init_once(res);
660 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
661 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
662 &ocfs2_rename_lops, osb);
665 static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res *res,
666 struct ocfs2_super *osb)
668 /* nfs_sync lockres doesn't come from a slab so we call init
669 * once on it manually. */
670 ocfs2_lock_res_init_once(res);
671 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC, 0, 0, res->l_name);
672 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_NFS_SYNC,
673 &ocfs2_nfs_sync_lops, osb);
676 static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res *res,
677 struct ocfs2_super *osb)
679 ocfs2_lock_res_init_once(res);
680 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN, 0, 0, res->l_name);
681 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_ORPHAN_SCAN,
682 &ocfs2_orphan_scan_lops, osb);
685 void ocfs2_file_lock_res_init(struct ocfs2_lock_res *lockres,
686 struct ocfs2_file_private *fp)
688 struct inode *inode = fp->fp_file->f_mapping->host;
689 struct ocfs2_inode_info *oi = OCFS2_I(inode);
691 ocfs2_lock_res_init_once(lockres);
692 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK, oi->ip_blkno,
693 inode->i_generation, lockres->l_name);
694 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
695 OCFS2_LOCK_TYPE_FLOCK, &ocfs2_flock_lops,
697 lockres->l_flags |= OCFS2_LOCK_NOCACHE;
700 void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res *lockres,
701 struct ocfs2_mem_dqinfo *info)
703 ocfs2_lock_res_init_once(lockres);
704 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO, info->dqi_gi.dqi_type,
706 ocfs2_lock_res_init_common(OCFS2_SB(info->dqi_gi.dqi_sb), lockres,
707 OCFS2_LOCK_TYPE_QINFO, &ocfs2_qinfo_lops,
711 void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res *lockres,
712 struct ocfs2_super *osb, u64 ref_blkno,
713 unsigned int generation)
715 ocfs2_lock_res_init_once(lockres);
716 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT, ref_blkno,
717 generation, lockres->l_name);
718 ocfs2_lock_res_init_common(osb, lockres, OCFS2_LOCK_TYPE_REFCOUNT,
719 &ocfs2_refcount_block_lops, osb);
722 void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
724 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
727 ocfs2_remove_lockres_tracking(res);
729 mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
730 "Lockres %s is on the blocked list\n",
732 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
733 "Lockres %s has mask waiters pending\n",
735 mlog_bug_on_msg(spin_is_locked(&res->l_lock),
736 "Lockres %s is locked\n",
738 mlog_bug_on_msg(res->l_ro_holders,
739 "Lockres %s has %u ro holders\n",
740 res->l_name, res->l_ro_holders);
741 mlog_bug_on_msg(res->l_ex_holders,
742 "Lockres %s has %u ex holders\n",
743 res->l_name, res->l_ex_holders);
745 /* Need to clear out the lock status block for the dlm */
746 memset(&res->l_lksb, 0, sizeof(res->l_lksb));
751 static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
758 lockres->l_ex_holders++;
761 lockres->l_ro_holders++;
768 static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
775 BUG_ON(!lockres->l_ex_holders);
776 lockres->l_ex_holders--;
779 BUG_ON(!lockres->l_ro_holders);
780 lockres->l_ro_holders--;
787 /* WARNING: This function lives in a world where the only three lock
788 * levels are EX, PR, and NL. It *will* have to be adjusted when more
789 * lock types are added. */
790 static inline int ocfs2_highest_compat_lock_level(int level)
792 int new_level = DLM_LOCK_EX;
794 if (level == DLM_LOCK_EX)
795 new_level = DLM_LOCK_NL;
796 else if (level == DLM_LOCK_PR)
797 new_level = DLM_LOCK_PR;
801 static void lockres_set_flags(struct ocfs2_lock_res *lockres,
802 unsigned long newflags)
804 struct ocfs2_mask_waiter *mw, *tmp;
806 assert_spin_locked(&lockres->l_lock);
808 lockres->l_flags = newflags;
810 list_for_each_entry_safe(mw, tmp, &lockres->l_mask_waiters, mw_item) {
811 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
814 list_del_init(&mw->mw_item);
816 complete(&mw->mw_complete);
819 static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
821 lockres_set_flags(lockres, lockres->l_flags | or);
823 static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
826 lockres_set_flags(lockres, lockres->l_flags & ~clear);
829 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
831 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
832 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
833 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
834 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
836 lockres->l_level = lockres->l_requested;
837 if (lockres->l_level <=
838 ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
839 lockres->l_blocking = DLM_LOCK_NL;
840 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
842 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
845 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
847 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
848 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
850 /* Convert from RO to EX doesn't really need anything as our
851 * information is already up to data. Convert from NL to
852 * *anything* however should mark ourselves as needing an
854 if (lockres->l_level == DLM_LOCK_NL &&
855 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
856 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
858 lockres->l_level = lockres->l_requested;
861 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
862 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
863 * downconverting the lock before the upconvert has fully completed.
864 * Do not prevent the dc thread from downconverting if NONBLOCK lock
865 * had already returned.
867 if (!(lockres->l_flags & OCFS2_LOCK_NONBLOCK_FINISHED))
868 lockres_or_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
870 lockres_clear_flags(lockres, OCFS2_LOCK_NONBLOCK_FINISHED);
872 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
875 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
877 BUG_ON((!(lockres->l_flags & OCFS2_LOCK_BUSY)));
878 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
880 if (lockres->l_requested > DLM_LOCK_NL &&
881 !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
882 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
883 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
885 lockres->l_level = lockres->l_requested;
886 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
887 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
890 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
893 int needs_downconvert = 0;
895 assert_spin_locked(&lockres->l_lock);
897 if (level > lockres->l_blocking) {
898 /* only schedule a downconvert if we haven't already scheduled
899 * one that goes low enough to satisfy the level we're
900 * blocking. this also catches the case where we get
902 if (ocfs2_highest_compat_lock_level(level) <
903 ocfs2_highest_compat_lock_level(lockres->l_blocking))
904 needs_downconvert = 1;
906 lockres->l_blocking = level;
909 mlog(ML_BASTS, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
910 lockres->l_name, level, lockres->l_level, lockres->l_blocking,
913 if (needs_downconvert)
914 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
915 mlog(0, "needs_downconvert = %d\n", needs_downconvert);
916 return needs_downconvert;
920 * OCFS2_LOCK_PENDING and l_pending_gen.
922 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
923 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
924 * for more details on the race.
926 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
927 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
928 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
929 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
930 * the caller is going to try to clear PENDING again. If nothing else is
931 * happening, __lockres_clear_pending() sees PENDING is unset and does
934 * But what if another path (eg downconvert thread) has just started a
935 * new locking action? The other path has re-set PENDING. Our path
936 * cannot clear PENDING, because that will re-open the original race
942 * ocfs2_cluster_lock()
947 * ocfs2_locking_ast() ocfs2_downconvert_thread()
948 * clear PENDING ocfs2_unblock_lock()
951 * ocfs2_prepare_downconvert()
961 * So as you can see, we now have a window where l_lock is not held,
962 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
964 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
965 * set by ocfs2_prepare_downconvert(). That wasn't nice.
967 * To solve this we introduce l_pending_gen. A call to
968 * lockres_clear_pending() will only do so when it is passed a generation
969 * number that matches the lockres. lockres_set_pending() will return the
970 * current generation number. When ocfs2_cluster_lock() goes to clear
971 * PENDING, it passes the generation it got from set_pending(). In our
972 * example above, the generation numbers will *not* match. Thus,
973 * ocfs2_cluster_lock() will not clear the PENDING set by
974 * ocfs2_prepare_downconvert().
977 /* Unlocked version for ocfs2_locking_ast() */
978 static void __lockres_clear_pending(struct ocfs2_lock_res *lockres,
979 unsigned int generation,
980 struct ocfs2_super *osb)
982 assert_spin_locked(&lockres->l_lock);
985 * The ast and locking functions can race us here. The winner
986 * will clear pending, the loser will not.
988 if (!(lockres->l_flags & OCFS2_LOCK_PENDING) ||
989 (lockres->l_pending_gen != generation))
992 lockres_clear_flags(lockres, OCFS2_LOCK_PENDING);
993 lockres->l_pending_gen++;
996 * The downconvert thread may have skipped us because we
997 * were PENDING. Wake it up.
999 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1000 ocfs2_wake_downconvert_thread(osb);
1003 /* Locked version for callers of ocfs2_dlm_lock() */
1004 static void lockres_clear_pending(struct ocfs2_lock_res *lockres,
1005 unsigned int generation,
1006 struct ocfs2_super *osb)
1008 unsigned long flags;
1010 spin_lock_irqsave(&lockres->l_lock, flags);
1011 __lockres_clear_pending(lockres, generation, osb);
1012 spin_unlock_irqrestore(&lockres->l_lock, flags);
1015 static unsigned int lockres_set_pending(struct ocfs2_lock_res *lockres)
1017 assert_spin_locked(&lockres->l_lock);
1018 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
1020 lockres_or_flags(lockres, OCFS2_LOCK_PENDING);
1022 return lockres->l_pending_gen;
1025 static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb *lksb, int level)
1027 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1028 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1029 int needs_downconvert;
1030 unsigned long flags;
1032 BUG_ON(level <= DLM_LOCK_NL);
1034 mlog(ML_BASTS, "BAST fired for lockres %s, blocking %d, level %d, "
1035 "type %s\n", lockres->l_name, level, lockres->l_level,
1036 ocfs2_lock_type_string(lockres->l_type));
1039 * We can skip the bast for locks which don't enable caching -
1040 * they'll be dropped at the earliest possible time anyway.
1042 if (lockres->l_flags & OCFS2_LOCK_NOCACHE)
1045 spin_lock_irqsave(&lockres->l_lock, flags);
1046 needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
1047 if (needs_downconvert)
1048 ocfs2_schedule_blocked_lock(osb, lockres);
1049 spin_unlock_irqrestore(&lockres->l_lock, flags);
1051 wake_up(&lockres->l_event);
1053 ocfs2_wake_downconvert_thread(osb);
1056 static void ocfs2_locking_ast(struct ocfs2_dlm_lksb *lksb)
1058 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1059 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1060 unsigned long flags;
1063 spin_lock_irqsave(&lockres->l_lock, flags);
1065 status = ocfs2_dlm_lock_status(&lockres->l_lksb);
1067 if (status == -EAGAIN) {
1068 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1073 mlog(ML_ERROR, "lockres %s: lksb status value of %d!\n",
1074 lockres->l_name, status);
1075 spin_unlock_irqrestore(&lockres->l_lock, flags);
1079 mlog(ML_BASTS, "AST fired for lockres %s, action %d, unlock %d, "
1080 "level %d => %d\n", lockres->l_name, lockres->l_action,
1081 lockres->l_unlock_action, lockres->l_level, lockres->l_requested);
1083 switch(lockres->l_action) {
1084 case OCFS2_AST_ATTACH:
1085 ocfs2_generic_handle_attach_action(lockres);
1086 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
1088 case OCFS2_AST_CONVERT:
1089 ocfs2_generic_handle_convert_action(lockres);
1091 case OCFS2_AST_DOWNCONVERT:
1092 ocfs2_generic_handle_downconvert_action(lockres);
1095 mlog(ML_ERROR, "lockres %s: AST fired with invalid action: %u, "
1096 "flags 0x%lx, unlock: %u\n",
1097 lockres->l_name, lockres->l_action, lockres->l_flags,
1098 lockres->l_unlock_action);
1102 /* set it to something invalid so if we get called again we
1104 lockres->l_action = OCFS2_AST_INVALID;
1106 /* Did we try to cancel this lock? Clear that state */
1107 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT)
1108 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1111 * We may have beaten the locking functions here. We certainly
1112 * know that dlm_lock() has been called :-)
1113 * Because we can't have two lock calls in flight at once, we
1114 * can use lockres->l_pending_gen.
1116 __lockres_clear_pending(lockres, lockres->l_pending_gen, osb);
1118 wake_up(&lockres->l_event);
1119 spin_unlock_irqrestore(&lockres->l_lock, flags);
1122 static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb *lksb, int error)
1124 struct ocfs2_lock_res *lockres = ocfs2_lksb_to_lock_res(lksb);
1125 unsigned long flags;
1127 mlog(ML_BASTS, "UNLOCK AST fired for lockres %s, action = %d\n",
1128 lockres->l_name, lockres->l_unlock_action);
1130 spin_lock_irqsave(&lockres->l_lock, flags);
1132 mlog(ML_ERROR, "Dlm passes error %d for lock %s, "
1133 "unlock_action %d\n", error, lockres->l_name,
1134 lockres->l_unlock_action);
1135 spin_unlock_irqrestore(&lockres->l_lock, flags);
1139 switch(lockres->l_unlock_action) {
1140 case OCFS2_UNLOCK_CANCEL_CONVERT:
1141 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
1142 lockres->l_action = OCFS2_AST_INVALID;
1143 /* Downconvert thread may have requeued this lock, we
1144 * need to wake it. */
1145 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
1146 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres));
1148 case OCFS2_UNLOCK_DROP_LOCK:
1149 lockres->l_level = DLM_LOCK_IV;
1155 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1156 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1157 wake_up(&lockres->l_event);
1158 spin_unlock_irqrestore(&lockres->l_lock, flags);
1162 * This is the filesystem locking protocol. It provides the lock handling
1163 * hooks for the underlying DLM. It has a maximum version number.
1164 * The version number allows interoperability with systems running at
1165 * the same major number and an equal or smaller minor number.
1167 * Whenever the filesystem does new things with locks (adds or removes a
1168 * lock, orders them differently, does different things underneath a lock),
1169 * the version must be changed. The protocol is negotiated when joining
1170 * the dlm domain. A node may join the domain if its major version is
1171 * identical to all other nodes and its minor version is greater than
1172 * or equal to all other nodes. When its minor version is greater than
1173 * the other nodes, it will run at the minor version specified by the
1176 * If a locking change is made that will not be compatible with older
1177 * versions, the major number must be increased and the minor version set
1178 * to zero. If a change merely adds a behavior that can be disabled when
1179 * speaking to older versions, the minor version must be increased. If a
1180 * change adds a fully backwards compatible change (eg, LVB changes that
1181 * are just ignored by older versions), the version does not need to be
1184 static struct ocfs2_locking_protocol lproto = {
1186 .pv_major = OCFS2_LOCKING_PROTOCOL_MAJOR,
1187 .pv_minor = OCFS2_LOCKING_PROTOCOL_MINOR,
1189 .lp_lock_ast = ocfs2_locking_ast,
1190 .lp_blocking_ast = ocfs2_blocking_ast,
1191 .lp_unlock_ast = ocfs2_unlock_ast,
1194 void ocfs2_set_locking_protocol(void)
1196 ocfs2_stack_glue_set_max_proto_version(&lproto.lp_max_version);
1199 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
1202 unsigned long flags;
1204 spin_lock_irqsave(&lockres->l_lock, flags);
1205 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
1206 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1208 lockres->l_action = OCFS2_AST_INVALID;
1210 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
1211 spin_unlock_irqrestore(&lockres->l_lock, flags);
1213 wake_up(&lockres->l_event);
1216 /* Note: If we detect another process working on the lock (i.e.,
1217 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1218 * to do the right thing in that case.
1220 static int ocfs2_lock_create(struct ocfs2_super *osb,
1221 struct ocfs2_lock_res *lockres,
1226 unsigned long flags;
1229 mlog(0, "lock %s, level = %d, flags = %u\n", lockres->l_name, level,
1232 spin_lock_irqsave(&lockres->l_lock, flags);
1233 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
1234 (lockres->l_flags & OCFS2_LOCK_BUSY)) {
1235 spin_unlock_irqrestore(&lockres->l_lock, flags);
1239 lockres->l_action = OCFS2_AST_ATTACH;
1240 lockres->l_requested = level;
1241 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1242 gen = lockres_set_pending(lockres);
1243 spin_unlock_irqrestore(&lockres->l_lock, flags);
1245 ret = ocfs2_dlm_lock(osb->cconn,
1250 OCFS2_LOCK_ID_MAX_LEN - 1);
1251 lockres_clear_pending(lockres, gen, osb);
1253 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1254 ocfs2_recover_from_dlm_error(lockres, 1);
1257 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres->l_name);
1263 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
1266 unsigned long flags;
1269 spin_lock_irqsave(&lockres->l_lock, flags);
1270 ret = lockres->l_flags & flag;
1271 spin_unlock_irqrestore(&lockres->l_lock, flags);
1276 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
1279 wait_event(lockres->l_event,
1280 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
1283 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
1286 wait_event(lockres->l_event,
1287 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
1290 /* predict what lock level we'll be dropping down to on behalf
1291 * of another node, and return true if the currently wanted
1292 * level will be compatible with it. */
1293 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
1296 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
1298 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
1301 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
1303 INIT_LIST_HEAD(&mw->mw_item);
1304 init_completion(&mw->mw_complete);
1305 ocfs2_init_start_time(mw);
1308 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
1310 wait_for_completion(&mw->mw_complete);
1311 /* Re-arm the completion in case we want to wait on it again */
1312 reinit_completion(&mw->mw_complete);
1313 return mw->mw_status;
1316 static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
1317 struct ocfs2_mask_waiter *mw,
1321 BUG_ON(!list_empty(&mw->mw_item));
1323 assert_spin_locked(&lockres->l_lock);
1325 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
1330 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
1331 * if the mask still hadn't reached its goal */
1332 static int __lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1333 struct ocfs2_mask_waiter *mw)
1337 assert_spin_locked(&lockres->l_lock);
1338 if (!list_empty(&mw->mw_item)) {
1339 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
1342 list_del_init(&mw->mw_item);
1343 init_completion(&mw->mw_complete);
1349 static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
1350 struct ocfs2_mask_waiter *mw)
1352 unsigned long flags;
1355 spin_lock_irqsave(&lockres->l_lock, flags);
1356 ret = __lockres_remove_mask_waiter(lockres, mw);
1357 spin_unlock_irqrestore(&lockres->l_lock, flags);
1363 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter *mw,
1364 struct ocfs2_lock_res *lockres)
1368 ret = wait_for_completion_interruptible(&mw->mw_complete);
1370 lockres_remove_mask_waiter(lockres, mw);
1372 ret = mw->mw_status;
1373 /* Re-arm the completion in case we want to wait on it again */
1374 reinit_completion(&mw->mw_complete);
1378 static int __ocfs2_cluster_lock(struct ocfs2_super *osb,
1379 struct ocfs2_lock_res *lockres,
1384 unsigned long caller_ip)
1386 struct ocfs2_mask_waiter mw;
1387 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
1388 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1389 unsigned long flags;
1391 int noqueue_attempted = 0;
1395 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED)) {
1396 mlog_errno(-EINVAL);
1400 ocfs2_init_mask_waiter(&mw);
1402 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
1403 lkm_flags |= DLM_LKF_VALBLK;
1408 spin_lock_irqsave(&lockres->l_lock, flags);
1410 if (catch_signals && signal_pending(current)) {
1415 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
1416 "Cluster lock called on freeing lockres %s! flags "
1417 "0x%lx\n", lockres->l_name, lockres->l_flags);
1419 /* We only compare against the currently granted level
1420 * here. If the lock is blocked waiting on a downconvert,
1421 * we'll get caught below. */
1422 if (lockres->l_flags & OCFS2_LOCK_BUSY &&
1423 level > lockres->l_level) {
1424 /* is someone sitting in dlm_lock? If so, wait on
1426 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1431 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING) {
1433 * We've upconverted. If the lock now has a level we can
1434 * work with, we take it. If, however, the lock is not at the
1435 * required level, we go thru the full cycle. One way this could
1436 * happen is if a process requesting an upconvert to PR is
1437 * closely followed by another requesting upconvert to an EX.
1438 * If the process requesting EX lands here, we want it to
1439 * continue attempting to upconvert and let the process
1440 * requesting PR take the lock.
1441 * If multiple processes request upconvert to PR, the first one
1442 * here will take the lock. The others will have to go thru the
1443 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1444 * downconvert request.
1446 if (level <= lockres->l_level)
1447 goto update_holders;
1450 if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
1451 !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
1452 /* is the lock is currently blocked on behalf of
1454 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
1459 if (level > lockres->l_level) {
1460 if (noqueue_attempted > 0) {
1464 if (lkm_flags & DLM_LKF_NOQUEUE)
1465 noqueue_attempted = 1;
1467 if (lockres->l_action != OCFS2_AST_INVALID)
1468 mlog(ML_ERROR, "lockres %s has action %u pending\n",
1469 lockres->l_name, lockres->l_action);
1471 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1472 lockres->l_action = OCFS2_AST_ATTACH;
1473 lkm_flags &= ~DLM_LKF_CONVERT;
1475 lockres->l_action = OCFS2_AST_CONVERT;
1476 lkm_flags |= DLM_LKF_CONVERT;
1479 lockres->l_requested = level;
1480 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1481 gen = lockres_set_pending(lockres);
1482 spin_unlock_irqrestore(&lockres->l_lock, flags);
1484 BUG_ON(level == DLM_LOCK_IV);
1485 BUG_ON(level == DLM_LOCK_NL);
1487 mlog(ML_BASTS, "lockres %s, convert from %d to %d\n",
1488 lockres->l_name, lockres->l_level, level);
1490 /* call dlm_lock to upgrade lock now */
1491 ret = ocfs2_dlm_lock(osb->cconn,
1496 OCFS2_LOCK_ID_MAX_LEN - 1);
1497 lockres_clear_pending(lockres, gen, osb);
1499 if (!(lkm_flags & DLM_LKF_NOQUEUE) ||
1501 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1504 ocfs2_recover_from_dlm_error(lockres, 1);
1509 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1512 /* At this point we've gone inside the dlm and need to
1513 * complete our work regardless. */
1516 /* wait for busy to clear and carry on */
1521 /* Ok, if we get here then we're good to go. */
1522 ocfs2_inc_holders(lockres, level);
1526 lockres_clear_flags(lockres, OCFS2_LOCK_UPCONVERT_FINISHING);
1528 /* ocfs2_unblock_lock reques on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1529 kick_dc = (lockres->l_flags & OCFS2_LOCK_BLOCKED);
1531 spin_unlock_irqrestore(&lockres->l_lock, flags);
1533 ocfs2_wake_downconvert_thread(osb);
1536 * This is helping work around a lock inversion between the page lock
1537 * and dlm locks. One path holds the page lock while calling aops
1538 * which block acquiring dlm locks. The voting thread holds dlm
1539 * locks while acquiring page locks while down converting data locks.
1540 * This block is helping an aop path notice the inversion and back
1541 * off to unlock its page lock before trying the dlm lock again.
1543 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1544 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1546 spin_lock_irqsave(&lockres->l_lock, flags);
1547 if (__lockres_remove_mask_waiter(lockres, &mw)) {
1549 lockres_or_flags(lockres,
1550 OCFS2_LOCK_NONBLOCK_FINISHED);
1551 spin_unlock_irqrestore(&lockres->l_lock, flags);
1554 spin_unlock_irqrestore(&lockres->l_lock, flags);
1559 ret = ocfs2_wait_for_mask(&mw);
1564 ocfs2_update_lock_stats(lockres, level, &mw, ret);
1566 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1567 if (!ret && lockres->l_lockdep_map.key != NULL) {
1568 if (level == DLM_LOCK_PR)
1569 rwsem_acquire_read(&lockres->l_lockdep_map, l_subclass,
1570 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1573 rwsem_acquire(&lockres->l_lockdep_map, l_subclass,
1574 !!(arg_flags & OCFS2_META_LOCK_NOQUEUE),
1581 static inline int ocfs2_cluster_lock(struct ocfs2_super *osb,
1582 struct ocfs2_lock_res *lockres,
1587 return __ocfs2_cluster_lock(osb, lockres, level, lkm_flags, arg_flags,
1592 static void __ocfs2_cluster_unlock(struct ocfs2_super *osb,
1593 struct ocfs2_lock_res *lockres,
1595 unsigned long caller_ip)
1597 unsigned long flags;
1599 spin_lock_irqsave(&lockres->l_lock, flags);
1600 ocfs2_dec_holders(lockres, level);
1601 ocfs2_downconvert_on_unlock(osb, lockres);
1602 spin_unlock_irqrestore(&lockres->l_lock, flags);
1603 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1604 if (lockres->l_lockdep_map.key != NULL)
1605 rwsem_release(&lockres->l_lockdep_map, 1, caller_ip);
1609 static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1610 struct ocfs2_lock_res *lockres,
1614 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1615 unsigned long flags;
1616 u32 lkm_flags = local ? DLM_LKF_LOCAL : 0;
1618 spin_lock_irqsave(&lockres->l_lock, flags);
1619 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1620 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1621 spin_unlock_irqrestore(&lockres->l_lock, flags);
1623 return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1626 /* Grants us an EX lock on the data and metadata resources, skipping
1627 * the normal cluster directory lookup. Use this ONLY on newly created
1628 * inodes which other nodes can't possibly see, and which haven't been
1629 * hashed in the inode hash yet. This can give us a good performance
1630 * increase as it'll skip the network broadcast normally associated
1631 * with creating a new lock resource. */
1632 int ocfs2_create_new_inode_locks(struct inode *inode)
1635 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1638 BUG_ON(!ocfs2_inode_is_new(inode));
1640 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1642 /* NOTE: That we don't increment any of the holder counts, nor
1643 * do we add anything to a journal handle. Since this is
1644 * supposed to be a new inode which the cluster doesn't know
1645 * about yet, there is no need to. As far as the LVB handling
1646 * is concerned, this is basically like acquiring an EX lock
1647 * on a resource which has an invalid one -- we'll set it
1648 * valid when we release the EX. */
1650 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1657 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1658 * don't use a generation in their lock names.
1660 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_inode_lockres, 1, 0);
1666 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1676 int ocfs2_rw_lock(struct inode *inode, int write)
1679 struct ocfs2_lock_res *lockres;
1680 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1684 mlog(0, "inode %llu take %s RW lock\n",
1685 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1686 write ? "EXMODE" : "PRMODE");
1688 if (ocfs2_mount_local(osb))
1691 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1693 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1695 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1703 void ocfs2_rw_unlock(struct inode *inode, int write)
1705 int level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1706 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1707 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1709 mlog(0, "inode %llu drop %s RW lock\n",
1710 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1711 write ? "EXMODE" : "PRMODE");
1713 if (!ocfs2_mount_local(osb))
1714 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1718 * ocfs2_open_lock always get PR mode lock.
1720 int ocfs2_open_lock(struct inode *inode)
1723 struct ocfs2_lock_res *lockres;
1724 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1728 mlog(0, "inode %llu take PRMODE open lock\n",
1729 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1731 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1734 lockres = &OCFS2_I(inode)->ip_open_lockres;
1736 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1745 int ocfs2_try_open_lock(struct inode *inode, int write)
1747 int status = 0, level;
1748 struct ocfs2_lock_res *lockres;
1749 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1753 mlog(0, "inode %llu try to take %s open lock\n",
1754 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1755 write ? "EXMODE" : "PRMODE");
1757 if (ocfs2_is_hard_readonly(osb)) {
1763 if (ocfs2_mount_local(osb))
1766 lockres = &OCFS2_I(inode)->ip_open_lockres;
1768 level = write ? DLM_LOCK_EX : DLM_LOCK_PR;
1771 * The file system may already holding a PRMODE/EXMODE open lock.
1772 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1773 * other nodes and the -EAGAIN will indicate to the caller that
1774 * this inode is still in use.
1776 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1777 level, DLM_LKF_NOQUEUE, 0);
1784 * ocfs2_open_unlock unlock PR and EX mode open locks.
1786 void ocfs2_open_unlock(struct inode *inode)
1788 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1789 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1791 mlog(0, "inode %llu drop open lock\n",
1792 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1794 if (ocfs2_mount_local(osb))
1797 if(lockres->l_ro_holders)
1798 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1800 if(lockres->l_ex_holders)
1801 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1808 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res *lockres,
1812 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
1813 unsigned long flags;
1814 struct ocfs2_mask_waiter mw;
1816 ocfs2_init_mask_waiter(&mw);
1819 spin_lock_irqsave(&lockres->l_lock, flags);
1820 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
1821 ret = ocfs2_prepare_cancel_convert(osb, lockres);
1823 spin_unlock_irqrestore(&lockres->l_lock, flags);
1824 ret = ocfs2_cancel_convert(osb, lockres);
1831 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1832 spin_unlock_irqrestore(&lockres->l_lock, flags);
1834 ocfs2_wait_for_mask(&mw);
1840 * We may still have gotten the lock, in which case there's no
1841 * point to restarting the syscall.
1843 if (lockres->l_level == level)
1846 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret,
1847 lockres->l_flags, lockres->l_level, lockres->l_action);
1849 spin_unlock_irqrestore(&lockres->l_lock, flags);
1856 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1857 * flock() calls. The locking approach this requires is sufficiently
1858 * different from all other cluster lock types that we implement a
1859 * separate path to the "low-level" dlm calls. In particular:
1861 * - No optimization of lock levels is done - we take at exactly
1862 * what's been requested.
1864 * - No lock caching is employed. We immediately downconvert to
1865 * no-lock at unlock time. This also means flock locks never go on
1866 * the blocking list).
1868 * - Since userspace can trivially deadlock itself with flock, we make
1869 * sure to allow cancellation of a misbehaving applications flock()
1872 * - Access to any flock lockres doesn't require concurrency, so we
1873 * can simplify the code by requiring the caller to guarantee
1874 * serialization of dlmglue flock calls.
1876 int ocfs2_file_lock(struct file *file, int ex, int trylock)
1878 int ret, level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
1879 unsigned int lkm_flags = trylock ? DLM_LKF_NOQUEUE : 0;
1880 unsigned long flags;
1881 struct ocfs2_file_private *fp = file->private_data;
1882 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1883 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1884 struct ocfs2_mask_waiter mw;
1886 ocfs2_init_mask_waiter(&mw);
1888 if ((lockres->l_flags & OCFS2_LOCK_BUSY) ||
1889 (lockres->l_level > DLM_LOCK_NL)) {
1891 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1892 "level: %u\n", lockres->l_name, lockres->l_flags,
1897 spin_lock_irqsave(&lockres->l_lock, flags);
1898 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
1899 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1900 spin_unlock_irqrestore(&lockres->l_lock, flags);
1903 * Get the lock at NLMODE to start - that way we
1904 * can cancel the upconvert request if need be.
1906 ret = ocfs2_lock_create(osb, lockres, DLM_LOCK_NL, 0);
1912 ret = ocfs2_wait_for_mask(&mw);
1917 spin_lock_irqsave(&lockres->l_lock, flags);
1920 lockres->l_action = OCFS2_AST_CONVERT;
1921 lkm_flags |= DLM_LKF_CONVERT;
1922 lockres->l_requested = level;
1923 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
1925 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
1926 spin_unlock_irqrestore(&lockres->l_lock, flags);
1928 ret = ocfs2_dlm_lock(osb->cconn, level, &lockres->l_lksb, lkm_flags,
1929 lockres->l_name, OCFS2_LOCK_ID_MAX_LEN - 1);
1931 if (!trylock || (ret != -EAGAIN)) {
1932 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
1936 ocfs2_recover_from_dlm_error(lockres, 1);
1937 lockres_remove_mask_waiter(lockres, &mw);
1941 ret = ocfs2_wait_for_mask_interruptible(&mw, lockres);
1942 if (ret == -ERESTARTSYS) {
1944 * Userspace can cause deadlock itself with
1945 * flock(). Current behavior locally is to allow the
1946 * deadlock, but abort the system call if a signal is
1947 * received. We follow this example, otherwise a
1948 * poorly written program could sit in kernel until
1951 * Handling this is a bit more complicated for Ocfs2
1952 * though. We can't exit this function with an
1953 * outstanding lock request, so a cancel convert is
1954 * required. We intentionally overwrite 'ret' - if the
1955 * cancel fails and the lock was granted, it's easier
1956 * to just bubble success back up to the user.
1958 ret = ocfs2_flock_handle_signal(lockres, level);
1959 } else if (!ret && (level > lockres->l_level)) {
1960 /* Trylock failed asynchronously */
1967 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1968 lockres->l_name, ex, trylock, ret);
1972 void ocfs2_file_unlock(struct file *file)
1976 unsigned long flags;
1977 struct ocfs2_file_private *fp = file->private_data;
1978 struct ocfs2_lock_res *lockres = &fp->fp_flock;
1979 struct ocfs2_super *osb = OCFS2_SB(file->f_mapping->host->i_sb);
1980 struct ocfs2_mask_waiter mw;
1982 ocfs2_init_mask_waiter(&mw);
1984 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED))
1987 if (lockres->l_level == DLM_LOCK_NL)
1990 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1991 lockres->l_name, lockres->l_flags, lockres->l_level,
1994 spin_lock_irqsave(&lockres->l_lock, flags);
1996 * Fake a blocking ast for the downconvert code.
1998 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
1999 lockres->l_blocking = DLM_LOCK_EX;
2001 gen = ocfs2_prepare_downconvert(lockres, DLM_LOCK_NL);
2002 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
2003 spin_unlock_irqrestore(&lockres->l_lock, flags);
2005 ret = ocfs2_downconvert_lock(osb, lockres, DLM_LOCK_NL, 0, gen);
2011 ret = ocfs2_wait_for_mask(&mw);
2016 static void ocfs2_downconvert_on_unlock(struct ocfs2_super *osb,
2017 struct ocfs2_lock_res *lockres)
2021 /* If we know that another node is waiting on our lock, kick
2022 * the downconvert thread * pre-emptively when we reach a release
2024 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
2025 switch(lockres->l_blocking) {
2027 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
2031 if (!lockres->l_ex_holders)
2040 ocfs2_wake_downconvert_thread(osb);
2043 #define OCFS2_SEC_BITS 34
2044 #define OCFS2_SEC_SHIFT (64 - 34)
2045 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
2047 /* LVB only has room for 64 bits of time here so we pack it for
2049 static u64 ocfs2_pack_timespec(struct timespec *spec)
2052 u64 sec = spec->tv_sec;
2053 u32 nsec = spec->tv_nsec;
2055 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
2060 /* Call this with the lockres locked. I am reasonably sure we don't
2061 * need ip_lock in this function as anyone who would be changing those
2062 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2063 static void __ocfs2_stuff_meta_lvb(struct inode *inode)
2065 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2066 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2067 struct ocfs2_meta_lvb *lvb;
2069 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2072 * Invalidate the LVB of a deleted inode - this way other
2073 * nodes are forced to go to disk and discover the new inode
2076 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2077 lvb->lvb_version = 0;
2081 lvb->lvb_version = OCFS2_LVB_VERSION;
2082 lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
2083 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
2084 lvb->lvb_iuid = cpu_to_be32(i_uid_read(inode));
2085 lvb->lvb_igid = cpu_to_be32(i_gid_read(inode));
2086 lvb->lvb_imode = cpu_to_be16(inode->i_mode);
2087 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink);
2088 lvb->lvb_iatime_packed =
2089 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
2090 lvb->lvb_ictime_packed =
2091 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
2092 lvb->lvb_imtime_packed =
2093 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
2094 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr);
2095 lvb->lvb_idynfeatures = cpu_to_be16(oi->ip_dyn_features);
2096 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
2099 mlog_meta_lvb(0, lockres);
2102 static void ocfs2_unpack_timespec(struct timespec *spec,
2105 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
2106 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
2109 static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
2111 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2112 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2113 struct ocfs2_meta_lvb *lvb;
2115 mlog_meta_lvb(0, lockres);
2117 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2119 /* We're safe here without the lockres lock... */
2120 spin_lock(&oi->ip_lock);
2121 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
2122 i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
2124 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
2125 oi->ip_dyn_features = be16_to_cpu(lvb->lvb_idynfeatures);
2126 ocfs2_set_inode_flags(inode);
2128 /* fast-symlinks are a special case */
2129 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
2130 inode->i_blocks = 0;
2132 inode->i_blocks = ocfs2_inode_sector_count(inode);
2134 i_uid_write(inode, be32_to_cpu(lvb->lvb_iuid));
2135 i_gid_write(inode, be32_to_cpu(lvb->lvb_igid));
2136 inode->i_mode = be16_to_cpu(lvb->lvb_imode);
2137 set_nlink(inode, be16_to_cpu(lvb->lvb_inlink));
2138 ocfs2_unpack_timespec(&inode->i_atime,
2139 be64_to_cpu(lvb->lvb_iatime_packed));
2140 ocfs2_unpack_timespec(&inode->i_mtime,
2141 be64_to_cpu(lvb->lvb_imtime_packed));
2142 ocfs2_unpack_timespec(&inode->i_ctime,
2143 be64_to_cpu(lvb->lvb_ictime_packed));
2144 spin_unlock(&oi->ip_lock);
2147 static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
2148 struct ocfs2_lock_res *lockres)
2150 struct ocfs2_meta_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2152 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb)
2153 && lvb->lvb_version == OCFS2_LVB_VERSION
2154 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
2159 /* Determine whether a lock resource needs to be refreshed, and
2160 * arbitrate who gets to refresh it.
2162 * 0 means no refresh needed.
2164 * > 0 means you need to refresh this and you MUST call
2165 * ocfs2_complete_lock_res_refresh afterwards. */
2166 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
2168 unsigned long flags;
2172 spin_lock_irqsave(&lockres->l_lock, flags);
2173 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
2174 spin_unlock_irqrestore(&lockres->l_lock, flags);
2178 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
2179 spin_unlock_irqrestore(&lockres->l_lock, flags);
2181 ocfs2_wait_on_refreshing_lock(lockres);
2185 /* Ok, I'll be the one to refresh this lock. */
2186 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
2187 spin_unlock_irqrestore(&lockres->l_lock, flags);
2191 mlog(0, "status %d\n", status);
2195 /* If status is non zero, I'll mark it as not being in refresh
2196 * anymroe, but i won't clear the needs refresh flag. */
2197 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
2200 unsigned long flags;
2202 spin_lock_irqsave(&lockres->l_lock, flags);
2203 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
2205 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
2206 spin_unlock_irqrestore(&lockres->l_lock, flags);
2208 wake_up(&lockres->l_event);
2211 /* may or may not return a bh if it went to disk. */
2212 static int ocfs2_inode_lock_update(struct inode *inode,
2213 struct buffer_head **bh)
2216 struct ocfs2_inode_info *oi = OCFS2_I(inode);
2217 struct ocfs2_lock_res *lockres = &oi->ip_inode_lockres;
2218 struct ocfs2_dinode *fe;
2219 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2221 if (ocfs2_mount_local(osb))
2224 spin_lock(&oi->ip_lock);
2225 if (oi->ip_flags & OCFS2_INODE_DELETED) {
2226 mlog(0, "Orphaned inode %llu was deleted while we "
2227 "were waiting on a lock. ip_flags = 0x%x\n",
2228 (unsigned long long)oi->ip_blkno, oi->ip_flags);
2229 spin_unlock(&oi->ip_lock);
2233 spin_unlock(&oi->ip_lock);
2235 if (!ocfs2_should_refresh_lock_res(lockres))
2238 /* This will discard any caching information we might have had
2239 * for the inode metadata. */
2240 ocfs2_metadata_cache_purge(INODE_CACHE(inode));
2242 ocfs2_extent_map_trunc(inode, 0);
2244 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
2245 mlog(0, "Trusting LVB on inode %llu\n",
2246 (unsigned long long)oi->ip_blkno);
2247 ocfs2_refresh_inode_from_lvb(inode);
2249 /* Boo, we have to go to disk. */
2250 /* read bh, cast, ocfs2_refresh_inode */
2251 status = ocfs2_read_inode_block(inode, bh);
2256 fe = (struct ocfs2_dinode *) (*bh)->b_data;
2258 /* This is a good chance to make sure we're not
2259 * locking an invalid object. ocfs2_read_inode_block()
2260 * already checked that the inode block is sane.
2262 * We bug on a stale inode here because we checked
2263 * above whether it was wiped from disk. The wiping
2264 * node provides a guarantee that we receive that
2265 * message and can mark the inode before dropping any
2266 * locks associated with it. */
2267 mlog_bug_on_msg(inode->i_generation !=
2268 le32_to_cpu(fe->i_generation),
2269 "Invalid dinode %llu disk generation: %u "
2270 "inode->i_generation: %u\n",
2271 (unsigned long long)oi->ip_blkno,
2272 le32_to_cpu(fe->i_generation),
2273 inode->i_generation);
2274 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
2275 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
2276 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2277 (unsigned long long)oi->ip_blkno,
2278 (unsigned long long)le64_to_cpu(fe->i_dtime),
2279 le32_to_cpu(fe->i_flags));
2281 ocfs2_refresh_inode(inode, fe);
2282 ocfs2_track_lock_refresh(lockres);
2287 ocfs2_complete_lock_res_refresh(lockres, status);
2292 static int ocfs2_assign_bh(struct inode *inode,
2293 struct buffer_head **ret_bh,
2294 struct buffer_head *passed_bh)
2299 /* Ok, the update went to disk for us, use the
2301 *ret_bh = passed_bh;
2307 status = ocfs2_read_inode_block(inode, ret_bh);
2315 * returns < 0 error if the callback will never be called, otherwise
2316 * the result of the lock will be communicated via the callback.
2318 int ocfs2_inode_lock_full_nested(struct inode *inode,
2319 struct buffer_head **ret_bh,
2324 int status, level, acquired;
2326 struct ocfs2_lock_res *lockres = NULL;
2327 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2328 struct buffer_head *local_bh = NULL;
2332 mlog(0, "inode %llu, take %s META lock\n",
2333 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2334 ex ? "EXMODE" : "PRMODE");
2338 /* We'll allow faking a readonly metadata lock for
2340 if (ocfs2_is_hard_readonly(osb)) {
2346 if (ocfs2_mount_local(osb))
2349 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2350 ocfs2_wait_for_recovery(osb);
2352 lockres = &OCFS2_I(inode)->ip_inode_lockres;
2353 level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2355 if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
2356 dlm_flags |= DLM_LKF_NOQUEUE;
2358 status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
2359 arg_flags, subclass, _RET_IP_);
2361 if (status != -EAGAIN)
2366 /* Notify the error cleanup path to drop the cluster lock. */
2369 /* We wait twice because a node may have died while we were in
2370 * the lower dlm layers. The second time though, we've
2371 * committed to owning this lock so we don't allow signals to
2372 * abort the operation. */
2373 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
2374 ocfs2_wait_for_recovery(osb);
2378 * We only see this flag if we're being called from
2379 * ocfs2_read_locked_inode(). It means we're locking an inode
2380 * which hasn't been populated yet, so clear the refresh flag
2381 * and let the caller handle it.
2383 if (inode->i_state & I_NEW) {
2386 ocfs2_complete_lock_res_refresh(lockres, 0);
2390 /* This is fun. The caller may want a bh back, or it may
2391 * not. ocfs2_inode_lock_update definitely wants one in, but
2392 * may or may not read one, depending on what's in the
2393 * LVB. The result of all of this is that we've *only* gone to
2394 * disk if we have to, so the complexity is worthwhile. */
2395 status = ocfs2_inode_lock_update(inode, &local_bh);
2397 if (status != -ENOENT)
2403 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
2412 if (ret_bh && (*ret_bh)) {
2417 ocfs2_inode_unlock(inode, ex);
2427 * This is working around a lock inversion between tasks acquiring DLM
2428 * locks while holding a page lock and the downconvert thread which
2429 * blocks dlm lock acquiry while acquiring page locks.
2431 * ** These _with_page variantes are only intended to be called from aop
2432 * methods that hold page locks and return a very specific *positive* error
2433 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2435 * The DLM is called such that it returns -EAGAIN if it would have
2436 * blocked waiting for the downconvert thread. In that case we unlock
2437 * our page so the downconvert thread can make progress. Once we've
2438 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2439 * that called us can bubble that back up into the VFS who will then
2440 * immediately retry the aop call.
2442 * We do a blocking lock and immediate unlock before returning, though, so that
2443 * the lock has a great chance of being cached on this node by the time the VFS
2444 * calls back to retry the aop. This has a potential to livelock as nodes
2445 * ping locks back and forth, but that's a risk we're willing to take to avoid
2446 * the lock inversion simply.
2448 int ocfs2_inode_lock_with_page(struct inode *inode,
2449 struct buffer_head **ret_bh,
2455 ret = ocfs2_inode_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
2456 if (ret == -EAGAIN) {
2458 if (ocfs2_inode_lock(inode, ret_bh, ex) == 0)
2459 ocfs2_inode_unlock(inode, ex);
2460 ret = AOP_TRUNCATED_PAGE;
2466 int ocfs2_inode_lock_atime(struct inode *inode,
2467 struct vfsmount *vfsmnt,
2472 ret = ocfs2_inode_lock(inode, NULL, 0);
2479 * If we should update atime, we will get EX lock,
2480 * otherwise we just get PR lock.
2482 if (ocfs2_should_update_atime(inode, vfsmnt)) {
2483 struct buffer_head *bh = NULL;
2485 ocfs2_inode_unlock(inode, 0);
2486 ret = ocfs2_inode_lock(inode, &bh, 1);
2492 if (ocfs2_should_update_atime(inode, vfsmnt))
2493 ocfs2_update_inode_atime(inode, bh);
2502 void ocfs2_inode_unlock(struct inode *inode,
2505 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2506 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_inode_lockres;
2507 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2509 mlog(0, "inode %llu drop %s META lock\n",
2510 (unsigned long long)OCFS2_I(inode)->ip_blkno,
2511 ex ? "EXMODE" : "PRMODE");
2513 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
2514 !ocfs2_mount_local(osb))
2515 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
2518 int ocfs2_orphan_scan_lock(struct ocfs2_super *osb, u32 *seqno)
2520 struct ocfs2_lock_res *lockres;
2521 struct ocfs2_orphan_scan_lvb *lvb;
2524 if (ocfs2_is_hard_readonly(osb))
2527 if (ocfs2_mount_local(osb))
2530 lockres = &osb->osb_orphan_scan.os_lockres;
2531 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2535 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2536 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
2537 lvb->lvb_version == OCFS2_ORPHAN_LVB_VERSION)
2538 *seqno = be32_to_cpu(lvb->lvb_os_seqno);
2540 *seqno = osb->osb_orphan_scan.os_seqno + 1;
2545 void ocfs2_orphan_scan_unlock(struct ocfs2_super *osb, u32 seqno)
2547 struct ocfs2_lock_res *lockres;
2548 struct ocfs2_orphan_scan_lvb *lvb;
2550 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb)) {
2551 lockres = &osb->osb_orphan_scan.os_lockres;
2552 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2553 lvb->lvb_version = OCFS2_ORPHAN_LVB_VERSION;
2554 lvb->lvb_os_seqno = cpu_to_be32(seqno);
2555 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2559 int ocfs2_super_lock(struct ocfs2_super *osb,
2563 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2564 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2566 if (ocfs2_is_hard_readonly(osb))
2569 if (ocfs2_mount_local(osb))
2572 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
2578 /* The super block lock path is really in the best position to
2579 * know when resources covered by the lock need to be
2580 * refreshed, so we do it here. Of course, making sense of
2581 * everything is up to the caller :) */
2582 status = ocfs2_should_refresh_lock_res(lockres);
2584 status = ocfs2_refresh_slot_info(osb);
2586 ocfs2_complete_lock_res_refresh(lockres, status);
2589 ocfs2_cluster_unlock(osb, lockres, level);
2592 ocfs2_track_lock_refresh(lockres);
2598 void ocfs2_super_unlock(struct ocfs2_super *osb,
2601 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2602 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
2604 if (!ocfs2_mount_local(osb))
2605 ocfs2_cluster_unlock(osb, lockres, level);
2608 int ocfs2_rename_lock(struct ocfs2_super *osb)
2611 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2613 if (ocfs2_is_hard_readonly(osb))
2616 if (ocfs2_mount_local(osb))
2619 status = ocfs2_cluster_lock(osb, lockres, DLM_LOCK_EX, 0, 0);
2626 void ocfs2_rename_unlock(struct ocfs2_super *osb)
2628 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
2630 if (!ocfs2_mount_local(osb))
2631 ocfs2_cluster_unlock(osb, lockres, DLM_LOCK_EX);
2634 int ocfs2_nfs_sync_lock(struct ocfs2_super *osb, int ex)
2637 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2639 if (ocfs2_is_hard_readonly(osb))
2642 if (ocfs2_mount_local(osb))
2645 status = ocfs2_cluster_lock(osb, lockres, ex ? LKM_EXMODE : LKM_PRMODE,
2648 mlog(ML_ERROR, "lock on nfs sync lock failed %d\n", status);
2653 void ocfs2_nfs_sync_unlock(struct ocfs2_super *osb, int ex)
2655 struct ocfs2_lock_res *lockres = &osb->osb_nfs_sync_lockres;
2657 if (!ocfs2_mount_local(osb))
2658 ocfs2_cluster_unlock(osb, lockres,
2659 ex ? LKM_EXMODE : LKM_PRMODE);
2662 int ocfs2_dentry_lock(struct dentry *dentry, int ex)
2665 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2666 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2667 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2671 if (ocfs2_is_hard_readonly(osb)) {
2677 if (ocfs2_mount_local(osb))
2680 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2687 void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2689 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
2690 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2691 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2693 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
2694 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2697 /* Reference counting of the dlm debug structure. We want this because
2698 * open references on the debug inodes can live on after a mount, so
2699 * we can't rely on the ocfs2_super to always exist. */
2700 static void ocfs2_dlm_debug_free(struct kref *kref)
2702 struct ocfs2_dlm_debug *dlm_debug;
2704 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2709 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2712 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2715 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2717 kref_get(&debug->d_refcnt);
2720 struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2722 struct ocfs2_dlm_debug *dlm_debug;
2724 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2726 mlog_errno(-ENOMEM);
2730 kref_init(&dlm_debug->d_refcnt);
2731 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2732 dlm_debug->d_locking_state = NULL;
2737 /* Access to this is arbitrated for us via seq_file->sem. */
2738 struct ocfs2_dlm_seq_priv {
2739 struct ocfs2_dlm_debug *p_dlm_debug;
2740 struct ocfs2_lock_res p_iter_res;
2741 struct ocfs2_lock_res p_tmp_res;
2744 static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2745 struct ocfs2_dlm_seq_priv *priv)
2747 struct ocfs2_lock_res *iter, *ret = NULL;
2748 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2750 assert_spin_locked(&ocfs2_dlm_tracking_lock);
2752 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2753 /* discover the head of the list */
2754 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2755 mlog(0, "End of list found, %p\n", ret);
2759 /* We track our "dummy" iteration lockres' by a NULL
2761 if (iter->l_ops != NULL) {
2770 static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2772 struct ocfs2_dlm_seq_priv *priv = m->private;
2773 struct ocfs2_lock_res *iter;
2775 spin_lock(&ocfs2_dlm_tracking_lock);
2776 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2778 /* Since lockres' have the lifetime of their container
2779 * (which can be inodes, ocfs2_supers, etc) we want to
2780 * copy this out to a temporary lockres while still
2781 * under the spinlock. Obviously after this we can't
2782 * trust any pointers on the copy returned, but that's
2783 * ok as the information we want isn't typically held
2785 priv->p_tmp_res = *iter;
2786 iter = &priv->p_tmp_res;
2788 spin_unlock(&ocfs2_dlm_tracking_lock);
2793 static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2797 static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2799 struct ocfs2_dlm_seq_priv *priv = m->private;
2800 struct ocfs2_lock_res *iter = v;
2801 struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2803 spin_lock(&ocfs2_dlm_tracking_lock);
2804 iter = ocfs2_dlm_next_res(iter, priv);
2805 list_del_init(&dummy->l_debug_list);
2807 list_add(&dummy->l_debug_list, &iter->l_debug_list);
2808 priv->p_tmp_res = *iter;
2809 iter = &priv->p_tmp_res;
2811 spin_unlock(&ocfs2_dlm_tracking_lock);
2817 * Version is used by debugfs.ocfs2 to determine the format being used
2820 * - Lock stats printed
2822 * - Max time in lock stats is in usecs (instead of nsecs)
2824 #define OCFS2_DLM_DEBUG_STR_VERSION 3
2825 static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2829 struct ocfs2_lock_res *lockres = v;
2834 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2836 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2837 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2839 (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2841 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2843 seq_printf(m, "%d\t"
2854 lockres->l_unlock_action,
2855 lockres->l_ro_holders,
2856 lockres->l_ex_holders,
2857 lockres->l_requested,
2858 lockres->l_blocking);
2860 /* Dump the raw LVB */
2861 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
2862 for(i = 0; i < DLM_LVB_LEN; i++)
2863 seq_printf(m, "0x%x\t", lvb[i]);
2865 #ifdef CONFIG_OCFS2_FS_STATS
2866 # define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets)
2867 # define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets)
2868 # define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail)
2869 # define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail)
2870 # define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total)
2871 # define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total)
2872 # define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max)
2873 # define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max)
2874 # define lock_refresh(_l) ((_l)->l_lock_refresh)
2876 # define lock_num_prmode(_l) (0)
2877 # define lock_num_exmode(_l) (0)
2878 # define lock_num_prmode_failed(_l) (0)
2879 # define lock_num_exmode_failed(_l) (0)
2880 # define lock_total_prmode(_l) (0ULL)
2881 # define lock_total_exmode(_l) (0ULL)
2882 # define lock_max_prmode(_l) (0)
2883 # define lock_max_exmode(_l) (0)
2884 # define lock_refresh(_l) (0)
2886 /* The following seq_print was added in version 2 of this output */
2887 seq_printf(m, "%u\t"
2896 lock_num_prmode(lockres),
2897 lock_num_exmode(lockres),
2898 lock_num_prmode_failed(lockres),
2899 lock_num_exmode_failed(lockres),
2900 lock_total_prmode(lockres),
2901 lock_total_exmode(lockres),
2902 lock_max_prmode(lockres),
2903 lock_max_exmode(lockres),
2904 lock_refresh(lockres));
2907 seq_printf(m, "\n");
2911 static const struct seq_operations ocfs2_dlm_seq_ops = {
2912 .start = ocfs2_dlm_seq_start,
2913 .stop = ocfs2_dlm_seq_stop,
2914 .next = ocfs2_dlm_seq_next,
2915 .show = ocfs2_dlm_seq_show,
2918 static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
2920 struct seq_file *seq = file->private_data;
2921 struct ocfs2_dlm_seq_priv *priv = seq->private;
2922 struct ocfs2_lock_res *res = &priv->p_iter_res;
2924 ocfs2_remove_lockres_tracking(res);
2925 ocfs2_put_dlm_debug(priv->p_dlm_debug);
2926 return seq_release_private(inode, file);
2929 static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
2931 struct ocfs2_dlm_seq_priv *priv;
2932 struct ocfs2_super *osb;
2934 priv = __seq_open_private(file, &ocfs2_dlm_seq_ops, sizeof(*priv));
2936 mlog_errno(-ENOMEM);
2940 osb = inode->i_private;
2941 ocfs2_get_dlm_debug(osb->osb_dlm_debug);
2942 priv->p_dlm_debug = osb->osb_dlm_debug;
2943 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
2945 ocfs2_add_lockres_tracking(&priv->p_iter_res,
2951 static const struct file_operations ocfs2_dlm_debug_fops = {
2952 .open = ocfs2_dlm_debug_open,
2953 .release = ocfs2_dlm_debug_release,
2955 .llseek = seq_lseek,
2958 static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
2961 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2963 dlm_debug->d_locking_state = debugfs_create_file("locking_state",
2965 osb->osb_debug_root,
2967 &ocfs2_dlm_debug_fops);
2968 if (!dlm_debug->d_locking_state) {
2971 "Unable to create locking state debugfs file.\n");
2975 ocfs2_get_dlm_debug(dlm_debug);
2980 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
2982 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2985 debugfs_remove(dlm_debug->d_locking_state);
2986 ocfs2_put_dlm_debug(dlm_debug);
2990 int ocfs2_dlm_init(struct ocfs2_super *osb)
2993 struct ocfs2_cluster_connection *conn = NULL;
2995 if (ocfs2_mount_local(osb)) {
3000 status = ocfs2_dlm_init_debug(osb);
3006 /* launch downconvert thread */
3007 osb->dc_task = kthread_run(ocfs2_downconvert_thread, osb, "ocfs2dc-%s",
3009 if (IS_ERR(osb->dc_task)) {
3010 status = PTR_ERR(osb->dc_task);
3011 osb->dc_task = NULL;
3016 /* for now, uuid == domain */
3017 status = ocfs2_cluster_connect(osb->osb_cluster_stack,
3018 osb->osb_cluster_name,
3019 strlen(osb->osb_cluster_name),
3021 strlen(osb->uuid_str),
3022 &lproto, ocfs2_do_node_down, osb,
3029 status = ocfs2_cluster_this_node(conn, &osb->node_num);
3033 "could not find this host's node number\n");
3034 ocfs2_cluster_disconnect(conn, 0);
3039 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
3040 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
3041 ocfs2_nfs_sync_lock_res_init(&osb->osb_nfs_sync_lockres, osb);
3042 ocfs2_orphan_scan_lock_res_init(&osb->osb_orphan_scan.os_lockres, osb);
3047 ocfs2_dlm_shutdown_debug(osb);
3049 kthread_stop(osb->dc_task);
3055 void ocfs2_dlm_shutdown(struct ocfs2_super *osb,
3058 ocfs2_drop_osb_locks(osb);
3061 * Now that we have dropped all locks and ocfs2_dismount_volume()
3062 * has disabled recovery, the DLM won't be talking to us. It's
3063 * safe to tear things down before disconnecting the cluster.
3067 kthread_stop(osb->dc_task);
3068 osb->dc_task = NULL;
3071 ocfs2_lock_res_free(&osb->osb_super_lockres);
3072 ocfs2_lock_res_free(&osb->osb_rename_lockres);
3073 ocfs2_lock_res_free(&osb->osb_nfs_sync_lockres);
3074 ocfs2_lock_res_free(&osb->osb_orphan_scan.os_lockres);
3076 ocfs2_cluster_disconnect(osb->cconn, hangup_pending);
3079 ocfs2_dlm_shutdown_debug(osb);
3082 static int ocfs2_drop_lock(struct ocfs2_super *osb,
3083 struct ocfs2_lock_res *lockres)
3086 unsigned long flags;
3089 /* We didn't get anywhere near actually using this lockres. */
3090 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
3093 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3094 lkm_flags |= DLM_LKF_VALBLK;
3096 spin_lock_irqsave(&lockres->l_lock, flags);
3098 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
3099 "lockres %s, flags 0x%lx\n",
3100 lockres->l_name, lockres->l_flags);
3102 while (lockres->l_flags & OCFS2_LOCK_BUSY) {
3103 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3104 "%u, unlock_action = %u\n",
3105 lockres->l_name, lockres->l_flags, lockres->l_action,
3106 lockres->l_unlock_action);
3108 spin_unlock_irqrestore(&lockres->l_lock, flags);
3110 /* XXX: Today we just wait on any busy
3111 * locks... Perhaps we need to cancel converts in the
3113 ocfs2_wait_on_busy_lock(lockres);
3115 spin_lock_irqsave(&lockres->l_lock, flags);
3118 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3119 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
3120 lockres->l_level == DLM_LOCK_EX &&
3121 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3122 lockres->l_ops->set_lvb(lockres);
3125 if (lockres->l_flags & OCFS2_LOCK_BUSY)
3126 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
3128 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
3129 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
3131 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
3132 spin_unlock_irqrestore(&lockres->l_lock, flags);
3136 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
3138 /* make sure we never get here while waiting for an ast to
3140 BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
3142 /* is this necessary? */
3143 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3144 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
3145 spin_unlock_irqrestore(&lockres->l_lock, flags);
3147 mlog(0, "lock %s\n", lockres->l_name);
3149 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb, lkm_flags);
3151 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3152 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
3153 ocfs2_dlm_dump_lksb(&lockres->l_lksb);
3156 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3159 ocfs2_wait_on_busy_lock(lockres);
3164 static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3165 struct ocfs2_lock_res *lockres);
3167 /* Mark the lockres as being dropped. It will no longer be
3168 * queued if blocking, but we still may have to wait on it
3169 * being dequeued from the downconvert thread before we can consider
3172 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3173 void ocfs2_mark_lockres_freeing(struct ocfs2_super *osb,
3174 struct ocfs2_lock_res *lockres)
3177 struct ocfs2_mask_waiter mw;
3178 unsigned long flags, flags2;
3180 ocfs2_init_mask_waiter(&mw);
3182 spin_lock_irqsave(&lockres->l_lock, flags);
3183 lockres->l_flags |= OCFS2_LOCK_FREEING;
3184 if (lockres->l_flags & OCFS2_LOCK_QUEUED && current == osb->dc_task) {
3186 * We know the downconvert is queued but not in progress
3187 * because we are the downconvert thread and processing
3188 * different lock. So we can just remove the lock from the
3189 * queue. This is not only an optimization but also a way
3190 * to avoid the following deadlock:
3191 * ocfs2_dentry_post_unlock()
3192 * ocfs2_dentry_lock_put()
3193 * ocfs2_drop_dentry_lock()
3195 * ocfs2_evict_inode()
3196 * ocfs2_clear_inode()
3197 * ocfs2_mark_lockres_freeing()
3198 * ... blocks waiting for OCFS2_LOCK_QUEUED
3199 * since we are the downconvert thread which
3200 * should clear the flag.
3202 spin_unlock_irqrestore(&lockres->l_lock, flags);
3203 spin_lock_irqsave(&osb->dc_task_lock, flags2);
3204 list_del_init(&lockres->l_blocked_list);
3205 osb->blocked_lock_count--;
3206 spin_unlock_irqrestore(&osb->dc_task_lock, flags2);
3208 * Warn if we recurse into another post_unlock call. Strictly
3209 * speaking it isn't a problem but we need to be careful if
3210 * that happens (stack overflow, deadlocks, ...) so warn if
3211 * ocfs2 grows a path for which this can happen.
3213 WARN_ON_ONCE(lockres->l_ops->post_unlock);
3214 /* Since the lock is freeing we don't do much in the fn below */
3215 ocfs2_process_blocked_lock(osb, lockres);
3218 while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
3219 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
3220 spin_unlock_irqrestore(&lockres->l_lock, flags);
3222 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
3224 status = ocfs2_wait_for_mask(&mw);
3228 spin_lock_irqsave(&lockres->l_lock, flags);
3230 spin_unlock_irqrestore(&lockres->l_lock, flags);
3233 void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
3234 struct ocfs2_lock_res *lockres)
3238 ocfs2_mark_lockres_freeing(osb, lockres);
3239 ret = ocfs2_drop_lock(osb, lockres);
3244 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
3246 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
3247 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
3248 ocfs2_simple_drop_lockres(osb, &osb->osb_nfs_sync_lockres);
3249 ocfs2_simple_drop_lockres(osb, &osb->osb_orphan_scan.os_lockres);
3252 int ocfs2_drop_inode_locks(struct inode *inode)
3256 /* No need to call ocfs2_mark_lockres_freeing here -
3257 * ocfs2_clear_inode has done it for us. */
3259 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3260 &OCFS2_I(inode)->ip_open_lockres);
3266 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3267 &OCFS2_I(inode)->ip_inode_lockres);
3270 if (err < 0 && !status)
3273 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
3274 &OCFS2_I(inode)->ip_rw_lockres);
3277 if (err < 0 && !status)
3283 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
3286 assert_spin_locked(&lockres->l_lock);
3288 BUG_ON(lockres->l_blocking <= DLM_LOCK_NL);
3290 if (lockres->l_level <= new_level) {
3291 mlog(ML_ERROR, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3292 "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3293 "block %d, pgen %d\n", lockres->l_name, lockres->l_level,
3294 new_level, list_empty(&lockres->l_blocked_list),
3295 list_empty(&lockres->l_mask_waiters), lockres->l_type,
3296 lockres->l_flags, lockres->l_ro_holders,
3297 lockres->l_ex_holders, lockres->l_action,
3298 lockres->l_unlock_action, lockres->l_requested,
3299 lockres->l_blocking, lockres->l_pending_gen);
3303 mlog(ML_BASTS, "lockres %s, level %d => %d, blocking %d\n",
3304 lockres->l_name, lockres->l_level, new_level, lockres->l_blocking);
3306 lockres->l_action = OCFS2_AST_DOWNCONVERT;
3307 lockres->l_requested = new_level;
3308 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
3309 return lockres_set_pending(lockres);
3312 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
3313 struct ocfs2_lock_res *lockres,
3316 unsigned int generation)
3319 u32 dlm_flags = DLM_LKF_CONVERT;
3321 mlog(ML_BASTS, "lockres %s, level %d => %d\n", lockres->l_name,
3322 lockres->l_level, new_level);
3325 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3326 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3327 * we can recover correctly from node failure. Otherwise, we may get
3328 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALIDÂ being set.
3330 if (!ocfs2_is_o2cb_active() &&
3331 lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
3335 dlm_flags |= DLM_LKF_VALBLK;
3337 ret = ocfs2_dlm_lock(osb->cconn,
3342 OCFS2_LOCK_ID_MAX_LEN - 1);
3343 lockres_clear_pending(lockres, generation, osb);
3345 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
3346 ocfs2_recover_from_dlm_error(lockres, 1);
3355 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3356 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
3357 struct ocfs2_lock_res *lockres)
3359 assert_spin_locked(&lockres->l_lock);
3361 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
3362 /* If we're already trying to cancel a lock conversion
3363 * then just drop the spinlock and allow the caller to
3364 * requeue this lock. */
3365 mlog(ML_BASTS, "lockres %s, skip convert\n", lockres->l_name);
3369 /* were we in a convert when we got the bast fire? */
3370 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
3371 lockres->l_action != OCFS2_AST_DOWNCONVERT);
3372 /* set things up for the unlockast to know to just
3373 * clear out the ast_action and unset busy, etc. */
3374 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
3376 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
3377 "lock %s, invalid flags: 0x%lx\n",
3378 lockres->l_name, lockres->l_flags);
3380 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3385 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
3386 struct ocfs2_lock_res *lockres)
3390 ret = ocfs2_dlm_unlock(osb->cconn, &lockres->l_lksb,
3393 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret, lockres);
3394 ocfs2_recover_from_dlm_error(lockres, 0);
3397 mlog(ML_BASTS, "lockres %s\n", lockres->l_name);
3402 static int ocfs2_unblock_lock(struct ocfs2_super *osb,
3403 struct ocfs2_lock_res *lockres,
3404 struct ocfs2_unblock_ctl *ctl)
3406 unsigned long flags;
3414 spin_lock_irqsave(&lockres->l_lock, flags);
3418 * Is it still blocking? If not, we have no more work to do.
3420 if (!(lockres->l_flags & OCFS2_LOCK_BLOCKED)) {
3421 BUG_ON(lockres->l_blocking != DLM_LOCK_NL);
3422 spin_unlock_irqrestore(&lockres->l_lock, flags);
3427 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
3429 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3430 * exists entirely for one reason - another thread has set
3431 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3433 * If we do ocfs2_cancel_convert() before the other thread
3434 * calls dlm_lock(), our cancel will do nothing. We will
3435 * get no ast, and we will have no way of knowing the
3436 * cancel failed. Meanwhile, the other thread will call
3437 * into dlm_lock() and wait...forever.
3439 * Why forever? Because another node has asked for the
3440 * lock first; that's why we're here in unblock_lock().
3442 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3443 * set, we just requeue the unblock. Only when the other
3444 * thread has called dlm_lock() and cleared PENDING will
3445 * we then cancel their request.
3447 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3448 * at the same time they set OCFS2_DLM_BUSY. They must
3449 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3451 if (lockres->l_flags & OCFS2_LOCK_PENDING) {
3452 mlog(ML_BASTS, "lockres %s, ReQ: Pending\n",
3458 ret = ocfs2_prepare_cancel_convert(osb, lockres);
3459 spin_unlock_irqrestore(&lockres->l_lock, flags);
3461 ret = ocfs2_cancel_convert(osb, lockres);
3469 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3470 * set when the ast is received for an upconvert just before the
3471 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3472 * on the heels of the ast, we want to delay the downconvert just
3473 * enough to allow the up requestor to do its task. Because this
3474 * lock is in the blocked queue, the lock will be downconverted
3475 * as soon as the requestor is done with the lock.
3477 if (lockres->l_flags & OCFS2_LOCK_UPCONVERT_FINISHING)
3481 * How can we block and yet be at NL? We were trying to upconvert
3482 * from NL and got canceled. The code comes back here, and now
3483 * we notice and clear BLOCKING.
3485 if (lockres->l_level == DLM_LOCK_NL) {
3486 BUG_ON(lockres->l_ex_holders || lockres->l_ro_holders);
3487 mlog(ML_BASTS, "lockres %s, Aborting dc\n", lockres->l_name);
3488 lockres->l_blocking = DLM_LOCK_NL;
3489 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
3490 spin_unlock_irqrestore(&lockres->l_lock, flags);
3494 /* if we're blocking an exclusive and we have *any* holders,
3496 if ((lockres->l_blocking == DLM_LOCK_EX)
3497 && (lockres->l_ex_holders || lockres->l_ro_holders)) {
3498 mlog(ML_BASTS, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3499 lockres->l_name, lockres->l_ex_holders,
3500 lockres->l_ro_holders);
3504 /* If it's a PR we're blocking, then only
3505 * requeue if we've got any EX holders */
3506 if (lockres->l_blocking == DLM_LOCK_PR &&
3507 lockres->l_ex_holders) {
3508 mlog(ML_BASTS, "lockres %s, ReQ: EX Holders %u\n",
3509 lockres->l_name, lockres->l_ex_holders);
3514 * Can we get a lock in this state if the holder counts are
3515 * zero? The meta data unblock code used to check this.
3517 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
3518 && (lockres->l_flags & OCFS2_LOCK_REFRESHING)) {
3519 mlog(ML_BASTS, "lockres %s, ReQ: Lock Refreshing\n",
3524 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
3526 if (lockres->l_ops->check_downconvert
3527 && !lockres->l_ops->check_downconvert(lockres, new_level)) {
3528 mlog(ML_BASTS, "lockres %s, ReQ: Checkpointing\n",
3533 /* If we get here, then we know that there are no more
3534 * incompatible holders (and anyone asking for an incompatible
3535 * lock is blocked). We can now downconvert the lock */
3536 if (!lockres->l_ops->downconvert_worker)
3539 /* Some lockres types want to do a bit of work before
3540 * downconverting a lock. Allow that here. The worker function
3541 * may sleep, so we save off a copy of what we're blocking as
3542 * it may change while we're not holding the spin lock. */
3543 blocking = lockres->l_blocking;
3544 level = lockres->l_level;
3545 spin_unlock_irqrestore(&lockres->l_lock, flags);
3547 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
3549 if (ctl->unblock_action == UNBLOCK_STOP_POST) {
3550 mlog(ML_BASTS, "lockres %s, UNBLOCK_STOP_POST\n",
3555 spin_lock_irqsave(&lockres->l_lock, flags);
3556 if ((blocking != lockres->l_blocking) || (level != lockres->l_level)) {
3557 /* If this changed underneath us, then we can't drop
3559 mlog(ML_BASTS, "lockres %s, block=%d:%d, level=%d:%d, "
3560 "Recheck\n", lockres->l_name, blocking,
3561 lockres->l_blocking, level, lockres->l_level);
3568 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
3569 if (lockres->l_level == DLM_LOCK_EX)
3573 * We only set the lvb if the lock has been fully
3574 * refreshed - otherwise we risk setting stale
3575 * data. Otherwise, there's no need to actually clear
3576 * out the lvb here as it's value is still valid.
3578 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
3579 lockres->l_ops->set_lvb(lockres);
3582 gen = ocfs2_prepare_downconvert(lockres, new_level);
3583 spin_unlock_irqrestore(&lockres->l_lock, flags);
3584 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
3593 spin_unlock_irqrestore(&lockres->l_lock, flags);
3599 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
3602 struct inode *inode;
3603 struct address_space *mapping;
3604 struct ocfs2_inode_info *oi;
3606 inode = ocfs2_lock_res_inode(lockres);
3607 mapping = inode->i_mapping;
3609 if (S_ISDIR(inode->i_mode)) {
3610 oi = OCFS2_I(inode);
3611 oi->ip_dir_lock_gen++;
3612 mlog(0, "generation: %u\n", oi->ip_dir_lock_gen);
3616 if (!S_ISREG(inode->i_mode))
3620 * We need this before the filemap_fdatawrite() so that it can
3621 * transfer the dirty bit from the PTE to the
3622 * page. Unfortunately this means that even for EX->PR
3623 * downconverts, we'll lose our mappings and have to build
3626 unmap_mapping_range(mapping, 0, 0, 0);
3628 if (filemap_fdatawrite(mapping)) {
3629 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
3630 (unsigned long long)OCFS2_I(inode)->ip_blkno);
3632 sync_mapping_buffers(mapping);
3633 if (blocking == DLM_LOCK_EX) {
3634 truncate_inode_pages(mapping, 0);
3636 /* We only need to wait on the I/O if we're not also
3637 * truncating pages because truncate_inode_pages waits
3638 * for us above. We don't truncate pages if we're
3639 * blocking anything < EXMODE because we want to keep
3640 * them around in that case. */
3641 filemap_fdatawait(mapping);
3645 return UNBLOCK_CONTINUE;
3648 static int ocfs2_ci_checkpointed(struct ocfs2_caching_info *ci,
3649 struct ocfs2_lock_res *lockres,
3652 int checkpointed = ocfs2_ci_fully_checkpointed(ci);
3654 BUG_ON(new_level != DLM_LOCK_NL && new_level != DLM_LOCK_PR);
3655 BUG_ON(lockres->l_level != DLM_LOCK_EX && !checkpointed);
3660 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci)));
3664 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
3667 struct inode *inode = ocfs2_lock_res_inode(lockres);
3669 return ocfs2_ci_checkpointed(INODE_CACHE(inode), lockres, new_level);
3672 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
3674 struct inode *inode = ocfs2_lock_res_inode(lockres);
3676 __ocfs2_stuff_meta_lvb(inode);
3680 * Does the final reference drop on our dentry lock. Right now this
3681 * happens in the downconvert thread, but we could choose to simplify the
3682 * dlmglue API and push these off to the ocfs2_wq in the future.
3684 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
3685 struct ocfs2_lock_res *lockres)
3687 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3688 ocfs2_dentry_lock_put(osb, dl);
3692 * d_delete() matching dentries before the lock downconvert.
3694 * At this point, any process waiting to destroy the
3695 * dentry_lock due to last ref count is stopped by the
3696 * OCFS2_LOCK_QUEUED flag.
3698 * We have two potential problems
3700 * 1) If we do the last reference drop on our dentry_lock (via dput)
3701 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3702 * the downconvert to finish. Instead we take an elevated
3703 * reference and push the drop until after we've completed our
3704 * unblock processing.
3706 * 2) There might be another process with a final reference,
3707 * waiting on us to finish processing. If this is the case, we
3708 * detect it and exit out - there's no more dentries anyway.
3710 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
3713 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
3714 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
3715 struct dentry *dentry;
3716 unsigned long flags;
3720 * This node is blocking another node from getting a read
3721 * lock. This happens when we've renamed within a
3722 * directory. We've forced the other nodes to d_delete(), but
3723 * we never actually dropped our lock because it's still
3724 * valid. The downconvert code will retain a PR for this node,
3725 * so there's no further work to do.
3727 if (blocking == DLM_LOCK_PR)
3728 return UNBLOCK_CONTINUE;
3731 * Mark this inode as potentially orphaned. The code in
3732 * ocfs2_delete_inode() will figure out whether it actually
3733 * needs to be freed or not.
3735 spin_lock(&oi->ip_lock);
3736 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
3737 spin_unlock(&oi->ip_lock);
3740 * Yuck. We need to make sure however that the check of
3741 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3742 * respect to a reference decrement or the setting of that
3745 spin_lock_irqsave(&lockres->l_lock, flags);
3746 spin_lock(&dentry_attach_lock);
3747 if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
3752 spin_unlock(&dentry_attach_lock);
3753 spin_unlock_irqrestore(&lockres->l_lock, flags);
3755 mlog(0, "extra_ref = %d\n", extra_ref);
3758 * We have a process waiting on us in ocfs2_dentry_iput(),
3759 * which means we can't have any more outstanding
3760 * aliases. There's no need to do any more work.
3763 return UNBLOCK_CONTINUE;
3765 spin_lock(&dentry_attach_lock);
3767 dentry = ocfs2_find_local_alias(dl->dl_inode,
3768 dl->dl_parent_blkno, 1);
3771 spin_unlock(&dentry_attach_lock);
3773 if (S_ISDIR(dl->dl_inode->i_mode))
3774 shrink_dcache_parent(dentry);
3776 mlog(0, "d_delete(%pd);\n", dentry);
3779 * The following dcache calls may do an
3780 * iput(). Normally we don't want that from the
3781 * downconverting thread, but in this case it's ok
3782 * because the requesting node already has an
3783 * exclusive lock on the inode, so it can't be queued
3784 * for a downconvert.
3789 spin_lock(&dentry_attach_lock);
3791 spin_unlock(&dentry_attach_lock);
3794 * If we are the last holder of this dentry lock, there is no
3795 * reason to downconvert so skip straight to the unlock.
3797 if (dl->dl_count == 1)
3798 return UNBLOCK_STOP_POST;
3800 return UNBLOCK_CONTINUE_POST;
3803 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res *lockres,
3806 struct ocfs2_refcount_tree *tree =
3807 ocfs2_lock_res_refcount_tree(lockres);
3809 return ocfs2_ci_checkpointed(&tree->rf_ci, lockres, new_level);
3812 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res *lockres,
3815 struct ocfs2_refcount_tree *tree =
3816 ocfs2_lock_res_refcount_tree(lockres);
3818 ocfs2_metadata_cache_purge(&tree->rf_ci);
3820 return UNBLOCK_CONTINUE;
3823 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res *lockres)
3825 struct ocfs2_qinfo_lvb *lvb;
3826 struct ocfs2_mem_dqinfo *oinfo = ocfs2_lock_res_qinfo(lockres);
3827 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3828 oinfo->dqi_gi.dqi_type);
3830 lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3831 lvb->lvb_version = OCFS2_QINFO_LVB_VERSION;
3832 lvb->lvb_bgrace = cpu_to_be32(info->dqi_bgrace);
3833 lvb->lvb_igrace = cpu_to_be32(info->dqi_igrace);
3834 lvb->lvb_syncms = cpu_to_be32(oinfo->dqi_syncms);
3835 lvb->lvb_blocks = cpu_to_be32(oinfo->dqi_gi.dqi_blocks);
3836 lvb->lvb_free_blk = cpu_to_be32(oinfo->dqi_gi.dqi_free_blk);
3837 lvb->lvb_free_entry = cpu_to_be32(oinfo->dqi_gi.dqi_free_entry);
3840 void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3842 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3843 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3844 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3846 if (!ocfs2_is_hard_readonly(osb) && !ocfs2_mount_local(osb))
3847 ocfs2_cluster_unlock(osb, lockres, level);
3850 static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo *oinfo)
3852 struct mem_dqinfo *info = sb_dqinfo(oinfo->dqi_gi.dqi_sb,
3853 oinfo->dqi_gi.dqi_type);
3854 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3855 struct ocfs2_qinfo_lvb *lvb = ocfs2_dlm_lvb(&lockres->l_lksb);
3856 struct buffer_head *bh = NULL;
3857 struct ocfs2_global_disk_dqinfo *gdinfo;
3860 if (ocfs2_dlm_lvb_valid(&lockres->l_lksb) &&
3861 lvb->lvb_version == OCFS2_QINFO_LVB_VERSION) {
3862 info->dqi_bgrace = be32_to_cpu(lvb->lvb_bgrace);
3863 info->dqi_igrace = be32_to_cpu(lvb->lvb_igrace);
3864 oinfo->dqi_syncms = be32_to_cpu(lvb->lvb_syncms);
3865 oinfo->dqi_gi.dqi_blocks = be32_to_cpu(lvb->lvb_blocks);
3866 oinfo->dqi_gi.dqi_free_blk = be32_to_cpu(lvb->lvb_free_blk);
3867 oinfo->dqi_gi.dqi_free_entry =
3868 be32_to_cpu(lvb->lvb_free_entry);
3870 status = ocfs2_read_quota_phys_block(oinfo->dqi_gqinode,
3871 oinfo->dqi_giblk, &bh);
3876 gdinfo = (struct ocfs2_global_disk_dqinfo *)
3877 (bh->b_data + OCFS2_GLOBAL_INFO_OFF);
3878 info->dqi_bgrace = le32_to_cpu(gdinfo->dqi_bgrace);
3879 info->dqi_igrace = le32_to_cpu(gdinfo->dqi_igrace);
3880 oinfo->dqi_syncms = le32_to_cpu(gdinfo->dqi_syncms);
3881 oinfo->dqi_gi.dqi_blocks = le32_to_cpu(gdinfo->dqi_blocks);
3882 oinfo->dqi_gi.dqi_free_blk = le32_to_cpu(gdinfo->dqi_free_blk);
3883 oinfo->dqi_gi.dqi_free_entry =
3884 le32_to_cpu(gdinfo->dqi_free_entry);
3886 ocfs2_track_lock_refresh(lockres);
3893 /* Lock quota info, this function expects at least shared lock on the quota file
3894 * so that we can safely refresh quota info from disk. */
3895 int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo *oinfo, int ex)
3897 struct ocfs2_lock_res *lockres = &oinfo->dqi_gqlock;
3898 struct ocfs2_super *osb = OCFS2_SB(oinfo->dqi_gi.dqi_sb);
3899 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3902 /* On RO devices, locking really isn't needed... */
3903 if (ocfs2_is_hard_readonly(osb)) {
3908 if (ocfs2_mount_local(osb))
3911 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3916 if (!ocfs2_should_refresh_lock_res(lockres))
3918 /* OK, we have the lock but we need to refresh the quota info */
3919 status = ocfs2_refresh_qinfo(oinfo);
3921 ocfs2_qinfo_unlock(oinfo, ex);
3922 ocfs2_complete_lock_res_refresh(lockres, status);
3927 int ocfs2_refcount_lock(struct ocfs2_refcount_tree *ref_tree, int ex)
3930 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3931 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3932 struct ocfs2_super *osb = lockres->l_priv;
3935 if (ocfs2_is_hard_readonly(osb))
3938 if (ocfs2_mount_local(osb))
3941 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
3948 void ocfs2_refcount_unlock(struct ocfs2_refcount_tree *ref_tree, int ex)
3950 int level = ex ? DLM_LOCK_EX : DLM_LOCK_PR;
3951 struct ocfs2_lock_res *lockres = &ref_tree->rf_lockres;
3952 struct ocfs2_super *osb = lockres->l_priv;
3954 if (!ocfs2_mount_local(osb))
3955 ocfs2_cluster_unlock(osb, lockres, level);
3958 static void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3959 struct ocfs2_lock_res *lockres)
3962 struct ocfs2_unblock_ctl ctl = {0, 0,};
3963 unsigned long flags;
3965 /* Our reference to the lockres in this function can be
3966 * considered valid until we remove the OCFS2_LOCK_QUEUED
3970 BUG_ON(!lockres->l_ops);
3972 mlog(ML_BASTS, "lockres %s blocked\n", lockres->l_name);
3974 /* Detect whether a lock has been marked as going away while
3975 * the downconvert thread was processing other things. A lock can
3976 * still be marked with OCFS2_LOCK_FREEING after this check,
3977 * but short circuiting here will still save us some
3979 spin_lock_irqsave(&lockres->l_lock, flags);
3980 if (lockres->l_flags & OCFS2_LOCK_FREEING)
3982 spin_unlock_irqrestore(&lockres->l_lock, flags);
3984 status = ocfs2_unblock_lock(osb, lockres, &ctl);
3988 spin_lock_irqsave(&lockres->l_lock, flags);
3990 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
3991 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
3993 ocfs2_schedule_blocked_lock(osb, lockres);
3995 mlog(ML_BASTS, "lockres %s, requeue = %s.\n", lockres->l_name,
3996 ctl.requeue ? "yes" : "no");
3997 spin_unlock_irqrestore(&lockres->l_lock, flags);
3999 if (ctl.unblock_action != UNBLOCK_CONTINUE
4000 && lockres->l_ops->post_unlock)
4001 lockres->l_ops->post_unlock(osb, lockres);
4004 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
4005 struct ocfs2_lock_res *lockres)
4007 unsigned long flags;
4009 assert_spin_locked(&lockres->l_lock);
4011 if (lockres->l_flags & OCFS2_LOCK_FREEING) {
4012 /* Do not schedule a lock for downconvert when it's on
4013 * the way to destruction - any nodes wanting access
4014 * to the resource will get it soon. */
4015 mlog(ML_BASTS, "lockres %s won't be scheduled: flags 0x%lx\n",
4016 lockres->l_name, lockres->l_flags);
4020 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
4022 spin_lock_irqsave(&osb->dc_task_lock, flags);
4023 if (list_empty(&lockres->l_blocked_list)) {
4024 list_add_tail(&lockres->l_blocked_list,
4025 &osb->blocked_lock_list);
4026 osb->blocked_lock_count++;
4028 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4031 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super *osb)
4033 unsigned long processed;
4034 unsigned long flags;
4035 struct ocfs2_lock_res *lockres;
4037 spin_lock_irqsave(&osb->dc_task_lock, flags);
4038 /* grab this early so we know to try again if a state change and
4039 * wake happens part-way through our work */
4040 osb->dc_work_sequence = osb->dc_wake_sequence;
4042 processed = osb->blocked_lock_count;
4044 * blocked lock processing in this loop might call iput which can
4045 * remove items off osb->blocked_lock_list. Downconvert up to
4046 * 'processed' number of locks, but stop short if we had some
4047 * removed in ocfs2_mark_lockres_freeing when downconverting.
4049 while (processed && !list_empty(&osb->blocked_lock_list)) {
4050 lockres = list_entry(osb->blocked_lock_list.next,
4051 struct ocfs2_lock_res, l_blocked_list);
4052 list_del_init(&lockres->l_blocked_list);
4053 osb->blocked_lock_count--;
4054 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4059 ocfs2_process_blocked_lock(osb, lockres);
4061 spin_lock_irqsave(&osb->dc_task_lock, flags);
4063 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4066 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super *osb)
4069 unsigned long flags;
4071 spin_lock_irqsave(&osb->dc_task_lock, flags);
4072 if (list_empty(&osb->blocked_lock_list))
4075 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4079 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super *osb)
4081 int should_wake = 0;
4082 unsigned long flags;
4084 spin_lock_irqsave(&osb->dc_task_lock, flags);
4085 if (osb->dc_work_sequence != osb->dc_wake_sequence)
4087 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4092 static int ocfs2_downconvert_thread(void *arg)
4095 struct ocfs2_super *osb = arg;
4097 /* only quit once we've been asked to stop and there is no more
4099 while (!(kthread_should_stop() &&
4100 ocfs2_downconvert_thread_lists_empty(osb))) {
4102 wait_event_interruptible(osb->dc_event,
4103 ocfs2_downconvert_thread_should_wake(osb) ||
4104 kthread_should_stop());
4106 mlog(0, "downconvert_thread: awoken\n");
4108 ocfs2_downconvert_thread_do_work(osb);
4111 osb->dc_task = NULL;
4115 void ocfs2_wake_downconvert_thread(struct ocfs2_super *osb)
4117 unsigned long flags;
4119 spin_lock_irqsave(&osb->dc_task_lock, flags);
4120 /* make sure the voting thread gets a swipe at whatever changes
4121 * the caller may have made to the voting state */
4122 osb->dc_wake_sequence++;
4123 spin_unlock_irqrestore(&osb->dc_task_lock, flags);
4124 wake_up(&osb->dc_event);