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/smp_lock.h>
31 #include <linux/crc32.h>
32 #include <linux/kthread.h>
33 #include <linux/pagemap.h>
34 #include <linux/debugfs.h>
35 #include <linux/seq_file.h>
37 #include <cluster/heartbeat.h>
38 #include <cluster/nodemanager.h>
39 #include <cluster/tcp.h>
41 #include <dlm/dlmapi.h>
43 #define MLOG_MASK_PREFIX ML_DLM_GLUE
44 #include <cluster/masklog.h>
51 #include "extent_map.h"
53 #include "heartbeat.h"
61 #include "buffer_head_io.h"
63 struct ocfs2_mask_waiter {
64 struct list_head mw_item;
66 struct completion mw_complete;
67 unsigned long mw_mask;
68 unsigned long mw_goal;
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);
75 * Return value from ->downconvert_worker functions.
77 * These control the precise actions of ocfs2_unblock_lock()
78 * and ocfs2_process_blocked_lock()
81 enum ocfs2_unblock_action {
82 UNBLOCK_CONTINUE = 0, /* Continue downconvert */
83 UNBLOCK_CONTINUE_POST = 1, /* Continue downconvert, fire
84 * ->post_unlock callback */
85 UNBLOCK_STOP_POST = 2, /* Do not downconvert, fire
86 * ->post_unlock() callback. */
89 struct ocfs2_unblock_ctl {
91 enum ocfs2_unblock_action unblock_action;
94 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
96 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres);
98 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
101 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
104 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
105 struct ocfs2_lock_res *lockres);
108 * OCFS2 Lock Resource Operations
110 * These fine tune the behavior of the generic dlmglue locking infrastructure.
112 * The most basic of lock types can point ->l_priv to their respective
113 * struct ocfs2_super and allow the default actions to manage things.
115 * Right now, each lock type also needs to implement an init function,
116 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
117 * should be called when the lock is no longer needed (i.e., object
120 struct ocfs2_lock_res_ops {
122 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
123 * this callback if ->l_priv is not an ocfs2_super pointer
125 struct ocfs2_super * (*get_osb)(struct ocfs2_lock_res *);
128 * Optionally called in the downconvert (or "vote") thread
129 * after a successful downconvert. The lockres will not be
130 * referenced after this callback is called, so it is safe to
133 * The exact semantics of when this is called are controlled
134 * by ->downconvert_worker()
136 void (*post_unlock)(struct ocfs2_super *, struct ocfs2_lock_res *);
139 * Allow a lock type to add checks to determine whether it is
140 * safe to downconvert a lock. Return 0 to re-queue the
141 * downconvert at a later time, nonzero to continue.
143 * For most locks, the default checks that there are no
144 * incompatible holders are sufficient.
146 * Called with the lockres spinlock held.
148 int (*check_downconvert)(struct ocfs2_lock_res *, int);
151 * Allows a lock type to populate the lock value block. This
152 * is called on downconvert, and when we drop a lock.
154 * Locks that want to use this should set LOCK_TYPE_USES_LVB
155 * in the flags field.
157 * Called with the lockres spinlock held.
159 void (*set_lvb)(struct ocfs2_lock_res *);
162 * Called from the downconvert thread when it is determined
163 * that a lock will be downconverted. This is called without
164 * any locks held so the function can do work that might
165 * schedule (syncing out data, etc).
167 * This should return any one of the ocfs2_unblock_action
168 * values, depending on what it wants the thread to do.
170 int (*downconvert_worker)(struct ocfs2_lock_res *, int);
173 * LOCK_TYPE_* flags which describe the specific requirements
174 * of a lock type. Descriptions of each individual flag follow.
180 * Some locks want to "refresh" potentially stale data when a
181 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
182 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
183 * individual lockres l_flags member from the ast function. It is
184 * expected that the locking wrapper will clear the
185 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
187 #define LOCK_TYPE_REQUIRES_REFRESH 0x1
190 * Indicate that a lock type makes use of the lock value block. The
191 * ->set_lvb lock type callback must be defined.
193 #define LOCK_TYPE_USES_LVB 0x2
195 static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops = {
196 .get_osb = ocfs2_get_inode_osb,
200 static struct ocfs2_lock_res_ops ocfs2_inode_meta_lops = {
201 .get_osb = ocfs2_get_inode_osb,
202 .check_downconvert = ocfs2_check_meta_downconvert,
203 .set_lvb = ocfs2_set_meta_lvb,
204 .flags = LOCK_TYPE_REQUIRES_REFRESH|LOCK_TYPE_USES_LVB,
207 static struct ocfs2_lock_res_ops ocfs2_inode_data_lops = {
208 .get_osb = ocfs2_get_inode_osb,
209 .downconvert_worker = ocfs2_data_convert_worker,
213 static struct ocfs2_lock_res_ops ocfs2_super_lops = {
214 .flags = LOCK_TYPE_REQUIRES_REFRESH,
217 static struct ocfs2_lock_res_ops ocfs2_rename_lops = {
221 static struct ocfs2_lock_res_ops ocfs2_dentry_lops = {
222 .get_osb = ocfs2_get_dentry_osb,
223 .post_unlock = ocfs2_dentry_post_unlock,
224 .downconvert_worker = ocfs2_dentry_convert_worker,
228 static struct ocfs2_lock_res_ops ocfs2_inode_open_lops = {
229 .get_osb = ocfs2_get_inode_osb,
233 static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res *lockres)
235 return lockres->l_type == OCFS2_LOCK_TYPE_META ||
236 lockres->l_type == OCFS2_LOCK_TYPE_DATA ||
237 lockres->l_type == OCFS2_LOCK_TYPE_RW ||
238 lockres->l_type == OCFS2_LOCK_TYPE_OPEN;
241 static inline struct inode *ocfs2_lock_res_inode(struct ocfs2_lock_res *lockres)
243 BUG_ON(!ocfs2_is_inode_lock(lockres));
245 return (struct inode *) lockres->l_priv;
248 static inline struct ocfs2_dentry_lock *ocfs2_lock_res_dl(struct ocfs2_lock_res *lockres)
250 BUG_ON(lockres->l_type != OCFS2_LOCK_TYPE_DENTRY);
252 return (struct ocfs2_dentry_lock *)lockres->l_priv;
255 static inline struct ocfs2_super *ocfs2_get_lockres_osb(struct ocfs2_lock_res *lockres)
257 if (lockres->l_ops->get_osb)
258 return lockres->l_ops->get_osb(lockres);
260 return (struct ocfs2_super *)lockres->l_priv;
263 static int ocfs2_lock_create(struct ocfs2_super *osb,
264 struct ocfs2_lock_res *lockres,
267 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
269 static void ocfs2_cluster_unlock(struct ocfs2_super *osb,
270 struct ocfs2_lock_res *lockres,
272 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres);
273 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres);
274 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres);
275 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres, int level);
276 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
277 struct ocfs2_lock_res *lockres);
278 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
280 #define ocfs2_log_dlm_error(_func, _stat, _lockres) do { \
281 mlog(ML_ERROR, "Dlm error \"%s\" while calling %s on " \
282 "resource %s: %s\n", dlm_errname(_stat), _func, \
283 _lockres->l_name, dlm_errmsg(_stat)); \
285 static void ocfs2_vote_on_unlock(struct ocfs2_super *osb,
286 struct ocfs2_lock_res *lockres);
287 static int ocfs2_meta_lock_update(struct inode *inode,
288 struct buffer_head **bh);
289 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb);
290 static inline int ocfs2_highest_compat_lock_level(int level);
292 static void ocfs2_build_lock_name(enum ocfs2_lock_type type,
301 BUG_ON(type >= OCFS2_NUM_LOCK_TYPES);
303 len = snprintf(name, OCFS2_LOCK_ID_MAX_LEN, "%c%s%016llx%08x",
304 ocfs2_lock_type_char(type), OCFS2_LOCK_ID_PAD,
305 (long long)blkno, generation);
307 BUG_ON(len != (OCFS2_LOCK_ID_MAX_LEN - 1));
309 mlog(0, "built lock resource with name: %s\n", name);
314 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock);
316 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res *res,
317 struct ocfs2_dlm_debug *dlm_debug)
319 mlog(0, "Add tracking for lockres %s\n", res->l_name);
321 spin_lock(&ocfs2_dlm_tracking_lock);
322 list_add(&res->l_debug_list, &dlm_debug->d_lockres_tracking);
323 spin_unlock(&ocfs2_dlm_tracking_lock);
326 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res *res)
328 spin_lock(&ocfs2_dlm_tracking_lock);
329 if (!list_empty(&res->l_debug_list))
330 list_del_init(&res->l_debug_list);
331 spin_unlock(&ocfs2_dlm_tracking_lock);
334 static void ocfs2_lock_res_init_common(struct ocfs2_super *osb,
335 struct ocfs2_lock_res *res,
336 enum ocfs2_lock_type type,
337 struct ocfs2_lock_res_ops *ops,
344 res->l_level = LKM_IVMODE;
345 res->l_requested = LKM_IVMODE;
346 res->l_blocking = LKM_IVMODE;
347 res->l_action = OCFS2_AST_INVALID;
348 res->l_unlock_action = OCFS2_UNLOCK_INVALID;
350 res->l_flags = OCFS2_LOCK_INITIALIZED;
352 ocfs2_add_lockres_tracking(res, osb->osb_dlm_debug);
355 void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
357 /* This also clears out the lock status block */
358 memset(res, 0, sizeof(struct ocfs2_lock_res));
359 spin_lock_init(&res->l_lock);
360 init_waitqueue_head(&res->l_event);
361 INIT_LIST_HEAD(&res->l_blocked_list);
362 INIT_LIST_HEAD(&res->l_mask_waiters);
365 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
366 enum ocfs2_lock_type type,
367 unsigned int generation,
370 struct ocfs2_lock_res_ops *ops;
373 case OCFS2_LOCK_TYPE_RW:
374 ops = &ocfs2_inode_rw_lops;
376 case OCFS2_LOCK_TYPE_META:
377 ops = &ocfs2_inode_meta_lops;
379 case OCFS2_LOCK_TYPE_DATA:
380 ops = &ocfs2_inode_data_lops;
382 case OCFS2_LOCK_TYPE_OPEN:
383 ops = &ocfs2_inode_open_lops;
386 mlog_bug_on_msg(1, "type: %d\n", type);
387 ops = NULL; /* thanks, gcc */
391 ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
392 generation, res->l_name);
393 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
396 static struct ocfs2_super *ocfs2_get_inode_osb(struct ocfs2_lock_res *lockres)
398 struct inode *inode = ocfs2_lock_res_inode(lockres);
400 return OCFS2_SB(inode->i_sb);
403 static __u64 ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res *lockres)
405 __be64 inode_blkno_be;
407 memcpy(&inode_blkno_be, &lockres->l_name[OCFS2_DENTRY_LOCK_INO_START],
410 return be64_to_cpu(inode_blkno_be);
413 static struct ocfs2_super *ocfs2_get_dentry_osb(struct ocfs2_lock_res *lockres)
415 struct ocfs2_dentry_lock *dl = lockres->l_priv;
417 return OCFS2_SB(dl->dl_inode->i_sb);
420 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
421 u64 parent, struct inode *inode)
424 u64 inode_blkno = OCFS2_I(inode)->ip_blkno;
425 __be64 inode_blkno_be = cpu_to_be64(inode_blkno);
426 struct ocfs2_lock_res *lockres = &dl->dl_lockres;
428 ocfs2_lock_res_init_once(lockres);
431 * Unfortunately, the standard lock naming scheme won't work
432 * here because we have two 16 byte values to use. Instead,
433 * we'll stuff the inode number as a binary value. We still
434 * want error prints to show something without garbling the
435 * display, so drop a null byte in there before the inode
436 * number. A future version of OCFS2 will likely use all
437 * binary lock names. The stringified names have been a
438 * tremendous aid in debugging, but now that the debugfs
439 * interface exists, we can mangle things there if need be.
441 * NOTE: We also drop the standard "pad" value (the total lock
442 * name size stays the same though - the last part is all
443 * zeros due to the memset in ocfs2_lock_res_init_once()
445 len = snprintf(lockres->l_name, OCFS2_DENTRY_LOCK_INO_START,
447 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY),
450 BUG_ON(len != (OCFS2_DENTRY_LOCK_INO_START - 1));
452 memcpy(&lockres->l_name[OCFS2_DENTRY_LOCK_INO_START], &inode_blkno_be,
455 ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), lockres,
456 OCFS2_LOCK_TYPE_DENTRY, &ocfs2_dentry_lops,
460 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res *res,
461 struct ocfs2_super *osb)
463 /* Superblock lockres doesn't come from a slab so we call init
464 * once on it manually. */
465 ocfs2_lock_res_init_once(res);
466 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER, OCFS2_SUPER_BLOCK_BLKNO,
468 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_SUPER,
469 &ocfs2_super_lops, osb);
472 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res *res,
473 struct ocfs2_super *osb)
475 /* Rename lockres doesn't come from a slab so we call init
476 * once on it manually. */
477 ocfs2_lock_res_init_once(res);
478 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME, 0, 0, res->l_name);
479 ocfs2_lock_res_init_common(osb, res, OCFS2_LOCK_TYPE_RENAME,
480 &ocfs2_rename_lops, osb);
483 void ocfs2_lock_res_free(struct ocfs2_lock_res *res)
487 if (!(res->l_flags & OCFS2_LOCK_INITIALIZED))
490 ocfs2_remove_lockres_tracking(res);
492 mlog_bug_on_msg(!list_empty(&res->l_blocked_list),
493 "Lockres %s is on the blocked list\n",
495 mlog_bug_on_msg(!list_empty(&res->l_mask_waiters),
496 "Lockres %s has mask waiters pending\n",
498 mlog_bug_on_msg(spin_is_locked(&res->l_lock),
499 "Lockres %s is locked\n",
501 mlog_bug_on_msg(res->l_ro_holders,
502 "Lockres %s has %u ro holders\n",
503 res->l_name, res->l_ro_holders);
504 mlog_bug_on_msg(res->l_ex_holders,
505 "Lockres %s has %u ex holders\n",
506 res->l_name, res->l_ex_holders);
508 /* Need to clear out the lock status block for the dlm */
509 memset(&res->l_lksb, 0, sizeof(res->l_lksb));
515 static inline void ocfs2_inc_holders(struct ocfs2_lock_res *lockres,
524 lockres->l_ex_holders++;
527 lockres->l_ro_holders++;
536 static inline void ocfs2_dec_holders(struct ocfs2_lock_res *lockres,
545 BUG_ON(!lockres->l_ex_holders);
546 lockres->l_ex_holders--;
549 BUG_ON(!lockres->l_ro_holders);
550 lockres->l_ro_holders--;
558 /* WARNING: This function lives in a world where the only three lock
559 * levels are EX, PR, and NL. It *will* have to be adjusted when more
560 * lock types are added. */
561 static inline int ocfs2_highest_compat_lock_level(int level)
563 int new_level = LKM_EXMODE;
565 if (level == LKM_EXMODE)
566 new_level = LKM_NLMODE;
567 else if (level == LKM_PRMODE)
568 new_level = LKM_PRMODE;
572 static void lockres_set_flags(struct ocfs2_lock_res *lockres,
573 unsigned long newflags)
575 struct list_head *pos, *tmp;
576 struct ocfs2_mask_waiter *mw;
578 assert_spin_locked(&lockres->l_lock);
580 lockres->l_flags = newflags;
582 list_for_each_safe(pos, tmp, &lockres->l_mask_waiters) {
583 mw = list_entry(pos, struct ocfs2_mask_waiter, mw_item);
584 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
587 list_del_init(&mw->mw_item);
589 complete(&mw->mw_complete);
592 static void lockres_or_flags(struct ocfs2_lock_res *lockres, unsigned long or)
594 lockres_set_flags(lockres, lockres->l_flags | or);
596 static void lockres_clear_flags(struct ocfs2_lock_res *lockres,
599 lockres_set_flags(lockres, lockres->l_flags & ~clear);
602 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res *lockres)
606 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
607 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
608 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
609 BUG_ON(lockres->l_blocking <= LKM_NLMODE);
611 lockres->l_level = lockres->l_requested;
612 if (lockres->l_level <=
613 ocfs2_highest_compat_lock_level(lockres->l_blocking)) {
614 lockres->l_blocking = LKM_NLMODE;
615 lockres_clear_flags(lockres, OCFS2_LOCK_BLOCKED);
617 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
622 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res *lockres)
626 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BUSY));
627 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_ATTACHED));
629 /* Convert from RO to EX doesn't really need anything as our
630 * information is already up to data. Convert from NL to
631 * *anything* however should mark ourselves as needing an
633 if (lockres->l_level == LKM_NLMODE &&
634 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
635 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
637 lockres->l_level = lockres->l_requested;
638 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
643 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res *lockres)
647 BUG_ON((!lockres->l_flags & OCFS2_LOCK_BUSY));
648 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
650 if (lockres->l_requested > LKM_NLMODE &&
651 !(lockres->l_flags & OCFS2_LOCK_LOCAL) &&
652 lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
653 lockres_or_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
655 lockres->l_level = lockres->l_requested;
656 lockres_or_flags(lockres, OCFS2_LOCK_ATTACHED);
657 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
662 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res *lockres,
665 int needs_downconvert = 0;
668 assert_spin_locked(&lockres->l_lock);
670 lockres_or_flags(lockres, OCFS2_LOCK_BLOCKED);
672 if (level > lockres->l_blocking) {
673 /* only schedule a downconvert if we haven't already scheduled
674 * one that goes low enough to satisfy the level we're
675 * blocking. this also catches the case where we get
677 if (ocfs2_highest_compat_lock_level(level) <
678 ocfs2_highest_compat_lock_level(lockres->l_blocking))
679 needs_downconvert = 1;
681 lockres->l_blocking = level;
684 mlog_exit(needs_downconvert);
685 return needs_downconvert;
688 static void ocfs2_blocking_ast(void *opaque, int level)
690 struct ocfs2_lock_res *lockres = opaque;
691 struct ocfs2_super *osb = ocfs2_get_lockres_osb(lockres);
692 int needs_downconvert;
695 BUG_ON(level <= LKM_NLMODE);
697 mlog(0, "BAST fired for lockres %s, blocking %d, level %d type %s\n",
698 lockres->l_name, level, lockres->l_level,
699 ocfs2_lock_type_string(lockres->l_type));
701 spin_lock_irqsave(&lockres->l_lock, flags);
702 needs_downconvert = ocfs2_generic_handle_bast(lockres, level);
703 if (needs_downconvert)
704 ocfs2_schedule_blocked_lock(osb, lockres);
705 spin_unlock_irqrestore(&lockres->l_lock, flags);
707 wake_up(&lockres->l_event);
709 ocfs2_kick_vote_thread(osb);
712 static void ocfs2_locking_ast(void *opaque)
714 struct ocfs2_lock_res *lockres = opaque;
715 struct dlm_lockstatus *lksb = &lockres->l_lksb;
718 spin_lock_irqsave(&lockres->l_lock, flags);
720 if (lksb->status != DLM_NORMAL) {
721 mlog(ML_ERROR, "lockres %s: lksb status value of %u!\n",
722 lockres->l_name, lksb->status);
723 spin_unlock_irqrestore(&lockres->l_lock, flags);
727 switch(lockres->l_action) {
728 case OCFS2_AST_ATTACH:
729 ocfs2_generic_handle_attach_action(lockres);
730 lockres_clear_flags(lockres, OCFS2_LOCK_LOCAL);
732 case OCFS2_AST_CONVERT:
733 ocfs2_generic_handle_convert_action(lockres);
735 case OCFS2_AST_DOWNCONVERT:
736 ocfs2_generic_handle_downconvert_action(lockres);
739 mlog(ML_ERROR, "lockres %s: ast fired with invalid action: %u "
740 "lockres flags = 0x%lx, unlock action: %u\n",
741 lockres->l_name, lockres->l_action, lockres->l_flags,
742 lockres->l_unlock_action);
746 /* set it to something invalid so if we get called again we
748 lockres->l_action = OCFS2_AST_INVALID;
750 wake_up(&lockres->l_event);
751 spin_unlock_irqrestore(&lockres->l_lock, flags);
754 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res *lockres,
760 spin_lock_irqsave(&lockres->l_lock, flags);
761 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
763 lockres->l_action = OCFS2_AST_INVALID;
765 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
766 spin_unlock_irqrestore(&lockres->l_lock, flags);
768 wake_up(&lockres->l_event);
772 /* Note: If we detect another process working on the lock (i.e.,
773 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
774 * to do the right thing in that case.
776 static int ocfs2_lock_create(struct ocfs2_super *osb,
777 struct ocfs2_lock_res *lockres,
782 enum dlm_status status = DLM_NORMAL;
787 mlog(0, "lock %s, level = %d, flags = %d\n", lockres->l_name, level,
790 spin_lock_irqsave(&lockres->l_lock, flags);
791 if ((lockres->l_flags & OCFS2_LOCK_ATTACHED) ||
792 (lockres->l_flags & OCFS2_LOCK_BUSY)) {
793 spin_unlock_irqrestore(&lockres->l_lock, flags);
797 lockres->l_action = OCFS2_AST_ATTACH;
798 lockres->l_requested = level;
799 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
800 spin_unlock_irqrestore(&lockres->l_lock, flags);
802 status = dlmlock(osb->dlm,
807 OCFS2_LOCK_ID_MAX_LEN - 1,
811 if (status != DLM_NORMAL) {
812 ocfs2_log_dlm_error("dlmlock", status, lockres);
814 ocfs2_recover_from_dlm_error(lockres, 1);
817 mlog(0, "lock %s, successfull return from dlmlock\n", lockres->l_name);
824 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res *lockres,
830 spin_lock_irqsave(&lockres->l_lock, flags);
831 ret = lockres->l_flags & flag;
832 spin_unlock_irqrestore(&lockres->l_lock, flags);
837 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res *lockres)
840 wait_event(lockres->l_event,
841 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_BUSY));
844 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res *lockres)
847 wait_event(lockres->l_event,
848 !ocfs2_check_wait_flag(lockres, OCFS2_LOCK_REFRESHING));
851 /* predict what lock level we'll be dropping down to on behalf
852 * of another node, and return true if the currently wanted
853 * level will be compatible with it. */
854 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res *lockres,
857 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
859 return wanted <= ocfs2_highest_compat_lock_level(lockres->l_blocking);
862 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter *mw)
864 INIT_LIST_HEAD(&mw->mw_item);
865 init_completion(&mw->mw_complete);
868 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter *mw)
870 wait_for_completion(&mw->mw_complete);
871 /* Re-arm the completion in case we want to wait on it again */
872 INIT_COMPLETION(mw->mw_complete);
873 return mw->mw_status;
876 static void lockres_add_mask_waiter(struct ocfs2_lock_res *lockres,
877 struct ocfs2_mask_waiter *mw,
881 BUG_ON(!list_empty(&mw->mw_item));
883 assert_spin_locked(&lockres->l_lock);
885 list_add_tail(&mw->mw_item, &lockres->l_mask_waiters);
890 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
891 * if the mask still hadn't reached its goal */
892 static int lockres_remove_mask_waiter(struct ocfs2_lock_res *lockres,
893 struct ocfs2_mask_waiter *mw)
898 spin_lock_irqsave(&lockres->l_lock, flags);
899 if (!list_empty(&mw->mw_item)) {
900 if ((lockres->l_flags & mw->mw_mask) != mw->mw_goal)
903 list_del_init(&mw->mw_item);
904 init_completion(&mw->mw_complete);
906 spin_unlock_irqrestore(&lockres->l_lock, flags);
912 static int ocfs2_cluster_lock(struct ocfs2_super *osb,
913 struct ocfs2_lock_res *lockres,
918 struct ocfs2_mask_waiter mw;
919 enum dlm_status status;
920 int wait, catch_signals = !(osb->s_mount_opt & OCFS2_MOUNT_NOINTR);
921 int ret = 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
926 ocfs2_init_mask_waiter(&mw);
928 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
929 lkm_flags |= LKM_VALBLK;
934 if (catch_signals && signal_pending(current)) {
939 spin_lock_irqsave(&lockres->l_lock, flags);
941 mlog_bug_on_msg(lockres->l_flags & OCFS2_LOCK_FREEING,
942 "Cluster lock called on freeing lockres %s! flags "
943 "0x%lx\n", lockres->l_name, lockres->l_flags);
945 /* We only compare against the currently granted level
946 * here. If the lock is blocked waiting on a downconvert,
947 * we'll get caught below. */
948 if (lockres->l_flags & OCFS2_LOCK_BUSY &&
949 level > lockres->l_level) {
950 /* is someone sitting in dlm_lock? If so, wait on
952 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BUSY, 0);
957 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
958 /* lock has not been created yet. */
959 spin_unlock_irqrestore(&lockres->l_lock, flags);
961 ret = ocfs2_lock_create(osb, lockres, LKM_NLMODE, 0);
969 if (lockres->l_flags & OCFS2_LOCK_BLOCKED &&
970 !ocfs2_may_continue_on_blocked_lock(lockres, level)) {
971 /* is the lock is currently blocked on behalf of
973 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_BLOCKED, 0);
978 if (level > lockres->l_level) {
979 if (lockres->l_action != OCFS2_AST_INVALID)
980 mlog(ML_ERROR, "lockres %s has action %u pending\n",
981 lockres->l_name, lockres->l_action);
983 lockres->l_action = OCFS2_AST_CONVERT;
984 lockres->l_requested = level;
985 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
986 spin_unlock_irqrestore(&lockres->l_lock, flags);
988 BUG_ON(level == LKM_IVMODE);
989 BUG_ON(level == LKM_NLMODE);
991 mlog(0, "lock %s, convert from %d to level = %d\n",
992 lockres->l_name, lockres->l_level, level);
994 /* call dlm_lock to upgrade lock now */
995 status = dlmlock(osb->dlm,
998 lkm_flags|LKM_CONVERT,
1000 OCFS2_LOCK_ID_MAX_LEN - 1,
1003 ocfs2_blocking_ast);
1004 if (status != DLM_NORMAL) {
1005 if ((lkm_flags & LKM_NOQUEUE) &&
1006 (status == DLM_NOTQUEUED))
1009 ocfs2_log_dlm_error("dlmlock", status,
1013 ocfs2_recover_from_dlm_error(lockres, 1);
1017 mlog(0, "lock %s, successfull return from dlmlock\n",
1020 /* At this point we've gone inside the dlm and need to
1021 * complete our work regardless. */
1024 /* wait for busy to clear and carry on */
1028 /* Ok, if we get here then we're good to go. */
1029 ocfs2_inc_holders(lockres, level);
1033 spin_unlock_irqrestore(&lockres->l_lock, flags);
1036 * This is helping work around a lock inversion between the page lock
1037 * and dlm locks. One path holds the page lock while calling aops
1038 * which block acquiring dlm locks. The voting thread holds dlm
1039 * locks while acquiring page locks while down converting data locks.
1040 * This block is helping an aop path notice the inversion and back
1041 * off to unlock its page lock before trying the dlm lock again.
1043 if (wait && arg_flags & OCFS2_LOCK_NONBLOCK &&
1044 mw.mw_mask & (OCFS2_LOCK_BUSY|OCFS2_LOCK_BLOCKED)) {
1046 if (lockres_remove_mask_waiter(lockres, &mw))
1052 ret = ocfs2_wait_for_mask(&mw);
1062 static void ocfs2_cluster_unlock(struct ocfs2_super *osb,
1063 struct ocfs2_lock_res *lockres,
1066 unsigned long flags;
1069 spin_lock_irqsave(&lockres->l_lock, flags);
1070 ocfs2_dec_holders(lockres, level);
1071 ocfs2_vote_on_unlock(osb, lockres);
1072 spin_unlock_irqrestore(&lockres->l_lock, flags);
1076 static int ocfs2_create_new_lock(struct ocfs2_super *osb,
1077 struct ocfs2_lock_res *lockres,
1081 int level = ex ? LKM_EXMODE : LKM_PRMODE;
1082 unsigned long flags;
1083 int lkm_flags = local ? LKM_LOCAL : 0;
1085 spin_lock_irqsave(&lockres->l_lock, flags);
1086 BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
1087 lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
1088 spin_unlock_irqrestore(&lockres->l_lock, flags);
1090 return ocfs2_lock_create(osb, lockres, level, lkm_flags);
1093 /* Grants us an EX lock on the data and metadata resources, skipping
1094 * the normal cluster directory lookup. Use this ONLY on newly created
1095 * inodes which other nodes can't possibly see, and which haven't been
1096 * hashed in the inode hash yet. This can give us a good performance
1097 * increase as it'll skip the network broadcast normally associated
1098 * with creating a new lock resource. */
1099 int ocfs2_create_new_inode_locks(struct inode *inode)
1102 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1105 BUG_ON(!ocfs2_inode_is_new(inode));
1109 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode)->ip_blkno);
1111 /* NOTE: That we don't increment any of the holder counts, nor
1112 * do we add anything to a journal handle. Since this is
1113 * supposed to be a new inode which the cluster doesn't know
1114 * about yet, there is no need to. As far as the LVB handling
1115 * is concerned, this is basically like acquiring an EX lock
1116 * on a resource which has an invalid one -- we'll set it
1117 * valid when we release the EX. */
1119 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
1126 * We don't want to use LKM_LOCAL on a meta data lock as they
1127 * don't use a generation in their lock names.
1129 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_meta_lockres, 1, 0);
1135 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_data_lockres, 1, 1);
1141 ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_open_lockres, 0, 0);
1152 int ocfs2_rw_lock(struct inode *inode, int write)
1155 struct ocfs2_lock_res *lockres;
1156 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1162 mlog(0, "inode %llu take %s RW lock\n",
1163 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1164 write ? "EXMODE" : "PRMODE");
1166 if (ocfs2_mount_local(osb))
1169 lockres = &OCFS2_I(inode)->ip_rw_lockres;
1171 level = write ? LKM_EXMODE : LKM_PRMODE;
1173 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level, 0,
1182 void ocfs2_rw_unlock(struct inode *inode, int write)
1184 int level = write ? LKM_EXMODE : LKM_PRMODE;
1185 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_rw_lockres;
1186 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1190 mlog(0, "inode %llu drop %s RW lock\n",
1191 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1192 write ? "EXMODE" : "PRMODE");
1194 if (!ocfs2_mount_local(osb))
1195 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1201 * ocfs2_open_lock always get PR mode lock.
1203 int ocfs2_open_lock(struct inode *inode)
1206 struct ocfs2_lock_res *lockres;
1207 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1213 mlog(0, "inode %llu take PRMODE open lock\n",
1214 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1216 if (ocfs2_mount_local(osb))
1219 lockres = &OCFS2_I(inode)->ip_open_lockres;
1221 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1231 int ocfs2_try_open_lock(struct inode *inode, int write)
1233 int status = 0, level;
1234 struct ocfs2_lock_res *lockres;
1235 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1241 mlog(0, "inode %llu try to take %s open lock\n",
1242 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1243 write ? "EXMODE" : "PRMODE");
1245 if (ocfs2_mount_local(osb))
1248 lockres = &OCFS2_I(inode)->ip_open_lockres;
1250 level = write ? LKM_EXMODE : LKM_PRMODE;
1253 * The file system may already holding a PRMODE/EXMODE open lock.
1254 * Since we pass LKM_NOQUEUE, the request won't block waiting on
1255 * other nodes and the -EAGAIN will indicate to the caller that
1256 * this inode is still in use.
1258 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres,
1259 level, LKM_NOQUEUE, 0);
1267 * ocfs2_open_unlock unlock PR and EX mode open locks.
1269 void ocfs2_open_unlock(struct inode *inode)
1271 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_open_lockres;
1272 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1276 mlog(0, "inode %llu drop open lock\n",
1277 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1279 if (ocfs2_mount_local(osb))
1282 if(lockres->l_ro_holders)
1283 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1285 if(lockres->l_ex_holders)
1286 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres,
1293 int ocfs2_data_lock_full(struct inode *inode,
1297 int status = 0, level;
1298 struct ocfs2_lock_res *lockres;
1299 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1305 mlog(0, "inode %llu take %s DATA lock\n",
1306 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1307 write ? "EXMODE" : "PRMODE");
1309 /* We'll allow faking a readonly data lock for
1311 if (ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb))) {
1319 if (ocfs2_mount_local(osb))
1322 lockres = &OCFS2_I(inode)->ip_data_lockres;
1324 level = write ? LKM_EXMODE : LKM_PRMODE;
1326 status = ocfs2_cluster_lock(OCFS2_SB(inode->i_sb), lockres, level,
1328 if (status < 0 && status != -EAGAIN)
1336 /* see ocfs2_meta_lock_with_page() */
1337 int ocfs2_data_lock_with_page(struct inode *inode,
1343 ret = ocfs2_data_lock_full(inode, write, OCFS2_LOCK_NONBLOCK);
1344 if (ret == -EAGAIN) {
1346 if (ocfs2_data_lock(inode, write) == 0)
1347 ocfs2_data_unlock(inode, write);
1348 ret = AOP_TRUNCATED_PAGE;
1354 static void ocfs2_vote_on_unlock(struct ocfs2_super *osb,
1355 struct ocfs2_lock_res *lockres)
1361 /* If we know that another node is waiting on our lock, kick
1362 * the vote thread * pre-emptively when we reach a release
1364 if (lockres->l_flags & OCFS2_LOCK_BLOCKED) {
1365 switch(lockres->l_blocking) {
1367 if (!lockres->l_ex_holders && !lockres->l_ro_holders)
1371 if (!lockres->l_ex_holders)
1380 ocfs2_kick_vote_thread(osb);
1385 void ocfs2_data_unlock(struct inode *inode,
1388 int level = write ? LKM_EXMODE : LKM_PRMODE;
1389 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_data_lockres;
1390 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1394 mlog(0, "inode %llu drop %s DATA lock\n",
1395 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1396 write ? "EXMODE" : "PRMODE");
1398 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
1399 !ocfs2_mount_local(osb))
1400 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1405 #define OCFS2_SEC_BITS 34
1406 #define OCFS2_SEC_SHIFT (64 - 34)
1407 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
1409 /* LVB only has room for 64 bits of time here so we pack it for
1411 static u64 ocfs2_pack_timespec(struct timespec *spec)
1414 u64 sec = spec->tv_sec;
1415 u32 nsec = spec->tv_nsec;
1417 res = (sec << OCFS2_SEC_SHIFT) | (nsec & OCFS2_NSEC_MASK);
1422 /* Call this with the lockres locked. I am reasonably sure we don't
1423 * need ip_lock in this function as anyone who would be changing those
1424 * values is supposed to be blocked in ocfs2_meta_lock right now. */
1425 static void __ocfs2_stuff_meta_lvb(struct inode *inode)
1427 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1428 struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
1429 struct ocfs2_meta_lvb *lvb;
1433 lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
1436 * Invalidate the LVB of a deleted inode - this way other
1437 * nodes are forced to go to disk and discover the new inode
1440 if (oi->ip_flags & OCFS2_INODE_DELETED) {
1441 lvb->lvb_version = 0;
1445 lvb->lvb_version = OCFS2_LVB_VERSION;
1446 lvb->lvb_isize = cpu_to_be64(i_size_read(inode));
1447 lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
1448 lvb->lvb_iuid = cpu_to_be32(inode->i_uid);
1449 lvb->lvb_igid = cpu_to_be32(inode->i_gid);
1450 lvb->lvb_imode = cpu_to_be16(inode->i_mode);
1451 lvb->lvb_inlink = cpu_to_be16(inode->i_nlink);
1452 lvb->lvb_iatime_packed =
1453 cpu_to_be64(ocfs2_pack_timespec(&inode->i_atime));
1454 lvb->lvb_ictime_packed =
1455 cpu_to_be64(ocfs2_pack_timespec(&inode->i_ctime));
1456 lvb->lvb_imtime_packed =
1457 cpu_to_be64(ocfs2_pack_timespec(&inode->i_mtime));
1458 lvb->lvb_iattr = cpu_to_be32(oi->ip_attr);
1459 lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
1462 mlog_meta_lvb(0, lockres);
1467 static void ocfs2_unpack_timespec(struct timespec *spec,
1470 spec->tv_sec = packed_time >> OCFS2_SEC_SHIFT;
1471 spec->tv_nsec = packed_time & OCFS2_NSEC_MASK;
1474 static void ocfs2_refresh_inode_from_lvb(struct inode *inode)
1476 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1477 struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
1478 struct ocfs2_meta_lvb *lvb;
1482 mlog_meta_lvb(0, lockres);
1484 lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
1486 /* We're safe here without the lockres lock... */
1487 spin_lock(&oi->ip_lock);
1488 oi->ip_clusters = be32_to_cpu(lvb->lvb_iclusters);
1489 i_size_write(inode, be64_to_cpu(lvb->lvb_isize));
1491 oi->ip_attr = be32_to_cpu(lvb->lvb_iattr);
1492 ocfs2_set_inode_flags(inode);
1494 /* fast-symlinks are a special case */
1495 if (S_ISLNK(inode->i_mode) && !oi->ip_clusters)
1496 inode->i_blocks = 0;
1498 inode->i_blocks = ocfs2_inode_sector_count(inode);
1500 inode->i_uid = be32_to_cpu(lvb->lvb_iuid);
1501 inode->i_gid = be32_to_cpu(lvb->lvb_igid);
1502 inode->i_mode = be16_to_cpu(lvb->lvb_imode);
1503 inode->i_nlink = be16_to_cpu(lvb->lvb_inlink);
1504 ocfs2_unpack_timespec(&inode->i_atime,
1505 be64_to_cpu(lvb->lvb_iatime_packed));
1506 ocfs2_unpack_timespec(&inode->i_mtime,
1507 be64_to_cpu(lvb->lvb_imtime_packed));
1508 ocfs2_unpack_timespec(&inode->i_ctime,
1509 be64_to_cpu(lvb->lvb_ictime_packed));
1510 spin_unlock(&oi->ip_lock);
1515 static inline int ocfs2_meta_lvb_is_trustable(struct inode *inode,
1516 struct ocfs2_lock_res *lockres)
1518 struct ocfs2_meta_lvb *lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
1520 if (lvb->lvb_version == OCFS2_LVB_VERSION
1521 && be32_to_cpu(lvb->lvb_igeneration) == inode->i_generation)
1526 /* Determine whether a lock resource needs to be refreshed, and
1527 * arbitrate who gets to refresh it.
1529 * 0 means no refresh needed.
1531 * > 0 means you need to refresh this and you MUST call
1532 * ocfs2_complete_lock_res_refresh afterwards. */
1533 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res *lockres)
1535 unsigned long flags;
1541 spin_lock_irqsave(&lockres->l_lock, flags);
1542 if (!(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH)) {
1543 spin_unlock_irqrestore(&lockres->l_lock, flags);
1547 if (lockres->l_flags & OCFS2_LOCK_REFRESHING) {
1548 spin_unlock_irqrestore(&lockres->l_lock, flags);
1550 ocfs2_wait_on_refreshing_lock(lockres);
1554 /* Ok, I'll be the one to refresh this lock. */
1555 lockres_or_flags(lockres, OCFS2_LOCK_REFRESHING);
1556 spin_unlock_irqrestore(&lockres->l_lock, flags);
1564 /* If status is non zero, I'll mark it as not being in refresh
1565 * anymroe, but i won't clear the needs refresh flag. */
1566 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res *lockres,
1569 unsigned long flags;
1572 spin_lock_irqsave(&lockres->l_lock, flags);
1573 lockres_clear_flags(lockres, OCFS2_LOCK_REFRESHING);
1575 lockres_clear_flags(lockres, OCFS2_LOCK_NEEDS_REFRESH);
1576 spin_unlock_irqrestore(&lockres->l_lock, flags);
1578 wake_up(&lockres->l_event);
1583 /* may or may not return a bh if it went to disk. */
1584 static int ocfs2_meta_lock_update(struct inode *inode,
1585 struct buffer_head **bh)
1588 struct ocfs2_inode_info *oi = OCFS2_I(inode);
1589 struct ocfs2_lock_res *lockres = &oi->ip_meta_lockres;
1590 struct ocfs2_dinode *fe;
1591 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1595 if (ocfs2_mount_local(osb))
1598 spin_lock(&oi->ip_lock);
1599 if (oi->ip_flags & OCFS2_INODE_DELETED) {
1600 mlog(0, "Orphaned inode %llu was deleted while we "
1601 "were waiting on a lock. ip_flags = 0x%x\n",
1602 (unsigned long long)oi->ip_blkno, oi->ip_flags);
1603 spin_unlock(&oi->ip_lock);
1607 spin_unlock(&oi->ip_lock);
1609 if (!ocfs2_should_refresh_lock_res(lockres))
1612 /* This will discard any caching information we might have had
1613 * for the inode metadata. */
1614 ocfs2_metadata_cache_purge(inode);
1616 if (ocfs2_meta_lvb_is_trustable(inode, lockres)) {
1617 mlog(0, "Trusting LVB on inode %llu\n",
1618 (unsigned long long)oi->ip_blkno);
1619 ocfs2_refresh_inode_from_lvb(inode);
1621 /* Boo, we have to go to disk. */
1622 /* read bh, cast, ocfs2_refresh_inode */
1623 status = ocfs2_read_block(OCFS2_SB(inode->i_sb), oi->ip_blkno,
1624 bh, OCFS2_BH_CACHED, inode);
1629 fe = (struct ocfs2_dinode *) (*bh)->b_data;
1631 /* This is a good chance to make sure we're not
1632 * locking an invalid object.
1634 * We bug on a stale inode here because we checked
1635 * above whether it was wiped from disk. The wiping
1636 * node provides a guarantee that we receive that
1637 * message and can mark the inode before dropping any
1638 * locks associated with it. */
1639 if (!OCFS2_IS_VALID_DINODE(fe)) {
1640 OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
1644 mlog_bug_on_msg(inode->i_generation !=
1645 le32_to_cpu(fe->i_generation),
1646 "Invalid dinode %llu disk generation: %u "
1647 "inode->i_generation: %u\n",
1648 (unsigned long long)oi->ip_blkno,
1649 le32_to_cpu(fe->i_generation),
1650 inode->i_generation);
1651 mlog_bug_on_msg(le64_to_cpu(fe->i_dtime) ||
1652 !(fe->i_flags & cpu_to_le32(OCFS2_VALID_FL)),
1653 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
1654 (unsigned long long)oi->ip_blkno,
1655 (unsigned long long)le64_to_cpu(fe->i_dtime),
1656 le32_to_cpu(fe->i_flags));
1658 ocfs2_refresh_inode(inode, fe);
1663 ocfs2_complete_lock_res_refresh(lockres, status);
1669 static int ocfs2_assign_bh(struct inode *inode,
1670 struct buffer_head **ret_bh,
1671 struct buffer_head *passed_bh)
1676 /* Ok, the update went to disk for us, use the
1678 *ret_bh = passed_bh;
1684 status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
1685 OCFS2_I(inode)->ip_blkno,
1696 * returns < 0 error if the callback will never be called, otherwise
1697 * the result of the lock will be communicated via the callback.
1699 int ocfs2_meta_lock_full(struct inode *inode,
1700 struct buffer_head **ret_bh,
1704 int status, level, dlm_flags, acquired;
1705 struct ocfs2_lock_res *lockres = NULL;
1706 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1707 struct buffer_head *local_bh = NULL;
1713 mlog(0, "inode %llu, take %s META lock\n",
1714 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1715 ex ? "EXMODE" : "PRMODE");
1719 /* We'll allow faking a readonly metadata lock for
1721 if (ocfs2_is_hard_readonly(osb)) {
1727 if (ocfs2_mount_local(osb))
1730 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
1731 wait_event(osb->recovery_event,
1732 ocfs2_node_map_is_empty(osb, &osb->recovery_map));
1734 lockres = &OCFS2_I(inode)->ip_meta_lockres;
1735 level = ex ? LKM_EXMODE : LKM_PRMODE;
1737 if (arg_flags & OCFS2_META_LOCK_NOQUEUE)
1738 dlm_flags |= LKM_NOQUEUE;
1740 status = ocfs2_cluster_lock(osb, lockres, level, dlm_flags, arg_flags);
1742 if (status != -EAGAIN && status != -EIOCBRETRY)
1747 /* Notify the error cleanup path to drop the cluster lock. */
1750 /* We wait twice because a node may have died while we were in
1751 * the lower dlm layers. The second time though, we've
1752 * committed to owning this lock so we don't allow signals to
1753 * abort the operation. */
1754 if (!(arg_flags & OCFS2_META_LOCK_RECOVERY))
1755 wait_event(osb->recovery_event,
1756 ocfs2_node_map_is_empty(osb, &osb->recovery_map));
1760 * We only see this flag if we're being called from
1761 * ocfs2_read_locked_inode(). It means we're locking an inode
1762 * which hasn't been populated yet, so clear the refresh flag
1763 * and let the caller handle it.
1765 if (inode->i_state & I_NEW) {
1768 ocfs2_complete_lock_res_refresh(lockres, 0);
1772 /* This is fun. The caller may want a bh back, or it may
1773 * not. ocfs2_meta_lock_update definitely wants one in, but
1774 * may or may not read one, depending on what's in the
1775 * LVB. The result of all of this is that we've *only* gone to
1776 * disk if we have to, so the complexity is worthwhile. */
1777 status = ocfs2_meta_lock_update(inode, &local_bh);
1779 if (status != -ENOENT)
1785 status = ocfs2_assign_bh(inode, ret_bh, local_bh);
1794 if (ret_bh && (*ret_bh)) {
1799 ocfs2_meta_unlock(inode, ex);
1810 * This is working around a lock inversion between tasks acquiring DLM locks
1811 * while holding a page lock and the vote thread which blocks dlm lock acquiry
1812 * while acquiring page locks.
1814 * ** These _with_page variantes are only intended to be called from aop
1815 * methods that hold page locks and return a very specific *positive* error
1816 * code that aop methods pass up to the VFS -- test for errors with != 0. **
1818 * The DLM is called such that it returns -EAGAIN if it would have blocked
1819 * waiting for the vote thread. In that case we unlock our page so the vote
1820 * thread can make progress. Once we've done this we have to return
1821 * AOP_TRUNCATED_PAGE so the aop method that called us can bubble that back up
1822 * into the VFS who will then immediately retry the aop call.
1824 * We do a blocking lock and immediate unlock before returning, though, so that
1825 * the lock has a great chance of being cached on this node by the time the VFS
1826 * calls back to retry the aop. This has a potential to livelock as nodes
1827 * ping locks back and forth, but that's a risk we're willing to take to avoid
1828 * the lock inversion simply.
1830 int ocfs2_meta_lock_with_page(struct inode *inode,
1831 struct buffer_head **ret_bh,
1837 ret = ocfs2_meta_lock_full(inode, ret_bh, ex, OCFS2_LOCK_NONBLOCK);
1838 if (ret == -EAGAIN) {
1840 if (ocfs2_meta_lock(inode, ret_bh, ex) == 0)
1841 ocfs2_meta_unlock(inode, ex);
1842 ret = AOP_TRUNCATED_PAGE;
1848 int ocfs2_meta_lock_atime(struct inode *inode,
1849 struct vfsmount *vfsmnt,
1855 ret = ocfs2_meta_lock(inode, NULL, 0);
1862 * If we should update atime, we will get EX lock,
1863 * otherwise we just get PR lock.
1865 if (ocfs2_should_update_atime(inode, vfsmnt)) {
1866 struct buffer_head *bh = NULL;
1868 ocfs2_meta_unlock(inode, 0);
1869 ret = ocfs2_meta_lock(inode, &bh, 1);
1875 if (ocfs2_should_update_atime(inode, vfsmnt))
1876 ocfs2_update_inode_atime(inode, bh);
1886 void ocfs2_meta_unlock(struct inode *inode,
1889 int level = ex ? LKM_EXMODE : LKM_PRMODE;
1890 struct ocfs2_lock_res *lockres = &OCFS2_I(inode)->ip_meta_lockres;
1891 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1895 mlog(0, "inode %llu drop %s META lock\n",
1896 (unsigned long long)OCFS2_I(inode)->ip_blkno,
1897 ex ? "EXMODE" : "PRMODE");
1899 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode->i_sb)) &&
1900 !ocfs2_mount_local(osb))
1901 ocfs2_cluster_unlock(OCFS2_SB(inode->i_sb), lockres, level);
1906 int ocfs2_super_lock(struct ocfs2_super *osb,
1910 int level = ex ? LKM_EXMODE : LKM_PRMODE;
1911 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
1912 struct buffer_head *bh;
1913 struct ocfs2_slot_info *si = osb->slot_info;
1917 if (ocfs2_is_hard_readonly(osb))
1920 if (ocfs2_mount_local(osb))
1923 status = ocfs2_cluster_lock(osb, lockres, level, 0, 0);
1929 /* The super block lock path is really in the best position to
1930 * know when resources covered by the lock need to be
1931 * refreshed, so we do it here. Of course, making sense of
1932 * everything is up to the caller :) */
1933 status = ocfs2_should_refresh_lock_res(lockres);
1940 status = ocfs2_read_block(osb, bh->b_blocknr, &bh, 0,
1943 ocfs2_update_slot_info(si);
1945 ocfs2_complete_lock_res_refresh(lockres, status);
1955 void ocfs2_super_unlock(struct ocfs2_super *osb,
1958 int level = ex ? LKM_EXMODE : LKM_PRMODE;
1959 struct ocfs2_lock_res *lockres = &osb->osb_super_lockres;
1961 if (!ocfs2_mount_local(osb))
1962 ocfs2_cluster_unlock(osb, lockres, level);
1965 int ocfs2_rename_lock(struct ocfs2_super *osb)
1968 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
1970 if (ocfs2_is_hard_readonly(osb))
1973 if (ocfs2_mount_local(osb))
1976 status = ocfs2_cluster_lock(osb, lockres, LKM_EXMODE, 0, 0);
1983 void ocfs2_rename_unlock(struct ocfs2_super *osb)
1985 struct ocfs2_lock_res *lockres = &osb->osb_rename_lockres;
1987 if (!ocfs2_mount_local(osb))
1988 ocfs2_cluster_unlock(osb, lockres, LKM_EXMODE);
1991 int ocfs2_dentry_lock(struct dentry *dentry, int ex)
1994 int level = ex ? LKM_EXMODE : LKM_PRMODE;
1995 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
1996 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2000 if (ocfs2_is_hard_readonly(osb))
2003 if (ocfs2_mount_local(osb))
2006 ret = ocfs2_cluster_lock(osb, &dl->dl_lockres, level, 0, 0);
2013 void ocfs2_dentry_unlock(struct dentry *dentry, int ex)
2015 int level = ex ? LKM_EXMODE : LKM_PRMODE;
2016 struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
2017 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
2019 if (!ocfs2_mount_local(osb))
2020 ocfs2_cluster_unlock(osb, &dl->dl_lockres, level);
2023 /* Reference counting of the dlm debug structure. We want this because
2024 * open references on the debug inodes can live on after a mount, so
2025 * we can't rely on the ocfs2_super to always exist. */
2026 static void ocfs2_dlm_debug_free(struct kref *kref)
2028 struct ocfs2_dlm_debug *dlm_debug;
2030 dlm_debug = container_of(kref, struct ocfs2_dlm_debug, d_refcnt);
2035 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug *dlm_debug)
2038 kref_put(&dlm_debug->d_refcnt, ocfs2_dlm_debug_free);
2041 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug *debug)
2043 kref_get(&debug->d_refcnt);
2046 struct ocfs2_dlm_debug *ocfs2_new_dlm_debug(void)
2048 struct ocfs2_dlm_debug *dlm_debug;
2050 dlm_debug = kmalloc(sizeof(struct ocfs2_dlm_debug), GFP_KERNEL);
2052 mlog_errno(-ENOMEM);
2056 kref_init(&dlm_debug->d_refcnt);
2057 INIT_LIST_HEAD(&dlm_debug->d_lockres_tracking);
2058 dlm_debug->d_locking_state = NULL;
2063 /* Access to this is arbitrated for us via seq_file->sem. */
2064 struct ocfs2_dlm_seq_priv {
2065 struct ocfs2_dlm_debug *p_dlm_debug;
2066 struct ocfs2_lock_res p_iter_res;
2067 struct ocfs2_lock_res p_tmp_res;
2070 static struct ocfs2_lock_res *ocfs2_dlm_next_res(struct ocfs2_lock_res *start,
2071 struct ocfs2_dlm_seq_priv *priv)
2073 struct ocfs2_lock_res *iter, *ret = NULL;
2074 struct ocfs2_dlm_debug *dlm_debug = priv->p_dlm_debug;
2076 assert_spin_locked(&ocfs2_dlm_tracking_lock);
2078 list_for_each_entry(iter, &start->l_debug_list, l_debug_list) {
2079 /* discover the head of the list */
2080 if (&iter->l_debug_list == &dlm_debug->d_lockres_tracking) {
2081 mlog(0, "End of list found, %p\n", ret);
2085 /* We track our "dummy" iteration lockres' by a NULL
2087 if (iter->l_ops != NULL) {
2096 static void *ocfs2_dlm_seq_start(struct seq_file *m, loff_t *pos)
2098 struct ocfs2_dlm_seq_priv *priv = m->private;
2099 struct ocfs2_lock_res *iter;
2101 spin_lock(&ocfs2_dlm_tracking_lock);
2102 iter = ocfs2_dlm_next_res(&priv->p_iter_res, priv);
2104 /* Since lockres' have the lifetime of their container
2105 * (which can be inodes, ocfs2_supers, etc) we want to
2106 * copy this out to a temporary lockres while still
2107 * under the spinlock. Obviously after this we can't
2108 * trust any pointers on the copy returned, but that's
2109 * ok as the information we want isn't typically held
2111 priv->p_tmp_res = *iter;
2112 iter = &priv->p_tmp_res;
2114 spin_unlock(&ocfs2_dlm_tracking_lock);
2119 static void ocfs2_dlm_seq_stop(struct seq_file *m, void *v)
2123 static void *ocfs2_dlm_seq_next(struct seq_file *m, void *v, loff_t *pos)
2125 struct ocfs2_dlm_seq_priv *priv = m->private;
2126 struct ocfs2_lock_res *iter = v;
2127 struct ocfs2_lock_res *dummy = &priv->p_iter_res;
2129 spin_lock(&ocfs2_dlm_tracking_lock);
2130 iter = ocfs2_dlm_next_res(iter, priv);
2131 list_del_init(&dummy->l_debug_list);
2133 list_add(&dummy->l_debug_list, &iter->l_debug_list);
2134 priv->p_tmp_res = *iter;
2135 iter = &priv->p_tmp_res;
2137 spin_unlock(&ocfs2_dlm_tracking_lock);
2142 /* So that debugfs.ocfs2 can determine which format is being used */
2143 #define OCFS2_DLM_DEBUG_STR_VERSION 1
2144 static int ocfs2_dlm_seq_show(struct seq_file *m, void *v)
2148 struct ocfs2_lock_res *lockres = v;
2153 seq_printf(m, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION);
2155 if (lockres->l_type == OCFS2_LOCK_TYPE_DENTRY)
2156 seq_printf(m, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START - 1,
2158 (unsigned int)ocfs2_get_dentry_lock_ino(lockres));
2160 seq_printf(m, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN, lockres->l_name);
2162 seq_printf(m, "%d\t"
2173 lockres->l_unlock_action,
2174 lockres->l_ro_holders,
2175 lockres->l_ex_holders,
2176 lockres->l_requested,
2177 lockres->l_blocking);
2179 /* Dump the raw LVB */
2180 lvb = lockres->l_lksb.lvb;
2181 for(i = 0; i < DLM_LVB_LEN; i++)
2182 seq_printf(m, "0x%x\t", lvb[i]);
2185 seq_printf(m, "\n");
2189 static struct seq_operations ocfs2_dlm_seq_ops = {
2190 .start = ocfs2_dlm_seq_start,
2191 .stop = ocfs2_dlm_seq_stop,
2192 .next = ocfs2_dlm_seq_next,
2193 .show = ocfs2_dlm_seq_show,
2196 static int ocfs2_dlm_debug_release(struct inode *inode, struct file *file)
2198 struct seq_file *seq = (struct seq_file *) file->private_data;
2199 struct ocfs2_dlm_seq_priv *priv = seq->private;
2200 struct ocfs2_lock_res *res = &priv->p_iter_res;
2202 ocfs2_remove_lockres_tracking(res);
2203 ocfs2_put_dlm_debug(priv->p_dlm_debug);
2204 return seq_release_private(inode, file);
2207 static int ocfs2_dlm_debug_open(struct inode *inode, struct file *file)
2210 struct ocfs2_dlm_seq_priv *priv;
2211 struct seq_file *seq;
2212 struct ocfs2_super *osb;
2214 priv = kzalloc(sizeof(struct ocfs2_dlm_seq_priv), GFP_KERNEL);
2220 osb = inode->i_private;
2221 ocfs2_get_dlm_debug(osb->osb_dlm_debug);
2222 priv->p_dlm_debug = osb->osb_dlm_debug;
2223 INIT_LIST_HEAD(&priv->p_iter_res.l_debug_list);
2225 ret = seq_open(file, &ocfs2_dlm_seq_ops);
2232 seq = (struct seq_file *) file->private_data;
2233 seq->private = priv;
2235 ocfs2_add_lockres_tracking(&priv->p_iter_res,
2242 static const struct file_operations ocfs2_dlm_debug_fops = {
2243 .open = ocfs2_dlm_debug_open,
2244 .release = ocfs2_dlm_debug_release,
2246 .llseek = seq_lseek,
2249 static int ocfs2_dlm_init_debug(struct ocfs2_super *osb)
2252 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2254 dlm_debug->d_locking_state = debugfs_create_file("locking_state",
2256 osb->osb_debug_root,
2258 &ocfs2_dlm_debug_fops);
2259 if (!dlm_debug->d_locking_state) {
2262 "Unable to create locking state debugfs file.\n");
2266 ocfs2_get_dlm_debug(dlm_debug);
2271 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super *osb)
2273 struct ocfs2_dlm_debug *dlm_debug = osb->osb_dlm_debug;
2276 debugfs_remove(dlm_debug->d_locking_state);
2277 ocfs2_put_dlm_debug(dlm_debug);
2281 int ocfs2_dlm_init(struct ocfs2_super *osb)
2285 struct dlm_ctxt *dlm = NULL;
2289 if (ocfs2_mount_local(osb))
2292 status = ocfs2_dlm_init_debug(osb);
2298 /* launch vote thread */
2299 osb->vote_task = kthread_run(ocfs2_vote_thread, osb, "ocfs2vote");
2300 if (IS_ERR(osb->vote_task)) {
2301 status = PTR_ERR(osb->vote_task);
2302 osb->vote_task = NULL;
2307 /* used by the dlm code to make message headers unique, each
2308 * node in this domain must agree on this. */
2309 dlm_key = crc32_le(0, osb->uuid_str, strlen(osb->uuid_str));
2311 /* for now, uuid == domain */
2312 dlm = dlm_register_domain(osb->uuid_str, dlm_key);
2314 status = PTR_ERR(dlm);
2319 dlm_register_eviction_cb(dlm, &osb->osb_eviction_cb);
2322 ocfs2_super_lock_res_init(&osb->osb_super_lockres, osb);
2323 ocfs2_rename_lock_res_init(&osb->osb_rename_lockres, osb);
2330 ocfs2_dlm_shutdown_debug(osb);
2332 kthread_stop(osb->vote_task);
2339 void ocfs2_dlm_shutdown(struct ocfs2_super *osb)
2343 dlm_unregister_eviction_cb(&osb->osb_eviction_cb);
2345 ocfs2_drop_osb_locks(osb);
2347 if (osb->vote_task) {
2348 kthread_stop(osb->vote_task);
2349 osb->vote_task = NULL;
2352 ocfs2_lock_res_free(&osb->osb_super_lockres);
2353 ocfs2_lock_res_free(&osb->osb_rename_lockres);
2355 dlm_unregister_domain(osb->dlm);
2358 ocfs2_dlm_shutdown_debug(osb);
2363 static void ocfs2_unlock_ast(void *opaque, enum dlm_status status)
2365 struct ocfs2_lock_res *lockres = opaque;
2366 unsigned long flags;
2370 mlog(0, "UNLOCK AST called on lock %s, action = %d\n", lockres->l_name,
2371 lockres->l_unlock_action);
2373 spin_lock_irqsave(&lockres->l_lock, flags);
2374 /* We tried to cancel a convert request, but it was already
2375 * granted. All we want to do here is clear our unlock
2376 * state. The wake_up call done at the bottom is redundant
2377 * (ocfs2_prepare_cancel_convert doesn't sleep on this) but doesn't
2378 * hurt anything anyway */
2379 if (status == DLM_CANCELGRANT &&
2380 lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
2381 mlog(0, "Got cancelgrant for %s\n", lockres->l_name);
2383 /* We don't clear the busy flag in this case as it
2384 * should have been cleared by the ast which the dlm
2386 goto complete_unlock;
2389 if (status != DLM_NORMAL) {
2390 mlog(ML_ERROR, "Dlm passes status %d for lock %s, "
2391 "unlock_action %d\n", status, lockres->l_name,
2392 lockres->l_unlock_action);
2393 spin_unlock_irqrestore(&lockres->l_lock, flags);
2397 switch(lockres->l_unlock_action) {
2398 case OCFS2_UNLOCK_CANCEL_CONVERT:
2399 mlog(0, "Cancel convert success for %s\n", lockres->l_name);
2400 lockres->l_action = OCFS2_AST_INVALID;
2402 case OCFS2_UNLOCK_DROP_LOCK:
2403 lockres->l_level = LKM_IVMODE;
2409 lockres_clear_flags(lockres, OCFS2_LOCK_BUSY);
2411 lockres->l_unlock_action = OCFS2_UNLOCK_INVALID;
2412 spin_unlock_irqrestore(&lockres->l_lock, flags);
2414 wake_up(&lockres->l_event);
2419 static int ocfs2_drop_lock(struct ocfs2_super *osb,
2420 struct ocfs2_lock_res *lockres)
2422 enum dlm_status status;
2423 unsigned long flags;
2426 /* We didn't get anywhere near actually using this lockres. */
2427 if (!(lockres->l_flags & OCFS2_LOCK_INITIALIZED))
2430 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB)
2431 lkm_flags |= LKM_VALBLK;
2433 spin_lock_irqsave(&lockres->l_lock, flags);
2435 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_FREEING),
2436 "lockres %s, flags 0x%lx\n",
2437 lockres->l_name, lockres->l_flags);
2439 while (lockres->l_flags & OCFS2_LOCK_BUSY) {
2440 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
2441 "%u, unlock_action = %u\n",
2442 lockres->l_name, lockres->l_flags, lockres->l_action,
2443 lockres->l_unlock_action);
2445 spin_unlock_irqrestore(&lockres->l_lock, flags);
2447 /* XXX: Today we just wait on any busy
2448 * locks... Perhaps we need to cancel converts in the
2450 ocfs2_wait_on_busy_lock(lockres);
2452 spin_lock_irqsave(&lockres->l_lock, flags);
2455 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
2456 if (lockres->l_flags & OCFS2_LOCK_ATTACHED &&
2457 lockres->l_level == LKM_EXMODE &&
2458 !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
2459 lockres->l_ops->set_lvb(lockres);
2462 if (lockres->l_flags & OCFS2_LOCK_BUSY)
2463 mlog(ML_ERROR, "destroying busy lock: \"%s\"\n",
2465 if (lockres->l_flags & OCFS2_LOCK_BLOCKED)
2466 mlog(0, "destroying blocked lock: \"%s\"\n", lockres->l_name);
2468 if (!(lockres->l_flags & OCFS2_LOCK_ATTACHED)) {
2469 spin_unlock_irqrestore(&lockres->l_lock, flags);
2473 lockres_clear_flags(lockres, OCFS2_LOCK_ATTACHED);
2475 /* make sure we never get here while waiting for an ast to
2477 BUG_ON(lockres->l_action != OCFS2_AST_INVALID);
2479 /* is this necessary? */
2480 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
2481 lockres->l_unlock_action = OCFS2_UNLOCK_DROP_LOCK;
2482 spin_unlock_irqrestore(&lockres->l_lock, flags);
2484 mlog(0, "lock %s\n", lockres->l_name);
2486 status = dlmunlock(osb->dlm, &lockres->l_lksb, lkm_flags,
2487 ocfs2_unlock_ast, lockres);
2488 if (status != DLM_NORMAL) {
2489 ocfs2_log_dlm_error("dlmunlock", status, lockres);
2490 mlog(ML_ERROR, "lockres flags: %lu\n", lockres->l_flags);
2491 dlm_print_one_lock(lockres->l_lksb.lockid);
2494 mlog(0, "lock %s, successfull return from dlmunlock\n",
2497 ocfs2_wait_on_busy_lock(lockres);
2503 /* Mark the lockres as being dropped. It will no longer be
2504 * queued if blocking, but we still may have to wait on it
2505 * being dequeued from the vote thread before we can consider
2508 * You can *not* attempt to call cluster_lock on this lockres anymore. */
2509 void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res *lockres)
2512 struct ocfs2_mask_waiter mw;
2513 unsigned long flags;
2515 ocfs2_init_mask_waiter(&mw);
2517 spin_lock_irqsave(&lockres->l_lock, flags);
2518 lockres->l_flags |= OCFS2_LOCK_FREEING;
2519 while (lockres->l_flags & OCFS2_LOCK_QUEUED) {
2520 lockres_add_mask_waiter(lockres, &mw, OCFS2_LOCK_QUEUED, 0);
2521 spin_unlock_irqrestore(&lockres->l_lock, flags);
2523 mlog(0, "Waiting on lockres %s\n", lockres->l_name);
2525 status = ocfs2_wait_for_mask(&mw);
2529 spin_lock_irqsave(&lockres->l_lock, flags);
2531 spin_unlock_irqrestore(&lockres->l_lock, flags);
2534 void ocfs2_simple_drop_lockres(struct ocfs2_super *osb,
2535 struct ocfs2_lock_res *lockres)
2539 ocfs2_mark_lockres_freeing(lockres);
2540 ret = ocfs2_drop_lock(osb, lockres);
2545 static void ocfs2_drop_osb_locks(struct ocfs2_super *osb)
2547 ocfs2_simple_drop_lockres(osb, &osb->osb_super_lockres);
2548 ocfs2_simple_drop_lockres(osb, &osb->osb_rename_lockres);
2551 int ocfs2_drop_inode_locks(struct inode *inode)
2557 /* No need to call ocfs2_mark_lockres_freeing here -
2558 * ocfs2_clear_inode has done it for us. */
2560 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2561 &OCFS2_I(inode)->ip_open_lockres);
2567 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2568 &OCFS2_I(inode)->ip_data_lockres);
2571 if (err < 0 && !status)
2574 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2575 &OCFS2_I(inode)->ip_meta_lockres);
2578 if (err < 0 && !status)
2581 err = ocfs2_drop_lock(OCFS2_SB(inode->i_sb),
2582 &OCFS2_I(inode)->ip_rw_lockres);
2585 if (err < 0 && !status)
2592 static void ocfs2_prepare_downconvert(struct ocfs2_lock_res *lockres,
2595 assert_spin_locked(&lockres->l_lock);
2597 BUG_ON(lockres->l_blocking <= LKM_NLMODE);
2599 if (lockres->l_level <= new_level) {
2600 mlog(ML_ERROR, "lockres->l_level (%u) <= new_level (%u)\n",
2601 lockres->l_level, new_level);
2605 mlog(0, "lock %s, new_level = %d, l_blocking = %d\n",
2606 lockres->l_name, new_level, lockres->l_blocking);
2608 lockres->l_action = OCFS2_AST_DOWNCONVERT;
2609 lockres->l_requested = new_level;
2610 lockres_or_flags(lockres, OCFS2_LOCK_BUSY);
2613 static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
2614 struct ocfs2_lock_res *lockres,
2618 int ret, dlm_flags = LKM_CONVERT;
2619 enum dlm_status status;
2624 dlm_flags |= LKM_VALBLK;
2626 status = dlmlock(osb->dlm,
2631 OCFS2_LOCK_ID_MAX_LEN - 1,
2634 ocfs2_blocking_ast);
2635 if (status != DLM_NORMAL) {
2636 ocfs2_log_dlm_error("dlmlock", status, lockres);
2638 ocfs2_recover_from_dlm_error(lockres, 1);
2648 /* returns 1 when the caller should unlock and call dlmunlock */
2649 static int ocfs2_prepare_cancel_convert(struct ocfs2_super *osb,
2650 struct ocfs2_lock_res *lockres)
2652 assert_spin_locked(&lockres->l_lock);
2655 mlog(0, "lock %s\n", lockres->l_name);
2657 if (lockres->l_unlock_action == OCFS2_UNLOCK_CANCEL_CONVERT) {
2658 /* If we're already trying to cancel a lock conversion
2659 * then just drop the spinlock and allow the caller to
2660 * requeue this lock. */
2662 mlog(0, "Lockres %s, skip convert\n", lockres->l_name);
2666 /* were we in a convert when we got the bast fire? */
2667 BUG_ON(lockres->l_action != OCFS2_AST_CONVERT &&
2668 lockres->l_action != OCFS2_AST_DOWNCONVERT);
2669 /* set things up for the unlockast to know to just
2670 * clear out the ast_action and unset busy, etc. */
2671 lockres->l_unlock_action = OCFS2_UNLOCK_CANCEL_CONVERT;
2673 mlog_bug_on_msg(!(lockres->l_flags & OCFS2_LOCK_BUSY),
2674 "lock %s, invalid flags: 0x%lx\n",
2675 lockres->l_name, lockres->l_flags);
2680 static int ocfs2_cancel_convert(struct ocfs2_super *osb,
2681 struct ocfs2_lock_res *lockres)
2684 enum dlm_status status;
2687 mlog(0, "lock %s\n", lockres->l_name);
2690 status = dlmunlock(osb->dlm,
2695 if (status != DLM_NORMAL) {
2696 ocfs2_log_dlm_error("dlmunlock", status, lockres);
2698 ocfs2_recover_from_dlm_error(lockres, 0);
2701 mlog(0, "lock %s return from dlmunlock\n", lockres->l_name);
2707 static int ocfs2_unblock_lock(struct ocfs2_super *osb,
2708 struct ocfs2_lock_res *lockres,
2709 struct ocfs2_unblock_ctl *ctl)
2711 unsigned long flags;
2719 spin_lock_irqsave(&lockres->l_lock, flags);
2721 BUG_ON(!(lockres->l_flags & OCFS2_LOCK_BLOCKED));
2724 if (lockres->l_flags & OCFS2_LOCK_BUSY) {
2726 ret = ocfs2_prepare_cancel_convert(osb, lockres);
2727 spin_unlock_irqrestore(&lockres->l_lock, flags);
2729 ret = ocfs2_cancel_convert(osb, lockres);
2736 /* if we're blocking an exclusive and we have *any* holders,
2738 if ((lockres->l_blocking == LKM_EXMODE)
2739 && (lockres->l_ex_holders || lockres->l_ro_holders))
2742 /* If it's a PR we're blocking, then only
2743 * requeue if we've got any EX holders */
2744 if (lockres->l_blocking == LKM_PRMODE &&
2745 lockres->l_ex_holders)
2749 * Can we get a lock in this state if the holder counts are
2750 * zero? The meta data unblock code used to check this.
2752 if ((lockres->l_ops->flags & LOCK_TYPE_REQUIRES_REFRESH)
2753 && (lockres->l_flags & OCFS2_LOCK_REFRESHING))
2756 new_level = ocfs2_highest_compat_lock_level(lockres->l_blocking);
2758 if (lockres->l_ops->check_downconvert
2759 && !lockres->l_ops->check_downconvert(lockres, new_level))
2762 /* If we get here, then we know that there are no more
2763 * incompatible holders (and anyone asking for an incompatible
2764 * lock is blocked). We can now downconvert the lock */
2765 if (!lockres->l_ops->downconvert_worker)
2768 /* Some lockres types want to do a bit of work before
2769 * downconverting a lock. Allow that here. The worker function
2770 * may sleep, so we save off a copy of what we're blocking as
2771 * it may change while we're not holding the spin lock. */
2772 blocking = lockres->l_blocking;
2773 spin_unlock_irqrestore(&lockres->l_lock, flags);
2775 ctl->unblock_action = lockres->l_ops->downconvert_worker(lockres, blocking);
2777 if (ctl->unblock_action == UNBLOCK_STOP_POST)
2780 spin_lock_irqsave(&lockres->l_lock, flags);
2781 if (blocking != lockres->l_blocking) {
2782 /* If this changed underneath us, then we can't drop
2790 if (lockres->l_ops->flags & LOCK_TYPE_USES_LVB) {
2791 if (lockres->l_level == LKM_EXMODE)
2795 * We only set the lvb if the lock has been fully
2796 * refreshed - otherwise we risk setting stale
2797 * data. Otherwise, there's no need to actually clear
2798 * out the lvb here as it's value is still valid.
2800 if (set_lvb && !(lockres->l_flags & OCFS2_LOCK_NEEDS_REFRESH))
2801 lockres->l_ops->set_lvb(lockres);
2804 ocfs2_prepare_downconvert(lockres, new_level);
2805 spin_unlock_irqrestore(&lockres->l_lock, flags);
2806 ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb);
2812 spin_unlock_irqrestore(&lockres->l_lock, flags);
2819 static int ocfs2_data_convert_worker(struct ocfs2_lock_res *lockres,
2822 struct inode *inode;
2823 struct address_space *mapping;
2825 inode = ocfs2_lock_res_inode(lockres);
2826 mapping = inode->i_mapping;
2829 * We need this before the filemap_fdatawrite() so that it can
2830 * transfer the dirty bit from the PTE to the
2831 * page. Unfortunately this means that even for EX->PR
2832 * downconverts, we'll lose our mappings and have to build
2835 unmap_mapping_range(mapping, 0, 0, 0);
2837 if (filemap_fdatawrite(mapping)) {
2838 mlog(ML_ERROR, "Could not sync inode %llu for downconvert!",
2839 (unsigned long long)OCFS2_I(inode)->ip_blkno);
2841 sync_mapping_buffers(mapping);
2842 if (blocking == LKM_EXMODE) {
2843 truncate_inode_pages(mapping, 0);
2845 /* We only need to wait on the I/O if we're not also
2846 * truncating pages because truncate_inode_pages waits
2847 * for us above. We don't truncate pages if we're
2848 * blocking anything < EXMODE because we want to keep
2849 * them around in that case. */
2850 filemap_fdatawait(mapping);
2853 return UNBLOCK_CONTINUE;
2856 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res *lockres,
2859 struct inode *inode = ocfs2_lock_res_inode(lockres);
2860 int checkpointed = ocfs2_inode_fully_checkpointed(inode);
2862 BUG_ON(new_level != LKM_NLMODE && new_level != LKM_PRMODE);
2863 BUG_ON(lockres->l_level != LKM_EXMODE && !checkpointed);
2868 ocfs2_start_checkpoint(OCFS2_SB(inode->i_sb));
2872 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res *lockres)
2874 struct inode *inode = ocfs2_lock_res_inode(lockres);
2876 __ocfs2_stuff_meta_lvb(inode);
2880 * Does the final reference drop on our dentry lock. Right now this
2881 * happens in the vote thread, but we could choose to simplify the
2882 * dlmglue API and push these off to the ocfs2_wq in the future.
2884 static void ocfs2_dentry_post_unlock(struct ocfs2_super *osb,
2885 struct ocfs2_lock_res *lockres)
2887 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
2888 ocfs2_dentry_lock_put(osb, dl);
2892 * d_delete() matching dentries before the lock downconvert.
2894 * At this point, any process waiting to destroy the
2895 * dentry_lock due to last ref count is stopped by the
2896 * OCFS2_LOCK_QUEUED flag.
2898 * We have two potential problems
2900 * 1) If we do the last reference drop on our dentry_lock (via dput)
2901 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
2902 * the downconvert to finish. Instead we take an elevated
2903 * reference and push the drop until after we've completed our
2904 * unblock processing.
2906 * 2) There might be another process with a final reference,
2907 * waiting on us to finish processing. If this is the case, we
2908 * detect it and exit out - there's no more dentries anyway.
2910 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res *lockres,
2913 struct ocfs2_dentry_lock *dl = ocfs2_lock_res_dl(lockres);
2914 struct ocfs2_inode_info *oi = OCFS2_I(dl->dl_inode);
2915 struct dentry *dentry;
2916 unsigned long flags;
2920 * This node is blocking another node from getting a read
2921 * lock. This happens when we've renamed within a
2922 * directory. We've forced the other nodes to d_delete(), but
2923 * we never actually dropped our lock because it's still
2924 * valid. The downconvert code will retain a PR for this node,
2925 * so there's no further work to do.
2927 if (blocking == LKM_PRMODE)
2928 return UNBLOCK_CONTINUE;
2931 * Mark this inode as potentially orphaned. The code in
2932 * ocfs2_delete_inode() will figure out whether it actually
2933 * needs to be freed or not.
2935 spin_lock(&oi->ip_lock);
2936 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2937 spin_unlock(&oi->ip_lock);
2940 * Yuck. We need to make sure however that the check of
2941 * OCFS2_LOCK_FREEING and the extra reference are atomic with
2942 * respect to a reference decrement or the setting of that
2945 spin_lock_irqsave(&lockres->l_lock, flags);
2946 spin_lock(&dentry_attach_lock);
2947 if (!(lockres->l_flags & OCFS2_LOCK_FREEING)
2952 spin_unlock(&dentry_attach_lock);
2953 spin_unlock_irqrestore(&lockres->l_lock, flags);
2955 mlog(0, "extra_ref = %d\n", extra_ref);
2958 * We have a process waiting on us in ocfs2_dentry_iput(),
2959 * which means we can't have any more outstanding
2960 * aliases. There's no need to do any more work.
2963 return UNBLOCK_CONTINUE;
2965 spin_lock(&dentry_attach_lock);
2967 dentry = ocfs2_find_local_alias(dl->dl_inode,
2968 dl->dl_parent_blkno, 1);
2971 spin_unlock(&dentry_attach_lock);
2973 mlog(0, "d_delete(%.*s);\n", dentry->d_name.len,
2974 dentry->d_name.name);
2977 * The following dcache calls may do an
2978 * iput(). Normally we don't want that from the
2979 * downconverting thread, but in this case it's ok
2980 * because the requesting node already has an
2981 * exclusive lock on the inode, so it can't be queued
2982 * for a downconvert.
2987 spin_lock(&dentry_attach_lock);
2989 spin_unlock(&dentry_attach_lock);
2992 * If we are the last holder of this dentry lock, there is no
2993 * reason to downconvert so skip straight to the unlock.
2995 if (dl->dl_count == 1)
2996 return UNBLOCK_STOP_POST;
2998 return UNBLOCK_CONTINUE_POST;
3001 void ocfs2_process_blocked_lock(struct ocfs2_super *osb,
3002 struct ocfs2_lock_res *lockres)
3005 struct ocfs2_unblock_ctl ctl = {0, 0,};
3006 unsigned long flags;
3008 /* Our reference to the lockres in this function can be
3009 * considered valid until we remove the OCFS2_LOCK_QUEUED
3015 BUG_ON(!lockres->l_ops);
3017 mlog(0, "lockres %s blocked.\n", lockres->l_name);
3019 /* Detect whether a lock has been marked as going away while
3020 * the vote thread was processing other things. A lock can
3021 * still be marked with OCFS2_LOCK_FREEING after this check,
3022 * but short circuiting here will still save us some
3024 spin_lock_irqsave(&lockres->l_lock, flags);
3025 if (lockres->l_flags & OCFS2_LOCK_FREEING)
3027 spin_unlock_irqrestore(&lockres->l_lock, flags);
3029 status = ocfs2_unblock_lock(osb, lockres, &ctl);
3033 spin_lock_irqsave(&lockres->l_lock, flags);
3035 if (lockres->l_flags & OCFS2_LOCK_FREEING || !ctl.requeue) {
3036 lockres_clear_flags(lockres, OCFS2_LOCK_QUEUED);
3038 ocfs2_schedule_blocked_lock(osb, lockres);
3040 mlog(0, "lockres %s, requeue = %s.\n", lockres->l_name,
3041 ctl.requeue ? "yes" : "no");
3042 spin_unlock_irqrestore(&lockres->l_lock, flags);
3044 if (ctl.unblock_action != UNBLOCK_CONTINUE
3045 && lockres->l_ops->post_unlock)
3046 lockres->l_ops->post_unlock(osb, lockres);
3051 static void ocfs2_schedule_blocked_lock(struct ocfs2_super *osb,
3052 struct ocfs2_lock_res *lockres)
3056 assert_spin_locked(&lockres->l_lock);
3058 if (lockres->l_flags & OCFS2_LOCK_FREEING) {
3059 /* Do not schedule a lock for downconvert when it's on
3060 * the way to destruction - any nodes wanting access
3061 * to the resource will get it soon. */
3062 mlog(0, "Lockres %s won't be scheduled: flags 0x%lx\n",
3063 lockres->l_name, lockres->l_flags);
3067 lockres_or_flags(lockres, OCFS2_LOCK_QUEUED);
3069 spin_lock(&osb->vote_task_lock);
3070 if (list_empty(&lockres->l_blocked_list)) {
3071 list_add_tail(&lockres->l_blocked_list,
3072 &osb->blocked_lock_list);
3073 osb->blocked_lock_count++;
3075 spin_unlock(&osb->vote_task_lock);
3080 /* This aids in debugging situations where a bad LVB might be involved. */
3081 void ocfs2_dump_meta_lvb_info(u64 level,
3082 const char *function,
3084 struct ocfs2_lock_res *lockres)
3086 struct ocfs2_meta_lvb *lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
3088 mlog(level, "LVB information for %s (called from %s:%u):\n",
3089 lockres->l_name, function, line);
3090 mlog(level, "version: %u, clusters: %u, generation: 0x%x\n",
3091 lvb->lvb_version, be32_to_cpu(lvb->lvb_iclusters),
3092 be32_to_cpu(lvb->lvb_igeneration));
3093 mlog(level, "size: %llu, uid %u, gid %u, mode 0x%x\n",
3094 (unsigned long long)be64_to_cpu(lvb->lvb_isize),
3095 be32_to_cpu(lvb->lvb_iuid), be32_to_cpu(lvb->lvb_igid),
3096 be16_to_cpu(lvb->lvb_imode));
3097 mlog(level, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
3098 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb->lvb_inlink),
3099 (long long)be64_to_cpu(lvb->lvb_iatime_packed),
3100 (long long)be64_to_cpu(lvb->lvb_ictime_packed),
3101 (long long)be64_to_cpu(lvb->lvb_imtime_packed),
3102 be32_to_cpu(lvb->lvb_iattr));