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>
35 #define MLOG_MASK_PREFIX ML_DLM_GLUE
36 #include <cluster/masklog.h>
39 #include "ocfs2_lockingver.h"
44 #include "extent_map.h"
46 #include "heartbeat.h"
49 #include "stackglue.h"
54 #include "buffer_head_io.h"
56 struct ocfs2_mask_waiter
{
57 struct list_head mw_item
;
59 struct completion mw_complete
;
60 unsigned long mw_mask
;
61 unsigned long mw_goal
;
64 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
);
65 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
);
66 static struct ocfs2_super
*ocfs2_get_file_osb(struct ocfs2_lock_res
*lockres
);
69 * Return value from ->downconvert_worker functions.
71 * These control the precise actions of ocfs2_unblock_lock()
72 * and ocfs2_process_blocked_lock()
75 enum ocfs2_unblock_action
{
76 UNBLOCK_CONTINUE
= 0, /* Continue downconvert */
77 UNBLOCK_CONTINUE_POST
= 1, /* Continue downconvert, fire
78 * ->post_unlock callback */
79 UNBLOCK_STOP_POST
= 2, /* Do not downconvert, fire
80 * ->post_unlock() callback. */
83 struct ocfs2_unblock_ctl
{
85 enum ocfs2_unblock_action unblock_action
;
88 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
90 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
);
92 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
95 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
98 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
99 struct ocfs2_lock_res
*lockres
);
102 #define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
104 /* This aids in debugging situations where a bad LVB might be involved. */
105 static void ocfs2_dump_meta_lvb_info(u64 level
,
106 const char *function
,
108 struct ocfs2_lock_res
*lockres
)
110 struct ocfs2_meta_lvb
*lvb
=
111 (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
113 mlog(level
, "LVB information for %s (called from %s:%u):\n",
114 lockres
->l_name
, function
, line
);
115 mlog(level
, "version: %u, clusters: %u, generation: 0x%x\n",
116 lvb
->lvb_version
, be32_to_cpu(lvb
->lvb_iclusters
),
117 be32_to_cpu(lvb
->lvb_igeneration
));
118 mlog(level
, "size: %llu, uid %u, gid %u, mode 0x%x\n",
119 (unsigned long long)be64_to_cpu(lvb
->lvb_isize
),
120 be32_to_cpu(lvb
->lvb_iuid
), be32_to_cpu(lvb
->lvb_igid
),
121 be16_to_cpu(lvb
->lvb_imode
));
122 mlog(level
, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
123 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb
->lvb_inlink
),
124 (long long)be64_to_cpu(lvb
->lvb_iatime_packed
),
125 (long long)be64_to_cpu(lvb
->lvb_ictime_packed
),
126 (long long)be64_to_cpu(lvb
->lvb_imtime_packed
),
127 be32_to_cpu(lvb
->lvb_iattr
));
132 * OCFS2 Lock Resource Operations
134 * These fine tune the behavior of the generic dlmglue locking infrastructure.
136 * The most basic of lock types can point ->l_priv to their respective
137 * struct ocfs2_super and allow the default actions to manage things.
139 * Right now, each lock type also needs to implement an init function,
140 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
141 * should be called when the lock is no longer needed (i.e., object
144 struct ocfs2_lock_res_ops
{
146 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
147 * this callback if ->l_priv is not an ocfs2_super pointer
149 struct ocfs2_super
* (*get_osb
)(struct ocfs2_lock_res
*);
152 * Optionally called in the downconvert thread after a
153 * successful downconvert. The lockres will not be referenced
154 * after this callback is called, so it is safe to free
157 * The exact semantics of when this is called are controlled
158 * by ->downconvert_worker()
160 void (*post_unlock
)(struct ocfs2_super
*, struct ocfs2_lock_res
*);
163 * Allow a lock type to add checks to determine whether it is
164 * safe to downconvert a lock. Return 0 to re-queue the
165 * downconvert at a later time, nonzero to continue.
167 * For most locks, the default checks that there are no
168 * incompatible holders are sufficient.
170 * Called with the lockres spinlock held.
172 int (*check_downconvert
)(struct ocfs2_lock_res
*, int);
175 * Allows a lock type to populate the lock value block. This
176 * is called on downconvert, and when we drop a lock.
178 * Locks that want to use this should set LOCK_TYPE_USES_LVB
179 * in the flags field.
181 * Called with the lockres spinlock held.
183 void (*set_lvb
)(struct ocfs2_lock_res
*);
186 * Called from the downconvert thread when it is determined
187 * that a lock will be downconverted. This is called without
188 * any locks held so the function can do work that might
189 * schedule (syncing out data, etc).
191 * This should return any one of the ocfs2_unblock_action
192 * values, depending on what it wants the thread to do.
194 int (*downconvert_worker
)(struct ocfs2_lock_res
*, int);
197 * LOCK_TYPE_* flags which describe the specific requirements
198 * of a lock type. Descriptions of each individual flag follow.
204 * Some locks want to "refresh" potentially stale data when a
205 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
206 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
207 * individual lockres l_flags member from the ast function. It is
208 * expected that the locking wrapper will clear the
209 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
211 #define LOCK_TYPE_REQUIRES_REFRESH 0x1
214 * Indicate that a lock type makes use of the lock value block. The
215 * ->set_lvb lock type callback must be defined.
217 #define LOCK_TYPE_USES_LVB 0x2
219 static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops
= {
220 .get_osb
= ocfs2_get_inode_osb
,
224 static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops
= {
225 .get_osb
= ocfs2_get_inode_osb
,
226 .check_downconvert
= ocfs2_check_meta_downconvert
,
227 .set_lvb
= ocfs2_set_meta_lvb
,
228 .downconvert_worker
= ocfs2_data_convert_worker
,
229 .flags
= LOCK_TYPE_REQUIRES_REFRESH
|LOCK_TYPE_USES_LVB
,
232 static struct ocfs2_lock_res_ops ocfs2_super_lops
= {
233 .flags
= LOCK_TYPE_REQUIRES_REFRESH
,
236 static struct ocfs2_lock_res_ops ocfs2_rename_lops
= {
240 static struct ocfs2_lock_res_ops ocfs2_dentry_lops
= {
241 .get_osb
= ocfs2_get_dentry_osb
,
242 .post_unlock
= ocfs2_dentry_post_unlock
,
243 .downconvert_worker
= ocfs2_dentry_convert_worker
,
247 static struct ocfs2_lock_res_ops ocfs2_inode_open_lops
= {
248 .get_osb
= ocfs2_get_inode_osb
,
252 static struct ocfs2_lock_res_ops ocfs2_flock_lops
= {
253 .get_osb
= ocfs2_get_file_osb
,
257 static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res
*lockres
)
259 return lockres
->l_type
== OCFS2_LOCK_TYPE_META
||
260 lockres
->l_type
== OCFS2_LOCK_TYPE_RW
||
261 lockres
->l_type
== OCFS2_LOCK_TYPE_OPEN
;
264 static inline struct inode
*ocfs2_lock_res_inode(struct ocfs2_lock_res
*lockres
)
266 BUG_ON(!ocfs2_is_inode_lock(lockres
));
268 return (struct inode
*) lockres
->l_priv
;
271 static inline struct ocfs2_dentry_lock
*ocfs2_lock_res_dl(struct ocfs2_lock_res
*lockres
)
273 BUG_ON(lockres
->l_type
!= OCFS2_LOCK_TYPE_DENTRY
);
275 return (struct ocfs2_dentry_lock
*)lockres
->l_priv
;
278 static inline struct ocfs2_super
*ocfs2_get_lockres_osb(struct ocfs2_lock_res
*lockres
)
280 if (lockres
->l_ops
->get_osb
)
281 return lockres
->l_ops
->get_osb(lockres
);
283 return (struct ocfs2_super
*)lockres
->l_priv
;
286 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
287 struct ocfs2_lock_res
*lockres
,
290 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
292 static void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
293 struct ocfs2_lock_res
*lockres
,
295 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
);
296 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
);
297 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
);
298 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
, int level
);
299 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
300 struct ocfs2_lock_res
*lockres
);
301 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
303 #define ocfs2_log_dlm_error(_func, _err, _lockres) do { \
304 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \
305 _err, _func, _lockres->l_name); \
307 static int ocfs2_downconvert_thread(void *arg
);
308 static void ocfs2_downconvert_on_unlock(struct ocfs2_super
*osb
,
309 struct ocfs2_lock_res
*lockres
);
310 static int ocfs2_inode_lock_update(struct inode
*inode
,
311 struct buffer_head
**bh
);
312 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
);
313 static inline int ocfs2_highest_compat_lock_level(int level
);
314 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
316 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
317 struct ocfs2_lock_res
*lockres
,
320 unsigned int generation
);
321 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
322 struct ocfs2_lock_res
*lockres
);
323 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
324 struct ocfs2_lock_res
*lockres
);
327 static void ocfs2_build_lock_name(enum ocfs2_lock_type type
,
336 BUG_ON(type
>= OCFS2_NUM_LOCK_TYPES
);
338 len
= snprintf(name
, OCFS2_LOCK_ID_MAX_LEN
, "%c%s%016llx%08x",
339 ocfs2_lock_type_char(type
), OCFS2_LOCK_ID_PAD
,
340 (long long)blkno
, generation
);
342 BUG_ON(len
!= (OCFS2_LOCK_ID_MAX_LEN
- 1));
344 mlog(0, "built lock resource with name: %s\n", name
);
349 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock
);
351 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res
*res
,
352 struct ocfs2_dlm_debug
*dlm_debug
)
354 mlog(0, "Add tracking for lockres %s\n", res
->l_name
);
356 spin_lock(&ocfs2_dlm_tracking_lock
);
357 list_add(&res
->l_debug_list
, &dlm_debug
->d_lockres_tracking
);
358 spin_unlock(&ocfs2_dlm_tracking_lock
);
361 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res
*res
)
363 spin_lock(&ocfs2_dlm_tracking_lock
);
364 if (!list_empty(&res
->l_debug_list
))
365 list_del_init(&res
->l_debug_list
);
366 spin_unlock(&ocfs2_dlm_tracking_lock
);
369 static void ocfs2_lock_res_init_common(struct ocfs2_super
*osb
,
370 struct ocfs2_lock_res
*res
,
371 enum ocfs2_lock_type type
,
372 struct ocfs2_lock_res_ops
*ops
,
379 res
->l_level
= DLM_LOCK_IV
;
380 res
->l_requested
= DLM_LOCK_IV
;
381 res
->l_blocking
= DLM_LOCK_IV
;
382 res
->l_action
= OCFS2_AST_INVALID
;
383 res
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
385 res
->l_flags
= OCFS2_LOCK_INITIALIZED
;
387 ocfs2_add_lockres_tracking(res
, osb
->osb_dlm_debug
);
390 void ocfs2_lock_res_init_once(struct ocfs2_lock_res
*res
)
392 /* This also clears out the lock status block */
393 memset(res
, 0, sizeof(struct ocfs2_lock_res
));
394 spin_lock_init(&res
->l_lock
);
395 init_waitqueue_head(&res
->l_event
);
396 INIT_LIST_HEAD(&res
->l_blocked_list
);
397 INIT_LIST_HEAD(&res
->l_mask_waiters
);
400 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res
*res
,
401 enum ocfs2_lock_type type
,
402 unsigned int generation
,
405 struct ocfs2_lock_res_ops
*ops
;
408 case OCFS2_LOCK_TYPE_RW
:
409 ops
= &ocfs2_inode_rw_lops
;
411 case OCFS2_LOCK_TYPE_META
:
412 ops
= &ocfs2_inode_inode_lops
;
414 case OCFS2_LOCK_TYPE_OPEN
:
415 ops
= &ocfs2_inode_open_lops
;
418 mlog_bug_on_msg(1, "type: %d\n", type
);
419 ops
= NULL
; /* thanks, gcc */
423 ocfs2_build_lock_name(type
, OCFS2_I(inode
)->ip_blkno
,
424 generation
, res
->l_name
);
425 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), res
, type
, ops
, inode
);
428 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
)
430 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
432 return OCFS2_SB(inode
->i_sb
);
435 static struct ocfs2_super
*ocfs2_get_file_osb(struct ocfs2_lock_res
*lockres
)
437 struct ocfs2_file_private
*fp
= lockres
->l_priv
;
439 return OCFS2_SB(fp
->fp_file
->f_mapping
->host
->i_sb
);
442 static __u64
ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res
*lockres
)
444 __be64 inode_blkno_be
;
446 memcpy(&inode_blkno_be
, &lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
],
449 return be64_to_cpu(inode_blkno_be
);
452 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
)
454 struct ocfs2_dentry_lock
*dl
= lockres
->l_priv
;
456 return OCFS2_SB(dl
->dl_inode
->i_sb
);
459 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock
*dl
,
460 u64 parent
, struct inode
*inode
)
463 u64 inode_blkno
= OCFS2_I(inode
)->ip_blkno
;
464 __be64 inode_blkno_be
= cpu_to_be64(inode_blkno
);
465 struct ocfs2_lock_res
*lockres
= &dl
->dl_lockres
;
467 ocfs2_lock_res_init_once(lockres
);
470 * Unfortunately, the standard lock naming scheme won't work
471 * here because we have two 16 byte values to use. Instead,
472 * we'll stuff the inode number as a binary value. We still
473 * want error prints to show something without garbling the
474 * display, so drop a null byte in there before the inode
475 * number. A future version of OCFS2 will likely use all
476 * binary lock names. The stringified names have been a
477 * tremendous aid in debugging, but now that the debugfs
478 * interface exists, we can mangle things there if need be.
480 * NOTE: We also drop the standard "pad" value (the total lock
481 * name size stays the same though - the last part is all
482 * zeros due to the memset in ocfs2_lock_res_init_once()
484 len
= snprintf(lockres
->l_name
, OCFS2_DENTRY_LOCK_INO_START
,
486 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY
),
489 BUG_ON(len
!= (OCFS2_DENTRY_LOCK_INO_START
- 1));
491 memcpy(&lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
], &inode_blkno_be
,
494 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
495 OCFS2_LOCK_TYPE_DENTRY
, &ocfs2_dentry_lops
,
499 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res
*res
,
500 struct ocfs2_super
*osb
)
502 /* Superblock lockres doesn't come from a slab so we call init
503 * once on it manually. */
504 ocfs2_lock_res_init_once(res
);
505 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER
, OCFS2_SUPER_BLOCK_BLKNO
,
507 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_SUPER
,
508 &ocfs2_super_lops
, osb
);
511 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res
*res
,
512 struct ocfs2_super
*osb
)
514 /* Rename lockres doesn't come from a slab so we call init
515 * once on it manually. */
516 ocfs2_lock_res_init_once(res
);
517 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME
, 0, 0, res
->l_name
);
518 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_RENAME
,
519 &ocfs2_rename_lops
, osb
);
522 void ocfs2_file_lock_res_init(struct ocfs2_lock_res
*lockres
,
523 struct ocfs2_file_private
*fp
)
525 struct inode
*inode
= fp
->fp_file
->f_mapping
->host
;
526 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
528 ocfs2_lock_res_init_once(lockres
);
529 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK
, oi
->ip_blkno
,
530 inode
->i_generation
, lockres
->l_name
);
531 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
532 OCFS2_LOCK_TYPE_FLOCK
, &ocfs2_flock_lops
,
534 lockres
->l_flags
|= OCFS2_LOCK_NOCACHE
;
537 void ocfs2_lock_res_free(struct ocfs2_lock_res
*res
)
541 if (!(res
->l_flags
& OCFS2_LOCK_INITIALIZED
))
544 ocfs2_remove_lockres_tracking(res
);
546 mlog_bug_on_msg(!list_empty(&res
->l_blocked_list
),
547 "Lockres %s is on the blocked list\n",
549 mlog_bug_on_msg(!list_empty(&res
->l_mask_waiters
),
550 "Lockres %s has mask waiters pending\n",
552 mlog_bug_on_msg(spin_is_locked(&res
->l_lock
),
553 "Lockres %s is locked\n",
555 mlog_bug_on_msg(res
->l_ro_holders
,
556 "Lockres %s has %u ro holders\n",
557 res
->l_name
, res
->l_ro_holders
);
558 mlog_bug_on_msg(res
->l_ex_holders
,
559 "Lockres %s has %u ex holders\n",
560 res
->l_name
, res
->l_ex_holders
);
562 /* Need to clear out the lock status block for the dlm */
563 memset(&res
->l_lksb
, 0, sizeof(res
->l_lksb
));
569 static inline void ocfs2_inc_holders(struct ocfs2_lock_res
*lockres
,
578 lockres
->l_ex_holders
++;
581 lockres
->l_ro_holders
++;
590 static inline void ocfs2_dec_holders(struct ocfs2_lock_res
*lockres
,
599 BUG_ON(!lockres
->l_ex_holders
);
600 lockres
->l_ex_holders
--;
603 BUG_ON(!lockres
->l_ro_holders
);
604 lockres
->l_ro_holders
--;
612 /* WARNING: This function lives in a world where the only three lock
613 * levels are EX, PR, and NL. It *will* have to be adjusted when more
614 * lock types are added. */
615 static inline int ocfs2_highest_compat_lock_level(int level
)
617 int new_level
= DLM_LOCK_EX
;
619 if (level
== DLM_LOCK_EX
)
620 new_level
= DLM_LOCK_NL
;
621 else if (level
== DLM_LOCK_PR
)
622 new_level
= DLM_LOCK_PR
;
626 static void lockres_set_flags(struct ocfs2_lock_res
*lockres
,
627 unsigned long newflags
)
629 struct ocfs2_mask_waiter
*mw
, *tmp
;
631 assert_spin_locked(&lockres
->l_lock
);
633 lockres
->l_flags
= newflags
;
635 list_for_each_entry_safe(mw
, tmp
, &lockres
->l_mask_waiters
, mw_item
) {
636 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
639 list_del_init(&mw
->mw_item
);
641 complete(&mw
->mw_complete
);
644 static void lockres_or_flags(struct ocfs2_lock_res
*lockres
, unsigned long or)
646 lockres_set_flags(lockres
, lockres
->l_flags
| or);
648 static void lockres_clear_flags(struct ocfs2_lock_res
*lockres
,
651 lockres_set_flags(lockres
, lockres
->l_flags
& ~clear
);
654 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
)
658 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
659 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
660 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
661 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
663 lockres
->l_level
= lockres
->l_requested
;
664 if (lockres
->l_level
<=
665 ocfs2_highest_compat_lock_level(lockres
->l_blocking
)) {
666 lockres
->l_blocking
= DLM_LOCK_NL
;
667 lockres_clear_flags(lockres
, OCFS2_LOCK_BLOCKED
);
669 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
674 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
)
678 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
679 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
681 /* Convert from RO to EX doesn't really need anything as our
682 * information is already up to data. Convert from NL to
683 * *anything* however should mark ourselves as needing an
685 if (lockres
->l_level
== DLM_LOCK_NL
&&
686 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
687 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
689 lockres
->l_level
= lockres
->l_requested
;
690 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
695 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
)
699 BUG_ON((!(lockres
->l_flags
& OCFS2_LOCK_BUSY
)));
700 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
702 if (lockres
->l_requested
> DLM_LOCK_NL
&&
703 !(lockres
->l_flags
& OCFS2_LOCK_LOCAL
) &&
704 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
705 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
707 lockres
->l_level
= lockres
->l_requested
;
708 lockres_or_flags(lockres
, OCFS2_LOCK_ATTACHED
);
709 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
714 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
,
717 int needs_downconvert
= 0;
720 assert_spin_locked(&lockres
->l_lock
);
722 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
724 if (level
> lockres
->l_blocking
) {
725 /* only schedule a downconvert if we haven't already scheduled
726 * one that goes low enough to satisfy the level we're
727 * blocking. this also catches the case where we get
729 if (ocfs2_highest_compat_lock_level(level
) <
730 ocfs2_highest_compat_lock_level(lockres
->l_blocking
))
731 needs_downconvert
= 1;
733 lockres
->l_blocking
= level
;
736 mlog_exit(needs_downconvert
);
737 return needs_downconvert
;
741 * OCFS2_LOCK_PENDING and l_pending_gen.
743 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
744 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
745 * for more details on the race.
747 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
748 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
749 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
750 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
751 * the caller is going to try to clear PENDING again. If nothing else is
752 * happening, __lockres_clear_pending() sees PENDING is unset and does
755 * But what if another path (eg downconvert thread) has just started a
756 * new locking action? The other path has re-set PENDING. Our path
757 * cannot clear PENDING, because that will re-open the original race
763 * ocfs2_cluster_lock()
768 * ocfs2_locking_ast() ocfs2_downconvert_thread()
769 * clear PENDING ocfs2_unblock_lock()
772 * ocfs2_prepare_downconvert()
782 * So as you can see, we now have a window where l_lock is not held,
783 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
785 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
786 * set by ocfs2_prepare_downconvert(). That wasn't nice.
788 * To solve this we introduce l_pending_gen. A call to
789 * lockres_clear_pending() will only do so when it is passed a generation
790 * number that matches the lockres. lockres_set_pending() will return the
791 * current generation number. When ocfs2_cluster_lock() goes to clear
792 * PENDING, it passes the generation it got from set_pending(). In our
793 * example above, the generation numbers will *not* match. Thus,
794 * ocfs2_cluster_lock() will not clear the PENDING set by
795 * ocfs2_prepare_downconvert().
798 /* Unlocked version for ocfs2_locking_ast() */
799 static void __lockres_clear_pending(struct ocfs2_lock_res
*lockres
,
800 unsigned int generation
,
801 struct ocfs2_super
*osb
)
803 assert_spin_locked(&lockres
->l_lock
);
806 * The ast and locking functions can race us here. The winner
807 * will clear pending, the loser will not.
809 if (!(lockres
->l_flags
& OCFS2_LOCK_PENDING
) ||
810 (lockres
->l_pending_gen
!= generation
))
813 lockres_clear_flags(lockres
, OCFS2_LOCK_PENDING
);
814 lockres
->l_pending_gen
++;
817 * The downconvert thread may have skipped us because we
818 * were PENDING. Wake it up.
820 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
821 ocfs2_wake_downconvert_thread(osb
);
824 /* Locked version for callers of ocfs2_dlm_lock() */
825 static void lockres_clear_pending(struct ocfs2_lock_res
*lockres
,
826 unsigned int generation
,
827 struct ocfs2_super
*osb
)
831 spin_lock_irqsave(&lockres
->l_lock
, flags
);
832 __lockres_clear_pending(lockres
, generation
, osb
);
833 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
836 static unsigned int lockres_set_pending(struct ocfs2_lock_res
*lockres
)
838 assert_spin_locked(&lockres
->l_lock
);
839 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
841 lockres_or_flags(lockres
, OCFS2_LOCK_PENDING
);
843 return lockres
->l_pending_gen
;
847 static void ocfs2_blocking_ast(void *opaque
, int level
)
849 struct ocfs2_lock_res
*lockres
= opaque
;
850 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
851 int needs_downconvert
;
854 BUG_ON(level
<= DLM_LOCK_NL
);
856 mlog(0, "BAST fired for lockres %s, blocking %d, level %d type %s\n",
857 lockres
->l_name
, level
, lockres
->l_level
,
858 ocfs2_lock_type_string(lockres
->l_type
));
861 * We can skip the bast for locks which don't enable caching -
862 * they'll be dropped at the earliest possible time anyway.
864 if (lockres
->l_flags
& OCFS2_LOCK_NOCACHE
)
867 spin_lock_irqsave(&lockres
->l_lock
, flags
);
868 needs_downconvert
= ocfs2_generic_handle_bast(lockres
, level
);
869 if (needs_downconvert
)
870 ocfs2_schedule_blocked_lock(osb
, lockres
);
871 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
873 wake_up(&lockres
->l_event
);
875 ocfs2_wake_downconvert_thread(osb
);
878 static void ocfs2_locking_ast(void *opaque
)
880 struct ocfs2_lock_res
*lockres
= opaque
;
881 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
884 spin_lock_irqsave(&lockres
->l_lock
, flags
);
886 if (ocfs2_dlm_lock_status(&lockres
->l_lksb
)) {
887 mlog(ML_ERROR
, "lockres %s: lksb status value of %d!\n",
889 ocfs2_dlm_lock_status(&lockres
->l_lksb
));
890 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
894 switch(lockres
->l_action
) {
895 case OCFS2_AST_ATTACH
:
896 ocfs2_generic_handle_attach_action(lockres
);
897 lockres_clear_flags(lockres
, OCFS2_LOCK_LOCAL
);
899 case OCFS2_AST_CONVERT
:
900 ocfs2_generic_handle_convert_action(lockres
);
902 case OCFS2_AST_DOWNCONVERT
:
903 ocfs2_generic_handle_downconvert_action(lockres
);
906 mlog(ML_ERROR
, "lockres %s: ast fired with invalid action: %u "
907 "lockres flags = 0x%lx, unlock action: %u\n",
908 lockres
->l_name
, lockres
->l_action
, lockres
->l_flags
,
909 lockres
->l_unlock_action
);
913 /* set it to something invalid so if we get called again we
915 lockres
->l_action
= OCFS2_AST_INVALID
;
917 /* Did we try to cancel this lock? Clear that state */
918 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
)
919 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
922 * We may have beaten the locking functions here. We certainly
923 * know that dlm_lock() has been called :-)
924 * Because we can't have two lock calls in flight at once, we
925 * can use lockres->l_pending_gen.
927 __lockres_clear_pending(lockres
, lockres
->l_pending_gen
, osb
);
929 wake_up(&lockres
->l_event
);
930 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
933 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
939 spin_lock_irqsave(&lockres
->l_lock
, flags
);
940 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
942 lockres
->l_action
= OCFS2_AST_INVALID
;
944 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
945 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
947 wake_up(&lockres
->l_event
);
951 /* Note: If we detect another process working on the lock (i.e.,
952 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
953 * to do the right thing in that case.
955 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
956 struct ocfs2_lock_res
*lockres
,
966 mlog(0, "lock %s, level = %d, flags = %u\n", lockres
->l_name
, level
,
969 spin_lock_irqsave(&lockres
->l_lock
, flags
);
970 if ((lockres
->l_flags
& OCFS2_LOCK_ATTACHED
) ||
971 (lockres
->l_flags
& OCFS2_LOCK_BUSY
)) {
972 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
976 lockres
->l_action
= OCFS2_AST_ATTACH
;
977 lockres
->l_requested
= level
;
978 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
979 gen
= lockres_set_pending(lockres
);
980 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
982 ret
= ocfs2_dlm_lock(osb
->cconn
,
987 OCFS2_LOCK_ID_MAX_LEN
- 1,
989 lockres_clear_pending(lockres
, gen
, osb
);
991 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
992 ocfs2_recover_from_dlm_error(lockres
, 1);
995 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres
->l_name
);
1002 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res
*lockres
,
1005 unsigned long flags
;
1008 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1009 ret
= lockres
->l_flags
& flag
;
1010 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1015 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res
*lockres
)
1018 wait_event(lockres
->l_event
,
1019 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_BUSY
));
1022 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res
*lockres
)
1025 wait_event(lockres
->l_event
,
1026 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_REFRESHING
));
1029 /* predict what lock level we'll be dropping down to on behalf
1030 * of another node, and return true if the currently wanted
1031 * level will be compatible with it. */
1032 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
1035 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
1037 return wanted
<= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
1040 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter
*mw
)
1042 INIT_LIST_HEAD(&mw
->mw_item
);
1043 init_completion(&mw
->mw_complete
);
1046 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter
*mw
)
1048 wait_for_completion(&mw
->mw_complete
);
1049 /* Re-arm the completion in case we want to wait on it again */
1050 INIT_COMPLETION(mw
->mw_complete
);
1051 return mw
->mw_status
;
1054 static void lockres_add_mask_waiter(struct ocfs2_lock_res
*lockres
,
1055 struct ocfs2_mask_waiter
*mw
,
1059 BUG_ON(!list_empty(&mw
->mw_item
));
1061 assert_spin_locked(&lockres
->l_lock
);
1063 list_add_tail(&mw
->mw_item
, &lockres
->l_mask_waiters
);
1068 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
1069 * if the mask still hadn't reached its goal */
1070 static int lockres_remove_mask_waiter(struct ocfs2_lock_res
*lockres
,
1071 struct ocfs2_mask_waiter
*mw
)
1073 unsigned long flags
;
1076 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1077 if (!list_empty(&mw
->mw_item
)) {
1078 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
1081 list_del_init(&mw
->mw_item
);
1082 init_completion(&mw
->mw_complete
);
1084 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1090 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter
*mw
,
1091 struct ocfs2_lock_res
*lockres
)
1095 ret
= wait_for_completion_interruptible(&mw
->mw_complete
);
1097 lockres_remove_mask_waiter(lockres
, mw
);
1099 ret
= mw
->mw_status
;
1100 /* Re-arm the completion in case we want to wait on it again */
1101 INIT_COMPLETION(mw
->mw_complete
);
1105 static int ocfs2_cluster_lock(struct ocfs2_super
*osb
,
1106 struct ocfs2_lock_res
*lockres
,
1111 struct ocfs2_mask_waiter mw
;
1112 int wait
, catch_signals
= !(osb
->s_mount_opt
& OCFS2_MOUNT_NOINTR
);
1113 int ret
= 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1114 unsigned long flags
;
1119 ocfs2_init_mask_waiter(&mw
);
1121 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
1122 lkm_flags
|= DLM_LKF_VALBLK
;
1127 if (catch_signals
&& signal_pending(current
)) {
1132 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1134 mlog_bug_on_msg(lockres
->l_flags
& OCFS2_LOCK_FREEING
,
1135 "Cluster lock called on freeing lockres %s! flags "
1136 "0x%lx\n", lockres
->l_name
, lockres
->l_flags
);
1138 /* We only compare against the currently granted level
1139 * here. If the lock is blocked waiting on a downconvert,
1140 * we'll get caught below. */
1141 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
&&
1142 level
> lockres
->l_level
) {
1143 /* is someone sitting in dlm_lock? If so, wait on
1145 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1150 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
&&
1151 !ocfs2_may_continue_on_blocked_lock(lockres
, level
)) {
1152 /* is the lock is currently blocked on behalf of
1154 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BLOCKED
, 0);
1159 if (level
> lockres
->l_level
) {
1160 if (lockres
->l_action
!= OCFS2_AST_INVALID
)
1161 mlog(ML_ERROR
, "lockres %s has action %u pending\n",
1162 lockres
->l_name
, lockres
->l_action
);
1164 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1165 lockres
->l_action
= OCFS2_AST_ATTACH
;
1166 lkm_flags
&= ~DLM_LKF_CONVERT
;
1168 lockres
->l_action
= OCFS2_AST_CONVERT
;
1169 lkm_flags
|= DLM_LKF_CONVERT
;
1172 lockres
->l_requested
= level
;
1173 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1174 gen
= lockres_set_pending(lockres
);
1175 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1177 BUG_ON(level
== DLM_LOCK_IV
);
1178 BUG_ON(level
== DLM_LOCK_NL
);
1180 mlog(0, "lock %s, convert from %d to level = %d\n",
1181 lockres
->l_name
, lockres
->l_level
, level
);
1183 /* call dlm_lock to upgrade lock now */
1184 ret
= ocfs2_dlm_lock(osb
->cconn
,
1189 OCFS2_LOCK_ID_MAX_LEN
- 1,
1191 lockres_clear_pending(lockres
, gen
, osb
);
1193 if (!(lkm_flags
& DLM_LKF_NOQUEUE
) ||
1195 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1198 ocfs2_recover_from_dlm_error(lockres
, 1);
1202 mlog(0, "lock %s, successfull return from ocfs2_dlm_lock\n",
1205 /* At this point we've gone inside the dlm and need to
1206 * complete our work regardless. */
1209 /* wait for busy to clear and carry on */
1213 /* Ok, if we get here then we're good to go. */
1214 ocfs2_inc_holders(lockres
, level
);
1218 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1221 * This is helping work around a lock inversion between the page lock
1222 * and dlm locks. One path holds the page lock while calling aops
1223 * which block acquiring dlm locks. The voting thread holds dlm
1224 * locks while acquiring page locks while down converting data locks.
1225 * This block is helping an aop path notice the inversion and back
1226 * off to unlock its page lock before trying the dlm lock again.
1228 if (wait
&& arg_flags
& OCFS2_LOCK_NONBLOCK
&&
1229 mw
.mw_mask
& (OCFS2_LOCK_BUSY
|OCFS2_LOCK_BLOCKED
)) {
1231 if (lockres_remove_mask_waiter(lockres
, &mw
))
1237 ret
= ocfs2_wait_for_mask(&mw
);
1247 static void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
1248 struct ocfs2_lock_res
*lockres
,
1251 unsigned long flags
;
1254 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1255 ocfs2_dec_holders(lockres
, level
);
1256 ocfs2_downconvert_on_unlock(osb
, lockres
);
1257 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1261 static int ocfs2_create_new_lock(struct ocfs2_super
*osb
,
1262 struct ocfs2_lock_res
*lockres
,
1266 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1267 unsigned long flags
;
1268 u32 lkm_flags
= local
? DLM_LKF_LOCAL
: 0;
1270 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1271 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
1272 lockres_or_flags(lockres
, OCFS2_LOCK_LOCAL
);
1273 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1275 return ocfs2_lock_create(osb
, lockres
, level
, lkm_flags
);
1278 /* Grants us an EX lock on the data and metadata resources, skipping
1279 * the normal cluster directory lookup. Use this ONLY on newly created
1280 * inodes which other nodes can't possibly see, and which haven't been
1281 * hashed in the inode hash yet. This can give us a good performance
1282 * increase as it'll skip the network broadcast normally associated
1283 * with creating a new lock resource. */
1284 int ocfs2_create_new_inode_locks(struct inode
*inode
)
1287 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1290 BUG_ON(!ocfs2_inode_is_new(inode
));
1294 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1296 /* NOTE: That we don't increment any of the holder counts, nor
1297 * do we add anything to a journal handle. Since this is
1298 * supposed to be a new inode which the cluster doesn't know
1299 * about yet, there is no need to. As far as the LVB handling
1300 * is concerned, this is basically like acquiring an EX lock
1301 * on a resource which has an invalid one -- we'll set it
1302 * valid when we release the EX. */
1304 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_rw_lockres
, 1, 1);
1311 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1312 * don't use a generation in their lock names.
1314 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_inode_lockres
, 1, 0);
1320 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_open_lockres
, 0, 0);
1331 int ocfs2_rw_lock(struct inode
*inode
, int write
)
1334 struct ocfs2_lock_res
*lockres
;
1335 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1341 mlog(0, "inode %llu take %s RW lock\n",
1342 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1343 write
? "EXMODE" : "PRMODE");
1345 if (ocfs2_mount_local(osb
))
1348 lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1350 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1352 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
, level
, 0,
1361 void ocfs2_rw_unlock(struct inode
*inode
, int write
)
1363 int level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1364 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1365 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1369 mlog(0, "inode %llu drop %s RW lock\n",
1370 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1371 write
? "EXMODE" : "PRMODE");
1373 if (!ocfs2_mount_local(osb
))
1374 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
1380 * ocfs2_open_lock always get PR mode lock.
1382 int ocfs2_open_lock(struct inode
*inode
)
1385 struct ocfs2_lock_res
*lockres
;
1386 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1392 mlog(0, "inode %llu take PRMODE open lock\n",
1393 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1395 if (ocfs2_mount_local(osb
))
1398 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1400 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
,
1410 int ocfs2_try_open_lock(struct inode
*inode
, int write
)
1412 int status
= 0, level
;
1413 struct ocfs2_lock_res
*lockres
;
1414 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1420 mlog(0, "inode %llu try to take %s open lock\n",
1421 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1422 write
? "EXMODE" : "PRMODE");
1424 if (ocfs2_mount_local(osb
))
1427 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1429 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1432 * The file system may already holding a PRMODE/EXMODE open lock.
1433 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1434 * other nodes and the -EAGAIN will indicate to the caller that
1435 * this inode is still in use.
1437 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
,
1438 level
, DLM_LKF_NOQUEUE
, 0);
1446 * ocfs2_open_unlock unlock PR and EX mode open locks.
1448 void ocfs2_open_unlock(struct inode
*inode
)
1450 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1451 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1455 mlog(0, "inode %llu drop open lock\n",
1456 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1458 if (ocfs2_mount_local(osb
))
1461 if(lockres
->l_ro_holders
)
1462 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
,
1464 if(lockres
->l_ex_holders
)
1465 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
,
1472 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res
*lockres
,
1476 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
1477 unsigned long flags
;
1478 struct ocfs2_mask_waiter mw
;
1480 ocfs2_init_mask_waiter(&mw
);
1483 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1484 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
1485 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
1487 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1488 ret
= ocfs2_cancel_convert(osb
, lockres
);
1495 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1496 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1498 ocfs2_wait_for_mask(&mw
);
1504 * We may still have gotten the lock, in which case there's no
1505 * point to restarting the syscall.
1507 if (lockres
->l_level
== level
)
1510 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret
,
1511 lockres
->l_flags
, lockres
->l_level
, lockres
->l_action
);
1513 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1520 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1521 * flock() calls. The locking approach this requires is sufficiently
1522 * different from all other cluster lock types that we implement a
1523 * seperate path to the "low-level" dlm calls. In particular:
1525 * - No optimization of lock levels is done - we take at exactly
1526 * what's been requested.
1528 * - No lock caching is employed. We immediately downconvert to
1529 * no-lock at unlock time. This also means flock locks never go on
1530 * the blocking list).
1532 * - Since userspace can trivially deadlock itself with flock, we make
1533 * sure to allow cancellation of a misbehaving applications flock()
1536 * - Access to any flock lockres doesn't require concurrency, so we
1537 * can simplify the code by requiring the caller to guarantee
1538 * serialization of dlmglue flock calls.
1540 int ocfs2_file_lock(struct file
*file
, int ex
, int trylock
)
1542 int ret
, level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1543 unsigned int lkm_flags
= trylock
? LKM_NOQUEUE
: 0;
1544 unsigned long flags
;
1545 struct ocfs2_file_private
*fp
= file
->private_data
;
1546 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
1547 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
1548 struct ocfs2_mask_waiter mw
;
1550 ocfs2_init_mask_waiter(&mw
);
1552 if ((lockres
->l_flags
& OCFS2_LOCK_BUSY
) ||
1553 (lockres
->l_level
> DLM_LOCK_NL
)) {
1555 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1556 "level: %u\n", lockres
->l_name
, lockres
->l_flags
,
1561 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1562 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1563 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1564 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1567 * Get the lock at NLMODE to start - that way we
1568 * can cancel the upconvert request if need be.
1570 ret
= ocfs2_lock_create(osb
, lockres
, LKM_NLMODE
, 0);
1576 ret
= ocfs2_wait_for_mask(&mw
);
1581 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1584 lockres
->l_action
= OCFS2_AST_CONVERT
;
1585 lkm_flags
|= LKM_CONVERT
;
1586 lockres
->l_requested
= level
;
1587 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1589 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1590 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1592 ret
= ocfs2_dlm_lock(osb
->cconn
, level
, &lockres
->l_lksb
, lkm_flags
,
1593 lockres
->l_name
, OCFS2_LOCK_ID_MAX_LEN
- 1,
1596 if (!trylock
|| (ret
!= -EAGAIN
)) {
1597 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
1601 ocfs2_recover_from_dlm_error(lockres
, 1);
1602 lockres_remove_mask_waiter(lockres
, &mw
);
1606 ret
= ocfs2_wait_for_mask_interruptible(&mw
, lockres
);
1607 if (ret
== -ERESTARTSYS
) {
1609 * Userspace can cause deadlock itself with
1610 * flock(). Current behavior locally is to allow the
1611 * deadlock, but abort the system call if a signal is
1612 * received. We follow this example, otherwise a
1613 * poorly written program could sit in kernel until
1616 * Handling this is a bit more complicated for Ocfs2
1617 * though. We can't exit this function with an
1618 * outstanding lock request, so a cancel convert is
1619 * required. We intentionally overwrite 'ret' - if the
1620 * cancel fails and the lock was granted, it's easier
1621 * to just bubble sucess back up to the user.
1623 ret
= ocfs2_flock_handle_signal(lockres
, level
);
1628 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1629 lockres
->l_name
, ex
, trylock
, ret
);
1633 void ocfs2_file_unlock(struct file
*file
)
1637 unsigned long flags
;
1638 struct ocfs2_file_private
*fp
= file
->private_data
;
1639 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
1640 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
1641 struct ocfs2_mask_waiter mw
;
1643 ocfs2_init_mask_waiter(&mw
);
1645 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
))
1648 if (lockres
->l_level
== LKM_NLMODE
)
1651 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1652 lockres
->l_name
, lockres
->l_flags
, lockres
->l_level
,
1655 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1657 * Fake a blocking ast for the downconvert code.
1659 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
1660 lockres
->l_blocking
= DLM_LOCK_EX
;
1662 gen
= ocfs2_prepare_downconvert(lockres
, LKM_NLMODE
);
1663 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1664 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1666 ret
= ocfs2_downconvert_lock(osb
, lockres
, LKM_NLMODE
, 0, gen
);
1672 ret
= ocfs2_wait_for_mask(&mw
);
1677 static void ocfs2_downconvert_on_unlock(struct ocfs2_super
*osb
,
1678 struct ocfs2_lock_res
*lockres
)
1684 /* If we know that another node is waiting on our lock, kick
1685 * the downconvert thread * pre-emptively when we reach a release
1687 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
) {
1688 switch(lockres
->l_blocking
) {
1690 if (!lockres
->l_ex_holders
&& !lockres
->l_ro_holders
)
1694 if (!lockres
->l_ex_holders
)
1703 ocfs2_wake_downconvert_thread(osb
);
1708 #define OCFS2_SEC_BITS 34
1709 #define OCFS2_SEC_SHIFT (64 - 34)
1710 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
1712 /* LVB only has room for 64 bits of time here so we pack it for
1714 static u64
ocfs2_pack_timespec(struct timespec
*spec
)
1717 u64 sec
= spec
->tv_sec
;
1718 u32 nsec
= spec
->tv_nsec
;
1720 res
= (sec
<< OCFS2_SEC_SHIFT
) | (nsec
& OCFS2_NSEC_MASK
);
1725 /* Call this with the lockres locked. I am reasonably sure we don't
1726 * need ip_lock in this function as anyone who would be changing those
1727 * values is supposed to be blocked in ocfs2_inode_lock right now. */
1728 static void __ocfs2_stuff_meta_lvb(struct inode
*inode
)
1730 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1731 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1732 struct ocfs2_meta_lvb
*lvb
;
1736 lvb
= (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1739 * Invalidate the LVB of a deleted inode - this way other
1740 * nodes are forced to go to disk and discover the new inode
1743 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1744 lvb
->lvb_version
= 0;
1748 lvb
->lvb_version
= OCFS2_LVB_VERSION
;
1749 lvb
->lvb_isize
= cpu_to_be64(i_size_read(inode
));
1750 lvb
->lvb_iclusters
= cpu_to_be32(oi
->ip_clusters
);
1751 lvb
->lvb_iuid
= cpu_to_be32(inode
->i_uid
);
1752 lvb
->lvb_igid
= cpu_to_be32(inode
->i_gid
);
1753 lvb
->lvb_imode
= cpu_to_be16(inode
->i_mode
);
1754 lvb
->lvb_inlink
= cpu_to_be16(inode
->i_nlink
);
1755 lvb
->lvb_iatime_packed
=
1756 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_atime
));
1757 lvb
->lvb_ictime_packed
=
1758 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_ctime
));
1759 lvb
->lvb_imtime_packed
=
1760 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_mtime
));
1761 lvb
->lvb_iattr
= cpu_to_be32(oi
->ip_attr
);
1762 lvb
->lvb_idynfeatures
= cpu_to_be16(oi
->ip_dyn_features
);
1763 lvb
->lvb_igeneration
= cpu_to_be32(inode
->i_generation
);
1766 mlog_meta_lvb(0, lockres
);
1771 static void ocfs2_unpack_timespec(struct timespec
*spec
,
1774 spec
->tv_sec
= packed_time
>> OCFS2_SEC_SHIFT
;
1775 spec
->tv_nsec
= packed_time
& OCFS2_NSEC_MASK
;
1778 static void ocfs2_refresh_inode_from_lvb(struct inode
*inode
)
1780 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1781 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1782 struct ocfs2_meta_lvb
*lvb
;
1786 mlog_meta_lvb(0, lockres
);
1788 lvb
= (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1790 /* We're safe here without the lockres lock... */
1791 spin_lock(&oi
->ip_lock
);
1792 oi
->ip_clusters
= be32_to_cpu(lvb
->lvb_iclusters
);
1793 i_size_write(inode
, be64_to_cpu(lvb
->lvb_isize
));
1795 oi
->ip_attr
= be32_to_cpu(lvb
->lvb_iattr
);
1796 oi
->ip_dyn_features
= be16_to_cpu(lvb
->lvb_idynfeatures
);
1797 ocfs2_set_inode_flags(inode
);
1799 /* fast-symlinks are a special case */
1800 if (S_ISLNK(inode
->i_mode
) && !oi
->ip_clusters
)
1801 inode
->i_blocks
= 0;
1803 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
1805 inode
->i_uid
= be32_to_cpu(lvb
->lvb_iuid
);
1806 inode
->i_gid
= be32_to_cpu(lvb
->lvb_igid
);
1807 inode
->i_mode
= be16_to_cpu(lvb
->lvb_imode
);
1808 inode
->i_nlink
= be16_to_cpu(lvb
->lvb_inlink
);
1809 ocfs2_unpack_timespec(&inode
->i_atime
,
1810 be64_to_cpu(lvb
->lvb_iatime_packed
));
1811 ocfs2_unpack_timespec(&inode
->i_mtime
,
1812 be64_to_cpu(lvb
->lvb_imtime_packed
));
1813 ocfs2_unpack_timespec(&inode
->i_ctime
,
1814 be64_to_cpu(lvb
->lvb_ictime_packed
));
1815 spin_unlock(&oi
->ip_lock
);
1820 static inline int ocfs2_meta_lvb_is_trustable(struct inode
*inode
,
1821 struct ocfs2_lock_res
*lockres
)
1823 struct ocfs2_meta_lvb
*lvb
=
1824 (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1826 if (lvb
->lvb_version
== OCFS2_LVB_VERSION
1827 && be32_to_cpu(lvb
->lvb_igeneration
) == inode
->i_generation
)
1832 /* Determine whether a lock resource needs to be refreshed, and
1833 * arbitrate who gets to refresh it.
1835 * 0 means no refresh needed.
1837 * > 0 means you need to refresh this and you MUST call
1838 * ocfs2_complete_lock_res_refresh afterwards. */
1839 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res
*lockres
)
1841 unsigned long flags
;
1847 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1848 if (!(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
)) {
1849 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1853 if (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
) {
1854 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1856 ocfs2_wait_on_refreshing_lock(lockres
);
1860 /* Ok, I'll be the one to refresh this lock. */
1861 lockres_or_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1862 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1870 /* If status is non zero, I'll mark it as not being in refresh
1871 * anymroe, but i won't clear the needs refresh flag. */
1872 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res
*lockres
,
1875 unsigned long flags
;
1878 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1879 lockres_clear_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1881 lockres_clear_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
1882 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1884 wake_up(&lockres
->l_event
);
1889 /* may or may not return a bh if it went to disk. */
1890 static int ocfs2_inode_lock_update(struct inode
*inode
,
1891 struct buffer_head
**bh
)
1894 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1895 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1896 struct ocfs2_dinode
*fe
;
1897 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1901 if (ocfs2_mount_local(osb
))
1904 spin_lock(&oi
->ip_lock
);
1905 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1906 mlog(0, "Orphaned inode %llu was deleted while we "
1907 "were waiting on a lock. ip_flags = 0x%x\n",
1908 (unsigned long long)oi
->ip_blkno
, oi
->ip_flags
);
1909 spin_unlock(&oi
->ip_lock
);
1913 spin_unlock(&oi
->ip_lock
);
1915 if (!ocfs2_should_refresh_lock_res(lockres
))
1918 /* This will discard any caching information we might have had
1919 * for the inode metadata. */
1920 ocfs2_metadata_cache_purge(inode
);
1922 ocfs2_extent_map_trunc(inode
, 0);
1924 if (ocfs2_meta_lvb_is_trustable(inode
, lockres
)) {
1925 mlog(0, "Trusting LVB on inode %llu\n",
1926 (unsigned long long)oi
->ip_blkno
);
1927 ocfs2_refresh_inode_from_lvb(inode
);
1929 /* Boo, we have to go to disk. */
1930 /* read bh, cast, ocfs2_refresh_inode */
1931 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), oi
->ip_blkno
,
1932 bh
, OCFS2_BH_CACHED
, inode
);
1937 fe
= (struct ocfs2_dinode
*) (*bh
)->b_data
;
1939 /* This is a good chance to make sure we're not
1940 * locking an invalid object.
1942 * We bug on a stale inode here because we checked
1943 * above whether it was wiped from disk. The wiping
1944 * node provides a guarantee that we receive that
1945 * message and can mark the inode before dropping any
1946 * locks associated with it. */
1947 if (!OCFS2_IS_VALID_DINODE(fe
)) {
1948 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
1952 mlog_bug_on_msg(inode
->i_generation
!=
1953 le32_to_cpu(fe
->i_generation
),
1954 "Invalid dinode %llu disk generation: %u "
1955 "inode->i_generation: %u\n",
1956 (unsigned long long)oi
->ip_blkno
,
1957 le32_to_cpu(fe
->i_generation
),
1958 inode
->i_generation
);
1959 mlog_bug_on_msg(le64_to_cpu(fe
->i_dtime
) ||
1960 !(fe
->i_flags
& cpu_to_le32(OCFS2_VALID_FL
)),
1961 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
1962 (unsigned long long)oi
->ip_blkno
,
1963 (unsigned long long)le64_to_cpu(fe
->i_dtime
),
1964 le32_to_cpu(fe
->i_flags
));
1966 ocfs2_refresh_inode(inode
, fe
);
1971 ocfs2_complete_lock_res_refresh(lockres
, status
);
1977 static int ocfs2_assign_bh(struct inode
*inode
,
1978 struct buffer_head
**ret_bh
,
1979 struct buffer_head
*passed_bh
)
1984 /* Ok, the update went to disk for us, use the
1986 *ret_bh
= passed_bh
;
1992 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
1993 OCFS2_I(inode
)->ip_blkno
,
2004 * returns < 0 error if the callback will never be called, otherwise
2005 * the result of the lock will be communicated via the callback.
2007 int ocfs2_inode_lock_full(struct inode
*inode
,
2008 struct buffer_head
**ret_bh
,
2012 int status
, level
, acquired
;
2014 struct ocfs2_lock_res
*lockres
= NULL
;
2015 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2016 struct buffer_head
*local_bh
= NULL
;
2022 mlog(0, "inode %llu, take %s META lock\n",
2023 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2024 ex
? "EXMODE" : "PRMODE");
2028 /* We'll allow faking a readonly metadata lock for
2030 if (ocfs2_is_hard_readonly(osb
)) {
2036 if (ocfs2_mount_local(osb
))
2039 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
2040 ocfs2_wait_for_recovery(osb
);
2042 lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2043 level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2045 if (arg_flags
& OCFS2_META_LOCK_NOQUEUE
)
2046 dlm_flags
|= DLM_LKF_NOQUEUE
;
2048 status
= ocfs2_cluster_lock(osb
, lockres
, level
, dlm_flags
, arg_flags
);
2050 if (status
!= -EAGAIN
&& status
!= -EIOCBRETRY
)
2055 /* Notify the error cleanup path to drop the cluster lock. */
2058 /* We wait twice because a node may have died while we were in
2059 * the lower dlm layers. The second time though, we've
2060 * committed to owning this lock so we don't allow signals to
2061 * abort the operation. */
2062 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
2063 ocfs2_wait_for_recovery(osb
);
2067 * We only see this flag if we're being called from
2068 * ocfs2_read_locked_inode(). It means we're locking an inode
2069 * which hasn't been populated yet, so clear the refresh flag
2070 * and let the caller handle it.
2072 if (inode
->i_state
& I_NEW
) {
2075 ocfs2_complete_lock_res_refresh(lockres
, 0);
2079 /* This is fun. The caller may want a bh back, or it may
2080 * not. ocfs2_inode_lock_update definitely wants one in, but
2081 * may or may not read one, depending on what's in the
2082 * LVB. The result of all of this is that we've *only* gone to
2083 * disk if we have to, so the complexity is worthwhile. */
2084 status
= ocfs2_inode_lock_update(inode
, &local_bh
);
2086 if (status
!= -ENOENT
)
2092 status
= ocfs2_assign_bh(inode
, ret_bh
, local_bh
);
2101 if (ret_bh
&& (*ret_bh
)) {
2106 ocfs2_inode_unlock(inode
, ex
);
2117 * This is working around a lock inversion between tasks acquiring DLM
2118 * locks while holding a page lock and the downconvert thread which
2119 * blocks dlm lock acquiry while acquiring page locks.
2121 * ** These _with_page variantes are only intended to be called from aop
2122 * methods that hold page locks and return a very specific *positive* error
2123 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2125 * The DLM is called such that it returns -EAGAIN if it would have
2126 * blocked waiting for the downconvert thread. In that case we unlock
2127 * our page so the downconvert thread can make progress. Once we've
2128 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2129 * that called us can bubble that back up into the VFS who will then
2130 * immediately retry the aop call.
2132 * We do a blocking lock and immediate unlock before returning, though, so that
2133 * the lock has a great chance of being cached on this node by the time the VFS
2134 * calls back to retry the aop. This has a potential to livelock as nodes
2135 * ping locks back and forth, but that's a risk we're willing to take to avoid
2136 * the lock inversion simply.
2138 int ocfs2_inode_lock_with_page(struct inode
*inode
,
2139 struct buffer_head
**ret_bh
,
2145 ret
= ocfs2_inode_lock_full(inode
, ret_bh
, ex
, OCFS2_LOCK_NONBLOCK
);
2146 if (ret
== -EAGAIN
) {
2148 if (ocfs2_inode_lock(inode
, ret_bh
, ex
) == 0)
2149 ocfs2_inode_unlock(inode
, ex
);
2150 ret
= AOP_TRUNCATED_PAGE
;
2156 int ocfs2_inode_lock_atime(struct inode
*inode
,
2157 struct vfsmount
*vfsmnt
,
2163 ret
= ocfs2_inode_lock(inode
, NULL
, 0);
2170 * If we should update atime, we will get EX lock,
2171 * otherwise we just get PR lock.
2173 if (ocfs2_should_update_atime(inode
, vfsmnt
)) {
2174 struct buffer_head
*bh
= NULL
;
2176 ocfs2_inode_unlock(inode
, 0);
2177 ret
= ocfs2_inode_lock(inode
, &bh
, 1);
2183 if (ocfs2_should_update_atime(inode
, vfsmnt
))
2184 ocfs2_update_inode_atime(inode
, bh
);
2194 void ocfs2_inode_unlock(struct inode
*inode
,
2197 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2198 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2199 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2203 mlog(0, "inode %llu drop %s META lock\n",
2204 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2205 ex
? "EXMODE" : "PRMODE");
2207 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode
->i_sb
)) &&
2208 !ocfs2_mount_local(osb
))
2209 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
2214 int ocfs2_super_lock(struct ocfs2_super
*osb
,
2218 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2219 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2223 if (ocfs2_is_hard_readonly(osb
))
2226 if (ocfs2_mount_local(osb
))
2229 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
2235 /* The super block lock path is really in the best position to
2236 * know when resources covered by the lock need to be
2237 * refreshed, so we do it here. Of course, making sense of
2238 * everything is up to the caller :) */
2239 status
= ocfs2_should_refresh_lock_res(lockres
);
2245 status
= ocfs2_refresh_slot_info(osb
);
2247 ocfs2_complete_lock_res_refresh(lockres
, status
);
2257 void ocfs2_super_unlock(struct ocfs2_super
*osb
,
2260 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2261 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2263 if (!ocfs2_mount_local(osb
))
2264 ocfs2_cluster_unlock(osb
, lockres
, level
);
2267 int ocfs2_rename_lock(struct ocfs2_super
*osb
)
2270 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2272 if (ocfs2_is_hard_readonly(osb
))
2275 if (ocfs2_mount_local(osb
))
2278 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_EX
, 0, 0);
2285 void ocfs2_rename_unlock(struct ocfs2_super
*osb
)
2287 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2289 if (!ocfs2_mount_local(osb
))
2290 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
2293 int ocfs2_dentry_lock(struct dentry
*dentry
, int ex
)
2296 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2297 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2298 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2302 if (ocfs2_is_hard_readonly(osb
))
2305 if (ocfs2_mount_local(osb
))
2308 ret
= ocfs2_cluster_lock(osb
, &dl
->dl_lockres
, level
, 0, 0);
2315 void ocfs2_dentry_unlock(struct dentry
*dentry
, int ex
)
2317 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2318 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2319 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2321 if (!ocfs2_mount_local(osb
))
2322 ocfs2_cluster_unlock(osb
, &dl
->dl_lockres
, level
);
2325 /* Reference counting of the dlm debug structure. We want this because
2326 * open references on the debug inodes can live on after a mount, so
2327 * we can't rely on the ocfs2_super to always exist. */
2328 static void ocfs2_dlm_debug_free(struct kref
*kref
)
2330 struct ocfs2_dlm_debug
*dlm_debug
;
2332 dlm_debug
= container_of(kref
, struct ocfs2_dlm_debug
, d_refcnt
);
2337 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug
*dlm_debug
)
2340 kref_put(&dlm_debug
->d_refcnt
, ocfs2_dlm_debug_free
);
2343 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug
*debug
)
2345 kref_get(&debug
->d_refcnt
);
2348 struct ocfs2_dlm_debug
*ocfs2_new_dlm_debug(void)
2350 struct ocfs2_dlm_debug
*dlm_debug
;
2352 dlm_debug
= kmalloc(sizeof(struct ocfs2_dlm_debug
), GFP_KERNEL
);
2354 mlog_errno(-ENOMEM
);
2358 kref_init(&dlm_debug
->d_refcnt
);
2359 INIT_LIST_HEAD(&dlm_debug
->d_lockres_tracking
);
2360 dlm_debug
->d_locking_state
= NULL
;
2365 /* Access to this is arbitrated for us via seq_file->sem. */
2366 struct ocfs2_dlm_seq_priv
{
2367 struct ocfs2_dlm_debug
*p_dlm_debug
;
2368 struct ocfs2_lock_res p_iter_res
;
2369 struct ocfs2_lock_res p_tmp_res
;
2372 static struct ocfs2_lock_res
*ocfs2_dlm_next_res(struct ocfs2_lock_res
*start
,
2373 struct ocfs2_dlm_seq_priv
*priv
)
2375 struct ocfs2_lock_res
*iter
, *ret
= NULL
;
2376 struct ocfs2_dlm_debug
*dlm_debug
= priv
->p_dlm_debug
;
2378 assert_spin_locked(&ocfs2_dlm_tracking_lock
);
2380 list_for_each_entry(iter
, &start
->l_debug_list
, l_debug_list
) {
2381 /* discover the head of the list */
2382 if (&iter
->l_debug_list
== &dlm_debug
->d_lockres_tracking
) {
2383 mlog(0, "End of list found, %p\n", ret
);
2387 /* We track our "dummy" iteration lockres' by a NULL
2389 if (iter
->l_ops
!= NULL
) {
2398 static void *ocfs2_dlm_seq_start(struct seq_file
*m
, loff_t
*pos
)
2400 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
2401 struct ocfs2_lock_res
*iter
;
2403 spin_lock(&ocfs2_dlm_tracking_lock
);
2404 iter
= ocfs2_dlm_next_res(&priv
->p_iter_res
, priv
);
2406 /* Since lockres' have the lifetime of their container
2407 * (which can be inodes, ocfs2_supers, etc) we want to
2408 * copy this out to a temporary lockres while still
2409 * under the spinlock. Obviously after this we can't
2410 * trust any pointers on the copy returned, but that's
2411 * ok as the information we want isn't typically held
2413 priv
->p_tmp_res
= *iter
;
2414 iter
= &priv
->p_tmp_res
;
2416 spin_unlock(&ocfs2_dlm_tracking_lock
);
2421 static void ocfs2_dlm_seq_stop(struct seq_file
*m
, void *v
)
2425 static void *ocfs2_dlm_seq_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
2427 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
2428 struct ocfs2_lock_res
*iter
= v
;
2429 struct ocfs2_lock_res
*dummy
= &priv
->p_iter_res
;
2431 spin_lock(&ocfs2_dlm_tracking_lock
);
2432 iter
= ocfs2_dlm_next_res(iter
, priv
);
2433 list_del_init(&dummy
->l_debug_list
);
2435 list_add(&dummy
->l_debug_list
, &iter
->l_debug_list
);
2436 priv
->p_tmp_res
= *iter
;
2437 iter
= &priv
->p_tmp_res
;
2439 spin_unlock(&ocfs2_dlm_tracking_lock
);
2444 /* So that debugfs.ocfs2 can determine which format is being used */
2445 #define OCFS2_DLM_DEBUG_STR_VERSION 1
2446 static int ocfs2_dlm_seq_show(struct seq_file
*m
, void *v
)
2450 struct ocfs2_lock_res
*lockres
= v
;
2455 seq_printf(m
, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION
);
2457 if (lockres
->l_type
== OCFS2_LOCK_TYPE_DENTRY
)
2458 seq_printf(m
, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START
- 1,
2460 (unsigned int)ocfs2_get_dentry_lock_ino(lockres
));
2462 seq_printf(m
, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN
, lockres
->l_name
);
2464 seq_printf(m
, "%d\t"
2475 lockres
->l_unlock_action
,
2476 lockres
->l_ro_holders
,
2477 lockres
->l_ex_holders
,
2478 lockres
->l_requested
,
2479 lockres
->l_blocking
);
2481 /* Dump the raw LVB */
2482 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2483 for(i
= 0; i
< DLM_LVB_LEN
; i
++)
2484 seq_printf(m
, "0x%x\t", lvb
[i
]);
2487 seq_printf(m
, "\n");
2491 static const struct seq_operations ocfs2_dlm_seq_ops
= {
2492 .start
= ocfs2_dlm_seq_start
,
2493 .stop
= ocfs2_dlm_seq_stop
,
2494 .next
= ocfs2_dlm_seq_next
,
2495 .show
= ocfs2_dlm_seq_show
,
2498 static int ocfs2_dlm_debug_release(struct inode
*inode
, struct file
*file
)
2500 struct seq_file
*seq
= (struct seq_file
*) file
->private_data
;
2501 struct ocfs2_dlm_seq_priv
*priv
= seq
->private;
2502 struct ocfs2_lock_res
*res
= &priv
->p_iter_res
;
2504 ocfs2_remove_lockres_tracking(res
);
2505 ocfs2_put_dlm_debug(priv
->p_dlm_debug
);
2506 return seq_release_private(inode
, file
);
2509 static int ocfs2_dlm_debug_open(struct inode
*inode
, struct file
*file
)
2512 struct ocfs2_dlm_seq_priv
*priv
;
2513 struct seq_file
*seq
;
2514 struct ocfs2_super
*osb
;
2516 priv
= kzalloc(sizeof(struct ocfs2_dlm_seq_priv
), GFP_KERNEL
);
2522 osb
= inode
->i_private
;
2523 ocfs2_get_dlm_debug(osb
->osb_dlm_debug
);
2524 priv
->p_dlm_debug
= osb
->osb_dlm_debug
;
2525 INIT_LIST_HEAD(&priv
->p_iter_res
.l_debug_list
);
2527 ret
= seq_open(file
, &ocfs2_dlm_seq_ops
);
2534 seq
= (struct seq_file
*) file
->private_data
;
2535 seq
->private = priv
;
2537 ocfs2_add_lockres_tracking(&priv
->p_iter_res
,
2544 static const struct file_operations ocfs2_dlm_debug_fops
= {
2545 .open
= ocfs2_dlm_debug_open
,
2546 .release
= ocfs2_dlm_debug_release
,
2548 .llseek
= seq_lseek
,
2551 static int ocfs2_dlm_init_debug(struct ocfs2_super
*osb
)
2554 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2556 dlm_debug
->d_locking_state
= debugfs_create_file("locking_state",
2558 osb
->osb_debug_root
,
2560 &ocfs2_dlm_debug_fops
);
2561 if (!dlm_debug
->d_locking_state
) {
2564 "Unable to create locking state debugfs file.\n");
2568 ocfs2_get_dlm_debug(dlm_debug
);
2573 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super
*osb
)
2575 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2578 debugfs_remove(dlm_debug
->d_locking_state
);
2579 ocfs2_put_dlm_debug(dlm_debug
);
2583 int ocfs2_dlm_init(struct ocfs2_super
*osb
)
2586 struct ocfs2_cluster_connection
*conn
= NULL
;
2590 if (ocfs2_mount_local(osb
)) {
2595 status
= ocfs2_dlm_init_debug(osb
);
2601 /* launch downconvert thread */
2602 osb
->dc_task
= kthread_run(ocfs2_downconvert_thread
, osb
, "ocfs2dc");
2603 if (IS_ERR(osb
->dc_task
)) {
2604 status
= PTR_ERR(osb
->dc_task
);
2605 osb
->dc_task
= NULL
;
2610 /* for now, uuid == domain */
2611 status
= ocfs2_cluster_connect(osb
->uuid_str
,
2612 strlen(osb
->uuid_str
),
2613 ocfs2_do_node_down
, osb
,
2620 status
= ocfs2_cluster_this_node(&osb
->node_num
);
2624 "could not find this host's node number\n");
2625 ocfs2_cluster_disconnect(conn
);
2630 ocfs2_super_lock_res_init(&osb
->osb_super_lockres
, osb
);
2631 ocfs2_rename_lock_res_init(&osb
->osb_rename_lockres
, osb
);
2638 ocfs2_dlm_shutdown_debug(osb
);
2640 kthread_stop(osb
->dc_task
);
2647 void ocfs2_dlm_shutdown(struct ocfs2_super
*osb
)
2651 ocfs2_drop_osb_locks(osb
);
2654 * Now that we have dropped all locks and ocfs2_dismount_volume()
2655 * has disabled recovery, the DLM won't be talking to us. It's
2656 * safe to tear things down before disconnecting the cluster.
2660 kthread_stop(osb
->dc_task
);
2661 osb
->dc_task
= NULL
;
2664 ocfs2_lock_res_free(&osb
->osb_super_lockres
);
2665 ocfs2_lock_res_free(&osb
->osb_rename_lockres
);
2667 ocfs2_cluster_disconnect(osb
->cconn
);
2670 ocfs2_dlm_shutdown_debug(osb
);
2675 static void ocfs2_unlock_ast(void *opaque
, int error
)
2677 struct ocfs2_lock_res
*lockres
= opaque
;
2678 unsigned long flags
;
2682 mlog(0, "UNLOCK AST called on lock %s, action = %d\n", lockres
->l_name
,
2683 lockres
->l_unlock_action
);
2685 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2687 mlog(ML_ERROR
, "Dlm passes error %d for lock %s, "
2688 "unlock_action %d\n", error
, lockres
->l_name
,
2689 lockres
->l_unlock_action
);
2690 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2694 switch(lockres
->l_unlock_action
) {
2695 case OCFS2_UNLOCK_CANCEL_CONVERT
:
2696 mlog(0, "Cancel convert success for %s\n", lockres
->l_name
);
2697 lockres
->l_action
= OCFS2_AST_INVALID
;
2699 case OCFS2_UNLOCK_DROP_LOCK
:
2700 lockres
->l_level
= DLM_LOCK_IV
;
2706 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
2707 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
2708 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2710 wake_up(&lockres
->l_event
);
2715 static int ocfs2_drop_lock(struct ocfs2_super
*osb
,
2716 struct ocfs2_lock_res
*lockres
)
2719 unsigned long flags
;
2722 /* We didn't get anywhere near actually using this lockres. */
2723 if (!(lockres
->l_flags
& OCFS2_LOCK_INITIALIZED
))
2726 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
2727 lkm_flags
|= DLM_LKF_VALBLK
;
2729 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2731 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_FREEING
),
2732 "lockres %s, flags 0x%lx\n",
2733 lockres
->l_name
, lockres
->l_flags
);
2735 while (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
2736 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
2737 "%u, unlock_action = %u\n",
2738 lockres
->l_name
, lockres
->l_flags
, lockres
->l_action
,
2739 lockres
->l_unlock_action
);
2741 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2743 /* XXX: Today we just wait on any busy
2744 * locks... Perhaps we need to cancel converts in the
2746 ocfs2_wait_on_busy_lock(lockres
);
2748 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2751 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
2752 if (lockres
->l_flags
& OCFS2_LOCK_ATTACHED
&&
2753 lockres
->l_level
== DLM_LOCK_EX
&&
2754 !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
2755 lockres
->l_ops
->set_lvb(lockres
);
2758 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
)
2759 mlog(ML_ERROR
, "destroying busy lock: \"%s\"\n",
2761 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
2762 mlog(0, "destroying blocked lock: \"%s\"\n", lockres
->l_name
);
2764 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
2765 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2769 lockres_clear_flags(lockres
, OCFS2_LOCK_ATTACHED
);
2771 /* make sure we never get here while waiting for an ast to
2773 BUG_ON(lockres
->l_action
!= OCFS2_AST_INVALID
);
2775 /* is this necessary? */
2776 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2777 lockres
->l_unlock_action
= OCFS2_UNLOCK_DROP_LOCK
;
2778 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2780 mlog(0, "lock %s\n", lockres
->l_name
);
2782 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
, lkm_flags
,
2785 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
2786 mlog(ML_ERROR
, "lockres flags: %lu\n", lockres
->l_flags
);
2787 /* XXX Need to abstract this */
2788 dlm_print_one_lock(lockres
->l_lksb
.lksb_o2dlm
.lockid
);
2791 mlog(0, "lock %s, successfull return from ocfs2_dlm_unlock\n",
2794 ocfs2_wait_on_busy_lock(lockres
);
2800 /* Mark the lockres as being dropped. It will no longer be
2801 * queued if blocking, but we still may have to wait on it
2802 * being dequeued from the downconvert thread before we can consider
2805 * You can *not* attempt to call cluster_lock on this lockres anymore. */
2806 void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res
*lockres
)
2809 struct ocfs2_mask_waiter mw
;
2810 unsigned long flags
;
2812 ocfs2_init_mask_waiter(&mw
);
2814 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2815 lockres
->l_flags
|= OCFS2_LOCK_FREEING
;
2816 while (lockres
->l_flags
& OCFS2_LOCK_QUEUED
) {
2817 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_QUEUED
, 0);
2818 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2820 mlog(0, "Waiting on lockres %s\n", lockres
->l_name
);
2822 status
= ocfs2_wait_for_mask(&mw
);
2826 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2828 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2831 void ocfs2_simple_drop_lockres(struct ocfs2_super
*osb
,
2832 struct ocfs2_lock_res
*lockres
)
2836 ocfs2_mark_lockres_freeing(lockres
);
2837 ret
= ocfs2_drop_lock(osb
, lockres
);
2842 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
)
2844 ocfs2_simple_drop_lockres(osb
, &osb
->osb_super_lockres
);
2845 ocfs2_simple_drop_lockres(osb
, &osb
->osb_rename_lockres
);
2848 int ocfs2_drop_inode_locks(struct inode
*inode
)
2854 /* No need to call ocfs2_mark_lockres_freeing here -
2855 * ocfs2_clear_inode has done it for us. */
2857 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2858 &OCFS2_I(inode
)->ip_open_lockres
);
2864 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2865 &OCFS2_I(inode
)->ip_inode_lockres
);
2868 if (err
< 0 && !status
)
2871 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2872 &OCFS2_I(inode
)->ip_rw_lockres
);
2875 if (err
< 0 && !status
)
2882 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
2885 assert_spin_locked(&lockres
->l_lock
);
2887 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
2889 if (lockres
->l_level
<= new_level
) {
2890 mlog(ML_ERROR
, "lockres->l_level (%d) <= new_level (%d)\n",
2891 lockres
->l_level
, new_level
);
2895 mlog(0, "lock %s, new_level = %d, l_blocking = %d\n",
2896 lockres
->l_name
, new_level
, lockres
->l_blocking
);
2898 lockres
->l_action
= OCFS2_AST_DOWNCONVERT
;
2899 lockres
->l_requested
= new_level
;
2900 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2901 return lockres_set_pending(lockres
);
2904 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
2905 struct ocfs2_lock_res
*lockres
,
2908 unsigned int generation
)
2911 u32 dlm_flags
= DLM_LKF_CONVERT
;
2916 dlm_flags
|= DLM_LKF_VALBLK
;
2918 ret
= ocfs2_dlm_lock(osb
->cconn
,
2923 OCFS2_LOCK_ID_MAX_LEN
- 1,
2925 lockres_clear_pending(lockres
, generation
, osb
);
2927 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
2928 ocfs2_recover_from_dlm_error(lockres
, 1);
2938 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
2939 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
2940 struct ocfs2_lock_res
*lockres
)
2942 assert_spin_locked(&lockres
->l_lock
);
2945 mlog(0, "lock %s\n", lockres
->l_name
);
2947 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
2948 /* If we're already trying to cancel a lock conversion
2949 * then just drop the spinlock and allow the caller to
2950 * requeue this lock. */
2952 mlog(0, "Lockres %s, skip convert\n", lockres
->l_name
);
2956 /* were we in a convert when we got the bast fire? */
2957 BUG_ON(lockres
->l_action
!= OCFS2_AST_CONVERT
&&
2958 lockres
->l_action
!= OCFS2_AST_DOWNCONVERT
);
2959 /* set things up for the unlockast to know to just
2960 * clear out the ast_action and unset busy, etc. */
2961 lockres
->l_unlock_action
= OCFS2_UNLOCK_CANCEL_CONVERT
;
2963 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
),
2964 "lock %s, invalid flags: 0x%lx\n",
2965 lockres
->l_name
, lockres
->l_flags
);
2970 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
2971 struct ocfs2_lock_res
*lockres
)
2976 mlog(0, "lock %s\n", lockres
->l_name
);
2978 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
,
2979 DLM_LKF_CANCEL
, lockres
);
2981 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
2982 ocfs2_recover_from_dlm_error(lockres
, 0);
2985 mlog(0, "lock %s return from ocfs2_dlm_unlock\n", lockres
->l_name
);
2991 static int ocfs2_unblock_lock(struct ocfs2_super
*osb
,
2992 struct ocfs2_lock_res
*lockres
,
2993 struct ocfs2_unblock_ctl
*ctl
)
2995 unsigned long flags
;
3004 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3006 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
3009 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
3011 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3012 * exists entirely for one reason - another thread has set
3013 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3015 * If we do ocfs2_cancel_convert() before the other thread
3016 * calls dlm_lock(), our cancel will do nothing. We will
3017 * get no ast, and we will have no way of knowing the
3018 * cancel failed. Meanwhile, the other thread will call
3019 * into dlm_lock() and wait...forever.
3021 * Why forever? Because another node has asked for the
3022 * lock first; that's why we're here in unblock_lock().
3024 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3025 * set, we just requeue the unblock. Only when the other
3026 * thread has called dlm_lock() and cleared PENDING will
3027 * we then cancel their request.
3029 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3030 * at the same time they set OCFS2_DLM_BUSY. They must
3031 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3033 if (lockres
->l_flags
& OCFS2_LOCK_PENDING
)
3037 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
3038 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3040 ret
= ocfs2_cancel_convert(osb
, lockres
);
3047 /* if we're blocking an exclusive and we have *any* holders,
3049 if ((lockres
->l_blocking
== DLM_LOCK_EX
)
3050 && (lockres
->l_ex_holders
|| lockres
->l_ro_holders
))
3053 /* If it's a PR we're blocking, then only
3054 * requeue if we've got any EX holders */
3055 if (lockres
->l_blocking
== DLM_LOCK_PR
&&
3056 lockres
->l_ex_holders
)
3060 * Can we get a lock in this state if the holder counts are
3061 * zero? The meta data unblock code used to check this.
3063 if ((lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
3064 && (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
))
3067 new_level
= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
3069 if (lockres
->l_ops
->check_downconvert
3070 && !lockres
->l_ops
->check_downconvert(lockres
, new_level
))
3073 /* If we get here, then we know that there are no more
3074 * incompatible holders (and anyone asking for an incompatible
3075 * lock is blocked). We can now downconvert the lock */
3076 if (!lockres
->l_ops
->downconvert_worker
)
3079 /* Some lockres types want to do a bit of work before
3080 * downconverting a lock. Allow that here. The worker function
3081 * may sleep, so we save off a copy of what we're blocking as
3082 * it may change while we're not holding the spin lock. */
3083 blocking
= lockres
->l_blocking
;
3084 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3086 ctl
->unblock_action
= lockres
->l_ops
->downconvert_worker(lockres
, blocking
);
3088 if (ctl
->unblock_action
== UNBLOCK_STOP_POST
)
3091 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3092 if (blocking
!= lockres
->l_blocking
) {
3093 /* If this changed underneath us, then we can't drop
3101 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
3102 if (lockres
->l_level
== DLM_LOCK_EX
)
3106 * We only set the lvb if the lock has been fully
3107 * refreshed - otherwise we risk setting stale
3108 * data. Otherwise, there's no need to actually clear
3109 * out the lvb here as it's value is still valid.
3111 if (set_lvb
&& !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
3112 lockres
->l_ops
->set_lvb(lockres
);
3115 gen
= ocfs2_prepare_downconvert(lockres
, new_level
);
3116 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3117 ret
= ocfs2_downconvert_lock(osb
, lockres
, new_level
, set_lvb
,
3125 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3132 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
3135 struct inode
*inode
;
3136 struct address_space
*mapping
;
3138 inode
= ocfs2_lock_res_inode(lockres
);
3139 mapping
= inode
->i_mapping
;
3141 if (!S_ISREG(inode
->i_mode
))
3145 * We need this before the filemap_fdatawrite() so that it can
3146 * transfer the dirty bit from the PTE to the
3147 * page. Unfortunately this means that even for EX->PR
3148 * downconverts, we'll lose our mappings and have to build
3151 unmap_mapping_range(mapping
, 0, 0, 0);
3153 if (filemap_fdatawrite(mapping
)) {
3154 mlog(ML_ERROR
, "Could not sync inode %llu for downconvert!",
3155 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3157 sync_mapping_buffers(mapping
);
3158 if (blocking
== DLM_LOCK_EX
) {
3159 truncate_inode_pages(mapping
, 0);
3161 /* We only need to wait on the I/O if we're not also
3162 * truncating pages because truncate_inode_pages waits
3163 * for us above. We don't truncate pages if we're
3164 * blocking anything < EXMODE because we want to keep
3165 * them around in that case. */
3166 filemap_fdatawait(mapping
);
3170 return UNBLOCK_CONTINUE
;
3173 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
3176 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3177 int checkpointed
= ocfs2_inode_fully_checkpointed(inode
);
3179 BUG_ON(new_level
!= DLM_LOCK_NL
&& new_level
!= DLM_LOCK_PR
);
3180 BUG_ON(lockres
->l_level
!= DLM_LOCK_EX
&& !checkpointed
);
3185 ocfs2_start_checkpoint(OCFS2_SB(inode
->i_sb
));
3189 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
)
3191 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3193 __ocfs2_stuff_meta_lvb(inode
);
3197 * Does the final reference drop on our dentry lock. Right now this
3198 * happens in the downconvert thread, but we could choose to simplify the
3199 * dlmglue API and push these off to the ocfs2_wq in the future.
3201 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
3202 struct ocfs2_lock_res
*lockres
)
3204 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3205 ocfs2_dentry_lock_put(osb
, dl
);
3209 * d_delete() matching dentries before the lock downconvert.
3211 * At this point, any process waiting to destroy the
3212 * dentry_lock due to last ref count is stopped by the
3213 * OCFS2_LOCK_QUEUED flag.
3215 * We have two potential problems
3217 * 1) If we do the last reference drop on our dentry_lock (via dput)
3218 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3219 * the downconvert to finish. Instead we take an elevated
3220 * reference and push the drop until after we've completed our
3221 * unblock processing.
3223 * 2) There might be another process with a final reference,
3224 * waiting on us to finish processing. If this is the case, we
3225 * detect it and exit out - there's no more dentries anyway.
3227 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
3230 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3231 struct ocfs2_inode_info
*oi
= OCFS2_I(dl
->dl_inode
);
3232 struct dentry
*dentry
;
3233 unsigned long flags
;
3237 * This node is blocking another node from getting a read
3238 * lock. This happens when we've renamed within a
3239 * directory. We've forced the other nodes to d_delete(), but
3240 * we never actually dropped our lock because it's still
3241 * valid. The downconvert code will retain a PR for this node,
3242 * so there's no further work to do.
3244 if (blocking
== DLM_LOCK_PR
)
3245 return UNBLOCK_CONTINUE
;
3248 * Mark this inode as potentially orphaned. The code in
3249 * ocfs2_delete_inode() will figure out whether it actually
3250 * needs to be freed or not.
3252 spin_lock(&oi
->ip_lock
);
3253 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
3254 spin_unlock(&oi
->ip_lock
);
3257 * Yuck. We need to make sure however that the check of
3258 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3259 * respect to a reference decrement or the setting of that
3262 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3263 spin_lock(&dentry_attach_lock
);
3264 if (!(lockres
->l_flags
& OCFS2_LOCK_FREEING
)
3269 spin_unlock(&dentry_attach_lock
);
3270 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3272 mlog(0, "extra_ref = %d\n", extra_ref
);
3275 * We have a process waiting on us in ocfs2_dentry_iput(),
3276 * which means we can't have any more outstanding
3277 * aliases. There's no need to do any more work.
3280 return UNBLOCK_CONTINUE
;
3282 spin_lock(&dentry_attach_lock
);
3284 dentry
= ocfs2_find_local_alias(dl
->dl_inode
,
3285 dl
->dl_parent_blkno
, 1);
3288 spin_unlock(&dentry_attach_lock
);
3290 mlog(0, "d_delete(%.*s);\n", dentry
->d_name
.len
,
3291 dentry
->d_name
.name
);
3294 * The following dcache calls may do an
3295 * iput(). Normally we don't want that from the
3296 * downconverting thread, but in this case it's ok
3297 * because the requesting node already has an
3298 * exclusive lock on the inode, so it can't be queued
3299 * for a downconvert.
3304 spin_lock(&dentry_attach_lock
);
3306 spin_unlock(&dentry_attach_lock
);
3309 * If we are the last holder of this dentry lock, there is no
3310 * reason to downconvert so skip straight to the unlock.
3312 if (dl
->dl_count
== 1)
3313 return UNBLOCK_STOP_POST
;
3315 return UNBLOCK_CONTINUE_POST
;
3319 * This is the filesystem locking protocol. It provides the lock handling
3320 * hooks for the underlying DLM. It has a maximum version number.
3321 * The version number allows interoperability with systems running at
3322 * the same major number and an equal or smaller minor number.
3324 * Whenever the filesystem does new things with locks (adds or removes a
3325 * lock, orders them differently, does different things underneath a lock),
3326 * the version must be changed. The protocol is negotiated when joining
3327 * the dlm domain. A node may join the domain if its major version is
3328 * identical to all other nodes and its minor version is greater than
3329 * or equal to all other nodes. When its minor version is greater than
3330 * the other nodes, it will run at the minor version specified by the
3333 * If a locking change is made that will not be compatible with older
3334 * versions, the major number must be increased and the minor version set
3335 * to zero. If a change merely adds a behavior that can be disabled when
3336 * speaking to older versions, the minor version must be increased. If a
3337 * change adds a fully backwards compatible change (eg, LVB changes that
3338 * are just ignored by older versions), the version does not need to be
3341 static struct ocfs2_locking_protocol lproto
= {
3343 .pv_major
= OCFS2_LOCKING_PROTOCOL_MAJOR
,
3344 .pv_minor
= OCFS2_LOCKING_PROTOCOL_MINOR
,
3346 .lp_lock_ast
= ocfs2_locking_ast
,
3347 .lp_blocking_ast
= ocfs2_blocking_ast
,
3348 .lp_unlock_ast
= ocfs2_unlock_ast
,
3351 /* This interface isn't the final one, hence the less-than-perfect names */
3352 void dlmglue_init_stack(void)
3354 o2cb_get_stack(&lproto
);
3357 void dlmglue_exit_stack(void)
3362 static void ocfs2_process_blocked_lock(struct ocfs2_super
*osb
,
3363 struct ocfs2_lock_res
*lockres
)
3366 struct ocfs2_unblock_ctl ctl
= {0, 0,};
3367 unsigned long flags
;
3369 /* Our reference to the lockres in this function can be
3370 * considered valid until we remove the OCFS2_LOCK_QUEUED
3376 BUG_ON(!lockres
->l_ops
);
3378 mlog(0, "lockres %s blocked.\n", lockres
->l_name
);
3380 /* Detect whether a lock has been marked as going away while
3381 * the downconvert thread was processing other things. A lock can
3382 * still be marked with OCFS2_LOCK_FREEING after this check,
3383 * but short circuiting here will still save us some
3385 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3386 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
)
3388 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3390 status
= ocfs2_unblock_lock(osb
, lockres
, &ctl
);
3394 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3396 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
|| !ctl
.requeue
) {
3397 lockres_clear_flags(lockres
, OCFS2_LOCK_QUEUED
);
3399 ocfs2_schedule_blocked_lock(osb
, lockres
);
3401 mlog(0, "lockres %s, requeue = %s.\n", lockres
->l_name
,
3402 ctl
.requeue
? "yes" : "no");
3403 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3405 if (ctl
.unblock_action
!= UNBLOCK_CONTINUE
3406 && lockres
->l_ops
->post_unlock
)
3407 lockres
->l_ops
->post_unlock(osb
, lockres
);
3412 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
3413 struct ocfs2_lock_res
*lockres
)
3417 assert_spin_locked(&lockres
->l_lock
);
3419 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
) {
3420 /* Do not schedule a lock for downconvert when it's on
3421 * the way to destruction - any nodes wanting access
3422 * to the resource will get it soon. */
3423 mlog(0, "Lockres %s won't be scheduled: flags 0x%lx\n",
3424 lockres
->l_name
, lockres
->l_flags
);
3428 lockres_or_flags(lockres
, OCFS2_LOCK_QUEUED
);
3430 spin_lock(&osb
->dc_task_lock
);
3431 if (list_empty(&lockres
->l_blocked_list
)) {
3432 list_add_tail(&lockres
->l_blocked_list
,
3433 &osb
->blocked_lock_list
);
3434 osb
->blocked_lock_count
++;
3436 spin_unlock(&osb
->dc_task_lock
);
3441 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super
*osb
)
3443 unsigned long processed
;
3444 struct ocfs2_lock_res
*lockres
;
3448 spin_lock(&osb
->dc_task_lock
);
3449 /* grab this early so we know to try again if a state change and
3450 * wake happens part-way through our work */
3451 osb
->dc_work_sequence
= osb
->dc_wake_sequence
;
3453 processed
= osb
->blocked_lock_count
;
3455 BUG_ON(list_empty(&osb
->blocked_lock_list
));
3457 lockres
= list_entry(osb
->blocked_lock_list
.next
,
3458 struct ocfs2_lock_res
, l_blocked_list
);
3459 list_del_init(&lockres
->l_blocked_list
);
3460 osb
->blocked_lock_count
--;
3461 spin_unlock(&osb
->dc_task_lock
);
3466 ocfs2_process_blocked_lock(osb
, lockres
);
3468 spin_lock(&osb
->dc_task_lock
);
3470 spin_unlock(&osb
->dc_task_lock
);
3475 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super
*osb
)
3479 spin_lock(&osb
->dc_task_lock
);
3480 if (list_empty(&osb
->blocked_lock_list
))
3483 spin_unlock(&osb
->dc_task_lock
);
3487 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super
*osb
)
3489 int should_wake
= 0;
3491 spin_lock(&osb
->dc_task_lock
);
3492 if (osb
->dc_work_sequence
!= osb
->dc_wake_sequence
)
3494 spin_unlock(&osb
->dc_task_lock
);
3499 static int ocfs2_downconvert_thread(void *arg
)
3502 struct ocfs2_super
*osb
= arg
;
3504 /* only quit once we've been asked to stop and there is no more
3506 while (!(kthread_should_stop() &&
3507 ocfs2_downconvert_thread_lists_empty(osb
))) {
3509 wait_event_interruptible(osb
->dc_event
,
3510 ocfs2_downconvert_thread_should_wake(osb
) ||
3511 kthread_should_stop());
3513 mlog(0, "downconvert_thread: awoken\n");
3515 ocfs2_downconvert_thread_do_work(osb
);
3518 osb
->dc_task
= NULL
;
3522 void ocfs2_wake_downconvert_thread(struct ocfs2_super
*osb
)
3524 spin_lock(&osb
->dc_task_lock
);
3525 /* make sure the voting thread gets a swipe at whatever changes
3526 * the caller may have made to the voting state */
3527 osb
->dc_wake_sequence
++;
3528 spin_unlock(&osb
->dc_task_lock
);
3529 wake_up(&osb
->dc_event
);