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 inline int ocfs2_is_inode_lock(struct ocfs2_lock_res
*lockres
)
230 return lockres
->l_type
== OCFS2_LOCK_TYPE_META
||
231 lockres
->l_type
== OCFS2_LOCK_TYPE_DATA
||
232 lockres
->l_type
== OCFS2_LOCK_TYPE_RW
;
235 static inline struct inode
*ocfs2_lock_res_inode(struct ocfs2_lock_res
*lockres
)
237 BUG_ON(!ocfs2_is_inode_lock(lockres
));
239 return (struct inode
*) lockres
->l_priv
;
242 static inline struct ocfs2_dentry_lock
*ocfs2_lock_res_dl(struct ocfs2_lock_res
*lockres
)
244 BUG_ON(lockres
->l_type
!= OCFS2_LOCK_TYPE_DENTRY
);
246 return (struct ocfs2_dentry_lock
*)lockres
->l_priv
;
249 static inline struct ocfs2_super
*ocfs2_get_lockres_osb(struct ocfs2_lock_res
*lockres
)
251 if (lockres
->l_ops
->get_osb
)
252 return lockres
->l_ops
->get_osb(lockres
);
254 return (struct ocfs2_super
*)lockres
->l_priv
;
257 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
258 struct ocfs2_lock_res
*lockres
,
261 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
263 static void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
264 struct ocfs2_lock_res
*lockres
,
266 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
);
267 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
);
268 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
);
269 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
, int level
);
270 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
271 struct ocfs2_lock_res
*lockres
);
272 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
274 #define ocfs2_log_dlm_error(_func, _stat, _lockres) do { \
275 mlog(ML_ERROR, "Dlm error \"%s\" while calling %s on " \
276 "resource %s: %s\n", dlm_errname(_stat), _func, \
277 _lockres->l_name, dlm_errmsg(_stat)); \
279 static void ocfs2_vote_on_unlock(struct ocfs2_super
*osb
,
280 struct ocfs2_lock_res
*lockres
);
281 static int ocfs2_meta_lock_update(struct inode
*inode
,
282 struct buffer_head
**bh
);
283 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
);
284 static inline int ocfs2_highest_compat_lock_level(int level
);
286 static void ocfs2_build_lock_name(enum ocfs2_lock_type type
,
295 BUG_ON(type
>= OCFS2_NUM_LOCK_TYPES
);
297 len
= snprintf(name
, OCFS2_LOCK_ID_MAX_LEN
, "%c%s%016llx%08x",
298 ocfs2_lock_type_char(type
), OCFS2_LOCK_ID_PAD
,
299 (long long)blkno
, generation
);
301 BUG_ON(len
!= (OCFS2_LOCK_ID_MAX_LEN
- 1));
303 mlog(0, "built lock resource with name: %s\n", name
);
308 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock
);
310 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res
*res
,
311 struct ocfs2_dlm_debug
*dlm_debug
)
313 mlog(0, "Add tracking for lockres %s\n", res
->l_name
);
315 spin_lock(&ocfs2_dlm_tracking_lock
);
316 list_add(&res
->l_debug_list
, &dlm_debug
->d_lockres_tracking
);
317 spin_unlock(&ocfs2_dlm_tracking_lock
);
320 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res
*res
)
322 spin_lock(&ocfs2_dlm_tracking_lock
);
323 if (!list_empty(&res
->l_debug_list
))
324 list_del_init(&res
->l_debug_list
);
325 spin_unlock(&ocfs2_dlm_tracking_lock
);
328 static void ocfs2_lock_res_init_common(struct ocfs2_super
*osb
,
329 struct ocfs2_lock_res
*res
,
330 enum ocfs2_lock_type type
,
331 struct ocfs2_lock_res_ops
*ops
,
338 res
->l_level
= LKM_IVMODE
;
339 res
->l_requested
= LKM_IVMODE
;
340 res
->l_blocking
= LKM_IVMODE
;
341 res
->l_action
= OCFS2_AST_INVALID
;
342 res
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
344 res
->l_flags
= OCFS2_LOCK_INITIALIZED
;
346 ocfs2_add_lockres_tracking(res
, osb
->osb_dlm_debug
);
349 void ocfs2_lock_res_init_once(struct ocfs2_lock_res
*res
)
351 /* This also clears out the lock status block */
352 memset(res
, 0, sizeof(struct ocfs2_lock_res
));
353 spin_lock_init(&res
->l_lock
);
354 init_waitqueue_head(&res
->l_event
);
355 INIT_LIST_HEAD(&res
->l_blocked_list
);
356 INIT_LIST_HEAD(&res
->l_mask_waiters
);
359 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res
*res
,
360 enum ocfs2_lock_type type
,
361 unsigned int generation
,
364 struct ocfs2_lock_res_ops
*ops
;
367 case OCFS2_LOCK_TYPE_RW
:
368 ops
= &ocfs2_inode_rw_lops
;
370 case OCFS2_LOCK_TYPE_META
:
371 ops
= &ocfs2_inode_meta_lops
;
373 case OCFS2_LOCK_TYPE_DATA
:
374 ops
= &ocfs2_inode_data_lops
;
377 mlog_bug_on_msg(1, "type: %d\n", type
);
378 ops
= NULL
; /* thanks, gcc */
382 ocfs2_build_lock_name(type
, OCFS2_I(inode
)->ip_blkno
,
383 generation
, res
->l_name
);
384 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), res
, type
, ops
, inode
);
387 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
)
389 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
391 return OCFS2_SB(inode
->i_sb
);
394 static __u64
ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res
*lockres
)
396 __be64 inode_blkno_be
;
398 memcpy(&inode_blkno_be
, &lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
],
401 return be64_to_cpu(inode_blkno_be
);
404 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
)
406 struct ocfs2_dentry_lock
*dl
= lockres
->l_priv
;
408 return OCFS2_SB(dl
->dl_inode
->i_sb
);
411 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock
*dl
,
412 u64 parent
, struct inode
*inode
)
415 u64 inode_blkno
= OCFS2_I(inode
)->ip_blkno
;
416 __be64 inode_blkno_be
= cpu_to_be64(inode_blkno
);
417 struct ocfs2_lock_res
*lockres
= &dl
->dl_lockres
;
419 ocfs2_lock_res_init_once(lockres
);
422 * Unfortunately, the standard lock naming scheme won't work
423 * here because we have two 16 byte values to use. Instead,
424 * we'll stuff the inode number as a binary value. We still
425 * want error prints to show something without garbling the
426 * display, so drop a null byte in there before the inode
427 * number. A future version of OCFS2 will likely use all
428 * binary lock names. The stringified names have been a
429 * tremendous aid in debugging, but now that the debugfs
430 * interface exists, we can mangle things there if need be.
432 * NOTE: We also drop the standard "pad" value (the total lock
433 * name size stays the same though - the last part is all
434 * zeros due to the memset in ocfs2_lock_res_init_once()
436 len
= snprintf(lockres
->l_name
, OCFS2_DENTRY_LOCK_INO_START
,
438 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY
),
441 BUG_ON(len
!= (OCFS2_DENTRY_LOCK_INO_START
- 1));
443 memcpy(&lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
], &inode_blkno_be
,
446 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
447 OCFS2_LOCK_TYPE_DENTRY
, &ocfs2_dentry_lops
,
451 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res
*res
,
452 struct ocfs2_super
*osb
)
454 /* Superblock lockres doesn't come from a slab so we call init
455 * once on it manually. */
456 ocfs2_lock_res_init_once(res
);
457 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER
, OCFS2_SUPER_BLOCK_BLKNO
,
459 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_SUPER
,
460 &ocfs2_super_lops
, osb
);
463 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res
*res
,
464 struct ocfs2_super
*osb
)
466 /* Rename lockres doesn't come from a slab so we call init
467 * once on it manually. */
468 ocfs2_lock_res_init_once(res
);
469 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME
, 0, 0, res
->l_name
);
470 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_RENAME
,
471 &ocfs2_rename_lops
, osb
);
474 void ocfs2_lock_res_free(struct ocfs2_lock_res
*res
)
478 if (!(res
->l_flags
& OCFS2_LOCK_INITIALIZED
))
481 ocfs2_remove_lockres_tracking(res
);
483 mlog_bug_on_msg(!list_empty(&res
->l_blocked_list
),
484 "Lockres %s is on the blocked list\n",
486 mlog_bug_on_msg(!list_empty(&res
->l_mask_waiters
),
487 "Lockres %s has mask waiters pending\n",
489 mlog_bug_on_msg(spin_is_locked(&res
->l_lock
),
490 "Lockres %s is locked\n",
492 mlog_bug_on_msg(res
->l_ro_holders
,
493 "Lockres %s has %u ro holders\n",
494 res
->l_name
, res
->l_ro_holders
);
495 mlog_bug_on_msg(res
->l_ex_holders
,
496 "Lockres %s has %u ex holders\n",
497 res
->l_name
, res
->l_ex_holders
);
499 /* Need to clear out the lock status block for the dlm */
500 memset(&res
->l_lksb
, 0, sizeof(res
->l_lksb
));
506 static inline void ocfs2_inc_holders(struct ocfs2_lock_res
*lockres
,
515 lockres
->l_ex_holders
++;
518 lockres
->l_ro_holders
++;
527 static inline void ocfs2_dec_holders(struct ocfs2_lock_res
*lockres
,
536 BUG_ON(!lockres
->l_ex_holders
);
537 lockres
->l_ex_holders
--;
540 BUG_ON(!lockres
->l_ro_holders
);
541 lockres
->l_ro_holders
--;
549 /* WARNING: This function lives in a world where the only three lock
550 * levels are EX, PR, and NL. It *will* have to be adjusted when more
551 * lock types are added. */
552 static inline int ocfs2_highest_compat_lock_level(int level
)
554 int new_level
= LKM_EXMODE
;
556 if (level
== LKM_EXMODE
)
557 new_level
= LKM_NLMODE
;
558 else if (level
== LKM_PRMODE
)
559 new_level
= LKM_PRMODE
;
563 static void lockres_set_flags(struct ocfs2_lock_res
*lockres
,
564 unsigned long newflags
)
566 struct list_head
*pos
, *tmp
;
567 struct ocfs2_mask_waiter
*mw
;
569 assert_spin_locked(&lockres
->l_lock
);
571 lockres
->l_flags
= newflags
;
573 list_for_each_safe(pos
, tmp
, &lockres
->l_mask_waiters
) {
574 mw
= list_entry(pos
, struct ocfs2_mask_waiter
, mw_item
);
575 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
578 list_del_init(&mw
->mw_item
);
580 complete(&mw
->mw_complete
);
583 static void lockres_or_flags(struct ocfs2_lock_res
*lockres
, unsigned long or)
585 lockres_set_flags(lockres
, lockres
->l_flags
| or);
587 static void lockres_clear_flags(struct ocfs2_lock_res
*lockres
,
590 lockres_set_flags(lockres
, lockres
->l_flags
& ~clear
);
593 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
)
597 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
598 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
599 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
600 BUG_ON(lockres
->l_blocking
<= LKM_NLMODE
);
602 lockres
->l_level
= lockres
->l_requested
;
603 if (lockres
->l_level
<=
604 ocfs2_highest_compat_lock_level(lockres
->l_blocking
)) {
605 lockres
->l_blocking
= LKM_NLMODE
;
606 lockres_clear_flags(lockres
, OCFS2_LOCK_BLOCKED
);
608 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
613 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
)
617 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
618 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
620 /* Convert from RO to EX doesn't really need anything as our
621 * information is already up to data. Convert from NL to
622 * *anything* however should mark ourselves as needing an
624 if (lockres
->l_level
== LKM_NLMODE
&&
625 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
626 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
628 lockres
->l_level
= lockres
->l_requested
;
629 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
634 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
)
638 BUG_ON((!lockres
->l_flags
& OCFS2_LOCK_BUSY
));
639 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
641 if (lockres
->l_requested
> LKM_NLMODE
&&
642 !(lockres
->l_flags
& OCFS2_LOCK_LOCAL
) &&
643 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
644 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
646 lockres
->l_level
= lockres
->l_requested
;
647 lockres_or_flags(lockres
, OCFS2_LOCK_ATTACHED
);
648 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
653 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
,
656 int needs_downconvert
= 0;
659 assert_spin_locked(&lockres
->l_lock
);
661 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
663 if (level
> lockres
->l_blocking
) {
664 /* only schedule a downconvert if we haven't already scheduled
665 * one that goes low enough to satisfy the level we're
666 * blocking. this also catches the case where we get
668 if (ocfs2_highest_compat_lock_level(level
) <
669 ocfs2_highest_compat_lock_level(lockres
->l_blocking
))
670 needs_downconvert
= 1;
672 lockres
->l_blocking
= level
;
675 mlog_exit(needs_downconvert
);
676 return needs_downconvert
;
679 static void ocfs2_blocking_ast(void *opaque
, int level
)
681 struct ocfs2_lock_res
*lockres
= opaque
;
682 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
683 int needs_downconvert
;
686 BUG_ON(level
<= LKM_NLMODE
);
688 mlog(0, "BAST fired for lockres %s, blocking %d, level %d type %s\n",
689 lockres
->l_name
, level
, lockres
->l_level
,
690 ocfs2_lock_type_string(lockres
->l_type
));
692 spin_lock_irqsave(&lockres
->l_lock
, flags
);
693 needs_downconvert
= ocfs2_generic_handle_bast(lockres
, level
);
694 if (needs_downconvert
)
695 ocfs2_schedule_blocked_lock(osb
, lockres
);
696 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
698 wake_up(&lockres
->l_event
);
700 ocfs2_kick_vote_thread(osb
);
703 static void ocfs2_locking_ast(void *opaque
)
705 struct ocfs2_lock_res
*lockres
= opaque
;
706 struct dlm_lockstatus
*lksb
= &lockres
->l_lksb
;
709 spin_lock_irqsave(&lockres
->l_lock
, flags
);
711 if (lksb
->status
!= DLM_NORMAL
) {
712 mlog(ML_ERROR
, "lockres %s: lksb status value of %u!\n",
713 lockres
->l_name
, lksb
->status
);
714 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
718 switch(lockres
->l_action
) {
719 case OCFS2_AST_ATTACH
:
720 ocfs2_generic_handle_attach_action(lockres
);
721 lockres_clear_flags(lockres
, OCFS2_LOCK_LOCAL
);
723 case OCFS2_AST_CONVERT
:
724 ocfs2_generic_handle_convert_action(lockres
);
726 case OCFS2_AST_DOWNCONVERT
:
727 ocfs2_generic_handle_downconvert_action(lockres
);
730 mlog(ML_ERROR
, "lockres %s: ast fired with invalid action: %u "
731 "lockres flags = 0x%lx, unlock action: %u\n",
732 lockres
->l_name
, lockres
->l_action
, lockres
->l_flags
,
733 lockres
->l_unlock_action
);
737 /* set it to something invalid so if we get called again we
739 lockres
->l_action
= OCFS2_AST_INVALID
;
741 wake_up(&lockres
->l_event
);
742 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
745 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
751 spin_lock_irqsave(&lockres
->l_lock
, flags
);
752 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
754 lockres
->l_action
= OCFS2_AST_INVALID
;
756 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
757 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
759 wake_up(&lockres
->l_event
);
763 /* Note: If we detect another process working on the lock (i.e.,
764 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
765 * to do the right thing in that case.
767 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
768 struct ocfs2_lock_res
*lockres
,
773 enum dlm_status status
;
778 mlog(0, "lock %s, level = %d, flags = %d\n", lockres
->l_name
, level
,
781 spin_lock_irqsave(&lockres
->l_lock
, flags
);
782 if ((lockres
->l_flags
& OCFS2_LOCK_ATTACHED
) ||
783 (lockres
->l_flags
& OCFS2_LOCK_BUSY
)) {
784 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
788 lockres
->l_action
= OCFS2_AST_ATTACH
;
789 lockres
->l_requested
= level
;
790 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
791 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
793 status
= dlmlock(osb
->dlm
,
798 OCFS2_LOCK_ID_MAX_LEN
- 1,
802 if (status
!= DLM_NORMAL
) {
803 ocfs2_log_dlm_error("dlmlock", status
, lockres
);
805 ocfs2_recover_from_dlm_error(lockres
, 1);
808 mlog(0, "lock %s, successfull return from dlmlock\n", lockres
->l_name
);
815 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res
*lockres
,
821 spin_lock_irqsave(&lockres
->l_lock
, flags
);
822 ret
= lockres
->l_flags
& flag
;
823 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
828 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res
*lockres
)
831 wait_event(lockres
->l_event
,
832 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_BUSY
));
835 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res
*lockres
)
838 wait_event(lockres
->l_event
,
839 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_REFRESHING
));
842 /* predict what lock level we'll be dropping down to on behalf
843 * of another node, and return true if the currently wanted
844 * level will be compatible with it. */
845 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
848 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
850 return wanted
<= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
853 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter
*mw
)
855 INIT_LIST_HEAD(&mw
->mw_item
);
856 init_completion(&mw
->mw_complete
);
859 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter
*mw
)
861 wait_for_completion(&mw
->mw_complete
);
862 /* Re-arm the completion in case we want to wait on it again */
863 INIT_COMPLETION(mw
->mw_complete
);
864 return mw
->mw_status
;
867 static void lockres_add_mask_waiter(struct ocfs2_lock_res
*lockres
,
868 struct ocfs2_mask_waiter
*mw
,
872 BUG_ON(!list_empty(&mw
->mw_item
));
874 assert_spin_locked(&lockres
->l_lock
);
876 list_add_tail(&mw
->mw_item
, &lockres
->l_mask_waiters
);
881 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
882 * if the mask still hadn't reached its goal */
883 static int lockres_remove_mask_waiter(struct ocfs2_lock_res
*lockres
,
884 struct ocfs2_mask_waiter
*mw
)
889 spin_lock_irqsave(&lockres
->l_lock
, flags
);
890 if (!list_empty(&mw
->mw_item
)) {
891 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
894 list_del_init(&mw
->mw_item
);
895 init_completion(&mw
->mw_complete
);
897 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
903 static int ocfs2_cluster_lock(struct ocfs2_super
*osb
,
904 struct ocfs2_lock_res
*lockres
,
909 struct ocfs2_mask_waiter mw
;
910 enum dlm_status status
;
911 int wait
, catch_signals
= !(osb
->s_mount_opt
& OCFS2_MOUNT_NOINTR
);
912 int ret
= 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
917 ocfs2_init_mask_waiter(&mw
);
919 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
920 lkm_flags
|= LKM_VALBLK
;
925 if (catch_signals
&& signal_pending(current
)) {
930 spin_lock_irqsave(&lockres
->l_lock
, flags
);
932 mlog_bug_on_msg(lockres
->l_flags
& OCFS2_LOCK_FREEING
,
933 "Cluster lock called on freeing lockres %s! flags "
934 "0x%lx\n", lockres
->l_name
, lockres
->l_flags
);
936 /* We only compare against the currently granted level
937 * here. If the lock is blocked waiting on a downconvert,
938 * we'll get caught below. */
939 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
&&
940 level
> lockres
->l_level
) {
941 /* is someone sitting in dlm_lock? If so, wait on
943 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
948 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
949 /* lock has not been created yet. */
950 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
952 ret
= ocfs2_lock_create(osb
, lockres
, LKM_NLMODE
, 0);
960 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
&&
961 !ocfs2_may_continue_on_blocked_lock(lockres
, level
)) {
962 /* is the lock is currently blocked on behalf of
964 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BLOCKED
, 0);
969 if (level
> lockres
->l_level
) {
970 if (lockres
->l_action
!= OCFS2_AST_INVALID
)
971 mlog(ML_ERROR
, "lockres %s has action %u pending\n",
972 lockres
->l_name
, lockres
->l_action
);
974 lockres
->l_action
= OCFS2_AST_CONVERT
;
975 lockres
->l_requested
= level
;
976 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
977 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
979 BUG_ON(level
== LKM_IVMODE
);
980 BUG_ON(level
== LKM_NLMODE
);
982 mlog(0, "lock %s, convert from %d to level = %d\n",
983 lockres
->l_name
, lockres
->l_level
, level
);
985 /* call dlm_lock to upgrade lock now */
986 status
= dlmlock(osb
->dlm
,
989 lkm_flags
|LKM_CONVERT
,
991 OCFS2_LOCK_ID_MAX_LEN
- 1,
995 if (status
!= DLM_NORMAL
) {
996 if ((lkm_flags
& LKM_NOQUEUE
) &&
997 (status
== DLM_NOTQUEUED
))
1000 ocfs2_log_dlm_error("dlmlock", status
,
1004 ocfs2_recover_from_dlm_error(lockres
, 1);
1008 mlog(0, "lock %s, successfull return from dlmlock\n",
1011 /* At this point we've gone inside the dlm and need to
1012 * complete our work regardless. */
1015 /* wait for busy to clear and carry on */
1019 /* Ok, if we get here then we're good to go. */
1020 ocfs2_inc_holders(lockres
, level
);
1024 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1027 * This is helping work around a lock inversion between the page lock
1028 * and dlm locks. One path holds the page lock while calling aops
1029 * which block acquiring dlm locks. The voting thread holds dlm
1030 * locks while acquiring page locks while down converting data locks.
1031 * This block is helping an aop path notice the inversion and back
1032 * off to unlock its page lock before trying the dlm lock again.
1034 if (wait
&& arg_flags
& OCFS2_LOCK_NONBLOCK
&&
1035 mw
.mw_mask
& (OCFS2_LOCK_BUSY
|OCFS2_LOCK_BLOCKED
)) {
1037 if (lockres_remove_mask_waiter(lockres
, &mw
))
1043 ret
= ocfs2_wait_for_mask(&mw
);
1053 static void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
1054 struct ocfs2_lock_res
*lockres
,
1057 unsigned long flags
;
1060 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1061 ocfs2_dec_holders(lockres
, level
);
1062 ocfs2_vote_on_unlock(osb
, lockres
);
1063 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1067 static int ocfs2_create_new_lock(struct ocfs2_super
*osb
,
1068 struct ocfs2_lock_res
*lockres
,
1072 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1073 unsigned long flags
;
1074 int lkm_flags
= local
? LKM_LOCAL
: 0;
1076 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1077 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
1078 lockres_or_flags(lockres
, OCFS2_LOCK_LOCAL
);
1079 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1081 return ocfs2_lock_create(osb
, lockres
, level
, lkm_flags
);
1084 /* Grants us an EX lock on the data and metadata resources, skipping
1085 * the normal cluster directory lookup. Use this ONLY on newly created
1086 * inodes which other nodes can't possibly see, and which haven't been
1087 * hashed in the inode hash yet. This can give us a good performance
1088 * increase as it'll skip the network broadcast normally associated
1089 * with creating a new lock resource. */
1090 int ocfs2_create_new_inode_locks(struct inode
*inode
)
1093 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1096 BUG_ON(!ocfs2_inode_is_new(inode
));
1100 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1102 /* NOTE: That we don't increment any of the holder counts, nor
1103 * do we add anything to a journal handle. Since this is
1104 * supposed to be a new inode which the cluster doesn't know
1105 * about yet, there is no need to. As far as the LVB handling
1106 * is concerned, this is basically like acquiring an EX lock
1107 * on a resource which has an invalid one -- we'll set it
1108 * valid when we release the EX. */
1110 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_rw_lockres
, 1, 1);
1117 * We don't want to use LKM_LOCAL on a meta data lock as they
1118 * don't use a generation in their lock names.
1120 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_meta_lockres
, 1, 0);
1126 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_data_lockres
, 1, 1);
1137 int ocfs2_rw_lock(struct inode
*inode
, int write
)
1140 struct ocfs2_lock_res
*lockres
;
1146 mlog(0, "inode %llu take %s RW lock\n",
1147 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1148 write
? "EXMODE" : "PRMODE");
1150 lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1152 level
= write
? LKM_EXMODE
: LKM_PRMODE
;
1154 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
, level
, 0,
1163 void ocfs2_rw_unlock(struct inode
*inode
, int write
)
1165 int level
= write
? LKM_EXMODE
: LKM_PRMODE
;
1166 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1170 mlog(0, "inode %llu drop %s RW lock\n",
1171 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1172 write
? "EXMODE" : "PRMODE");
1174 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
1179 int ocfs2_data_lock_full(struct inode
*inode
,
1183 int status
= 0, level
;
1184 struct ocfs2_lock_res
*lockres
;
1190 mlog(0, "inode %llu take %s DATA lock\n",
1191 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1192 write
? "EXMODE" : "PRMODE");
1194 /* We'll allow faking a readonly data lock for
1196 if (ocfs2_is_hard_readonly(OCFS2_SB(inode
->i_sb
))) {
1204 lockres
= &OCFS2_I(inode
)->ip_data_lockres
;
1206 level
= write
? LKM_EXMODE
: LKM_PRMODE
;
1208 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
, level
,
1210 if (status
< 0 && status
!= -EAGAIN
)
1218 /* see ocfs2_meta_lock_with_page() */
1219 int ocfs2_data_lock_with_page(struct inode
*inode
,
1225 ret
= ocfs2_data_lock_full(inode
, write
, OCFS2_LOCK_NONBLOCK
);
1226 if (ret
== -EAGAIN
) {
1228 if (ocfs2_data_lock(inode
, write
) == 0)
1229 ocfs2_data_unlock(inode
, write
);
1230 ret
= AOP_TRUNCATED_PAGE
;
1236 static void ocfs2_vote_on_unlock(struct ocfs2_super
*osb
,
1237 struct ocfs2_lock_res
*lockres
)
1243 /* If we know that another node is waiting on our lock, kick
1244 * the vote thread * pre-emptively when we reach a release
1246 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
) {
1247 switch(lockres
->l_blocking
) {
1249 if (!lockres
->l_ex_holders
&& !lockres
->l_ro_holders
)
1253 if (!lockres
->l_ex_holders
)
1262 ocfs2_kick_vote_thread(osb
);
1267 void ocfs2_data_unlock(struct inode
*inode
,
1270 int level
= write
? LKM_EXMODE
: LKM_PRMODE
;
1271 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_data_lockres
;
1275 mlog(0, "inode %llu drop %s DATA lock\n",
1276 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1277 write
? "EXMODE" : "PRMODE");
1279 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode
->i_sb
)))
1280 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
1285 #define OCFS2_SEC_BITS 34
1286 #define OCFS2_SEC_SHIFT (64 - 34)
1287 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
1289 /* LVB only has room for 64 bits of time here so we pack it for
1291 static u64
ocfs2_pack_timespec(struct timespec
*spec
)
1294 u64 sec
= spec
->tv_sec
;
1295 u32 nsec
= spec
->tv_nsec
;
1297 res
= (sec
<< OCFS2_SEC_SHIFT
) | (nsec
& OCFS2_NSEC_MASK
);
1302 /* Call this with the lockres locked. I am reasonably sure we don't
1303 * need ip_lock in this function as anyone who would be changing those
1304 * values is supposed to be blocked in ocfs2_meta_lock right now. */
1305 static void __ocfs2_stuff_meta_lvb(struct inode
*inode
)
1307 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1308 struct ocfs2_lock_res
*lockres
= &oi
->ip_meta_lockres
;
1309 struct ocfs2_meta_lvb
*lvb
;
1313 lvb
= (struct ocfs2_meta_lvb
*) lockres
->l_lksb
.lvb
;
1316 * Invalidate the LVB of a deleted inode - this way other
1317 * nodes are forced to go to disk and discover the new inode
1320 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1321 lvb
->lvb_version
= 0;
1325 lvb
->lvb_version
= OCFS2_LVB_VERSION
;
1326 lvb
->lvb_isize
= cpu_to_be64(i_size_read(inode
));
1327 lvb
->lvb_iclusters
= cpu_to_be32(oi
->ip_clusters
);
1328 lvb
->lvb_iuid
= cpu_to_be32(inode
->i_uid
);
1329 lvb
->lvb_igid
= cpu_to_be32(inode
->i_gid
);
1330 lvb
->lvb_imode
= cpu_to_be16(inode
->i_mode
);
1331 lvb
->lvb_inlink
= cpu_to_be16(inode
->i_nlink
);
1332 lvb
->lvb_iatime_packed
=
1333 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_atime
));
1334 lvb
->lvb_ictime_packed
=
1335 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_ctime
));
1336 lvb
->lvb_imtime_packed
=
1337 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_mtime
));
1338 lvb
->lvb_iattr
= cpu_to_be32(oi
->ip_attr
);
1339 lvb
->lvb_igeneration
= cpu_to_be32(inode
->i_generation
);
1342 mlog_meta_lvb(0, lockres
);
1347 static void ocfs2_unpack_timespec(struct timespec
*spec
,
1350 spec
->tv_sec
= packed_time
>> OCFS2_SEC_SHIFT
;
1351 spec
->tv_nsec
= packed_time
& OCFS2_NSEC_MASK
;
1354 static void ocfs2_refresh_inode_from_lvb(struct inode
*inode
)
1356 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1357 struct ocfs2_lock_res
*lockres
= &oi
->ip_meta_lockres
;
1358 struct ocfs2_meta_lvb
*lvb
;
1362 mlog_meta_lvb(0, lockres
);
1364 lvb
= (struct ocfs2_meta_lvb
*) lockres
->l_lksb
.lvb
;
1366 /* We're safe here without the lockres lock... */
1367 spin_lock(&oi
->ip_lock
);
1368 oi
->ip_clusters
= be32_to_cpu(lvb
->lvb_iclusters
);
1369 i_size_write(inode
, be64_to_cpu(lvb
->lvb_isize
));
1371 oi
->ip_attr
= be32_to_cpu(lvb
->lvb_iattr
);
1372 ocfs2_set_inode_flags(inode
);
1374 /* fast-symlinks are a special case */
1375 if (S_ISLNK(inode
->i_mode
) && !oi
->ip_clusters
)
1376 inode
->i_blocks
= 0;
1379 ocfs2_align_bytes_to_sectors(i_size_read(inode
));
1381 inode
->i_uid
= be32_to_cpu(lvb
->lvb_iuid
);
1382 inode
->i_gid
= be32_to_cpu(lvb
->lvb_igid
);
1383 inode
->i_mode
= be16_to_cpu(lvb
->lvb_imode
);
1384 inode
->i_nlink
= be16_to_cpu(lvb
->lvb_inlink
);
1385 ocfs2_unpack_timespec(&inode
->i_atime
,
1386 be64_to_cpu(lvb
->lvb_iatime_packed
));
1387 ocfs2_unpack_timespec(&inode
->i_mtime
,
1388 be64_to_cpu(lvb
->lvb_imtime_packed
));
1389 ocfs2_unpack_timespec(&inode
->i_ctime
,
1390 be64_to_cpu(lvb
->lvb_ictime_packed
));
1391 spin_unlock(&oi
->ip_lock
);
1396 static inline int ocfs2_meta_lvb_is_trustable(struct inode
*inode
,
1397 struct ocfs2_lock_res
*lockres
)
1399 struct ocfs2_meta_lvb
*lvb
= (struct ocfs2_meta_lvb
*) lockres
->l_lksb
.lvb
;
1401 if (lvb
->lvb_version
== OCFS2_LVB_VERSION
1402 && be32_to_cpu(lvb
->lvb_igeneration
) == inode
->i_generation
)
1407 /* Determine whether a lock resource needs to be refreshed, and
1408 * arbitrate who gets to refresh it.
1410 * 0 means no refresh needed.
1412 * > 0 means you need to refresh this and you MUST call
1413 * ocfs2_complete_lock_res_refresh afterwards. */
1414 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res
*lockres
)
1416 unsigned long flags
;
1422 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1423 if (!(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
)) {
1424 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1428 if (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
) {
1429 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1431 ocfs2_wait_on_refreshing_lock(lockres
);
1435 /* Ok, I'll be the one to refresh this lock. */
1436 lockres_or_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1437 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1445 /* If status is non zero, I'll mark it as not being in refresh
1446 * anymroe, but i won't clear the needs refresh flag. */
1447 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res
*lockres
,
1450 unsigned long flags
;
1453 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1454 lockres_clear_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1456 lockres_clear_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
1457 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1459 wake_up(&lockres
->l_event
);
1464 /* may or may not return a bh if it went to disk. */
1465 static int ocfs2_meta_lock_update(struct inode
*inode
,
1466 struct buffer_head
**bh
)
1469 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1470 struct ocfs2_lock_res
*lockres
;
1471 struct ocfs2_dinode
*fe
;
1475 spin_lock(&oi
->ip_lock
);
1476 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1477 mlog(0, "Orphaned inode %llu was deleted while we "
1478 "were waiting on a lock. ip_flags = 0x%x\n",
1479 (unsigned long long)oi
->ip_blkno
, oi
->ip_flags
);
1480 spin_unlock(&oi
->ip_lock
);
1484 spin_unlock(&oi
->ip_lock
);
1486 lockres
= &oi
->ip_meta_lockres
;
1488 if (!ocfs2_should_refresh_lock_res(lockres
))
1491 /* This will discard any caching information we might have had
1492 * for the inode metadata. */
1493 ocfs2_metadata_cache_purge(inode
);
1495 /* will do nothing for inode types that don't use the extent
1496 * map (directories, bitmap files, etc) */
1497 ocfs2_extent_map_trunc(inode
, 0);
1499 if (ocfs2_meta_lvb_is_trustable(inode
, lockres
)) {
1500 mlog(0, "Trusting LVB on inode %llu\n",
1501 (unsigned long long)oi
->ip_blkno
);
1502 ocfs2_refresh_inode_from_lvb(inode
);
1504 /* Boo, we have to go to disk. */
1505 /* read bh, cast, ocfs2_refresh_inode */
1506 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), oi
->ip_blkno
,
1507 bh
, OCFS2_BH_CACHED
, inode
);
1512 fe
= (struct ocfs2_dinode
*) (*bh
)->b_data
;
1514 /* This is a good chance to make sure we're not
1515 * locking an invalid object.
1517 * We bug on a stale inode here because we checked
1518 * above whether it was wiped from disk. The wiping
1519 * node provides a guarantee that we receive that
1520 * message and can mark the inode before dropping any
1521 * locks associated with it. */
1522 if (!OCFS2_IS_VALID_DINODE(fe
)) {
1523 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
1527 mlog_bug_on_msg(inode
->i_generation
!=
1528 le32_to_cpu(fe
->i_generation
),
1529 "Invalid dinode %llu disk generation: %u "
1530 "inode->i_generation: %u\n",
1531 (unsigned long long)oi
->ip_blkno
,
1532 le32_to_cpu(fe
->i_generation
),
1533 inode
->i_generation
);
1534 mlog_bug_on_msg(le64_to_cpu(fe
->i_dtime
) ||
1535 !(fe
->i_flags
& cpu_to_le32(OCFS2_VALID_FL
)),
1536 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
1537 (unsigned long long)oi
->ip_blkno
,
1538 (unsigned long long)le64_to_cpu(fe
->i_dtime
),
1539 le32_to_cpu(fe
->i_flags
));
1541 ocfs2_refresh_inode(inode
, fe
);
1546 ocfs2_complete_lock_res_refresh(lockres
, status
);
1552 static int ocfs2_assign_bh(struct inode
*inode
,
1553 struct buffer_head
**ret_bh
,
1554 struct buffer_head
*passed_bh
)
1559 /* Ok, the update went to disk for us, use the
1561 *ret_bh
= passed_bh
;
1567 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
1568 OCFS2_I(inode
)->ip_blkno
,
1579 * returns < 0 error if the callback will never be called, otherwise
1580 * the result of the lock will be communicated via the callback.
1582 int ocfs2_meta_lock_full(struct inode
*inode
,
1583 struct buffer_head
**ret_bh
,
1587 int status
, level
, dlm_flags
, acquired
;
1588 struct ocfs2_lock_res
*lockres
;
1589 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1590 struct buffer_head
*local_bh
= NULL
;
1596 mlog(0, "inode %llu, take %s META lock\n",
1597 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1598 ex
? "EXMODE" : "PRMODE");
1602 /* We'll allow faking a readonly metadata lock for
1604 if (ocfs2_is_hard_readonly(osb
)) {
1610 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
1611 wait_event(osb
->recovery_event
,
1612 ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
));
1615 lockres
= &OCFS2_I(inode
)->ip_meta_lockres
;
1616 level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1618 if (arg_flags
& OCFS2_META_LOCK_NOQUEUE
)
1619 dlm_flags
|= LKM_NOQUEUE
;
1621 status
= ocfs2_cluster_lock(osb
, lockres
, level
, dlm_flags
, arg_flags
);
1623 if (status
!= -EAGAIN
&& status
!= -EIOCBRETRY
)
1628 /* Notify the error cleanup path to drop the cluster lock. */
1631 /* We wait twice because a node may have died while we were in
1632 * the lower dlm layers. The second time though, we've
1633 * committed to owning this lock so we don't allow signals to
1634 * abort the operation. */
1635 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
1636 wait_event(osb
->recovery_event
,
1637 ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
));
1640 * We only see this flag if we're being called from
1641 * ocfs2_read_locked_inode(). It means we're locking an inode
1642 * which hasn't been populated yet, so clear the refresh flag
1643 * and let the caller handle it.
1645 if (inode
->i_state
& I_NEW
) {
1647 ocfs2_complete_lock_res_refresh(lockres
, 0);
1651 /* This is fun. The caller may want a bh back, or it may
1652 * not. ocfs2_meta_lock_update definitely wants one in, but
1653 * may or may not read one, depending on what's in the
1654 * LVB. The result of all of this is that we've *only* gone to
1655 * disk if we have to, so the complexity is worthwhile. */
1656 status
= ocfs2_meta_lock_update(inode
, &local_bh
);
1658 if (status
!= -ENOENT
)
1664 status
= ocfs2_assign_bh(inode
, ret_bh
, local_bh
);
1673 if (ret_bh
&& (*ret_bh
)) {
1678 ocfs2_meta_unlock(inode
, ex
);
1689 * This is working around a lock inversion between tasks acquiring DLM locks
1690 * while holding a page lock and the vote thread which blocks dlm lock acquiry
1691 * while acquiring page locks.
1693 * ** These _with_page variantes are only intended to be called from aop
1694 * methods that hold page locks and return a very specific *positive* error
1695 * code that aop methods pass up to the VFS -- test for errors with != 0. **
1697 * The DLM is called such that it returns -EAGAIN if it would have blocked
1698 * waiting for the vote thread. In that case we unlock our page so the vote
1699 * thread can make progress. Once we've done this we have to return
1700 * AOP_TRUNCATED_PAGE so the aop method that called us can bubble that back up
1701 * into the VFS who will then immediately retry the aop call.
1703 * We do a blocking lock and immediate unlock before returning, though, so that
1704 * the lock has a great chance of being cached on this node by the time the VFS
1705 * calls back to retry the aop. This has a potential to livelock as nodes
1706 * ping locks back and forth, but that's a risk we're willing to take to avoid
1707 * the lock inversion simply.
1709 int ocfs2_meta_lock_with_page(struct inode
*inode
,
1710 struct buffer_head
**ret_bh
,
1716 ret
= ocfs2_meta_lock_full(inode
, ret_bh
, ex
, OCFS2_LOCK_NONBLOCK
);
1717 if (ret
== -EAGAIN
) {
1719 if (ocfs2_meta_lock(inode
, ret_bh
, ex
) == 0)
1720 ocfs2_meta_unlock(inode
, ex
);
1721 ret
= AOP_TRUNCATED_PAGE
;
1727 int ocfs2_meta_lock_atime(struct inode
*inode
,
1728 struct vfsmount
*vfsmnt
,
1734 ret
= ocfs2_meta_lock(inode
, NULL
, 0);
1741 * If we should update atime, we will get EX lock,
1742 * otherwise we just get PR lock.
1744 if (ocfs2_should_update_atime(inode
, vfsmnt
)) {
1745 struct buffer_head
*bh
= NULL
;
1747 ocfs2_meta_unlock(inode
, 0);
1748 ret
= ocfs2_meta_lock(inode
, &bh
, 1);
1754 if (ocfs2_should_update_atime(inode
, vfsmnt
))
1755 ocfs2_update_inode_atime(inode
, bh
);
1765 void ocfs2_meta_unlock(struct inode
*inode
,
1768 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1769 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_meta_lockres
;
1773 mlog(0, "inode %llu drop %s META lock\n",
1774 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1775 ex
? "EXMODE" : "PRMODE");
1777 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode
->i_sb
)))
1778 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
1783 int ocfs2_super_lock(struct ocfs2_super
*osb
,
1787 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1788 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
1789 struct buffer_head
*bh
;
1790 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1794 if (ocfs2_is_hard_readonly(osb
))
1797 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
1803 /* The super block lock path is really in the best position to
1804 * know when resources covered by the lock need to be
1805 * refreshed, so we do it here. Of course, making sense of
1806 * everything is up to the caller :) */
1807 status
= ocfs2_should_refresh_lock_res(lockres
);
1814 status
= ocfs2_read_block(osb
, bh
->b_blocknr
, &bh
, 0,
1817 ocfs2_update_slot_info(si
);
1819 ocfs2_complete_lock_res_refresh(lockres
, status
);
1829 void ocfs2_super_unlock(struct ocfs2_super
*osb
,
1832 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1833 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
1835 ocfs2_cluster_unlock(osb
, lockres
, level
);
1838 int ocfs2_rename_lock(struct ocfs2_super
*osb
)
1841 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
1843 if (ocfs2_is_hard_readonly(osb
))
1846 status
= ocfs2_cluster_lock(osb
, lockres
, LKM_EXMODE
, 0, 0);
1853 void ocfs2_rename_unlock(struct ocfs2_super
*osb
)
1855 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
1857 ocfs2_cluster_unlock(osb
, lockres
, LKM_EXMODE
);
1860 int ocfs2_dentry_lock(struct dentry
*dentry
, int ex
)
1863 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1864 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
1865 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
1869 if (ocfs2_is_hard_readonly(osb
))
1872 ret
= ocfs2_cluster_lock(osb
, &dl
->dl_lockres
, level
, 0, 0);
1879 void ocfs2_dentry_unlock(struct dentry
*dentry
, int ex
)
1881 int level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1882 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
1883 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
1885 ocfs2_cluster_unlock(osb
, &dl
->dl_lockres
, level
);
1888 /* Reference counting of the dlm debug structure. We want this because
1889 * open references on the debug inodes can live on after a mount, so
1890 * we can't rely on the ocfs2_super to always exist. */
1891 static void ocfs2_dlm_debug_free(struct kref
*kref
)
1893 struct ocfs2_dlm_debug
*dlm_debug
;
1895 dlm_debug
= container_of(kref
, struct ocfs2_dlm_debug
, d_refcnt
);
1900 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug
*dlm_debug
)
1903 kref_put(&dlm_debug
->d_refcnt
, ocfs2_dlm_debug_free
);
1906 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug
*debug
)
1908 kref_get(&debug
->d_refcnt
);
1911 struct ocfs2_dlm_debug
*ocfs2_new_dlm_debug(void)
1913 struct ocfs2_dlm_debug
*dlm_debug
;
1915 dlm_debug
= kmalloc(sizeof(struct ocfs2_dlm_debug
), GFP_KERNEL
);
1917 mlog_errno(-ENOMEM
);
1921 kref_init(&dlm_debug
->d_refcnt
);
1922 INIT_LIST_HEAD(&dlm_debug
->d_lockres_tracking
);
1923 dlm_debug
->d_locking_state
= NULL
;
1928 /* Access to this is arbitrated for us via seq_file->sem. */
1929 struct ocfs2_dlm_seq_priv
{
1930 struct ocfs2_dlm_debug
*p_dlm_debug
;
1931 struct ocfs2_lock_res p_iter_res
;
1932 struct ocfs2_lock_res p_tmp_res
;
1935 static struct ocfs2_lock_res
*ocfs2_dlm_next_res(struct ocfs2_lock_res
*start
,
1936 struct ocfs2_dlm_seq_priv
*priv
)
1938 struct ocfs2_lock_res
*iter
, *ret
= NULL
;
1939 struct ocfs2_dlm_debug
*dlm_debug
= priv
->p_dlm_debug
;
1941 assert_spin_locked(&ocfs2_dlm_tracking_lock
);
1943 list_for_each_entry(iter
, &start
->l_debug_list
, l_debug_list
) {
1944 /* discover the head of the list */
1945 if (&iter
->l_debug_list
== &dlm_debug
->d_lockres_tracking
) {
1946 mlog(0, "End of list found, %p\n", ret
);
1950 /* We track our "dummy" iteration lockres' by a NULL
1952 if (iter
->l_ops
!= NULL
) {
1961 static void *ocfs2_dlm_seq_start(struct seq_file
*m
, loff_t
*pos
)
1963 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
1964 struct ocfs2_lock_res
*iter
;
1966 spin_lock(&ocfs2_dlm_tracking_lock
);
1967 iter
= ocfs2_dlm_next_res(&priv
->p_iter_res
, priv
);
1969 /* Since lockres' have the lifetime of their container
1970 * (which can be inodes, ocfs2_supers, etc) we want to
1971 * copy this out to a temporary lockres while still
1972 * under the spinlock. Obviously after this we can't
1973 * trust any pointers on the copy returned, but that's
1974 * ok as the information we want isn't typically held
1976 priv
->p_tmp_res
= *iter
;
1977 iter
= &priv
->p_tmp_res
;
1979 spin_unlock(&ocfs2_dlm_tracking_lock
);
1984 static void ocfs2_dlm_seq_stop(struct seq_file
*m
, void *v
)
1988 static void *ocfs2_dlm_seq_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
1990 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
1991 struct ocfs2_lock_res
*iter
= v
;
1992 struct ocfs2_lock_res
*dummy
= &priv
->p_iter_res
;
1994 spin_lock(&ocfs2_dlm_tracking_lock
);
1995 iter
= ocfs2_dlm_next_res(iter
, priv
);
1996 list_del_init(&dummy
->l_debug_list
);
1998 list_add(&dummy
->l_debug_list
, &iter
->l_debug_list
);
1999 priv
->p_tmp_res
= *iter
;
2000 iter
= &priv
->p_tmp_res
;
2002 spin_unlock(&ocfs2_dlm_tracking_lock
);
2007 /* So that debugfs.ocfs2 can determine which format is being used */
2008 #define OCFS2_DLM_DEBUG_STR_VERSION 1
2009 static int ocfs2_dlm_seq_show(struct seq_file
*m
, void *v
)
2013 struct ocfs2_lock_res
*lockres
= v
;
2018 seq_printf(m
, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION
);
2020 if (lockres
->l_type
== OCFS2_LOCK_TYPE_DENTRY
)
2021 seq_printf(m
, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START
- 1,
2023 (unsigned int)ocfs2_get_dentry_lock_ino(lockres
));
2025 seq_printf(m
, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN
, lockres
->l_name
);
2027 seq_printf(m
, "%d\t"
2038 lockres
->l_unlock_action
,
2039 lockres
->l_ro_holders
,
2040 lockres
->l_ex_holders
,
2041 lockres
->l_requested
,
2042 lockres
->l_blocking
);
2044 /* Dump the raw LVB */
2045 lvb
= lockres
->l_lksb
.lvb
;
2046 for(i
= 0; i
< DLM_LVB_LEN
; i
++)
2047 seq_printf(m
, "0x%x\t", lvb
[i
]);
2050 seq_printf(m
, "\n");
2054 static struct seq_operations ocfs2_dlm_seq_ops
= {
2055 .start
= ocfs2_dlm_seq_start
,
2056 .stop
= ocfs2_dlm_seq_stop
,
2057 .next
= ocfs2_dlm_seq_next
,
2058 .show
= ocfs2_dlm_seq_show
,
2061 static int ocfs2_dlm_debug_release(struct inode
*inode
, struct file
*file
)
2063 struct seq_file
*seq
= (struct seq_file
*) file
->private_data
;
2064 struct ocfs2_dlm_seq_priv
*priv
= seq
->private;
2065 struct ocfs2_lock_res
*res
= &priv
->p_iter_res
;
2067 ocfs2_remove_lockres_tracking(res
);
2068 ocfs2_put_dlm_debug(priv
->p_dlm_debug
);
2069 return seq_release_private(inode
, file
);
2072 static int ocfs2_dlm_debug_open(struct inode
*inode
, struct file
*file
)
2075 struct ocfs2_dlm_seq_priv
*priv
;
2076 struct seq_file
*seq
;
2077 struct ocfs2_super
*osb
;
2079 priv
= kzalloc(sizeof(struct ocfs2_dlm_seq_priv
), GFP_KERNEL
);
2085 osb
= inode
->i_private
;
2086 ocfs2_get_dlm_debug(osb
->osb_dlm_debug
);
2087 priv
->p_dlm_debug
= osb
->osb_dlm_debug
;
2088 INIT_LIST_HEAD(&priv
->p_iter_res
.l_debug_list
);
2090 ret
= seq_open(file
, &ocfs2_dlm_seq_ops
);
2097 seq
= (struct seq_file
*) file
->private_data
;
2098 seq
->private = priv
;
2100 ocfs2_add_lockres_tracking(&priv
->p_iter_res
,
2107 static const struct file_operations ocfs2_dlm_debug_fops
= {
2108 .open
= ocfs2_dlm_debug_open
,
2109 .release
= ocfs2_dlm_debug_release
,
2111 .llseek
= seq_lseek
,
2114 static int ocfs2_dlm_init_debug(struct ocfs2_super
*osb
)
2117 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2119 dlm_debug
->d_locking_state
= debugfs_create_file("locking_state",
2121 osb
->osb_debug_root
,
2123 &ocfs2_dlm_debug_fops
);
2124 if (!dlm_debug
->d_locking_state
) {
2127 "Unable to create locking state debugfs file.\n");
2131 ocfs2_get_dlm_debug(dlm_debug
);
2136 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super
*osb
)
2138 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2141 debugfs_remove(dlm_debug
->d_locking_state
);
2142 ocfs2_put_dlm_debug(dlm_debug
);
2146 int ocfs2_dlm_init(struct ocfs2_super
*osb
)
2150 struct dlm_ctxt
*dlm
;
2154 status
= ocfs2_dlm_init_debug(osb
);
2160 /* launch vote thread */
2161 osb
->vote_task
= kthread_run(ocfs2_vote_thread
, osb
, "ocfs2vote");
2162 if (IS_ERR(osb
->vote_task
)) {
2163 status
= PTR_ERR(osb
->vote_task
);
2164 osb
->vote_task
= NULL
;
2169 /* used by the dlm code to make message headers unique, each
2170 * node in this domain must agree on this. */
2171 dlm_key
= crc32_le(0, osb
->uuid_str
, strlen(osb
->uuid_str
));
2173 /* for now, uuid == domain */
2174 dlm
= dlm_register_domain(osb
->uuid_str
, dlm_key
);
2176 status
= PTR_ERR(dlm
);
2181 ocfs2_super_lock_res_init(&osb
->osb_super_lockres
, osb
);
2182 ocfs2_rename_lock_res_init(&osb
->osb_rename_lockres
, osb
);
2184 dlm_register_eviction_cb(dlm
, &osb
->osb_eviction_cb
);
2191 ocfs2_dlm_shutdown_debug(osb
);
2193 kthread_stop(osb
->vote_task
);
2200 void ocfs2_dlm_shutdown(struct ocfs2_super
*osb
)
2204 dlm_unregister_eviction_cb(&osb
->osb_eviction_cb
);
2206 ocfs2_drop_osb_locks(osb
);
2208 if (osb
->vote_task
) {
2209 kthread_stop(osb
->vote_task
);
2210 osb
->vote_task
= NULL
;
2213 ocfs2_lock_res_free(&osb
->osb_super_lockres
);
2214 ocfs2_lock_res_free(&osb
->osb_rename_lockres
);
2216 dlm_unregister_domain(osb
->dlm
);
2219 ocfs2_dlm_shutdown_debug(osb
);
2224 static void ocfs2_unlock_ast(void *opaque
, enum dlm_status status
)
2226 struct ocfs2_lock_res
*lockres
= opaque
;
2227 unsigned long flags
;
2231 mlog(0, "UNLOCK AST called on lock %s, action = %d\n", lockres
->l_name
,
2232 lockres
->l_unlock_action
);
2234 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2235 /* We tried to cancel a convert request, but it was already
2236 * granted. All we want to do here is clear our unlock
2237 * state. The wake_up call done at the bottom is redundant
2238 * (ocfs2_prepare_cancel_convert doesn't sleep on this) but doesn't
2239 * hurt anything anyway */
2240 if (status
== DLM_CANCELGRANT
&&
2241 lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
2242 mlog(0, "Got cancelgrant for %s\n", lockres
->l_name
);
2244 /* We don't clear the busy flag in this case as it
2245 * should have been cleared by the ast which the dlm
2247 goto complete_unlock
;
2250 if (status
!= DLM_NORMAL
) {
2251 mlog(ML_ERROR
, "Dlm passes status %d for lock %s, "
2252 "unlock_action %d\n", status
, lockres
->l_name
,
2253 lockres
->l_unlock_action
);
2254 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2258 switch(lockres
->l_unlock_action
) {
2259 case OCFS2_UNLOCK_CANCEL_CONVERT
:
2260 mlog(0, "Cancel convert success for %s\n", lockres
->l_name
);
2261 lockres
->l_action
= OCFS2_AST_INVALID
;
2263 case OCFS2_UNLOCK_DROP_LOCK
:
2264 lockres
->l_level
= LKM_IVMODE
;
2270 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
2272 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
2273 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2275 wake_up(&lockres
->l_event
);
2280 static int ocfs2_drop_lock(struct ocfs2_super
*osb
,
2281 struct ocfs2_lock_res
*lockres
)
2283 enum dlm_status status
;
2284 unsigned long flags
;
2287 /* We didn't get anywhere near actually using this lockres. */
2288 if (!(lockres
->l_flags
& OCFS2_LOCK_INITIALIZED
))
2291 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
2292 lkm_flags
|= LKM_VALBLK
;
2294 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2296 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_FREEING
),
2297 "lockres %s, flags 0x%lx\n",
2298 lockres
->l_name
, lockres
->l_flags
);
2300 while (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
2301 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
2302 "%u, unlock_action = %u\n",
2303 lockres
->l_name
, lockres
->l_flags
, lockres
->l_action
,
2304 lockres
->l_unlock_action
);
2306 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2308 /* XXX: Today we just wait on any busy
2309 * locks... Perhaps we need to cancel converts in the
2311 ocfs2_wait_on_busy_lock(lockres
);
2313 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2316 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
2317 if (lockres
->l_flags
& OCFS2_LOCK_ATTACHED
&&
2318 lockres
->l_level
== LKM_EXMODE
&&
2319 !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
2320 lockres
->l_ops
->set_lvb(lockres
);
2323 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
)
2324 mlog(ML_ERROR
, "destroying busy lock: \"%s\"\n",
2326 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
2327 mlog(0, "destroying blocked lock: \"%s\"\n", lockres
->l_name
);
2329 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
2330 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2334 lockres_clear_flags(lockres
, OCFS2_LOCK_ATTACHED
);
2336 /* make sure we never get here while waiting for an ast to
2338 BUG_ON(lockres
->l_action
!= OCFS2_AST_INVALID
);
2340 /* is this necessary? */
2341 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2342 lockres
->l_unlock_action
= OCFS2_UNLOCK_DROP_LOCK
;
2343 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2345 mlog(0, "lock %s\n", lockres
->l_name
);
2347 status
= dlmunlock(osb
->dlm
, &lockres
->l_lksb
, lkm_flags
,
2348 ocfs2_unlock_ast
, lockres
);
2349 if (status
!= DLM_NORMAL
) {
2350 ocfs2_log_dlm_error("dlmunlock", status
, lockres
);
2351 mlog(ML_ERROR
, "lockres flags: %lu\n", lockres
->l_flags
);
2352 dlm_print_one_lock(lockres
->l_lksb
.lockid
);
2355 mlog(0, "lock %s, successfull return from dlmunlock\n",
2358 ocfs2_wait_on_busy_lock(lockres
);
2364 /* Mark the lockres as being dropped. It will no longer be
2365 * queued if blocking, but we still may have to wait on it
2366 * being dequeued from the vote thread before we can consider
2369 * You can *not* attempt to call cluster_lock on this lockres anymore. */
2370 void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res
*lockres
)
2373 struct ocfs2_mask_waiter mw
;
2374 unsigned long flags
;
2376 ocfs2_init_mask_waiter(&mw
);
2378 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2379 lockres
->l_flags
|= OCFS2_LOCK_FREEING
;
2380 while (lockres
->l_flags
& OCFS2_LOCK_QUEUED
) {
2381 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_QUEUED
, 0);
2382 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2384 mlog(0, "Waiting on lockres %s\n", lockres
->l_name
);
2386 status
= ocfs2_wait_for_mask(&mw
);
2390 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2392 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2395 void ocfs2_simple_drop_lockres(struct ocfs2_super
*osb
,
2396 struct ocfs2_lock_res
*lockres
)
2400 ocfs2_mark_lockres_freeing(lockres
);
2401 ret
= ocfs2_drop_lock(osb
, lockres
);
2406 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
)
2408 ocfs2_simple_drop_lockres(osb
, &osb
->osb_super_lockres
);
2409 ocfs2_simple_drop_lockres(osb
, &osb
->osb_rename_lockres
);
2412 int ocfs2_drop_inode_locks(struct inode
*inode
)
2418 /* No need to call ocfs2_mark_lockres_freeing here -
2419 * ocfs2_clear_inode has done it for us. */
2421 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2422 &OCFS2_I(inode
)->ip_data_lockres
);
2428 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2429 &OCFS2_I(inode
)->ip_meta_lockres
);
2432 if (err
< 0 && !status
)
2435 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2436 &OCFS2_I(inode
)->ip_rw_lockres
);
2439 if (err
< 0 && !status
)
2446 static void ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
2449 assert_spin_locked(&lockres
->l_lock
);
2451 BUG_ON(lockres
->l_blocking
<= LKM_NLMODE
);
2453 if (lockres
->l_level
<= new_level
) {
2454 mlog(ML_ERROR
, "lockres->l_level (%u) <= new_level (%u)\n",
2455 lockres
->l_level
, new_level
);
2459 mlog(0, "lock %s, new_level = %d, l_blocking = %d\n",
2460 lockres
->l_name
, new_level
, lockres
->l_blocking
);
2462 lockres
->l_action
= OCFS2_AST_DOWNCONVERT
;
2463 lockres
->l_requested
= new_level
;
2464 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2467 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
2468 struct ocfs2_lock_res
*lockres
,
2472 int ret
, dlm_flags
= LKM_CONVERT
;
2473 enum dlm_status status
;
2478 dlm_flags
|= LKM_VALBLK
;
2480 status
= dlmlock(osb
->dlm
,
2485 OCFS2_LOCK_ID_MAX_LEN
- 1,
2488 ocfs2_blocking_ast
);
2489 if (status
!= DLM_NORMAL
) {
2490 ocfs2_log_dlm_error("dlmlock", status
, lockres
);
2492 ocfs2_recover_from_dlm_error(lockres
, 1);
2502 /* returns 1 when the caller should unlock and call dlmunlock */
2503 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
2504 struct ocfs2_lock_res
*lockres
)
2506 assert_spin_locked(&lockres
->l_lock
);
2509 mlog(0, "lock %s\n", lockres
->l_name
);
2511 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
2512 /* If we're already trying to cancel a lock conversion
2513 * then just drop the spinlock and allow the caller to
2514 * requeue this lock. */
2516 mlog(0, "Lockres %s, skip convert\n", lockres
->l_name
);
2520 /* were we in a convert when we got the bast fire? */
2521 BUG_ON(lockres
->l_action
!= OCFS2_AST_CONVERT
&&
2522 lockres
->l_action
!= OCFS2_AST_DOWNCONVERT
);
2523 /* set things up for the unlockast to know to just
2524 * clear out the ast_action and unset busy, etc. */
2525 lockres
->l_unlock_action
= OCFS2_UNLOCK_CANCEL_CONVERT
;
2527 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
),
2528 "lock %s, invalid flags: 0x%lx\n",
2529 lockres
->l_name
, lockres
->l_flags
);
2534 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
2535 struct ocfs2_lock_res
*lockres
)
2538 enum dlm_status status
;
2541 mlog(0, "lock %s\n", lockres
->l_name
);
2544 status
= dlmunlock(osb
->dlm
,
2549 if (status
!= DLM_NORMAL
) {
2550 ocfs2_log_dlm_error("dlmunlock", status
, lockres
);
2552 ocfs2_recover_from_dlm_error(lockres
, 0);
2555 mlog(0, "lock %s return from dlmunlock\n", lockres
->l_name
);
2561 static int ocfs2_unblock_lock(struct ocfs2_super
*osb
,
2562 struct ocfs2_lock_res
*lockres
,
2563 struct ocfs2_unblock_ctl
*ctl
)
2565 unsigned long flags
;
2573 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2575 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
2578 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
2580 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
2581 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2583 ret
= ocfs2_cancel_convert(osb
, lockres
);
2590 /* if we're blocking an exclusive and we have *any* holders,
2592 if ((lockres
->l_blocking
== LKM_EXMODE
)
2593 && (lockres
->l_ex_holders
|| lockres
->l_ro_holders
))
2596 /* If it's a PR we're blocking, then only
2597 * requeue if we've got any EX holders */
2598 if (lockres
->l_blocking
== LKM_PRMODE
&&
2599 lockres
->l_ex_holders
)
2603 * Can we get a lock in this state if the holder counts are
2604 * zero? The meta data unblock code used to check this.
2606 if ((lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
2607 && (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
))
2610 new_level
= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
2612 if (lockres
->l_ops
->check_downconvert
2613 && !lockres
->l_ops
->check_downconvert(lockres
, new_level
))
2616 /* If we get here, then we know that there are no more
2617 * incompatible holders (and anyone asking for an incompatible
2618 * lock is blocked). We can now downconvert the lock */
2619 if (!lockres
->l_ops
->downconvert_worker
)
2622 /* Some lockres types want to do a bit of work before
2623 * downconverting a lock. Allow that here. The worker function
2624 * may sleep, so we save off a copy of what we're blocking as
2625 * it may change while we're not holding the spin lock. */
2626 blocking
= lockres
->l_blocking
;
2627 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2629 ctl
->unblock_action
= lockres
->l_ops
->downconvert_worker(lockres
, blocking
);
2631 if (ctl
->unblock_action
== UNBLOCK_STOP_POST
)
2634 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2635 if (blocking
!= lockres
->l_blocking
) {
2636 /* If this changed underneath us, then we can't drop
2644 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
2645 if (lockres
->l_level
== LKM_EXMODE
)
2649 * We only set the lvb if the lock has been fully
2650 * refreshed - otherwise we risk setting stale
2651 * data. Otherwise, there's no need to actually clear
2652 * out the lvb here as it's value is still valid.
2654 if (set_lvb
&& !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
2655 lockres
->l_ops
->set_lvb(lockres
);
2658 ocfs2_prepare_downconvert(lockres
, new_level
);
2659 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2660 ret
= ocfs2_downconvert_lock(osb
, lockres
, new_level
, set_lvb
);
2666 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2673 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
2676 struct inode
*inode
;
2677 struct address_space
*mapping
;
2679 inode
= ocfs2_lock_res_inode(lockres
);
2680 mapping
= inode
->i_mapping
;
2682 if (filemap_fdatawrite(mapping
)) {
2683 mlog(ML_ERROR
, "Could not sync inode %llu for downconvert!",
2684 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
2686 sync_mapping_buffers(mapping
);
2687 if (blocking
== LKM_EXMODE
) {
2688 truncate_inode_pages(mapping
, 0);
2689 unmap_mapping_range(mapping
, 0, 0, 0);
2691 /* We only need to wait on the I/O if we're not also
2692 * truncating pages because truncate_inode_pages waits
2693 * for us above. We don't truncate pages if we're
2694 * blocking anything < EXMODE because we want to keep
2695 * them around in that case. */
2696 filemap_fdatawait(mapping
);
2699 return UNBLOCK_CONTINUE
;
2702 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
2705 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
2706 int checkpointed
= ocfs2_inode_fully_checkpointed(inode
);
2708 BUG_ON(new_level
!= LKM_NLMODE
&& new_level
!= LKM_PRMODE
);
2709 BUG_ON(lockres
->l_level
!= LKM_EXMODE
&& !checkpointed
);
2714 ocfs2_start_checkpoint(OCFS2_SB(inode
->i_sb
));
2718 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
)
2720 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
2722 __ocfs2_stuff_meta_lvb(inode
);
2726 * Does the final reference drop on our dentry lock. Right now this
2727 * happens in the vote thread, but we could choose to simplify the
2728 * dlmglue API and push these off to the ocfs2_wq in the future.
2730 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
2731 struct ocfs2_lock_res
*lockres
)
2733 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
2734 ocfs2_dentry_lock_put(osb
, dl
);
2738 * d_delete() matching dentries before the lock downconvert.
2740 * At this point, any process waiting to destroy the
2741 * dentry_lock due to last ref count is stopped by the
2742 * OCFS2_LOCK_QUEUED flag.
2744 * We have two potential problems
2746 * 1) If we do the last reference drop on our dentry_lock (via dput)
2747 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
2748 * the downconvert to finish. Instead we take an elevated
2749 * reference and push the drop until after we've completed our
2750 * unblock processing.
2752 * 2) There might be another process with a final reference,
2753 * waiting on us to finish processing. If this is the case, we
2754 * detect it and exit out - there's no more dentries anyway.
2756 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
2759 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
2760 struct ocfs2_inode_info
*oi
= OCFS2_I(dl
->dl_inode
);
2761 struct dentry
*dentry
;
2762 unsigned long flags
;
2766 * This node is blocking another node from getting a read
2767 * lock. This happens when we've renamed within a
2768 * directory. We've forced the other nodes to d_delete(), but
2769 * we never actually dropped our lock because it's still
2770 * valid. The downconvert code will retain a PR for this node,
2771 * so there's no further work to do.
2773 if (blocking
== LKM_PRMODE
)
2774 return UNBLOCK_CONTINUE
;
2777 * Mark this inode as potentially orphaned. The code in
2778 * ocfs2_delete_inode() will figure out whether it actually
2779 * needs to be freed or not.
2781 spin_lock(&oi
->ip_lock
);
2782 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
2783 spin_unlock(&oi
->ip_lock
);
2786 * Yuck. We need to make sure however that the check of
2787 * OCFS2_LOCK_FREEING and the extra reference are atomic with
2788 * respect to a reference decrement or the setting of that
2791 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2792 spin_lock(&dentry_attach_lock
);
2793 if (!(lockres
->l_flags
& OCFS2_LOCK_FREEING
)
2798 spin_unlock(&dentry_attach_lock
);
2799 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2801 mlog(0, "extra_ref = %d\n", extra_ref
);
2804 * We have a process waiting on us in ocfs2_dentry_iput(),
2805 * which means we can't have any more outstanding
2806 * aliases. There's no need to do any more work.
2809 return UNBLOCK_CONTINUE
;
2811 spin_lock(&dentry_attach_lock
);
2813 dentry
= ocfs2_find_local_alias(dl
->dl_inode
,
2814 dl
->dl_parent_blkno
, 1);
2817 spin_unlock(&dentry_attach_lock
);
2819 mlog(0, "d_delete(%.*s);\n", dentry
->d_name
.len
,
2820 dentry
->d_name
.name
);
2823 * The following dcache calls may do an
2824 * iput(). Normally we don't want that from the
2825 * downconverting thread, but in this case it's ok
2826 * because the requesting node already has an
2827 * exclusive lock on the inode, so it can't be queued
2828 * for a downconvert.
2833 spin_lock(&dentry_attach_lock
);
2835 spin_unlock(&dentry_attach_lock
);
2838 * If we are the last holder of this dentry lock, there is no
2839 * reason to downconvert so skip straight to the unlock.
2841 if (dl
->dl_count
== 1)
2842 return UNBLOCK_STOP_POST
;
2844 return UNBLOCK_CONTINUE_POST
;
2847 void ocfs2_process_blocked_lock(struct ocfs2_super
*osb
,
2848 struct ocfs2_lock_res
*lockres
)
2851 struct ocfs2_unblock_ctl ctl
= {0, 0,};
2852 unsigned long flags
;
2854 /* Our reference to the lockres in this function can be
2855 * considered valid until we remove the OCFS2_LOCK_QUEUED
2861 BUG_ON(!lockres
->l_ops
);
2863 mlog(0, "lockres %s blocked.\n", lockres
->l_name
);
2865 /* Detect whether a lock has been marked as going away while
2866 * the vote thread was processing other things. A lock can
2867 * still be marked with OCFS2_LOCK_FREEING after this check,
2868 * but short circuiting here will still save us some
2870 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2871 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
)
2873 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2875 status
= ocfs2_unblock_lock(osb
, lockres
, &ctl
);
2879 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2881 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
|| !ctl
.requeue
) {
2882 lockres_clear_flags(lockres
, OCFS2_LOCK_QUEUED
);
2884 ocfs2_schedule_blocked_lock(osb
, lockres
);
2886 mlog(0, "lockres %s, requeue = %s.\n", lockres
->l_name
,
2887 ctl
.requeue
? "yes" : "no");
2888 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2890 if (ctl
.unblock_action
!= UNBLOCK_CONTINUE
2891 && lockres
->l_ops
->post_unlock
)
2892 lockres
->l_ops
->post_unlock(osb
, lockres
);
2897 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
2898 struct ocfs2_lock_res
*lockres
)
2902 assert_spin_locked(&lockres
->l_lock
);
2904 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
) {
2905 /* Do not schedule a lock for downconvert when it's on
2906 * the way to destruction - any nodes wanting access
2907 * to the resource will get it soon. */
2908 mlog(0, "Lockres %s won't be scheduled: flags 0x%lx\n",
2909 lockres
->l_name
, lockres
->l_flags
);
2913 lockres_or_flags(lockres
, OCFS2_LOCK_QUEUED
);
2915 spin_lock(&osb
->vote_task_lock
);
2916 if (list_empty(&lockres
->l_blocked_list
)) {
2917 list_add_tail(&lockres
->l_blocked_list
,
2918 &osb
->blocked_lock_list
);
2919 osb
->blocked_lock_count
++;
2921 spin_unlock(&osb
->vote_task_lock
);
2926 /* This aids in debugging situations where a bad LVB might be involved. */
2927 void ocfs2_dump_meta_lvb_info(u64 level
,
2928 const char *function
,
2930 struct ocfs2_lock_res
*lockres
)
2932 struct ocfs2_meta_lvb
*lvb
= (struct ocfs2_meta_lvb
*) lockres
->l_lksb
.lvb
;
2934 mlog(level
, "LVB information for %s (called from %s:%u):\n",
2935 lockres
->l_name
, function
, line
);
2936 mlog(level
, "version: %u, clusters: %u, generation: 0x%x\n",
2937 lvb
->lvb_version
, be32_to_cpu(lvb
->lvb_iclusters
),
2938 be32_to_cpu(lvb
->lvb_igeneration
));
2939 mlog(level
, "size: %llu, uid %u, gid %u, mode 0x%x\n",
2940 (unsigned long long)be64_to_cpu(lvb
->lvb_isize
),
2941 be32_to_cpu(lvb
->lvb_iuid
), be32_to_cpu(lvb
->lvb_igid
),
2942 be16_to_cpu(lvb
->lvb_imode
));
2943 mlog(level
, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
2944 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb
->lvb_inlink
),
2945 (long long)be64_to_cpu(lvb
->lvb_iatime_packed
),
2946 (long long)be64_to_cpu(lvb
->lvb_ictime_packed
),
2947 (long long)be64_to_cpu(lvb
->lvb_imtime_packed
),
2948 be32_to_cpu(lvb
->lvb_iattr
));