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 void 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 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
321 struct ocfs2_lock_res
*lockres
);
322 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
323 struct ocfs2_lock_res
*lockres
);
326 static void ocfs2_build_lock_name(enum ocfs2_lock_type type
,
335 BUG_ON(type
>= OCFS2_NUM_LOCK_TYPES
);
337 len
= snprintf(name
, OCFS2_LOCK_ID_MAX_LEN
, "%c%s%016llx%08x",
338 ocfs2_lock_type_char(type
), OCFS2_LOCK_ID_PAD
,
339 (long long)blkno
, generation
);
341 BUG_ON(len
!= (OCFS2_LOCK_ID_MAX_LEN
- 1));
343 mlog(0, "built lock resource with name: %s\n", name
);
348 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock
);
350 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res
*res
,
351 struct ocfs2_dlm_debug
*dlm_debug
)
353 mlog(0, "Add tracking for lockres %s\n", res
->l_name
);
355 spin_lock(&ocfs2_dlm_tracking_lock
);
356 list_add(&res
->l_debug_list
, &dlm_debug
->d_lockres_tracking
);
357 spin_unlock(&ocfs2_dlm_tracking_lock
);
360 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res
*res
)
362 spin_lock(&ocfs2_dlm_tracking_lock
);
363 if (!list_empty(&res
->l_debug_list
))
364 list_del_init(&res
->l_debug_list
);
365 spin_unlock(&ocfs2_dlm_tracking_lock
);
368 static void ocfs2_lock_res_init_common(struct ocfs2_super
*osb
,
369 struct ocfs2_lock_res
*res
,
370 enum ocfs2_lock_type type
,
371 struct ocfs2_lock_res_ops
*ops
,
378 res
->l_level
= DLM_LOCK_IV
;
379 res
->l_requested
= DLM_LOCK_IV
;
380 res
->l_blocking
= DLM_LOCK_IV
;
381 res
->l_action
= OCFS2_AST_INVALID
;
382 res
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
384 res
->l_flags
= OCFS2_LOCK_INITIALIZED
;
386 ocfs2_add_lockres_tracking(res
, osb
->osb_dlm_debug
);
389 void ocfs2_lock_res_init_once(struct ocfs2_lock_res
*res
)
391 /* This also clears out the lock status block */
392 memset(res
, 0, sizeof(struct ocfs2_lock_res
));
393 spin_lock_init(&res
->l_lock
);
394 init_waitqueue_head(&res
->l_event
);
395 INIT_LIST_HEAD(&res
->l_blocked_list
);
396 INIT_LIST_HEAD(&res
->l_mask_waiters
);
399 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res
*res
,
400 enum ocfs2_lock_type type
,
401 unsigned int generation
,
404 struct ocfs2_lock_res_ops
*ops
;
407 case OCFS2_LOCK_TYPE_RW
:
408 ops
= &ocfs2_inode_rw_lops
;
410 case OCFS2_LOCK_TYPE_META
:
411 ops
= &ocfs2_inode_inode_lops
;
413 case OCFS2_LOCK_TYPE_OPEN
:
414 ops
= &ocfs2_inode_open_lops
;
417 mlog_bug_on_msg(1, "type: %d\n", type
);
418 ops
= NULL
; /* thanks, gcc */
422 ocfs2_build_lock_name(type
, OCFS2_I(inode
)->ip_blkno
,
423 generation
, res
->l_name
);
424 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), res
, type
, ops
, inode
);
427 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
)
429 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
431 return OCFS2_SB(inode
->i_sb
);
434 static struct ocfs2_super
*ocfs2_get_file_osb(struct ocfs2_lock_res
*lockres
)
436 struct ocfs2_file_private
*fp
= lockres
->l_priv
;
438 return OCFS2_SB(fp
->fp_file
->f_mapping
->host
->i_sb
);
441 static __u64
ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res
*lockres
)
443 __be64 inode_blkno_be
;
445 memcpy(&inode_blkno_be
, &lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
],
448 return be64_to_cpu(inode_blkno_be
);
451 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
)
453 struct ocfs2_dentry_lock
*dl
= lockres
->l_priv
;
455 return OCFS2_SB(dl
->dl_inode
->i_sb
);
458 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock
*dl
,
459 u64 parent
, struct inode
*inode
)
462 u64 inode_blkno
= OCFS2_I(inode
)->ip_blkno
;
463 __be64 inode_blkno_be
= cpu_to_be64(inode_blkno
);
464 struct ocfs2_lock_res
*lockres
= &dl
->dl_lockres
;
466 ocfs2_lock_res_init_once(lockres
);
469 * Unfortunately, the standard lock naming scheme won't work
470 * here because we have two 16 byte values to use. Instead,
471 * we'll stuff the inode number as a binary value. We still
472 * want error prints to show something without garbling the
473 * display, so drop a null byte in there before the inode
474 * number. A future version of OCFS2 will likely use all
475 * binary lock names. The stringified names have been a
476 * tremendous aid in debugging, but now that the debugfs
477 * interface exists, we can mangle things there if need be.
479 * NOTE: We also drop the standard "pad" value (the total lock
480 * name size stays the same though - the last part is all
481 * zeros due to the memset in ocfs2_lock_res_init_once()
483 len
= snprintf(lockres
->l_name
, OCFS2_DENTRY_LOCK_INO_START
,
485 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY
),
488 BUG_ON(len
!= (OCFS2_DENTRY_LOCK_INO_START
- 1));
490 memcpy(&lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
], &inode_blkno_be
,
493 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
494 OCFS2_LOCK_TYPE_DENTRY
, &ocfs2_dentry_lops
,
498 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res
*res
,
499 struct ocfs2_super
*osb
)
501 /* Superblock lockres doesn't come from a slab so we call init
502 * once on it manually. */
503 ocfs2_lock_res_init_once(res
);
504 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER
, OCFS2_SUPER_BLOCK_BLKNO
,
506 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_SUPER
,
507 &ocfs2_super_lops
, osb
);
510 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res
*res
,
511 struct ocfs2_super
*osb
)
513 /* Rename lockres doesn't come from a slab so we call init
514 * once on it manually. */
515 ocfs2_lock_res_init_once(res
);
516 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME
, 0, 0, res
->l_name
);
517 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_RENAME
,
518 &ocfs2_rename_lops
, osb
);
521 void ocfs2_file_lock_res_init(struct ocfs2_lock_res
*lockres
,
522 struct ocfs2_file_private
*fp
)
524 struct inode
*inode
= fp
->fp_file
->f_mapping
->host
;
525 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
527 ocfs2_lock_res_init_once(lockres
);
528 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK
, oi
->ip_blkno
,
529 inode
->i_generation
, lockres
->l_name
);
530 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
531 OCFS2_LOCK_TYPE_FLOCK
, &ocfs2_flock_lops
,
533 lockres
->l_flags
|= OCFS2_LOCK_NOCACHE
;
536 void ocfs2_lock_res_free(struct ocfs2_lock_res
*res
)
540 if (!(res
->l_flags
& OCFS2_LOCK_INITIALIZED
))
543 ocfs2_remove_lockres_tracking(res
);
545 mlog_bug_on_msg(!list_empty(&res
->l_blocked_list
),
546 "Lockres %s is on the blocked list\n",
548 mlog_bug_on_msg(!list_empty(&res
->l_mask_waiters
),
549 "Lockres %s has mask waiters pending\n",
551 mlog_bug_on_msg(spin_is_locked(&res
->l_lock
),
552 "Lockres %s is locked\n",
554 mlog_bug_on_msg(res
->l_ro_holders
,
555 "Lockres %s has %u ro holders\n",
556 res
->l_name
, res
->l_ro_holders
);
557 mlog_bug_on_msg(res
->l_ex_holders
,
558 "Lockres %s has %u ex holders\n",
559 res
->l_name
, res
->l_ex_holders
);
561 /* Need to clear out the lock status block for the dlm */
562 memset(&res
->l_lksb
, 0, sizeof(res
->l_lksb
));
568 static inline void ocfs2_inc_holders(struct ocfs2_lock_res
*lockres
,
577 lockres
->l_ex_holders
++;
580 lockres
->l_ro_holders
++;
589 static inline void ocfs2_dec_holders(struct ocfs2_lock_res
*lockres
,
598 BUG_ON(!lockres
->l_ex_holders
);
599 lockres
->l_ex_holders
--;
602 BUG_ON(!lockres
->l_ro_holders
);
603 lockres
->l_ro_holders
--;
611 /* WARNING: This function lives in a world where the only three lock
612 * levels are EX, PR, and NL. It *will* have to be adjusted when more
613 * lock types are added. */
614 static inline int ocfs2_highest_compat_lock_level(int level
)
616 int new_level
= DLM_LOCK_EX
;
618 if (level
== DLM_LOCK_EX
)
619 new_level
= DLM_LOCK_NL
;
620 else if (level
== DLM_LOCK_PR
)
621 new_level
= DLM_LOCK_PR
;
625 static void lockres_set_flags(struct ocfs2_lock_res
*lockres
,
626 unsigned long newflags
)
628 struct ocfs2_mask_waiter
*mw
, *tmp
;
630 assert_spin_locked(&lockres
->l_lock
);
632 lockres
->l_flags
= newflags
;
634 list_for_each_entry_safe(mw
, tmp
, &lockres
->l_mask_waiters
, mw_item
) {
635 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
638 list_del_init(&mw
->mw_item
);
640 complete(&mw
->mw_complete
);
643 static void lockres_or_flags(struct ocfs2_lock_res
*lockres
, unsigned long or)
645 lockres_set_flags(lockres
, lockres
->l_flags
| or);
647 static void lockres_clear_flags(struct ocfs2_lock_res
*lockres
,
650 lockres_set_flags(lockres
, lockres
->l_flags
& ~clear
);
653 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
)
657 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
658 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
659 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
660 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
662 lockres
->l_level
= lockres
->l_requested
;
663 if (lockres
->l_level
<=
664 ocfs2_highest_compat_lock_level(lockres
->l_blocking
)) {
665 lockres
->l_blocking
= DLM_LOCK_NL
;
666 lockres_clear_flags(lockres
, OCFS2_LOCK_BLOCKED
);
668 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
673 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
)
677 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
678 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
680 /* Convert from RO to EX doesn't really need anything as our
681 * information is already up to data. Convert from NL to
682 * *anything* however should mark ourselves as needing an
684 if (lockres
->l_level
== DLM_LOCK_NL
&&
685 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
686 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
688 lockres
->l_level
= lockres
->l_requested
;
689 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
694 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
)
698 BUG_ON((!(lockres
->l_flags
& OCFS2_LOCK_BUSY
)));
699 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
701 if (lockres
->l_requested
> DLM_LOCK_NL
&&
702 !(lockres
->l_flags
& OCFS2_LOCK_LOCAL
) &&
703 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
704 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
706 lockres
->l_level
= lockres
->l_requested
;
707 lockres_or_flags(lockres
, OCFS2_LOCK_ATTACHED
);
708 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
713 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
,
716 int needs_downconvert
= 0;
719 assert_spin_locked(&lockres
->l_lock
);
721 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
723 if (level
> lockres
->l_blocking
) {
724 /* only schedule a downconvert if we haven't already scheduled
725 * one that goes low enough to satisfy the level we're
726 * blocking. this also catches the case where we get
728 if (ocfs2_highest_compat_lock_level(level
) <
729 ocfs2_highest_compat_lock_level(lockres
->l_blocking
))
730 needs_downconvert
= 1;
732 lockres
->l_blocking
= level
;
735 mlog_exit(needs_downconvert
);
736 return needs_downconvert
;
739 static void ocfs2_blocking_ast(void *opaque
, int level
)
741 struct ocfs2_lock_res
*lockres
= opaque
;
742 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
743 int needs_downconvert
;
746 BUG_ON(level
<= DLM_LOCK_NL
);
748 mlog(0, "BAST fired for lockres %s, blocking %d, level %d type %s\n",
749 lockres
->l_name
, level
, lockres
->l_level
,
750 ocfs2_lock_type_string(lockres
->l_type
));
753 * We can skip the bast for locks which don't enable caching -
754 * they'll be dropped at the earliest possible time anyway.
756 if (lockres
->l_flags
& OCFS2_LOCK_NOCACHE
)
759 spin_lock_irqsave(&lockres
->l_lock
, flags
);
760 needs_downconvert
= ocfs2_generic_handle_bast(lockres
, level
);
761 if (needs_downconvert
)
762 ocfs2_schedule_blocked_lock(osb
, lockres
);
763 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
765 wake_up(&lockres
->l_event
);
767 ocfs2_wake_downconvert_thread(osb
);
770 static void ocfs2_locking_ast(void *opaque
)
772 struct ocfs2_lock_res
*lockres
= opaque
;
775 spin_lock_irqsave(&lockres
->l_lock
, flags
);
777 if (ocfs2_dlm_lock_status(&lockres
->l_lksb
)) {
778 mlog(ML_ERROR
, "lockres %s: lksb status value of %d!\n",
780 ocfs2_dlm_lock_status(&lockres
->l_lksb
));
781 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
785 switch(lockres
->l_action
) {
786 case OCFS2_AST_ATTACH
:
787 ocfs2_generic_handle_attach_action(lockres
);
788 lockres_clear_flags(lockres
, OCFS2_LOCK_LOCAL
);
790 case OCFS2_AST_CONVERT
:
791 ocfs2_generic_handle_convert_action(lockres
);
793 case OCFS2_AST_DOWNCONVERT
:
794 ocfs2_generic_handle_downconvert_action(lockres
);
797 mlog(ML_ERROR
, "lockres %s: ast fired with invalid action: %u "
798 "lockres flags = 0x%lx, unlock action: %u\n",
799 lockres
->l_name
, lockres
->l_action
, lockres
->l_flags
,
800 lockres
->l_unlock_action
);
804 /* set it to something invalid so if we get called again we
806 lockres
->l_action
= OCFS2_AST_INVALID
;
808 wake_up(&lockres
->l_event
);
809 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
812 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
818 spin_lock_irqsave(&lockres
->l_lock
, flags
);
819 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
821 lockres
->l_action
= OCFS2_AST_INVALID
;
823 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
824 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
826 wake_up(&lockres
->l_event
);
830 /* Note: If we detect another process working on the lock (i.e.,
831 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
832 * to do the right thing in that case.
834 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
835 struct ocfs2_lock_res
*lockres
,
844 mlog(0, "lock %s, level = %d, flags = %u\n", lockres
->l_name
, level
,
847 spin_lock_irqsave(&lockres
->l_lock
, flags
);
848 if ((lockres
->l_flags
& OCFS2_LOCK_ATTACHED
) ||
849 (lockres
->l_flags
& OCFS2_LOCK_BUSY
)) {
850 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
854 lockres
->l_action
= OCFS2_AST_ATTACH
;
855 lockres
->l_requested
= level
;
856 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
857 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
859 ret
= ocfs2_dlm_lock(osb
->cconn
,
864 OCFS2_LOCK_ID_MAX_LEN
- 1,
867 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
868 ocfs2_recover_from_dlm_error(lockres
, 1);
871 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres
->l_name
);
878 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res
*lockres
,
884 spin_lock_irqsave(&lockres
->l_lock
, flags
);
885 ret
= lockres
->l_flags
& flag
;
886 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
891 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res
*lockres
)
894 wait_event(lockres
->l_event
,
895 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_BUSY
));
898 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res
*lockres
)
901 wait_event(lockres
->l_event
,
902 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_REFRESHING
));
905 /* predict what lock level we'll be dropping down to on behalf
906 * of another node, and return true if the currently wanted
907 * level will be compatible with it. */
908 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
911 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
913 return wanted
<= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
916 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter
*mw
)
918 INIT_LIST_HEAD(&mw
->mw_item
);
919 init_completion(&mw
->mw_complete
);
922 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter
*mw
)
924 wait_for_completion(&mw
->mw_complete
);
925 /* Re-arm the completion in case we want to wait on it again */
926 INIT_COMPLETION(mw
->mw_complete
);
927 return mw
->mw_status
;
930 static void lockres_add_mask_waiter(struct ocfs2_lock_res
*lockres
,
931 struct ocfs2_mask_waiter
*mw
,
935 BUG_ON(!list_empty(&mw
->mw_item
));
937 assert_spin_locked(&lockres
->l_lock
);
939 list_add_tail(&mw
->mw_item
, &lockres
->l_mask_waiters
);
944 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
945 * if the mask still hadn't reached its goal */
946 static int lockres_remove_mask_waiter(struct ocfs2_lock_res
*lockres
,
947 struct ocfs2_mask_waiter
*mw
)
952 spin_lock_irqsave(&lockres
->l_lock
, flags
);
953 if (!list_empty(&mw
->mw_item
)) {
954 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
957 list_del_init(&mw
->mw_item
);
958 init_completion(&mw
->mw_complete
);
960 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
966 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter
*mw
,
967 struct ocfs2_lock_res
*lockres
)
971 ret
= wait_for_completion_interruptible(&mw
->mw_complete
);
973 lockres_remove_mask_waiter(lockres
, mw
);
976 /* Re-arm the completion in case we want to wait on it again */
977 INIT_COMPLETION(mw
->mw_complete
);
981 static int ocfs2_cluster_lock(struct ocfs2_super
*osb
,
982 struct ocfs2_lock_res
*lockres
,
987 struct ocfs2_mask_waiter mw
;
988 int wait
, catch_signals
= !(osb
->s_mount_opt
& OCFS2_MOUNT_NOINTR
);
989 int ret
= 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
994 ocfs2_init_mask_waiter(&mw
);
996 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
997 lkm_flags
|= DLM_LKF_VALBLK
;
1002 if (catch_signals
&& signal_pending(current
)) {
1007 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1009 mlog_bug_on_msg(lockres
->l_flags
& OCFS2_LOCK_FREEING
,
1010 "Cluster lock called on freeing lockres %s! flags "
1011 "0x%lx\n", lockres
->l_name
, lockres
->l_flags
);
1013 /* We only compare against the currently granted level
1014 * here. If the lock is blocked waiting on a downconvert,
1015 * we'll get caught below. */
1016 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
&&
1017 level
> lockres
->l_level
) {
1018 /* is someone sitting in dlm_lock? If so, wait on
1020 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1025 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
&&
1026 !ocfs2_may_continue_on_blocked_lock(lockres
, level
)) {
1027 /* is the lock is currently blocked on behalf of
1029 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BLOCKED
, 0);
1034 if (level
> lockres
->l_level
) {
1035 if (lockres
->l_action
!= OCFS2_AST_INVALID
)
1036 mlog(ML_ERROR
, "lockres %s has action %u pending\n",
1037 lockres
->l_name
, lockres
->l_action
);
1039 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1040 lockres
->l_action
= OCFS2_AST_ATTACH
;
1041 lkm_flags
&= ~DLM_LKF_CONVERT
;
1043 lockres
->l_action
= OCFS2_AST_CONVERT
;
1044 lkm_flags
|= DLM_LKF_CONVERT
;
1047 lockres
->l_requested
= level
;
1048 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1049 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1051 BUG_ON(level
== DLM_LOCK_IV
);
1052 BUG_ON(level
== DLM_LOCK_NL
);
1054 mlog(0, "lock %s, convert from %d to level = %d\n",
1055 lockres
->l_name
, lockres
->l_level
, level
);
1057 /* call dlm_lock to upgrade lock now */
1058 ret
= ocfs2_dlm_lock(osb
->cconn
,
1063 OCFS2_LOCK_ID_MAX_LEN
- 1,
1066 if (!(lkm_flags
& DLM_LKF_NOQUEUE
) ||
1068 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1071 ocfs2_recover_from_dlm_error(lockres
, 1);
1075 mlog(0, "lock %s, successfull return from ocfs2_dlm_lock\n",
1078 /* At this point we've gone inside the dlm and need to
1079 * complete our work regardless. */
1082 /* wait for busy to clear and carry on */
1086 /* Ok, if we get here then we're good to go. */
1087 ocfs2_inc_holders(lockres
, level
);
1091 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1094 * This is helping work around a lock inversion between the page lock
1095 * and dlm locks. One path holds the page lock while calling aops
1096 * which block acquiring dlm locks. The voting thread holds dlm
1097 * locks while acquiring page locks while down converting data locks.
1098 * This block is helping an aop path notice the inversion and back
1099 * off to unlock its page lock before trying the dlm lock again.
1101 if (wait
&& arg_flags
& OCFS2_LOCK_NONBLOCK
&&
1102 mw
.mw_mask
& (OCFS2_LOCK_BUSY
|OCFS2_LOCK_BLOCKED
)) {
1104 if (lockres_remove_mask_waiter(lockres
, &mw
))
1110 ret
= ocfs2_wait_for_mask(&mw
);
1120 static void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
1121 struct ocfs2_lock_res
*lockres
,
1124 unsigned long flags
;
1127 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1128 ocfs2_dec_holders(lockres
, level
);
1129 ocfs2_downconvert_on_unlock(osb
, lockres
);
1130 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1134 static int ocfs2_create_new_lock(struct ocfs2_super
*osb
,
1135 struct ocfs2_lock_res
*lockres
,
1139 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1140 unsigned long flags
;
1141 u32 lkm_flags
= local
? DLM_LKF_LOCAL
: 0;
1143 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1144 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
1145 lockres_or_flags(lockres
, OCFS2_LOCK_LOCAL
);
1146 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1148 return ocfs2_lock_create(osb
, lockres
, level
, lkm_flags
);
1151 /* Grants us an EX lock on the data and metadata resources, skipping
1152 * the normal cluster directory lookup. Use this ONLY on newly created
1153 * inodes which other nodes can't possibly see, and which haven't been
1154 * hashed in the inode hash yet. This can give us a good performance
1155 * increase as it'll skip the network broadcast normally associated
1156 * with creating a new lock resource. */
1157 int ocfs2_create_new_inode_locks(struct inode
*inode
)
1160 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1163 BUG_ON(!ocfs2_inode_is_new(inode
));
1167 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1169 /* NOTE: That we don't increment any of the holder counts, nor
1170 * do we add anything to a journal handle. Since this is
1171 * supposed to be a new inode which the cluster doesn't know
1172 * about yet, there is no need to. As far as the LVB handling
1173 * is concerned, this is basically like acquiring an EX lock
1174 * on a resource which has an invalid one -- we'll set it
1175 * valid when we release the EX. */
1177 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_rw_lockres
, 1, 1);
1184 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1185 * don't use a generation in their lock names.
1187 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_inode_lockres
, 1, 0);
1193 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_open_lockres
, 0, 0);
1204 int ocfs2_rw_lock(struct inode
*inode
, int write
)
1207 struct ocfs2_lock_res
*lockres
;
1208 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1214 mlog(0, "inode %llu take %s RW lock\n",
1215 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1216 write
? "EXMODE" : "PRMODE");
1218 if (ocfs2_mount_local(osb
))
1221 lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1223 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1225 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
, level
, 0,
1234 void ocfs2_rw_unlock(struct inode
*inode
, int write
)
1236 int level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1237 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1238 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1242 mlog(0, "inode %llu drop %s RW lock\n",
1243 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1244 write
? "EXMODE" : "PRMODE");
1246 if (!ocfs2_mount_local(osb
))
1247 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
1253 * ocfs2_open_lock always get PR mode lock.
1255 int ocfs2_open_lock(struct inode
*inode
)
1258 struct ocfs2_lock_res
*lockres
;
1259 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1265 mlog(0, "inode %llu take PRMODE open lock\n",
1266 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1268 if (ocfs2_mount_local(osb
))
1271 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1273 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
,
1283 int ocfs2_try_open_lock(struct inode
*inode
, int write
)
1285 int status
= 0, level
;
1286 struct ocfs2_lock_res
*lockres
;
1287 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1293 mlog(0, "inode %llu try to take %s open lock\n",
1294 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1295 write
? "EXMODE" : "PRMODE");
1297 if (ocfs2_mount_local(osb
))
1300 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1302 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1305 * The file system may already holding a PRMODE/EXMODE open lock.
1306 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1307 * other nodes and the -EAGAIN will indicate to the caller that
1308 * this inode is still in use.
1310 status
= ocfs2_cluster_lock(OCFS2_SB(inode
->i_sb
), lockres
,
1311 level
, DLM_LKF_NOQUEUE
, 0);
1319 * ocfs2_open_unlock unlock PR and EX mode open locks.
1321 void ocfs2_open_unlock(struct inode
*inode
)
1323 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1324 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1328 mlog(0, "inode %llu drop open lock\n",
1329 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1331 if (ocfs2_mount_local(osb
))
1334 if(lockres
->l_ro_holders
)
1335 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
,
1337 if(lockres
->l_ex_holders
)
1338 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
,
1345 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res
*lockres
,
1349 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
1350 unsigned long flags
;
1351 struct ocfs2_mask_waiter mw
;
1353 ocfs2_init_mask_waiter(&mw
);
1356 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1357 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
1358 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
1360 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1361 ret
= ocfs2_cancel_convert(osb
, lockres
);
1368 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1369 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1371 ocfs2_wait_for_mask(&mw
);
1377 * We may still have gotten the lock, in which case there's no
1378 * point to restarting the syscall.
1380 if (lockres
->l_level
== level
)
1383 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret
,
1384 lockres
->l_flags
, lockres
->l_level
, lockres
->l_action
);
1386 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1393 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1394 * flock() calls. The locking approach this requires is sufficiently
1395 * different from all other cluster lock types that we implement a
1396 * seperate path to the "low-level" dlm calls. In particular:
1398 * - No optimization of lock levels is done - we take at exactly
1399 * what's been requested.
1401 * - No lock caching is employed. We immediately downconvert to
1402 * no-lock at unlock time. This also means flock locks never go on
1403 * the blocking list).
1405 * - Since userspace can trivially deadlock itself with flock, we make
1406 * sure to allow cancellation of a misbehaving applications flock()
1409 * - Access to any flock lockres doesn't require concurrency, so we
1410 * can simplify the code by requiring the caller to guarantee
1411 * serialization of dlmglue flock calls.
1413 int ocfs2_file_lock(struct file
*file
, int ex
, int trylock
)
1415 int ret
, level
= ex
? LKM_EXMODE
: LKM_PRMODE
;
1416 unsigned int lkm_flags
= trylock
? LKM_NOQUEUE
: 0;
1417 unsigned long flags
;
1418 struct ocfs2_file_private
*fp
= file
->private_data
;
1419 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
1420 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
1421 struct ocfs2_mask_waiter mw
;
1423 ocfs2_init_mask_waiter(&mw
);
1425 if ((lockres
->l_flags
& OCFS2_LOCK_BUSY
) ||
1426 (lockres
->l_level
> DLM_LOCK_NL
)) {
1428 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1429 "level: %u\n", lockres
->l_name
, lockres
->l_flags
,
1434 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1435 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1436 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1437 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1440 * Get the lock at NLMODE to start - that way we
1441 * can cancel the upconvert request if need be.
1443 ret
= ocfs2_lock_create(osb
, lockres
, LKM_NLMODE
, 0);
1449 ret
= ocfs2_wait_for_mask(&mw
);
1454 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1457 lockres
->l_action
= OCFS2_AST_CONVERT
;
1458 lkm_flags
|= LKM_CONVERT
;
1459 lockres
->l_requested
= level
;
1460 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1462 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1463 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1465 ret
= ocfs2_dlm_lock(osb
->cconn
, level
, &lockres
->l_lksb
, lkm_flags
,
1466 lockres
->l_name
, OCFS2_LOCK_ID_MAX_LEN
- 1,
1469 if (!trylock
|| (ret
!= -EAGAIN
)) {
1470 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
1474 ocfs2_recover_from_dlm_error(lockres
, 1);
1475 lockres_remove_mask_waiter(lockres
, &mw
);
1479 ret
= ocfs2_wait_for_mask_interruptible(&mw
, lockres
);
1480 if (ret
== -ERESTARTSYS
) {
1482 * Userspace can cause deadlock itself with
1483 * flock(). Current behavior locally is to allow the
1484 * deadlock, but abort the system call if a signal is
1485 * received. We follow this example, otherwise a
1486 * poorly written program could sit in kernel until
1489 * Handling this is a bit more complicated for Ocfs2
1490 * though. We can't exit this function with an
1491 * outstanding lock request, so a cancel convert is
1492 * required. We intentionally overwrite 'ret' - if the
1493 * cancel fails and the lock was granted, it's easier
1494 * to just bubble sucess back up to the user.
1496 ret
= ocfs2_flock_handle_signal(lockres
, level
);
1501 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
1502 lockres
->l_name
, ex
, trylock
, ret
);
1506 void ocfs2_file_unlock(struct file
*file
)
1509 unsigned long flags
;
1510 struct ocfs2_file_private
*fp
= file
->private_data
;
1511 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
1512 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
1513 struct ocfs2_mask_waiter mw
;
1515 ocfs2_init_mask_waiter(&mw
);
1517 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
))
1520 if (lockres
->l_level
== LKM_NLMODE
)
1523 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
1524 lockres
->l_name
, lockres
->l_flags
, lockres
->l_level
,
1527 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1529 * Fake a blocking ast for the downconvert code.
1531 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
1532 lockres
->l_blocking
= DLM_LOCK_EX
;
1534 ocfs2_prepare_downconvert(lockres
, LKM_NLMODE
);
1535 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1536 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1538 ret
= ocfs2_downconvert_lock(osb
, lockres
, LKM_NLMODE
, 0);
1544 ret
= ocfs2_wait_for_mask(&mw
);
1549 static void ocfs2_downconvert_on_unlock(struct ocfs2_super
*osb
,
1550 struct ocfs2_lock_res
*lockres
)
1556 /* If we know that another node is waiting on our lock, kick
1557 * the downconvert thread * pre-emptively when we reach a release
1559 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
) {
1560 switch(lockres
->l_blocking
) {
1562 if (!lockres
->l_ex_holders
&& !lockres
->l_ro_holders
)
1566 if (!lockres
->l_ex_holders
)
1575 ocfs2_wake_downconvert_thread(osb
);
1580 #define OCFS2_SEC_BITS 34
1581 #define OCFS2_SEC_SHIFT (64 - 34)
1582 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
1584 /* LVB only has room for 64 bits of time here so we pack it for
1586 static u64
ocfs2_pack_timespec(struct timespec
*spec
)
1589 u64 sec
= spec
->tv_sec
;
1590 u32 nsec
= spec
->tv_nsec
;
1592 res
= (sec
<< OCFS2_SEC_SHIFT
) | (nsec
& OCFS2_NSEC_MASK
);
1597 /* Call this with the lockres locked. I am reasonably sure we don't
1598 * need ip_lock in this function as anyone who would be changing those
1599 * values is supposed to be blocked in ocfs2_inode_lock right now. */
1600 static void __ocfs2_stuff_meta_lvb(struct inode
*inode
)
1602 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1603 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1604 struct ocfs2_meta_lvb
*lvb
;
1608 lvb
= (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1611 * Invalidate the LVB of a deleted inode - this way other
1612 * nodes are forced to go to disk and discover the new inode
1615 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1616 lvb
->lvb_version
= 0;
1620 lvb
->lvb_version
= OCFS2_LVB_VERSION
;
1621 lvb
->lvb_isize
= cpu_to_be64(i_size_read(inode
));
1622 lvb
->lvb_iclusters
= cpu_to_be32(oi
->ip_clusters
);
1623 lvb
->lvb_iuid
= cpu_to_be32(inode
->i_uid
);
1624 lvb
->lvb_igid
= cpu_to_be32(inode
->i_gid
);
1625 lvb
->lvb_imode
= cpu_to_be16(inode
->i_mode
);
1626 lvb
->lvb_inlink
= cpu_to_be16(inode
->i_nlink
);
1627 lvb
->lvb_iatime_packed
=
1628 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_atime
));
1629 lvb
->lvb_ictime_packed
=
1630 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_ctime
));
1631 lvb
->lvb_imtime_packed
=
1632 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_mtime
));
1633 lvb
->lvb_iattr
= cpu_to_be32(oi
->ip_attr
);
1634 lvb
->lvb_idynfeatures
= cpu_to_be16(oi
->ip_dyn_features
);
1635 lvb
->lvb_igeneration
= cpu_to_be32(inode
->i_generation
);
1638 mlog_meta_lvb(0, lockres
);
1643 static void ocfs2_unpack_timespec(struct timespec
*spec
,
1646 spec
->tv_sec
= packed_time
>> OCFS2_SEC_SHIFT
;
1647 spec
->tv_nsec
= packed_time
& OCFS2_NSEC_MASK
;
1650 static void ocfs2_refresh_inode_from_lvb(struct inode
*inode
)
1652 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1653 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1654 struct ocfs2_meta_lvb
*lvb
;
1658 mlog_meta_lvb(0, lockres
);
1660 lvb
= (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1662 /* We're safe here without the lockres lock... */
1663 spin_lock(&oi
->ip_lock
);
1664 oi
->ip_clusters
= be32_to_cpu(lvb
->lvb_iclusters
);
1665 i_size_write(inode
, be64_to_cpu(lvb
->lvb_isize
));
1667 oi
->ip_attr
= be32_to_cpu(lvb
->lvb_iattr
);
1668 oi
->ip_dyn_features
= be16_to_cpu(lvb
->lvb_idynfeatures
);
1669 ocfs2_set_inode_flags(inode
);
1671 /* fast-symlinks are a special case */
1672 if (S_ISLNK(inode
->i_mode
) && !oi
->ip_clusters
)
1673 inode
->i_blocks
= 0;
1675 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
1677 inode
->i_uid
= be32_to_cpu(lvb
->lvb_iuid
);
1678 inode
->i_gid
= be32_to_cpu(lvb
->lvb_igid
);
1679 inode
->i_mode
= be16_to_cpu(lvb
->lvb_imode
);
1680 inode
->i_nlink
= be16_to_cpu(lvb
->lvb_inlink
);
1681 ocfs2_unpack_timespec(&inode
->i_atime
,
1682 be64_to_cpu(lvb
->lvb_iatime_packed
));
1683 ocfs2_unpack_timespec(&inode
->i_mtime
,
1684 be64_to_cpu(lvb
->lvb_imtime_packed
));
1685 ocfs2_unpack_timespec(&inode
->i_ctime
,
1686 be64_to_cpu(lvb
->lvb_ictime_packed
));
1687 spin_unlock(&oi
->ip_lock
);
1692 static inline int ocfs2_meta_lvb_is_trustable(struct inode
*inode
,
1693 struct ocfs2_lock_res
*lockres
)
1695 struct ocfs2_meta_lvb
*lvb
=
1696 (struct ocfs2_meta_lvb
*)ocfs2_dlm_lvb(&lockres
->l_lksb
);
1698 if (lvb
->lvb_version
== OCFS2_LVB_VERSION
1699 && be32_to_cpu(lvb
->lvb_igeneration
) == inode
->i_generation
)
1704 /* Determine whether a lock resource needs to be refreshed, and
1705 * arbitrate who gets to refresh it.
1707 * 0 means no refresh needed.
1709 * > 0 means you need to refresh this and you MUST call
1710 * ocfs2_complete_lock_res_refresh afterwards. */
1711 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res
*lockres
)
1713 unsigned long flags
;
1719 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1720 if (!(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
)) {
1721 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1725 if (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
) {
1726 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1728 ocfs2_wait_on_refreshing_lock(lockres
);
1732 /* Ok, I'll be the one to refresh this lock. */
1733 lockres_or_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1734 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1742 /* If status is non zero, I'll mark it as not being in refresh
1743 * anymroe, but i won't clear the needs refresh flag. */
1744 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res
*lockres
,
1747 unsigned long flags
;
1750 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1751 lockres_clear_flags(lockres
, OCFS2_LOCK_REFRESHING
);
1753 lockres_clear_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
1754 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1756 wake_up(&lockres
->l_event
);
1761 /* may or may not return a bh if it went to disk. */
1762 static int ocfs2_inode_lock_update(struct inode
*inode
,
1763 struct buffer_head
**bh
)
1766 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
1767 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
1768 struct ocfs2_dinode
*fe
;
1769 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1773 if (ocfs2_mount_local(osb
))
1776 spin_lock(&oi
->ip_lock
);
1777 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
1778 mlog(0, "Orphaned inode %llu was deleted while we "
1779 "were waiting on a lock. ip_flags = 0x%x\n",
1780 (unsigned long long)oi
->ip_blkno
, oi
->ip_flags
);
1781 spin_unlock(&oi
->ip_lock
);
1785 spin_unlock(&oi
->ip_lock
);
1787 if (!ocfs2_should_refresh_lock_res(lockres
))
1790 /* This will discard any caching information we might have had
1791 * for the inode metadata. */
1792 ocfs2_metadata_cache_purge(inode
);
1794 ocfs2_extent_map_trunc(inode
, 0);
1796 if (ocfs2_meta_lvb_is_trustable(inode
, lockres
)) {
1797 mlog(0, "Trusting LVB on inode %llu\n",
1798 (unsigned long long)oi
->ip_blkno
);
1799 ocfs2_refresh_inode_from_lvb(inode
);
1801 /* Boo, we have to go to disk. */
1802 /* read bh, cast, ocfs2_refresh_inode */
1803 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
), oi
->ip_blkno
,
1804 bh
, OCFS2_BH_CACHED
, inode
);
1809 fe
= (struct ocfs2_dinode
*) (*bh
)->b_data
;
1811 /* This is a good chance to make sure we're not
1812 * locking an invalid object.
1814 * We bug on a stale inode here because we checked
1815 * above whether it was wiped from disk. The wiping
1816 * node provides a guarantee that we receive that
1817 * message and can mark the inode before dropping any
1818 * locks associated with it. */
1819 if (!OCFS2_IS_VALID_DINODE(fe
)) {
1820 OCFS2_RO_ON_INVALID_DINODE(inode
->i_sb
, fe
);
1824 mlog_bug_on_msg(inode
->i_generation
!=
1825 le32_to_cpu(fe
->i_generation
),
1826 "Invalid dinode %llu disk generation: %u "
1827 "inode->i_generation: %u\n",
1828 (unsigned long long)oi
->ip_blkno
,
1829 le32_to_cpu(fe
->i_generation
),
1830 inode
->i_generation
);
1831 mlog_bug_on_msg(le64_to_cpu(fe
->i_dtime
) ||
1832 !(fe
->i_flags
& cpu_to_le32(OCFS2_VALID_FL
)),
1833 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
1834 (unsigned long long)oi
->ip_blkno
,
1835 (unsigned long long)le64_to_cpu(fe
->i_dtime
),
1836 le32_to_cpu(fe
->i_flags
));
1838 ocfs2_refresh_inode(inode
, fe
);
1843 ocfs2_complete_lock_res_refresh(lockres
, status
);
1849 static int ocfs2_assign_bh(struct inode
*inode
,
1850 struct buffer_head
**ret_bh
,
1851 struct buffer_head
*passed_bh
)
1856 /* Ok, the update went to disk for us, use the
1858 *ret_bh
= passed_bh
;
1864 status
= ocfs2_read_block(OCFS2_SB(inode
->i_sb
),
1865 OCFS2_I(inode
)->ip_blkno
,
1876 * returns < 0 error if the callback will never be called, otherwise
1877 * the result of the lock will be communicated via the callback.
1879 int ocfs2_inode_lock_full(struct inode
*inode
,
1880 struct buffer_head
**ret_bh
,
1884 int status
, level
, acquired
;
1886 struct ocfs2_lock_res
*lockres
= NULL
;
1887 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1888 struct buffer_head
*local_bh
= NULL
;
1894 mlog(0, "inode %llu, take %s META lock\n",
1895 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1896 ex
? "EXMODE" : "PRMODE");
1900 /* We'll allow faking a readonly metadata lock for
1902 if (ocfs2_is_hard_readonly(osb
)) {
1908 if (ocfs2_mount_local(osb
))
1911 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
1912 ocfs2_wait_for_recovery(osb
);
1914 lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
1915 level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1917 if (arg_flags
& OCFS2_META_LOCK_NOQUEUE
)
1918 dlm_flags
|= DLM_LKF_NOQUEUE
;
1920 status
= ocfs2_cluster_lock(osb
, lockres
, level
, dlm_flags
, arg_flags
);
1922 if (status
!= -EAGAIN
&& status
!= -EIOCBRETRY
)
1927 /* Notify the error cleanup path to drop the cluster lock. */
1930 /* We wait twice because a node may have died while we were in
1931 * the lower dlm layers. The second time though, we've
1932 * committed to owning this lock so we don't allow signals to
1933 * abort the operation. */
1934 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
1935 ocfs2_wait_for_recovery(osb
);
1939 * We only see this flag if we're being called from
1940 * ocfs2_read_locked_inode(). It means we're locking an inode
1941 * which hasn't been populated yet, so clear the refresh flag
1942 * and let the caller handle it.
1944 if (inode
->i_state
& I_NEW
) {
1947 ocfs2_complete_lock_res_refresh(lockres
, 0);
1951 /* This is fun. The caller may want a bh back, or it may
1952 * not. ocfs2_inode_lock_update definitely wants one in, but
1953 * may or may not read one, depending on what's in the
1954 * LVB. The result of all of this is that we've *only* gone to
1955 * disk if we have to, so the complexity is worthwhile. */
1956 status
= ocfs2_inode_lock_update(inode
, &local_bh
);
1958 if (status
!= -ENOENT
)
1964 status
= ocfs2_assign_bh(inode
, ret_bh
, local_bh
);
1973 if (ret_bh
&& (*ret_bh
)) {
1978 ocfs2_inode_unlock(inode
, ex
);
1989 * This is working around a lock inversion between tasks acquiring DLM
1990 * locks while holding a page lock and the downconvert thread which
1991 * blocks dlm lock acquiry while acquiring page locks.
1993 * ** These _with_page variantes are only intended to be called from aop
1994 * methods that hold page locks and return a very specific *positive* error
1995 * code that aop methods pass up to the VFS -- test for errors with != 0. **
1997 * The DLM is called such that it returns -EAGAIN if it would have
1998 * blocked waiting for the downconvert thread. In that case we unlock
1999 * our page so the downconvert thread can make progress. Once we've
2000 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2001 * that called us can bubble that back up into the VFS who will then
2002 * immediately retry the aop call.
2004 * We do a blocking lock and immediate unlock before returning, though, so that
2005 * the lock has a great chance of being cached on this node by the time the VFS
2006 * calls back to retry the aop. This has a potential to livelock as nodes
2007 * ping locks back and forth, but that's a risk we're willing to take to avoid
2008 * the lock inversion simply.
2010 int ocfs2_inode_lock_with_page(struct inode
*inode
,
2011 struct buffer_head
**ret_bh
,
2017 ret
= ocfs2_inode_lock_full(inode
, ret_bh
, ex
, OCFS2_LOCK_NONBLOCK
);
2018 if (ret
== -EAGAIN
) {
2020 if (ocfs2_inode_lock(inode
, ret_bh
, ex
) == 0)
2021 ocfs2_inode_unlock(inode
, ex
);
2022 ret
= AOP_TRUNCATED_PAGE
;
2028 int ocfs2_inode_lock_atime(struct inode
*inode
,
2029 struct vfsmount
*vfsmnt
,
2035 ret
= ocfs2_inode_lock(inode
, NULL
, 0);
2042 * If we should update atime, we will get EX lock,
2043 * otherwise we just get PR lock.
2045 if (ocfs2_should_update_atime(inode
, vfsmnt
)) {
2046 struct buffer_head
*bh
= NULL
;
2048 ocfs2_inode_unlock(inode
, 0);
2049 ret
= ocfs2_inode_lock(inode
, &bh
, 1);
2055 if (ocfs2_should_update_atime(inode
, vfsmnt
))
2056 ocfs2_update_inode_atime(inode
, bh
);
2066 void ocfs2_inode_unlock(struct inode
*inode
,
2069 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2070 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2071 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2075 mlog(0, "inode %llu drop %s META lock\n",
2076 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2077 ex
? "EXMODE" : "PRMODE");
2079 if (!ocfs2_is_hard_readonly(OCFS2_SB(inode
->i_sb
)) &&
2080 !ocfs2_mount_local(osb
))
2081 ocfs2_cluster_unlock(OCFS2_SB(inode
->i_sb
), lockres
, level
);
2086 int ocfs2_super_lock(struct ocfs2_super
*osb
,
2090 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2091 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2095 if (ocfs2_is_hard_readonly(osb
))
2098 if (ocfs2_mount_local(osb
))
2101 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
2107 /* The super block lock path is really in the best position to
2108 * know when resources covered by the lock need to be
2109 * refreshed, so we do it here. Of course, making sense of
2110 * everything is up to the caller :) */
2111 status
= ocfs2_should_refresh_lock_res(lockres
);
2117 status
= ocfs2_refresh_slot_info(osb
);
2119 ocfs2_complete_lock_res_refresh(lockres
, status
);
2129 void ocfs2_super_unlock(struct ocfs2_super
*osb
,
2132 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2133 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2135 if (!ocfs2_mount_local(osb
))
2136 ocfs2_cluster_unlock(osb
, lockres
, level
);
2139 int ocfs2_rename_lock(struct ocfs2_super
*osb
)
2142 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2144 if (ocfs2_is_hard_readonly(osb
))
2147 if (ocfs2_mount_local(osb
))
2150 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_EX
, 0, 0);
2157 void ocfs2_rename_unlock(struct ocfs2_super
*osb
)
2159 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2161 if (!ocfs2_mount_local(osb
))
2162 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
2165 int ocfs2_dentry_lock(struct dentry
*dentry
, int ex
)
2168 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2169 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2170 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2174 if (ocfs2_is_hard_readonly(osb
))
2177 if (ocfs2_mount_local(osb
))
2180 ret
= ocfs2_cluster_lock(osb
, &dl
->dl_lockres
, level
, 0, 0);
2187 void ocfs2_dentry_unlock(struct dentry
*dentry
, int ex
)
2189 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2190 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2191 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2193 if (!ocfs2_mount_local(osb
))
2194 ocfs2_cluster_unlock(osb
, &dl
->dl_lockres
, level
);
2197 /* Reference counting of the dlm debug structure. We want this because
2198 * open references on the debug inodes can live on after a mount, so
2199 * we can't rely on the ocfs2_super to always exist. */
2200 static void ocfs2_dlm_debug_free(struct kref
*kref
)
2202 struct ocfs2_dlm_debug
*dlm_debug
;
2204 dlm_debug
= container_of(kref
, struct ocfs2_dlm_debug
, d_refcnt
);
2209 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug
*dlm_debug
)
2212 kref_put(&dlm_debug
->d_refcnt
, ocfs2_dlm_debug_free
);
2215 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug
*debug
)
2217 kref_get(&debug
->d_refcnt
);
2220 struct ocfs2_dlm_debug
*ocfs2_new_dlm_debug(void)
2222 struct ocfs2_dlm_debug
*dlm_debug
;
2224 dlm_debug
= kmalloc(sizeof(struct ocfs2_dlm_debug
), GFP_KERNEL
);
2226 mlog_errno(-ENOMEM
);
2230 kref_init(&dlm_debug
->d_refcnt
);
2231 INIT_LIST_HEAD(&dlm_debug
->d_lockres_tracking
);
2232 dlm_debug
->d_locking_state
= NULL
;
2237 /* Access to this is arbitrated for us via seq_file->sem. */
2238 struct ocfs2_dlm_seq_priv
{
2239 struct ocfs2_dlm_debug
*p_dlm_debug
;
2240 struct ocfs2_lock_res p_iter_res
;
2241 struct ocfs2_lock_res p_tmp_res
;
2244 static struct ocfs2_lock_res
*ocfs2_dlm_next_res(struct ocfs2_lock_res
*start
,
2245 struct ocfs2_dlm_seq_priv
*priv
)
2247 struct ocfs2_lock_res
*iter
, *ret
= NULL
;
2248 struct ocfs2_dlm_debug
*dlm_debug
= priv
->p_dlm_debug
;
2250 assert_spin_locked(&ocfs2_dlm_tracking_lock
);
2252 list_for_each_entry(iter
, &start
->l_debug_list
, l_debug_list
) {
2253 /* discover the head of the list */
2254 if (&iter
->l_debug_list
== &dlm_debug
->d_lockres_tracking
) {
2255 mlog(0, "End of list found, %p\n", ret
);
2259 /* We track our "dummy" iteration lockres' by a NULL
2261 if (iter
->l_ops
!= NULL
) {
2270 static void *ocfs2_dlm_seq_start(struct seq_file
*m
, loff_t
*pos
)
2272 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
2273 struct ocfs2_lock_res
*iter
;
2275 spin_lock(&ocfs2_dlm_tracking_lock
);
2276 iter
= ocfs2_dlm_next_res(&priv
->p_iter_res
, priv
);
2278 /* Since lockres' have the lifetime of their container
2279 * (which can be inodes, ocfs2_supers, etc) we want to
2280 * copy this out to a temporary lockres while still
2281 * under the spinlock. Obviously after this we can't
2282 * trust any pointers on the copy returned, but that's
2283 * ok as the information we want isn't typically held
2285 priv
->p_tmp_res
= *iter
;
2286 iter
= &priv
->p_tmp_res
;
2288 spin_unlock(&ocfs2_dlm_tracking_lock
);
2293 static void ocfs2_dlm_seq_stop(struct seq_file
*m
, void *v
)
2297 static void *ocfs2_dlm_seq_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
2299 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
2300 struct ocfs2_lock_res
*iter
= v
;
2301 struct ocfs2_lock_res
*dummy
= &priv
->p_iter_res
;
2303 spin_lock(&ocfs2_dlm_tracking_lock
);
2304 iter
= ocfs2_dlm_next_res(iter
, priv
);
2305 list_del_init(&dummy
->l_debug_list
);
2307 list_add(&dummy
->l_debug_list
, &iter
->l_debug_list
);
2308 priv
->p_tmp_res
= *iter
;
2309 iter
= &priv
->p_tmp_res
;
2311 spin_unlock(&ocfs2_dlm_tracking_lock
);
2316 /* So that debugfs.ocfs2 can determine which format is being used */
2317 #define OCFS2_DLM_DEBUG_STR_VERSION 1
2318 static int ocfs2_dlm_seq_show(struct seq_file
*m
, void *v
)
2322 struct ocfs2_lock_res
*lockres
= v
;
2327 seq_printf(m
, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION
);
2329 if (lockres
->l_type
== OCFS2_LOCK_TYPE_DENTRY
)
2330 seq_printf(m
, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START
- 1,
2332 (unsigned int)ocfs2_get_dentry_lock_ino(lockres
));
2334 seq_printf(m
, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN
, lockres
->l_name
);
2336 seq_printf(m
, "%d\t"
2347 lockres
->l_unlock_action
,
2348 lockres
->l_ro_holders
,
2349 lockres
->l_ex_holders
,
2350 lockres
->l_requested
,
2351 lockres
->l_blocking
);
2353 /* Dump the raw LVB */
2354 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2355 for(i
= 0; i
< DLM_LVB_LEN
; i
++)
2356 seq_printf(m
, "0x%x\t", lvb
[i
]);
2359 seq_printf(m
, "\n");
2363 static const struct seq_operations ocfs2_dlm_seq_ops
= {
2364 .start
= ocfs2_dlm_seq_start
,
2365 .stop
= ocfs2_dlm_seq_stop
,
2366 .next
= ocfs2_dlm_seq_next
,
2367 .show
= ocfs2_dlm_seq_show
,
2370 static int ocfs2_dlm_debug_release(struct inode
*inode
, struct file
*file
)
2372 struct seq_file
*seq
= (struct seq_file
*) file
->private_data
;
2373 struct ocfs2_dlm_seq_priv
*priv
= seq
->private;
2374 struct ocfs2_lock_res
*res
= &priv
->p_iter_res
;
2376 ocfs2_remove_lockres_tracking(res
);
2377 ocfs2_put_dlm_debug(priv
->p_dlm_debug
);
2378 return seq_release_private(inode
, file
);
2381 static int ocfs2_dlm_debug_open(struct inode
*inode
, struct file
*file
)
2384 struct ocfs2_dlm_seq_priv
*priv
;
2385 struct seq_file
*seq
;
2386 struct ocfs2_super
*osb
;
2388 priv
= kzalloc(sizeof(struct ocfs2_dlm_seq_priv
), GFP_KERNEL
);
2394 osb
= inode
->i_private
;
2395 ocfs2_get_dlm_debug(osb
->osb_dlm_debug
);
2396 priv
->p_dlm_debug
= osb
->osb_dlm_debug
;
2397 INIT_LIST_HEAD(&priv
->p_iter_res
.l_debug_list
);
2399 ret
= seq_open(file
, &ocfs2_dlm_seq_ops
);
2406 seq
= (struct seq_file
*) file
->private_data
;
2407 seq
->private = priv
;
2409 ocfs2_add_lockres_tracking(&priv
->p_iter_res
,
2416 static const struct file_operations ocfs2_dlm_debug_fops
= {
2417 .open
= ocfs2_dlm_debug_open
,
2418 .release
= ocfs2_dlm_debug_release
,
2420 .llseek
= seq_lseek
,
2423 static int ocfs2_dlm_init_debug(struct ocfs2_super
*osb
)
2426 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2428 dlm_debug
->d_locking_state
= debugfs_create_file("locking_state",
2430 osb
->osb_debug_root
,
2432 &ocfs2_dlm_debug_fops
);
2433 if (!dlm_debug
->d_locking_state
) {
2436 "Unable to create locking state debugfs file.\n");
2440 ocfs2_get_dlm_debug(dlm_debug
);
2445 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super
*osb
)
2447 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
2450 debugfs_remove(dlm_debug
->d_locking_state
);
2451 ocfs2_put_dlm_debug(dlm_debug
);
2455 int ocfs2_dlm_init(struct ocfs2_super
*osb
)
2458 struct ocfs2_cluster_connection
*conn
= NULL
;
2462 if (ocfs2_mount_local(osb
)) {
2467 status
= ocfs2_dlm_init_debug(osb
);
2473 /* launch downconvert thread */
2474 osb
->dc_task
= kthread_run(ocfs2_downconvert_thread
, osb
, "ocfs2dc");
2475 if (IS_ERR(osb
->dc_task
)) {
2476 status
= PTR_ERR(osb
->dc_task
);
2477 osb
->dc_task
= NULL
;
2482 /* for now, uuid == domain */
2483 status
= ocfs2_cluster_connect(osb
->uuid_str
,
2484 strlen(osb
->uuid_str
),
2485 ocfs2_do_node_down
, osb
,
2492 status
= ocfs2_cluster_this_node(&osb
->node_num
);
2496 "could not find this host's node number\n");
2497 ocfs2_cluster_disconnect(conn
);
2502 ocfs2_super_lock_res_init(&osb
->osb_super_lockres
, osb
);
2503 ocfs2_rename_lock_res_init(&osb
->osb_rename_lockres
, osb
);
2510 ocfs2_dlm_shutdown_debug(osb
);
2512 kthread_stop(osb
->dc_task
);
2519 void ocfs2_dlm_shutdown(struct ocfs2_super
*osb
)
2523 ocfs2_drop_osb_locks(osb
);
2526 * Now that we have dropped all locks and ocfs2_dismount_volume()
2527 * has disabled recovery, the DLM won't be talking to us. It's
2528 * safe to tear things down before disconnecting the cluster.
2532 kthread_stop(osb
->dc_task
);
2533 osb
->dc_task
= NULL
;
2536 ocfs2_lock_res_free(&osb
->osb_super_lockres
);
2537 ocfs2_lock_res_free(&osb
->osb_rename_lockres
);
2539 ocfs2_cluster_disconnect(osb
->cconn
);
2542 ocfs2_dlm_shutdown_debug(osb
);
2547 static void ocfs2_unlock_ast(void *opaque
, int error
)
2549 struct ocfs2_lock_res
*lockres
= opaque
;
2550 unsigned long flags
;
2554 mlog(0, "UNLOCK AST called on lock %s, action = %d\n", lockres
->l_name
,
2555 lockres
->l_unlock_action
);
2557 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2558 /* We tried to cancel a convert request, but it was already
2559 * granted. All we want to do here is clear our unlock
2560 * state. The wake_up call done at the bottom is redundant
2561 * (ocfs2_prepare_cancel_convert doesn't sleep on this) but doesn't
2562 * hurt anything anyway */
2563 if (error
== -DLM_ECANCEL
&&
2564 lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
2565 mlog(0, "Got cancelgrant for %s\n", lockres
->l_name
);
2567 /* We don't clear the busy flag in this case as it
2568 * should have been cleared by the ast which the dlm
2570 goto complete_unlock
;
2573 /* DLM_EUNLOCK is the success code for unlock */
2574 if (error
!= -DLM_EUNLOCK
) {
2575 mlog(ML_ERROR
, "Dlm passes error %d for lock %s, "
2576 "unlock_action %d\n", error
, lockres
->l_name
,
2577 lockres
->l_unlock_action
);
2578 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2582 switch(lockres
->l_unlock_action
) {
2583 case OCFS2_UNLOCK_CANCEL_CONVERT
:
2584 mlog(0, "Cancel convert success for %s\n", lockres
->l_name
);
2585 lockres
->l_action
= OCFS2_AST_INVALID
;
2587 case OCFS2_UNLOCK_DROP_LOCK
:
2588 lockres
->l_level
= DLM_LOCK_IV
;
2594 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
2596 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
2597 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2599 wake_up(&lockres
->l_event
);
2604 static int ocfs2_drop_lock(struct ocfs2_super
*osb
,
2605 struct ocfs2_lock_res
*lockres
)
2608 unsigned long flags
;
2611 /* We didn't get anywhere near actually using this lockres. */
2612 if (!(lockres
->l_flags
& OCFS2_LOCK_INITIALIZED
))
2615 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
2616 lkm_flags
|= DLM_LKF_VALBLK
;
2618 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2620 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_FREEING
),
2621 "lockres %s, flags 0x%lx\n",
2622 lockres
->l_name
, lockres
->l_flags
);
2624 while (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
2625 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
2626 "%u, unlock_action = %u\n",
2627 lockres
->l_name
, lockres
->l_flags
, lockres
->l_action
,
2628 lockres
->l_unlock_action
);
2630 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2632 /* XXX: Today we just wait on any busy
2633 * locks... Perhaps we need to cancel converts in the
2635 ocfs2_wait_on_busy_lock(lockres
);
2637 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2640 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
2641 if (lockres
->l_flags
& OCFS2_LOCK_ATTACHED
&&
2642 lockres
->l_level
== DLM_LOCK_EX
&&
2643 !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
2644 lockres
->l_ops
->set_lvb(lockres
);
2647 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
)
2648 mlog(ML_ERROR
, "destroying busy lock: \"%s\"\n",
2650 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
2651 mlog(0, "destroying blocked lock: \"%s\"\n", lockres
->l_name
);
2653 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
2654 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2658 lockres_clear_flags(lockres
, OCFS2_LOCK_ATTACHED
);
2660 /* make sure we never get here while waiting for an ast to
2662 BUG_ON(lockres
->l_action
!= OCFS2_AST_INVALID
);
2664 /* is this necessary? */
2665 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2666 lockres
->l_unlock_action
= OCFS2_UNLOCK_DROP_LOCK
;
2667 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2669 mlog(0, "lock %s\n", lockres
->l_name
);
2671 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
, lkm_flags
,
2674 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
2675 mlog(ML_ERROR
, "lockres flags: %lu\n", lockres
->l_flags
);
2676 /* XXX Need to abstract this */
2677 dlm_print_one_lock(lockres
->l_lksb
.lksb_o2dlm
.lockid
);
2680 mlog(0, "lock %s, successfull return from ocfs2_dlm_unlock\n",
2683 ocfs2_wait_on_busy_lock(lockres
);
2689 /* Mark the lockres as being dropped. It will no longer be
2690 * queued if blocking, but we still may have to wait on it
2691 * being dequeued from the downconvert thread before we can consider
2694 * You can *not* attempt to call cluster_lock on this lockres anymore. */
2695 void ocfs2_mark_lockres_freeing(struct ocfs2_lock_res
*lockres
)
2698 struct ocfs2_mask_waiter mw
;
2699 unsigned long flags
;
2701 ocfs2_init_mask_waiter(&mw
);
2703 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2704 lockres
->l_flags
|= OCFS2_LOCK_FREEING
;
2705 while (lockres
->l_flags
& OCFS2_LOCK_QUEUED
) {
2706 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_QUEUED
, 0);
2707 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2709 mlog(0, "Waiting on lockres %s\n", lockres
->l_name
);
2711 status
= ocfs2_wait_for_mask(&mw
);
2715 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2717 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2720 void ocfs2_simple_drop_lockres(struct ocfs2_super
*osb
,
2721 struct ocfs2_lock_res
*lockres
)
2725 ocfs2_mark_lockres_freeing(lockres
);
2726 ret
= ocfs2_drop_lock(osb
, lockres
);
2731 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
)
2733 ocfs2_simple_drop_lockres(osb
, &osb
->osb_super_lockres
);
2734 ocfs2_simple_drop_lockres(osb
, &osb
->osb_rename_lockres
);
2737 int ocfs2_drop_inode_locks(struct inode
*inode
)
2743 /* No need to call ocfs2_mark_lockres_freeing here -
2744 * ocfs2_clear_inode has done it for us. */
2746 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2747 &OCFS2_I(inode
)->ip_open_lockres
);
2753 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2754 &OCFS2_I(inode
)->ip_inode_lockres
);
2757 if (err
< 0 && !status
)
2760 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
2761 &OCFS2_I(inode
)->ip_rw_lockres
);
2764 if (err
< 0 && !status
)
2771 static void ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
2774 assert_spin_locked(&lockres
->l_lock
);
2776 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
2778 if (lockres
->l_level
<= new_level
) {
2779 mlog(ML_ERROR
, "lockres->l_level (%d) <= new_level (%d)\n",
2780 lockres
->l_level
, new_level
);
2784 mlog(0, "lock %s, new_level = %d, l_blocking = %d\n",
2785 lockres
->l_name
, new_level
, lockres
->l_blocking
);
2787 lockres
->l_action
= OCFS2_AST_DOWNCONVERT
;
2788 lockres
->l_requested
= new_level
;
2789 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2792 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
2793 struct ocfs2_lock_res
*lockres
,
2798 u32 dlm_flags
= DLM_LKF_CONVERT
;
2803 dlm_flags
|= DLM_LKF_VALBLK
;
2805 ret
= ocfs2_dlm_lock(osb
->cconn
,
2810 OCFS2_LOCK_ID_MAX_LEN
- 1,
2813 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
2814 ocfs2_recover_from_dlm_error(lockres
, 1);
2824 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
2825 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
2826 struct ocfs2_lock_res
*lockres
)
2828 assert_spin_locked(&lockres
->l_lock
);
2831 mlog(0, "lock %s\n", lockres
->l_name
);
2833 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
2834 /* If we're already trying to cancel a lock conversion
2835 * then just drop the spinlock and allow the caller to
2836 * requeue this lock. */
2838 mlog(0, "Lockres %s, skip convert\n", lockres
->l_name
);
2842 /* were we in a convert when we got the bast fire? */
2843 BUG_ON(lockres
->l_action
!= OCFS2_AST_CONVERT
&&
2844 lockres
->l_action
!= OCFS2_AST_DOWNCONVERT
);
2845 /* set things up for the unlockast to know to just
2846 * clear out the ast_action and unset busy, etc. */
2847 lockres
->l_unlock_action
= OCFS2_UNLOCK_CANCEL_CONVERT
;
2849 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
),
2850 "lock %s, invalid flags: 0x%lx\n",
2851 lockres
->l_name
, lockres
->l_flags
);
2856 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
2857 struct ocfs2_lock_res
*lockres
)
2862 mlog(0, "lock %s\n", lockres
->l_name
);
2864 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
,
2865 DLM_LKF_CANCEL
, lockres
);
2867 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
2868 ocfs2_recover_from_dlm_error(lockres
, 0);
2871 mlog(0, "lock %s return from ocfs2_dlm_unlock\n", lockres
->l_name
);
2877 static int ocfs2_unblock_lock(struct ocfs2_super
*osb
,
2878 struct ocfs2_lock_res
*lockres
,
2879 struct ocfs2_unblock_ctl
*ctl
)
2881 unsigned long flags
;
2889 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2891 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
2894 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
2896 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
2897 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2899 ret
= ocfs2_cancel_convert(osb
, lockres
);
2906 /* if we're blocking an exclusive and we have *any* holders,
2908 if ((lockres
->l_blocking
== DLM_LOCK_EX
)
2909 && (lockres
->l_ex_holders
|| lockres
->l_ro_holders
))
2912 /* If it's a PR we're blocking, then only
2913 * requeue if we've got any EX holders */
2914 if (lockres
->l_blocking
== DLM_LOCK_PR
&&
2915 lockres
->l_ex_holders
)
2919 * Can we get a lock in this state if the holder counts are
2920 * zero? The meta data unblock code used to check this.
2922 if ((lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
2923 && (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
))
2926 new_level
= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
2928 if (lockres
->l_ops
->check_downconvert
2929 && !lockres
->l_ops
->check_downconvert(lockres
, new_level
))
2932 /* If we get here, then we know that there are no more
2933 * incompatible holders (and anyone asking for an incompatible
2934 * lock is blocked). We can now downconvert the lock */
2935 if (!lockres
->l_ops
->downconvert_worker
)
2938 /* Some lockres types want to do a bit of work before
2939 * downconverting a lock. Allow that here. The worker function
2940 * may sleep, so we save off a copy of what we're blocking as
2941 * it may change while we're not holding the spin lock. */
2942 blocking
= lockres
->l_blocking
;
2943 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2945 ctl
->unblock_action
= lockres
->l_ops
->downconvert_worker(lockres
, blocking
);
2947 if (ctl
->unblock_action
== UNBLOCK_STOP_POST
)
2950 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2951 if (blocking
!= lockres
->l_blocking
) {
2952 /* If this changed underneath us, then we can't drop
2960 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
2961 if (lockres
->l_level
== DLM_LOCK_EX
)
2965 * We only set the lvb if the lock has been fully
2966 * refreshed - otherwise we risk setting stale
2967 * data. Otherwise, there's no need to actually clear
2968 * out the lvb here as it's value is still valid.
2970 if (set_lvb
&& !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
2971 lockres
->l_ops
->set_lvb(lockres
);
2974 ocfs2_prepare_downconvert(lockres
, new_level
);
2975 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2976 ret
= ocfs2_downconvert_lock(osb
, lockres
, new_level
, set_lvb
);
2982 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2989 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
2992 struct inode
*inode
;
2993 struct address_space
*mapping
;
2995 inode
= ocfs2_lock_res_inode(lockres
);
2996 mapping
= inode
->i_mapping
;
2998 if (!S_ISREG(inode
->i_mode
))
3002 * We need this before the filemap_fdatawrite() so that it can
3003 * transfer the dirty bit from the PTE to the
3004 * page. Unfortunately this means that even for EX->PR
3005 * downconverts, we'll lose our mappings and have to build
3008 unmap_mapping_range(mapping
, 0, 0, 0);
3010 if (filemap_fdatawrite(mapping
)) {
3011 mlog(ML_ERROR
, "Could not sync inode %llu for downconvert!",
3012 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3014 sync_mapping_buffers(mapping
);
3015 if (blocking
== DLM_LOCK_EX
) {
3016 truncate_inode_pages(mapping
, 0);
3018 /* We only need to wait on the I/O if we're not also
3019 * truncating pages because truncate_inode_pages waits
3020 * for us above. We don't truncate pages if we're
3021 * blocking anything < EXMODE because we want to keep
3022 * them around in that case. */
3023 filemap_fdatawait(mapping
);
3027 return UNBLOCK_CONTINUE
;
3030 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
3033 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3034 int checkpointed
= ocfs2_inode_fully_checkpointed(inode
);
3036 BUG_ON(new_level
!= DLM_LOCK_NL
&& new_level
!= DLM_LOCK_PR
);
3037 BUG_ON(lockres
->l_level
!= DLM_LOCK_EX
&& !checkpointed
);
3042 ocfs2_start_checkpoint(OCFS2_SB(inode
->i_sb
));
3046 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
)
3048 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3050 __ocfs2_stuff_meta_lvb(inode
);
3054 * Does the final reference drop on our dentry lock. Right now this
3055 * happens in the downconvert thread, but we could choose to simplify the
3056 * dlmglue API and push these off to the ocfs2_wq in the future.
3058 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
3059 struct ocfs2_lock_res
*lockres
)
3061 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3062 ocfs2_dentry_lock_put(osb
, dl
);
3066 * d_delete() matching dentries before the lock downconvert.
3068 * At this point, any process waiting to destroy the
3069 * dentry_lock due to last ref count is stopped by the
3070 * OCFS2_LOCK_QUEUED flag.
3072 * We have two potential problems
3074 * 1) If we do the last reference drop on our dentry_lock (via dput)
3075 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3076 * the downconvert to finish. Instead we take an elevated
3077 * reference and push the drop until after we've completed our
3078 * unblock processing.
3080 * 2) There might be another process with a final reference,
3081 * waiting on us to finish processing. If this is the case, we
3082 * detect it and exit out - there's no more dentries anyway.
3084 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
3087 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3088 struct ocfs2_inode_info
*oi
= OCFS2_I(dl
->dl_inode
);
3089 struct dentry
*dentry
;
3090 unsigned long flags
;
3094 * This node is blocking another node from getting a read
3095 * lock. This happens when we've renamed within a
3096 * directory. We've forced the other nodes to d_delete(), but
3097 * we never actually dropped our lock because it's still
3098 * valid. The downconvert code will retain a PR for this node,
3099 * so there's no further work to do.
3101 if (blocking
== DLM_LOCK_PR
)
3102 return UNBLOCK_CONTINUE
;
3105 * Mark this inode as potentially orphaned. The code in
3106 * ocfs2_delete_inode() will figure out whether it actually
3107 * needs to be freed or not.
3109 spin_lock(&oi
->ip_lock
);
3110 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
3111 spin_unlock(&oi
->ip_lock
);
3114 * Yuck. We need to make sure however that the check of
3115 * OCFS2_LOCK_FREEING and the extra reference are atomic with
3116 * respect to a reference decrement or the setting of that
3119 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3120 spin_lock(&dentry_attach_lock
);
3121 if (!(lockres
->l_flags
& OCFS2_LOCK_FREEING
)
3126 spin_unlock(&dentry_attach_lock
);
3127 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3129 mlog(0, "extra_ref = %d\n", extra_ref
);
3132 * We have a process waiting on us in ocfs2_dentry_iput(),
3133 * which means we can't have any more outstanding
3134 * aliases. There's no need to do any more work.
3137 return UNBLOCK_CONTINUE
;
3139 spin_lock(&dentry_attach_lock
);
3141 dentry
= ocfs2_find_local_alias(dl
->dl_inode
,
3142 dl
->dl_parent_blkno
, 1);
3145 spin_unlock(&dentry_attach_lock
);
3147 mlog(0, "d_delete(%.*s);\n", dentry
->d_name
.len
,
3148 dentry
->d_name
.name
);
3151 * The following dcache calls may do an
3152 * iput(). Normally we don't want that from the
3153 * downconverting thread, but in this case it's ok
3154 * because the requesting node already has an
3155 * exclusive lock on the inode, so it can't be queued
3156 * for a downconvert.
3161 spin_lock(&dentry_attach_lock
);
3163 spin_unlock(&dentry_attach_lock
);
3166 * If we are the last holder of this dentry lock, there is no
3167 * reason to downconvert so skip straight to the unlock.
3169 if (dl
->dl_count
== 1)
3170 return UNBLOCK_STOP_POST
;
3172 return UNBLOCK_CONTINUE_POST
;
3176 * This is the filesystem locking protocol. It provides the lock handling
3177 * hooks for the underlying DLM. It has a maximum version number.
3178 * The version number allows interoperability with systems running at
3179 * the same major number and an equal or smaller minor number.
3181 * Whenever the filesystem does new things with locks (adds or removes a
3182 * lock, orders them differently, does different things underneath a lock),
3183 * the version must be changed. The protocol is negotiated when joining
3184 * the dlm domain. A node may join the domain if its major version is
3185 * identical to all other nodes and its minor version is greater than
3186 * or equal to all other nodes. When its minor version is greater than
3187 * the other nodes, it will run at the minor version specified by the
3190 * If a locking change is made that will not be compatible with older
3191 * versions, the major number must be increased and the minor version set
3192 * to zero. If a change merely adds a behavior that can be disabled when
3193 * speaking to older versions, the minor version must be increased. If a
3194 * change adds a fully backwards compatible change (eg, LVB changes that
3195 * are just ignored by older versions), the version does not need to be
3198 static struct ocfs2_locking_protocol lproto
= {
3200 .pv_major
= OCFS2_LOCKING_PROTOCOL_MAJOR
,
3201 .pv_minor
= OCFS2_LOCKING_PROTOCOL_MINOR
,
3203 .lp_lock_ast
= ocfs2_locking_ast
,
3204 .lp_blocking_ast
= ocfs2_blocking_ast
,
3205 .lp_unlock_ast
= ocfs2_unlock_ast
,
3208 /* This interface isn't the final one, hence the less-than-perfect names */
3209 void dlmglue_init_stack(void)
3211 o2cb_get_stack(&lproto
);
3214 void dlmglue_exit_stack(void)
3219 static void ocfs2_process_blocked_lock(struct ocfs2_super
*osb
,
3220 struct ocfs2_lock_res
*lockres
)
3223 struct ocfs2_unblock_ctl ctl
= {0, 0,};
3224 unsigned long flags
;
3226 /* Our reference to the lockres in this function can be
3227 * considered valid until we remove the OCFS2_LOCK_QUEUED
3233 BUG_ON(!lockres
->l_ops
);
3235 mlog(0, "lockres %s blocked.\n", lockres
->l_name
);
3237 /* Detect whether a lock has been marked as going away while
3238 * the downconvert thread was processing other things. A lock can
3239 * still be marked with OCFS2_LOCK_FREEING after this check,
3240 * but short circuiting here will still save us some
3242 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3243 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
)
3245 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3247 status
= ocfs2_unblock_lock(osb
, lockres
, &ctl
);
3251 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3253 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
|| !ctl
.requeue
) {
3254 lockres_clear_flags(lockres
, OCFS2_LOCK_QUEUED
);
3256 ocfs2_schedule_blocked_lock(osb
, lockres
);
3258 mlog(0, "lockres %s, requeue = %s.\n", lockres
->l_name
,
3259 ctl
.requeue
? "yes" : "no");
3260 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3262 if (ctl
.unblock_action
!= UNBLOCK_CONTINUE
3263 && lockres
->l_ops
->post_unlock
)
3264 lockres
->l_ops
->post_unlock(osb
, lockres
);
3269 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
3270 struct ocfs2_lock_res
*lockres
)
3274 assert_spin_locked(&lockres
->l_lock
);
3276 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
) {
3277 /* Do not schedule a lock for downconvert when it's on
3278 * the way to destruction - any nodes wanting access
3279 * to the resource will get it soon. */
3280 mlog(0, "Lockres %s won't be scheduled: flags 0x%lx\n",
3281 lockres
->l_name
, lockres
->l_flags
);
3285 lockres_or_flags(lockres
, OCFS2_LOCK_QUEUED
);
3287 spin_lock(&osb
->dc_task_lock
);
3288 if (list_empty(&lockres
->l_blocked_list
)) {
3289 list_add_tail(&lockres
->l_blocked_list
,
3290 &osb
->blocked_lock_list
);
3291 osb
->blocked_lock_count
++;
3293 spin_unlock(&osb
->dc_task_lock
);
3298 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super
*osb
)
3300 unsigned long processed
;
3301 struct ocfs2_lock_res
*lockres
;
3305 spin_lock(&osb
->dc_task_lock
);
3306 /* grab this early so we know to try again if a state change and
3307 * wake happens part-way through our work */
3308 osb
->dc_work_sequence
= osb
->dc_wake_sequence
;
3310 processed
= osb
->blocked_lock_count
;
3312 BUG_ON(list_empty(&osb
->blocked_lock_list
));
3314 lockres
= list_entry(osb
->blocked_lock_list
.next
,
3315 struct ocfs2_lock_res
, l_blocked_list
);
3316 list_del_init(&lockres
->l_blocked_list
);
3317 osb
->blocked_lock_count
--;
3318 spin_unlock(&osb
->dc_task_lock
);
3323 ocfs2_process_blocked_lock(osb
, lockres
);
3325 spin_lock(&osb
->dc_task_lock
);
3327 spin_unlock(&osb
->dc_task_lock
);
3332 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super
*osb
)
3336 spin_lock(&osb
->dc_task_lock
);
3337 if (list_empty(&osb
->blocked_lock_list
))
3340 spin_unlock(&osb
->dc_task_lock
);
3344 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super
*osb
)
3346 int should_wake
= 0;
3348 spin_lock(&osb
->dc_task_lock
);
3349 if (osb
->dc_work_sequence
!= osb
->dc_wake_sequence
)
3351 spin_unlock(&osb
->dc_task_lock
);
3356 static int ocfs2_downconvert_thread(void *arg
)
3359 struct ocfs2_super
*osb
= arg
;
3361 /* only quit once we've been asked to stop and there is no more
3363 while (!(kthread_should_stop() &&
3364 ocfs2_downconvert_thread_lists_empty(osb
))) {
3366 wait_event_interruptible(osb
->dc_event
,
3367 ocfs2_downconvert_thread_should_wake(osb
) ||
3368 kthread_should_stop());
3370 mlog(0, "downconvert_thread: awoken\n");
3372 ocfs2_downconvert_thread_do_work(osb
);
3375 osb
->dc_task
= NULL
;
3379 void ocfs2_wake_downconvert_thread(struct ocfs2_super
*osb
)
3381 spin_lock(&osb
->dc_task_lock
);
3382 /* make sure the voting thread gets a swipe at whatever changes
3383 * the caller may have made to the voting state */
3384 osb
->dc_wake_sequence
++;
3385 spin_unlock(&osb
->dc_task_lock
);
3386 wake_up(&osb
->dc_event
);