1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Code which implements an OCFS2 specific interface to our DLM.
8 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
26 #include <linux/types.h>
27 #include <linux/slab.h>
28 #include <linux/highmem.h>
30 #include <linux/kthread.h>
31 #include <linux/pagemap.h>
32 #include <linux/debugfs.h>
33 #include <linux/seq_file.h>
34 #include <linux/time.h>
35 #include <linux/quotaops.h>
36 #include <linux/sched/signal.h>
38 #define MLOG_MASK_PREFIX ML_DLM_GLUE
39 #include <cluster/masklog.h>
42 #include "ocfs2_lockingver.h"
47 #include "extent_map.h"
49 #include "heartbeat.h"
52 #include "stackglue.h"
57 #include "refcounttree.h"
60 #include "buffer_head_io.h"
62 struct ocfs2_mask_waiter
{
63 struct list_head mw_item
;
65 struct completion mw_complete
;
66 unsigned long mw_mask
;
67 unsigned long mw_goal
;
68 #ifdef CONFIG_OCFS2_FS_STATS
69 ktime_t mw_lock_start
;
73 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
);
74 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
);
75 static struct ocfs2_super
*ocfs2_get_file_osb(struct ocfs2_lock_res
*lockres
);
76 static struct ocfs2_super
*ocfs2_get_qinfo_osb(struct ocfs2_lock_res
*lockres
);
79 * Return value from ->downconvert_worker functions.
81 * These control the precise actions of ocfs2_unblock_lock()
82 * and ocfs2_process_blocked_lock()
85 enum ocfs2_unblock_action
{
86 UNBLOCK_CONTINUE
= 0, /* Continue downconvert */
87 UNBLOCK_CONTINUE_POST
= 1, /* Continue downconvert, fire
88 * ->post_unlock callback */
89 UNBLOCK_STOP_POST
= 2, /* Do not downconvert, fire
90 * ->post_unlock() callback. */
93 struct ocfs2_unblock_ctl
{
95 enum ocfs2_unblock_action unblock_action
;
98 /* Lockdep class keys */
99 struct lock_class_key lockdep_keys
[OCFS2_NUM_LOCK_TYPES
];
101 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
103 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
);
105 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
108 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
111 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
112 struct ocfs2_lock_res
*lockres
);
114 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res
*lockres
);
116 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res
*lockres
,
118 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res
*lockres
,
121 #define mlog_meta_lvb(__level, __lockres) ocfs2_dump_meta_lvb_info(__level, __PRETTY_FUNCTION__, __LINE__, __lockres)
123 /* This aids in debugging situations where a bad LVB might be involved. */
124 static void ocfs2_dump_meta_lvb_info(u64 level
,
125 const char *function
,
127 struct ocfs2_lock_res
*lockres
)
129 struct ocfs2_meta_lvb
*lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
131 mlog(level
, "LVB information for %s (called from %s:%u):\n",
132 lockres
->l_name
, function
, line
);
133 mlog(level
, "version: %u, clusters: %u, generation: 0x%x\n",
134 lvb
->lvb_version
, be32_to_cpu(lvb
->lvb_iclusters
),
135 be32_to_cpu(lvb
->lvb_igeneration
));
136 mlog(level
, "size: %llu, uid %u, gid %u, mode 0x%x\n",
137 (unsigned long long)be64_to_cpu(lvb
->lvb_isize
),
138 be32_to_cpu(lvb
->lvb_iuid
), be32_to_cpu(lvb
->lvb_igid
),
139 be16_to_cpu(lvb
->lvb_imode
));
140 mlog(level
, "nlink %u, atime_packed 0x%llx, ctime_packed 0x%llx, "
141 "mtime_packed 0x%llx iattr 0x%x\n", be16_to_cpu(lvb
->lvb_inlink
),
142 (long long)be64_to_cpu(lvb
->lvb_iatime_packed
),
143 (long long)be64_to_cpu(lvb
->lvb_ictime_packed
),
144 (long long)be64_to_cpu(lvb
->lvb_imtime_packed
),
145 be32_to_cpu(lvb
->lvb_iattr
));
150 * OCFS2 Lock Resource Operations
152 * These fine tune the behavior of the generic dlmglue locking infrastructure.
154 * The most basic of lock types can point ->l_priv to their respective
155 * struct ocfs2_super and allow the default actions to manage things.
157 * Right now, each lock type also needs to implement an init function,
158 * and trivial lock/unlock wrappers. ocfs2_simple_drop_lockres()
159 * should be called when the lock is no longer needed (i.e., object
162 struct ocfs2_lock_res_ops
{
164 * Translate an ocfs2_lock_res * into an ocfs2_super *. Define
165 * this callback if ->l_priv is not an ocfs2_super pointer
167 struct ocfs2_super
* (*get_osb
)(struct ocfs2_lock_res
*);
170 * Optionally called in the downconvert thread after a
171 * successful downconvert. The lockres will not be referenced
172 * after this callback is called, so it is safe to free
175 * The exact semantics of when this is called are controlled
176 * by ->downconvert_worker()
178 void (*post_unlock
)(struct ocfs2_super
*, struct ocfs2_lock_res
*);
181 * Allow a lock type to add checks to determine whether it is
182 * safe to downconvert a lock. Return 0 to re-queue the
183 * downconvert at a later time, nonzero to continue.
185 * For most locks, the default checks that there are no
186 * incompatible holders are sufficient.
188 * Called with the lockres spinlock held.
190 int (*check_downconvert
)(struct ocfs2_lock_res
*, int);
193 * Allows a lock type to populate the lock value block. This
194 * is called on downconvert, and when we drop a lock.
196 * Locks that want to use this should set LOCK_TYPE_USES_LVB
197 * in the flags field.
199 * Called with the lockres spinlock held.
201 void (*set_lvb
)(struct ocfs2_lock_res
*);
204 * Called from the downconvert thread when it is determined
205 * that a lock will be downconverted. This is called without
206 * any locks held so the function can do work that might
207 * schedule (syncing out data, etc).
209 * This should return any one of the ocfs2_unblock_action
210 * values, depending on what it wants the thread to do.
212 int (*downconvert_worker
)(struct ocfs2_lock_res
*, int);
215 * LOCK_TYPE_* flags which describe the specific requirements
216 * of a lock type. Descriptions of each individual flag follow.
222 * Some locks want to "refresh" potentially stale data when a
223 * meaningful (PRMODE or EXMODE) lock level is first obtained. If this
224 * flag is set, the OCFS2_LOCK_NEEDS_REFRESH flag will be set on the
225 * individual lockres l_flags member from the ast function. It is
226 * expected that the locking wrapper will clear the
227 * OCFS2_LOCK_NEEDS_REFRESH flag when done.
229 #define LOCK_TYPE_REQUIRES_REFRESH 0x1
232 * Indicate that a lock type makes use of the lock value block. The
233 * ->set_lvb lock type callback must be defined.
235 #define LOCK_TYPE_USES_LVB 0x2
237 static struct ocfs2_lock_res_ops ocfs2_inode_rw_lops
= {
238 .get_osb
= ocfs2_get_inode_osb
,
242 static struct ocfs2_lock_res_ops ocfs2_inode_inode_lops
= {
243 .get_osb
= ocfs2_get_inode_osb
,
244 .check_downconvert
= ocfs2_check_meta_downconvert
,
245 .set_lvb
= ocfs2_set_meta_lvb
,
246 .downconvert_worker
= ocfs2_data_convert_worker
,
247 .flags
= LOCK_TYPE_REQUIRES_REFRESH
|LOCK_TYPE_USES_LVB
,
250 static struct ocfs2_lock_res_ops ocfs2_super_lops
= {
251 .flags
= LOCK_TYPE_REQUIRES_REFRESH
,
254 static struct ocfs2_lock_res_ops ocfs2_rename_lops
= {
258 static struct ocfs2_lock_res_ops ocfs2_nfs_sync_lops
= {
262 static struct ocfs2_lock_res_ops ocfs2_trim_fs_lops
= {
263 .flags
= LOCK_TYPE_REQUIRES_REFRESH
|LOCK_TYPE_USES_LVB
,
266 static struct ocfs2_lock_res_ops ocfs2_orphan_scan_lops
= {
267 .flags
= LOCK_TYPE_REQUIRES_REFRESH
|LOCK_TYPE_USES_LVB
,
270 static struct ocfs2_lock_res_ops ocfs2_dentry_lops
= {
271 .get_osb
= ocfs2_get_dentry_osb
,
272 .post_unlock
= ocfs2_dentry_post_unlock
,
273 .downconvert_worker
= ocfs2_dentry_convert_worker
,
277 static struct ocfs2_lock_res_ops ocfs2_inode_open_lops
= {
278 .get_osb
= ocfs2_get_inode_osb
,
282 static struct ocfs2_lock_res_ops ocfs2_flock_lops
= {
283 .get_osb
= ocfs2_get_file_osb
,
287 static struct ocfs2_lock_res_ops ocfs2_qinfo_lops
= {
288 .set_lvb
= ocfs2_set_qinfo_lvb
,
289 .get_osb
= ocfs2_get_qinfo_osb
,
290 .flags
= LOCK_TYPE_REQUIRES_REFRESH
| LOCK_TYPE_USES_LVB
,
293 static struct ocfs2_lock_res_ops ocfs2_refcount_block_lops
= {
294 .check_downconvert
= ocfs2_check_refcount_downconvert
,
295 .downconvert_worker
= ocfs2_refcount_convert_worker
,
299 static inline int ocfs2_is_inode_lock(struct ocfs2_lock_res
*lockres
)
301 return lockres
->l_type
== OCFS2_LOCK_TYPE_META
||
302 lockres
->l_type
== OCFS2_LOCK_TYPE_RW
||
303 lockres
->l_type
== OCFS2_LOCK_TYPE_OPEN
;
306 static inline struct ocfs2_lock_res
*ocfs2_lksb_to_lock_res(struct ocfs2_dlm_lksb
*lksb
)
308 return container_of(lksb
, struct ocfs2_lock_res
, l_lksb
);
311 static inline struct inode
*ocfs2_lock_res_inode(struct ocfs2_lock_res
*lockres
)
313 BUG_ON(!ocfs2_is_inode_lock(lockres
));
315 return (struct inode
*) lockres
->l_priv
;
318 static inline struct ocfs2_dentry_lock
*ocfs2_lock_res_dl(struct ocfs2_lock_res
*lockres
)
320 BUG_ON(lockres
->l_type
!= OCFS2_LOCK_TYPE_DENTRY
);
322 return (struct ocfs2_dentry_lock
*)lockres
->l_priv
;
325 static inline struct ocfs2_mem_dqinfo
*ocfs2_lock_res_qinfo(struct ocfs2_lock_res
*lockres
)
327 BUG_ON(lockres
->l_type
!= OCFS2_LOCK_TYPE_QINFO
);
329 return (struct ocfs2_mem_dqinfo
*)lockres
->l_priv
;
332 static inline struct ocfs2_refcount_tree
*
333 ocfs2_lock_res_refcount_tree(struct ocfs2_lock_res
*res
)
335 return container_of(res
, struct ocfs2_refcount_tree
, rf_lockres
);
338 static inline struct ocfs2_super
*ocfs2_get_lockres_osb(struct ocfs2_lock_res
*lockres
)
340 if (lockres
->l_ops
->get_osb
)
341 return lockres
->l_ops
->get_osb(lockres
);
343 return (struct ocfs2_super
*)lockres
->l_priv
;
346 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
347 struct ocfs2_lock_res
*lockres
,
350 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
352 static void __ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
353 struct ocfs2_lock_res
*lockres
,
354 int level
, unsigned long caller_ip
);
355 static inline void ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
356 struct ocfs2_lock_res
*lockres
,
359 __ocfs2_cluster_unlock(osb
, lockres
, level
, _RET_IP_
);
362 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
);
363 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
);
364 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
);
365 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
, int level
);
366 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
367 struct ocfs2_lock_res
*lockres
);
368 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
370 #define ocfs2_log_dlm_error(_func, _err, _lockres) do { \
371 if ((_lockres)->l_type != OCFS2_LOCK_TYPE_DENTRY) \
372 mlog(ML_ERROR, "DLM error %d while calling %s on resource %s\n", \
373 _err, _func, _lockres->l_name); \
375 mlog(ML_ERROR, "DLM error %d while calling %s on resource %.*s%08x\n", \
376 _err, _func, OCFS2_DENTRY_LOCK_INO_START - 1, (_lockres)->l_name, \
377 (unsigned int)ocfs2_get_dentry_lock_ino(_lockres)); \
379 static int ocfs2_downconvert_thread(void *arg
);
380 static void ocfs2_downconvert_on_unlock(struct ocfs2_super
*osb
,
381 struct ocfs2_lock_res
*lockres
);
382 static int ocfs2_inode_lock_update(struct inode
*inode
,
383 struct buffer_head
**bh
);
384 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
);
385 static inline int ocfs2_highest_compat_lock_level(int level
);
386 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
388 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
389 struct ocfs2_lock_res
*lockres
,
392 unsigned int generation
);
393 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
394 struct ocfs2_lock_res
*lockres
);
395 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
396 struct ocfs2_lock_res
*lockres
);
399 static void ocfs2_build_lock_name(enum ocfs2_lock_type type
,
406 BUG_ON(type
>= OCFS2_NUM_LOCK_TYPES
);
408 len
= snprintf(name
, OCFS2_LOCK_ID_MAX_LEN
, "%c%s%016llx%08x",
409 ocfs2_lock_type_char(type
), OCFS2_LOCK_ID_PAD
,
410 (long long)blkno
, generation
);
412 BUG_ON(len
!= (OCFS2_LOCK_ID_MAX_LEN
- 1));
414 mlog(0, "built lock resource with name: %s\n", name
);
417 static DEFINE_SPINLOCK(ocfs2_dlm_tracking_lock
);
419 static void ocfs2_add_lockres_tracking(struct ocfs2_lock_res
*res
,
420 struct ocfs2_dlm_debug
*dlm_debug
)
422 mlog(0, "Add tracking for lockres %s\n", res
->l_name
);
424 spin_lock(&ocfs2_dlm_tracking_lock
);
425 list_add(&res
->l_debug_list
, &dlm_debug
->d_lockres_tracking
);
426 spin_unlock(&ocfs2_dlm_tracking_lock
);
429 static void ocfs2_remove_lockres_tracking(struct ocfs2_lock_res
*res
)
431 spin_lock(&ocfs2_dlm_tracking_lock
);
432 if (!list_empty(&res
->l_debug_list
))
433 list_del_init(&res
->l_debug_list
);
434 spin_unlock(&ocfs2_dlm_tracking_lock
);
437 #ifdef CONFIG_OCFS2_FS_STATS
438 static void ocfs2_init_lock_stats(struct ocfs2_lock_res
*res
)
440 res
->l_lock_refresh
= 0;
441 memset(&res
->l_lock_prmode
, 0, sizeof(struct ocfs2_lock_stats
));
442 memset(&res
->l_lock_exmode
, 0, sizeof(struct ocfs2_lock_stats
));
445 static void ocfs2_update_lock_stats(struct ocfs2_lock_res
*res
, int level
,
446 struct ocfs2_mask_waiter
*mw
, int ret
)
450 struct ocfs2_lock_stats
*stats
;
452 if (level
== LKM_PRMODE
)
453 stats
= &res
->l_lock_prmode
;
454 else if (level
== LKM_EXMODE
)
455 stats
= &res
->l_lock_exmode
;
459 kt
= ktime_sub(ktime_get(), mw
->mw_lock_start
);
460 usec
= ktime_to_us(kt
);
463 stats
->ls_total
+= ktime_to_ns(kt
);
465 if (unlikely(stats
->ls_gets
== 0)) {
467 stats
->ls_total
= ktime_to_ns(kt
);
470 if (stats
->ls_max
< usec
)
471 stats
->ls_max
= usec
;
477 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res
*lockres
)
479 lockres
->l_lock_refresh
++;
482 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter
*mw
)
484 mw
->mw_lock_start
= ktime_get();
487 static inline void ocfs2_init_lock_stats(struct ocfs2_lock_res
*res
)
490 static inline void ocfs2_update_lock_stats(struct ocfs2_lock_res
*res
,
491 int level
, struct ocfs2_mask_waiter
*mw
, int ret
)
494 static inline void ocfs2_track_lock_refresh(struct ocfs2_lock_res
*lockres
)
497 static inline void ocfs2_init_start_time(struct ocfs2_mask_waiter
*mw
)
502 static void ocfs2_lock_res_init_common(struct ocfs2_super
*osb
,
503 struct ocfs2_lock_res
*res
,
504 enum ocfs2_lock_type type
,
505 struct ocfs2_lock_res_ops
*ops
,
512 res
->l_level
= DLM_LOCK_IV
;
513 res
->l_requested
= DLM_LOCK_IV
;
514 res
->l_blocking
= DLM_LOCK_IV
;
515 res
->l_action
= OCFS2_AST_INVALID
;
516 res
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
518 res
->l_flags
= OCFS2_LOCK_INITIALIZED
;
520 ocfs2_add_lockres_tracking(res
, osb
->osb_dlm_debug
);
522 ocfs2_init_lock_stats(res
);
523 #ifdef CONFIG_DEBUG_LOCK_ALLOC
524 if (type
!= OCFS2_LOCK_TYPE_OPEN
)
525 lockdep_init_map(&res
->l_lockdep_map
, ocfs2_lock_type_strings
[type
],
526 &lockdep_keys
[type
], 0);
528 res
->l_lockdep_map
.key
= NULL
;
532 void ocfs2_lock_res_init_once(struct ocfs2_lock_res
*res
)
534 /* This also clears out the lock status block */
535 memset(res
, 0, sizeof(struct ocfs2_lock_res
));
536 spin_lock_init(&res
->l_lock
);
537 init_waitqueue_head(&res
->l_event
);
538 INIT_LIST_HEAD(&res
->l_blocked_list
);
539 INIT_LIST_HEAD(&res
->l_mask_waiters
);
540 INIT_LIST_HEAD(&res
->l_holders
);
543 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res
*res
,
544 enum ocfs2_lock_type type
,
545 unsigned int generation
,
548 struct ocfs2_lock_res_ops
*ops
;
551 case OCFS2_LOCK_TYPE_RW
:
552 ops
= &ocfs2_inode_rw_lops
;
554 case OCFS2_LOCK_TYPE_META
:
555 ops
= &ocfs2_inode_inode_lops
;
557 case OCFS2_LOCK_TYPE_OPEN
:
558 ops
= &ocfs2_inode_open_lops
;
561 mlog_bug_on_msg(1, "type: %d\n", type
);
562 ops
= NULL
; /* thanks, gcc */
566 ocfs2_build_lock_name(type
, OCFS2_I(inode
)->ip_blkno
,
567 generation
, res
->l_name
);
568 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), res
, type
, ops
, inode
);
571 static struct ocfs2_super
*ocfs2_get_inode_osb(struct ocfs2_lock_res
*lockres
)
573 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
575 return OCFS2_SB(inode
->i_sb
);
578 static struct ocfs2_super
*ocfs2_get_qinfo_osb(struct ocfs2_lock_res
*lockres
)
580 struct ocfs2_mem_dqinfo
*info
= lockres
->l_priv
;
582 return OCFS2_SB(info
->dqi_gi
.dqi_sb
);
585 static struct ocfs2_super
*ocfs2_get_file_osb(struct ocfs2_lock_res
*lockres
)
587 struct ocfs2_file_private
*fp
= lockres
->l_priv
;
589 return OCFS2_SB(fp
->fp_file
->f_mapping
->host
->i_sb
);
592 static __u64
ocfs2_get_dentry_lock_ino(struct ocfs2_lock_res
*lockres
)
594 __be64 inode_blkno_be
;
596 memcpy(&inode_blkno_be
, &lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
],
599 return be64_to_cpu(inode_blkno_be
);
602 static struct ocfs2_super
*ocfs2_get_dentry_osb(struct ocfs2_lock_res
*lockres
)
604 struct ocfs2_dentry_lock
*dl
= lockres
->l_priv
;
606 return OCFS2_SB(dl
->dl_inode
->i_sb
);
609 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock
*dl
,
610 u64 parent
, struct inode
*inode
)
613 u64 inode_blkno
= OCFS2_I(inode
)->ip_blkno
;
614 __be64 inode_blkno_be
= cpu_to_be64(inode_blkno
);
615 struct ocfs2_lock_res
*lockres
= &dl
->dl_lockres
;
617 ocfs2_lock_res_init_once(lockres
);
620 * Unfortunately, the standard lock naming scheme won't work
621 * here because we have two 16 byte values to use. Instead,
622 * we'll stuff the inode number as a binary value. We still
623 * want error prints to show something without garbling the
624 * display, so drop a null byte in there before the inode
625 * number. A future version of OCFS2 will likely use all
626 * binary lock names. The stringified names have been a
627 * tremendous aid in debugging, but now that the debugfs
628 * interface exists, we can mangle things there if need be.
630 * NOTE: We also drop the standard "pad" value (the total lock
631 * name size stays the same though - the last part is all
632 * zeros due to the memset in ocfs2_lock_res_init_once()
634 len
= snprintf(lockres
->l_name
, OCFS2_DENTRY_LOCK_INO_START
,
636 ocfs2_lock_type_char(OCFS2_LOCK_TYPE_DENTRY
),
639 BUG_ON(len
!= (OCFS2_DENTRY_LOCK_INO_START
- 1));
641 memcpy(&lockres
->l_name
[OCFS2_DENTRY_LOCK_INO_START
], &inode_blkno_be
,
644 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
645 OCFS2_LOCK_TYPE_DENTRY
, &ocfs2_dentry_lops
,
649 static void ocfs2_super_lock_res_init(struct ocfs2_lock_res
*res
,
650 struct ocfs2_super
*osb
)
652 /* Superblock lockres doesn't come from a slab so we call init
653 * once on it manually. */
654 ocfs2_lock_res_init_once(res
);
655 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_SUPER
, OCFS2_SUPER_BLOCK_BLKNO
,
657 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_SUPER
,
658 &ocfs2_super_lops
, osb
);
661 static void ocfs2_rename_lock_res_init(struct ocfs2_lock_res
*res
,
662 struct ocfs2_super
*osb
)
664 /* Rename lockres doesn't come from a slab so we call init
665 * once on it manually. */
666 ocfs2_lock_res_init_once(res
);
667 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_RENAME
, 0, 0, res
->l_name
);
668 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_RENAME
,
669 &ocfs2_rename_lops
, osb
);
672 static void ocfs2_nfs_sync_lock_res_init(struct ocfs2_lock_res
*res
,
673 struct ocfs2_super
*osb
)
675 /* nfs_sync lockres doesn't come from a slab so we call init
676 * once on it manually. */
677 ocfs2_lock_res_init_once(res
);
678 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_NFS_SYNC
, 0, 0, res
->l_name
);
679 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_NFS_SYNC
,
680 &ocfs2_nfs_sync_lops
, osb
);
683 void ocfs2_trim_fs_lock_res_init(struct ocfs2_super
*osb
)
685 struct ocfs2_lock_res
*lockres
= &osb
->osb_trim_fs_lockres
;
687 ocfs2_lock_res_init_once(lockres
);
688 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_TRIM_FS
, 0, 0, lockres
->l_name
);
689 ocfs2_lock_res_init_common(osb
, lockres
, OCFS2_LOCK_TYPE_TRIM_FS
,
690 &ocfs2_trim_fs_lops
, osb
);
693 void ocfs2_trim_fs_lock_res_uninit(struct ocfs2_super
*osb
)
695 struct ocfs2_lock_res
*lockres
= &osb
->osb_trim_fs_lockres
;
697 ocfs2_simple_drop_lockres(osb
, lockres
);
698 ocfs2_lock_res_free(lockres
);
701 static void ocfs2_orphan_scan_lock_res_init(struct ocfs2_lock_res
*res
,
702 struct ocfs2_super
*osb
)
704 ocfs2_lock_res_init_once(res
);
705 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_ORPHAN_SCAN
, 0, 0, res
->l_name
);
706 ocfs2_lock_res_init_common(osb
, res
, OCFS2_LOCK_TYPE_ORPHAN_SCAN
,
707 &ocfs2_orphan_scan_lops
, osb
);
710 void ocfs2_file_lock_res_init(struct ocfs2_lock_res
*lockres
,
711 struct ocfs2_file_private
*fp
)
713 struct inode
*inode
= fp
->fp_file
->f_mapping
->host
;
714 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
716 ocfs2_lock_res_init_once(lockres
);
717 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_FLOCK
, oi
->ip_blkno
,
718 inode
->i_generation
, lockres
->l_name
);
719 ocfs2_lock_res_init_common(OCFS2_SB(inode
->i_sb
), lockres
,
720 OCFS2_LOCK_TYPE_FLOCK
, &ocfs2_flock_lops
,
722 lockres
->l_flags
|= OCFS2_LOCK_NOCACHE
;
725 void ocfs2_qinfo_lock_res_init(struct ocfs2_lock_res
*lockres
,
726 struct ocfs2_mem_dqinfo
*info
)
728 ocfs2_lock_res_init_once(lockres
);
729 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_QINFO
, info
->dqi_gi
.dqi_type
,
731 ocfs2_lock_res_init_common(OCFS2_SB(info
->dqi_gi
.dqi_sb
), lockres
,
732 OCFS2_LOCK_TYPE_QINFO
, &ocfs2_qinfo_lops
,
736 void ocfs2_refcount_lock_res_init(struct ocfs2_lock_res
*lockres
,
737 struct ocfs2_super
*osb
, u64 ref_blkno
,
738 unsigned int generation
)
740 ocfs2_lock_res_init_once(lockres
);
741 ocfs2_build_lock_name(OCFS2_LOCK_TYPE_REFCOUNT
, ref_blkno
,
742 generation
, lockres
->l_name
);
743 ocfs2_lock_res_init_common(osb
, lockres
, OCFS2_LOCK_TYPE_REFCOUNT
,
744 &ocfs2_refcount_block_lops
, osb
);
747 void ocfs2_lock_res_free(struct ocfs2_lock_res
*res
)
749 if (!(res
->l_flags
& OCFS2_LOCK_INITIALIZED
))
752 ocfs2_remove_lockres_tracking(res
);
754 mlog_bug_on_msg(!list_empty(&res
->l_blocked_list
),
755 "Lockres %s is on the blocked list\n",
757 mlog_bug_on_msg(!list_empty(&res
->l_mask_waiters
),
758 "Lockres %s has mask waiters pending\n",
760 mlog_bug_on_msg(spin_is_locked(&res
->l_lock
),
761 "Lockres %s is locked\n",
763 mlog_bug_on_msg(res
->l_ro_holders
,
764 "Lockres %s has %u ro holders\n",
765 res
->l_name
, res
->l_ro_holders
);
766 mlog_bug_on_msg(res
->l_ex_holders
,
767 "Lockres %s has %u ex holders\n",
768 res
->l_name
, res
->l_ex_holders
);
770 /* Need to clear out the lock status block for the dlm */
771 memset(&res
->l_lksb
, 0, sizeof(res
->l_lksb
));
777 * Keep a list of processes who have interest in a lockres.
778 * Note: this is now only uesed for check recursive cluster locking.
780 static inline void ocfs2_add_holder(struct ocfs2_lock_res
*lockres
,
781 struct ocfs2_lock_holder
*oh
)
783 INIT_LIST_HEAD(&oh
->oh_list
);
784 oh
->oh_owner_pid
= get_pid(task_pid(current
));
786 spin_lock(&lockres
->l_lock
);
787 list_add_tail(&oh
->oh_list
, &lockres
->l_holders
);
788 spin_unlock(&lockres
->l_lock
);
791 static struct ocfs2_lock_holder
*
792 ocfs2_pid_holder(struct ocfs2_lock_res
*lockres
,
795 struct ocfs2_lock_holder
*oh
;
797 spin_lock(&lockres
->l_lock
);
798 list_for_each_entry(oh
, &lockres
->l_holders
, oh_list
) {
799 if (oh
->oh_owner_pid
== pid
) {
800 spin_unlock(&lockres
->l_lock
);
804 spin_unlock(&lockres
->l_lock
);
808 static inline void ocfs2_remove_holder(struct ocfs2_lock_res
*lockres
,
809 struct ocfs2_lock_holder
*oh
)
811 spin_lock(&lockres
->l_lock
);
812 list_del(&oh
->oh_list
);
813 spin_unlock(&lockres
->l_lock
);
815 put_pid(oh
->oh_owner_pid
);
819 static inline void ocfs2_inc_holders(struct ocfs2_lock_res
*lockres
,
826 lockres
->l_ex_holders
++;
829 lockres
->l_ro_holders
++;
836 static inline void ocfs2_dec_holders(struct ocfs2_lock_res
*lockres
,
843 BUG_ON(!lockres
->l_ex_holders
);
844 lockres
->l_ex_holders
--;
847 BUG_ON(!lockres
->l_ro_holders
);
848 lockres
->l_ro_holders
--;
855 /* WARNING: This function lives in a world where the only three lock
856 * levels are EX, PR, and NL. It *will* have to be adjusted when more
857 * lock types are added. */
858 static inline int ocfs2_highest_compat_lock_level(int level
)
860 int new_level
= DLM_LOCK_EX
;
862 if (level
== DLM_LOCK_EX
)
863 new_level
= DLM_LOCK_NL
;
864 else if (level
== DLM_LOCK_PR
)
865 new_level
= DLM_LOCK_PR
;
869 static void lockres_set_flags(struct ocfs2_lock_res
*lockres
,
870 unsigned long newflags
)
872 struct ocfs2_mask_waiter
*mw
, *tmp
;
874 assert_spin_locked(&lockres
->l_lock
);
876 lockres
->l_flags
= newflags
;
878 list_for_each_entry_safe(mw
, tmp
, &lockres
->l_mask_waiters
, mw_item
) {
879 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
882 list_del_init(&mw
->mw_item
);
884 complete(&mw
->mw_complete
);
887 static void lockres_or_flags(struct ocfs2_lock_res
*lockres
, unsigned long or)
889 lockres_set_flags(lockres
, lockres
->l_flags
| or);
891 static void lockres_clear_flags(struct ocfs2_lock_res
*lockres
,
894 lockres_set_flags(lockres
, lockres
->l_flags
& ~clear
);
897 static inline void ocfs2_generic_handle_downconvert_action(struct ocfs2_lock_res
*lockres
)
899 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
900 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
901 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
902 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
904 lockres
->l_level
= lockres
->l_requested
;
905 if (lockres
->l_level
<=
906 ocfs2_highest_compat_lock_level(lockres
->l_blocking
)) {
907 lockres
->l_blocking
= DLM_LOCK_NL
;
908 lockres_clear_flags(lockres
, OCFS2_LOCK_BLOCKED
);
910 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
913 static inline void ocfs2_generic_handle_convert_action(struct ocfs2_lock_res
*lockres
)
915 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
916 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
));
918 /* Convert from RO to EX doesn't really need anything as our
919 * information is already up to data. Convert from NL to
920 * *anything* however should mark ourselves as needing an
922 if (lockres
->l_level
== DLM_LOCK_NL
&&
923 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
924 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
926 lockres
->l_level
= lockres
->l_requested
;
929 * We set the OCFS2_LOCK_UPCONVERT_FINISHING flag before clearing
930 * the OCFS2_LOCK_BUSY flag to prevent the dc thread from
931 * downconverting the lock before the upconvert has fully completed.
932 * Do not prevent the dc thread from downconverting if NONBLOCK lock
933 * had already returned.
935 if (!(lockres
->l_flags
& OCFS2_LOCK_NONBLOCK_FINISHED
))
936 lockres_or_flags(lockres
, OCFS2_LOCK_UPCONVERT_FINISHING
);
938 lockres_clear_flags(lockres
, OCFS2_LOCK_NONBLOCK_FINISHED
);
940 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
943 static inline void ocfs2_generic_handle_attach_action(struct ocfs2_lock_res
*lockres
)
945 BUG_ON((!(lockres
->l_flags
& OCFS2_LOCK_BUSY
)));
946 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
948 if (lockres
->l_requested
> DLM_LOCK_NL
&&
949 !(lockres
->l_flags
& OCFS2_LOCK_LOCAL
) &&
950 lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
951 lockres_or_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
953 lockres
->l_level
= lockres
->l_requested
;
954 lockres_or_flags(lockres
, OCFS2_LOCK_ATTACHED
);
955 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
958 static int ocfs2_generic_handle_bast(struct ocfs2_lock_res
*lockres
,
961 int needs_downconvert
= 0;
963 assert_spin_locked(&lockres
->l_lock
);
965 if (level
> lockres
->l_blocking
) {
966 /* only schedule a downconvert if we haven't already scheduled
967 * one that goes low enough to satisfy the level we're
968 * blocking. this also catches the case where we get
970 if (ocfs2_highest_compat_lock_level(level
) <
971 ocfs2_highest_compat_lock_level(lockres
->l_blocking
))
972 needs_downconvert
= 1;
974 lockres
->l_blocking
= level
;
977 mlog(ML_BASTS
, "lockres %s, block %d, level %d, l_block %d, dwn %d\n",
978 lockres
->l_name
, level
, lockres
->l_level
, lockres
->l_blocking
,
981 if (needs_downconvert
)
982 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
983 mlog(0, "needs_downconvert = %d\n", needs_downconvert
);
984 return needs_downconvert
;
988 * OCFS2_LOCK_PENDING and l_pending_gen.
990 * Why does OCFS2_LOCK_PENDING exist? To close a race between setting
991 * OCFS2_LOCK_BUSY and calling ocfs2_dlm_lock(). See ocfs2_unblock_lock()
992 * for more details on the race.
994 * OCFS2_LOCK_PENDING closes the race quite nicely. However, it introduces
995 * a race on itself. In o2dlm, we can get the ast before ocfs2_dlm_lock()
996 * returns. The ast clears OCFS2_LOCK_BUSY, and must therefore clear
997 * OCFS2_LOCK_PENDING at the same time. When ocfs2_dlm_lock() returns,
998 * the caller is going to try to clear PENDING again. If nothing else is
999 * happening, __lockres_clear_pending() sees PENDING is unset and does
1002 * But what if another path (eg downconvert thread) has just started a
1003 * new locking action? The other path has re-set PENDING. Our path
1004 * cannot clear PENDING, because that will re-open the original race
1010 * ocfs2_cluster_lock()
1015 * ocfs2_locking_ast() ocfs2_downconvert_thread()
1016 * clear PENDING ocfs2_unblock_lock()
1019 * ocfs2_prepare_downconvert()
1029 * So as you can see, we now have a window where l_lock is not held,
1030 * PENDING is not set, and ocfs2_dlm_lock() has not been called.
1032 * The core problem is that ocfs2_cluster_lock() has cleared the PENDING
1033 * set by ocfs2_prepare_downconvert(). That wasn't nice.
1035 * To solve this we introduce l_pending_gen. A call to
1036 * lockres_clear_pending() will only do so when it is passed a generation
1037 * number that matches the lockres. lockres_set_pending() will return the
1038 * current generation number. When ocfs2_cluster_lock() goes to clear
1039 * PENDING, it passes the generation it got from set_pending(). In our
1040 * example above, the generation numbers will *not* match. Thus,
1041 * ocfs2_cluster_lock() will not clear the PENDING set by
1042 * ocfs2_prepare_downconvert().
1045 /* Unlocked version for ocfs2_locking_ast() */
1046 static void __lockres_clear_pending(struct ocfs2_lock_res
*lockres
,
1047 unsigned int generation
,
1048 struct ocfs2_super
*osb
)
1050 assert_spin_locked(&lockres
->l_lock
);
1053 * The ast and locking functions can race us here. The winner
1054 * will clear pending, the loser will not.
1056 if (!(lockres
->l_flags
& OCFS2_LOCK_PENDING
) ||
1057 (lockres
->l_pending_gen
!= generation
))
1060 lockres_clear_flags(lockres
, OCFS2_LOCK_PENDING
);
1061 lockres
->l_pending_gen
++;
1064 * The downconvert thread may have skipped us because we
1065 * were PENDING. Wake it up.
1067 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
1068 ocfs2_wake_downconvert_thread(osb
);
1071 /* Locked version for callers of ocfs2_dlm_lock() */
1072 static void lockres_clear_pending(struct ocfs2_lock_res
*lockres
,
1073 unsigned int generation
,
1074 struct ocfs2_super
*osb
)
1076 unsigned long flags
;
1078 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1079 __lockres_clear_pending(lockres
, generation
, osb
);
1080 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1083 static unsigned int lockres_set_pending(struct ocfs2_lock_res
*lockres
)
1085 assert_spin_locked(&lockres
->l_lock
);
1086 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
));
1088 lockres_or_flags(lockres
, OCFS2_LOCK_PENDING
);
1090 return lockres
->l_pending_gen
;
1093 static void ocfs2_blocking_ast(struct ocfs2_dlm_lksb
*lksb
, int level
)
1095 struct ocfs2_lock_res
*lockres
= ocfs2_lksb_to_lock_res(lksb
);
1096 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
1097 int needs_downconvert
;
1098 unsigned long flags
;
1100 BUG_ON(level
<= DLM_LOCK_NL
);
1102 mlog(ML_BASTS
, "BAST fired for lockres %s, blocking %d, level %d, "
1103 "type %s\n", lockres
->l_name
, level
, lockres
->l_level
,
1104 ocfs2_lock_type_string(lockres
->l_type
));
1107 * We can skip the bast for locks which don't enable caching -
1108 * they'll be dropped at the earliest possible time anyway.
1110 if (lockres
->l_flags
& OCFS2_LOCK_NOCACHE
)
1113 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1114 needs_downconvert
= ocfs2_generic_handle_bast(lockres
, level
);
1115 if (needs_downconvert
)
1116 ocfs2_schedule_blocked_lock(osb
, lockres
);
1117 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1119 wake_up(&lockres
->l_event
);
1121 ocfs2_wake_downconvert_thread(osb
);
1124 static void ocfs2_locking_ast(struct ocfs2_dlm_lksb
*lksb
)
1126 struct ocfs2_lock_res
*lockres
= ocfs2_lksb_to_lock_res(lksb
);
1127 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
1128 unsigned long flags
;
1131 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1133 status
= ocfs2_dlm_lock_status(&lockres
->l_lksb
);
1135 if (status
== -EAGAIN
) {
1136 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
1141 mlog(ML_ERROR
, "lockres %s: lksb status value of %d!\n",
1142 lockres
->l_name
, status
);
1143 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1147 mlog(ML_BASTS
, "AST fired for lockres %s, action %d, unlock %d, "
1148 "level %d => %d\n", lockres
->l_name
, lockres
->l_action
,
1149 lockres
->l_unlock_action
, lockres
->l_level
, lockres
->l_requested
);
1151 switch(lockres
->l_action
) {
1152 case OCFS2_AST_ATTACH
:
1153 ocfs2_generic_handle_attach_action(lockres
);
1154 lockres_clear_flags(lockres
, OCFS2_LOCK_LOCAL
);
1156 case OCFS2_AST_CONVERT
:
1157 ocfs2_generic_handle_convert_action(lockres
);
1159 case OCFS2_AST_DOWNCONVERT
:
1160 ocfs2_generic_handle_downconvert_action(lockres
);
1163 mlog(ML_ERROR
, "lockres %s: AST fired with invalid action: %u, "
1164 "flags 0x%lx, unlock: %u\n",
1165 lockres
->l_name
, lockres
->l_action
, lockres
->l_flags
,
1166 lockres
->l_unlock_action
);
1170 /* set it to something invalid so if we get called again we
1172 lockres
->l_action
= OCFS2_AST_INVALID
;
1174 /* Did we try to cancel this lock? Clear that state */
1175 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
)
1176 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
1179 * We may have beaten the locking functions here. We certainly
1180 * know that dlm_lock() has been called :-)
1181 * Because we can't have two lock calls in flight at once, we
1182 * can use lockres->l_pending_gen.
1184 __lockres_clear_pending(lockres
, lockres
->l_pending_gen
, osb
);
1186 wake_up(&lockres
->l_event
);
1187 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1190 static void ocfs2_unlock_ast(struct ocfs2_dlm_lksb
*lksb
, int error
)
1192 struct ocfs2_lock_res
*lockres
= ocfs2_lksb_to_lock_res(lksb
);
1193 unsigned long flags
;
1195 mlog(ML_BASTS
, "UNLOCK AST fired for lockres %s, action = %d\n",
1196 lockres
->l_name
, lockres
->l_unlock_action
);
1198 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1200 mlog(ML_ERROR
, "Dlm passes error %d for lock %s, "
1201 "unlock_action %d\n", error
, lockres
->l_name
,
1202 lockres
->l_unlock_action
);
1203 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1207 switch(lockres
->l_unlock_action
) {
1208 case OCFS2_UNLOCK_CANCEL_CONVERT
:
1209 mlog(0, "Cancel convert success for %s\n", lockres
->l_name
);
1210 lockres
->l_action
= OCFS2_AST_INVALID
;
1211 /* Downconvert thread may have requeued this lock, we
1212 * need to wake it. */
1213 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
1214 ocfs2_wake_downconvert_thread(ocfs2_get_lockres_osb(lockres
));
1216 case OCFS2_UNLOCK_DROP_LOCK
:
1217 lockres
->l_level
= DLM_LOCK_IV
;
1223 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
1224 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
1225 wake_up(&lockres
->l_event
);
1226 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1230 * This is the filesystem locking protocol. It provides the lock handling
1231 * hooks for the underlying DLM. It has a maximum version number.
1232 * The version number allows interoperability with systems running at
1233 * the same major number and an equal or smaller minor number.
1235 * Whenever the filesystem does new things with locks (adds or removes a
1236 * lock, orders them differently, does different things underneath a lock),
1237 * the version must be changed. The protocol is negotiated when joining
1238 * the dlm domain. A node may join the domain if its major version is
1239 * identical to all other nodes and its minor version is greater than
1240 * or equal to all other nodes. When its minor version is greater than
1241 * the other nodes, it will run at the minor version specified by the
1244 * If a locking change is made that will not be compatible with older
1245 * versions, the major number must be increased and the minor version set
1246 * to zero. If a change merely adds a behavior that can be disabled when
1247 * speaking to older versions, the minor version must be increased. If a
1248 * change adds a fully backwards compatible change (eg, LVB changes that
1249 * are just ignored by older versions), the version does not need to be
1252 static struct ocfs2_locking_protocol lproto
= {
1254 .pv_major
= OCFS2_LOCKING_PROTOCOL_MAJOR
,
1255 .pv_minor
= OCFS2_LOCKING_PROTOCOL_MINOR
,
1257 .lp_lock_ast
= ocfs2_locking_ast
,
1258 .lp_blocking_ast
= ocfs2_blocking_ast
,
1259 .lp_unlock_ast
= ocfs2_unlock_ast
,
1262 void ocfs2_set_locking_protocol(void)
1264 ocfs2_stack_glue_set_max_proto_version(&lproto
.lp_max_version
);
1267 static inline void ocfs2_recover_from_dlm_error(struct ocfs2_lock_res
*lockres
,
1270 unsigned long flags
;
1272 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1273 lockres_clear_flags(lockres
, OCFS2_LOCK_BUSY
);
1274 lockres_clear_flags(lockres
, OCFS2_LOCK_UPCONVERT_FINISHING
);
1276 lockres
->l_action
= OCFS2_AST_INVALID
;
1278 lockres
->l_unlock_action
= OCFS2_UNLOCK_INVALID
;
1279 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1281 wake_up(&lockres
->l_event
);
1284 /* Note: If we detect another process working on the lock (i.e.,
1285 * OCFS2_LOCK_BUSY), we'll bail out returning 0. It's up to the caller
1286 * to do the right thing in that case.
1288 static int ocfs2_lock_create(struct ocfs2_super
*osb
,
1289 struct ocfs2_lock_res
*lockres
,
1294 unsigned long flags
;
1297 mlog(0, "lock %s, level = %d, flags = %u\n", lockres
->l_name
, level
,
1300 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1301 if ((lockres
->l_flags
& OCFS2_LOCK_ATTACHED
) ||
1302 (lockres
->l_flags
& OCFS2_LOCK_BUSY
)) {
1303 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1307 lockres
->l_action
= OCFS2_AST_ATTACH
;
1308 lockres
->l_requested
= level
;
1309 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1310 gen
= lockres_set_pending(lockres
);
1311 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1313 ret
= ocfs2_dlm_lock(osb
->cconn
,
1318 OCFS2_LOCK_ID_MAX_LEN
- 1);
1319 lockres_clear_pending(lockres
, gen
, osb
);
1321 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
1322 ocfs2_recover_from_dlm_error(lockres
, 1);
1325 mlog(0, "lock %s, return from ocfs2_dlm_lock\n", lockres
->l_name
);
1331 static inline int ocfs2_check_wait_flag(struct ocfs2_lock_res
*lockres
,
1334 unsigned long flags
;
1337 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1338 ret
= lockres
->l_flags
& flag
;
1339 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1344 static inline void ocfs2_wait_on_busy_lock(struct ocfs2_lock_res
*lockres
)
1347 wait_event(lockres
->l_event
,
1348 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_BUSY
));
1351 static inline void ocfs2_wait_on_refreshing_lock(struct ocfs2_lock_res
*lockres
)
1354 wait_event(lockres
->l_event
,
1355 !ocfs2_check_wait_flag(lockres
, OCFS2_LOCK_REFRESHING
));
1358 /* predict what lock level we'll be dropping down to on behalf
1359 * of another node, and return true if the currently wanted
1360 * level will be compatible with it. */
1361 static inline int ocfs2_may_continue_on_blocked_lock(struct ocfs2_lock_res
*lockres
,
1364 BUG_ON(!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
));
1366 return wanted
<= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
1369 static void ocfs2_init_mask_waiter(struct ocfs2_mask_waiter
*mw
)
1371 INIT_LIST_HEAD(&mw
->mw_item
);
1372 init_completion(&mw
->mw_complete
);
1373 ocfs2_init_start_time(mw
);
1376 static int ocfs2_wait_for_mask(struct ocfs2_mask_waiter
*mw
)
1378 wait_for_completion(&mw
->mw_complete
);
1379 /* Re-arm the completion in case we want to wait on it again */
1380 reinit_completion(&mw
->mw_complete
);
1381 return mw
->mw_status
;
1384 static void lockres_add_mask_waiter(struct ocfs2_lock_res
*lockres
,
1385 struct ocfs2_mask_waiter
*mw
,
1389 BUG_ON(!list_empty(&mw
->mw_item
));
1391 assert_spin_locked(&lockres
->l_lock
);
1393 list_add_tail(&mw
->mw_item
, &lockres
->l_mask_waiters
);
1398 /* returns 0 if the mw that was removed was already satisfied, -EBUSY
1399 * if the mask still hadn't reached its goal */
1400 static int __lockres_remove_mask_waiter(struct ocfs2_lock_res
*lockres
,
1401 struct ocfs2_mask_waiter
*mw
)
1405 assert_spin_locked(&lockres
->l_lock
);
1406 if (!list_empty(&mw
->mw_item
)) {
1407 if ((lockres
->l_flags
& mw
->mw_mask
) != mw
->mw_goal
)
1410 list_del_init(&mw
->mw_item
);
1411 init_completion(&mw
->mw_complete
);
1417 static int lockres_remove_mask_waiter(struct ocfs2_lock_res
*lockres
,
1418 struct ocfs2_mask_waiter
*mw
)
1420 unsigned long flags
;
1423 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1424 ret
= __lockres_remove_mask_waiter(lockres
, mw
);
1425 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1431 static int ocfs2_wait_for_mask_interruptible(struct ocfs2_mask_waiter
*mw
,
1432 struct ocfs2_lock_res
*lockres
)
1436 ret
= wait_for_completion_interruptible(&mw
->mw_complete
);
1438 lockres_remove_mask_waiter(lockres
, mw
);
1440 ret
= mw
->mw_status
;
1441 /* Re-arm the completion in case we want to wait on it again */
1442 reinit_completion(&mw
->mw_complete
);
1446 static int __ocfs2_cluster_lock(struct ocfs2_super
*osb
,
1447 struct ocfs2_lock_res
*lockres
,
1452 unsigned long caller_ip
)
1454 struct ocfs2_mask_waiter mw
;
1455 int wait
, catch_signals
= !(osb
->s_mount_opt
& OCFS2_MOUNT_NOINTR
);
1456 int ret
= 0; /* gcc doesn't realize wait = 1 guarantees ret is set */
1457 unsigned long flags
;
1459 int noqueue_attempted
= 0;
1463 if (!(lockres
->l_flags
& OCFS2_LOCK_INITIALIZED
)) {
1464 mlog_errno(-EINVAL
);
1468 ocfs2_init_mask_waiter(&mw
);
1470 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
1471 lkm_flags
|= DLM_LKF_VALBLK
;
1476 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1478 if (catch_signals
&& signal_pending(current
)) {
1483 mlog_bug_on_msg(lockres
->l_flags
& OCFS2_LOCK_FREEING
,
1484 "Cluster lock called on freeing lockres %s! flags "
1485 "0x%lx\n", lockres
->l_name
, lockres
->l_flags
);
1487 /* We only compare against the currently granted level
1488 * here. If the lock is blocked waiting on a downconvert,
1489 * we'll get caught below. */
1490 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
&&
1491 level
> lockres
->l_level
) {
1492 /* is someone sitting in dlm_lock? If so, wait on
1494 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1499 if (lockres
->l_flags
& OCFS2_LOCK_UPCONVERT_FINISHING
) {
1501 * We've upconverted. If the lock now has a level we can
1502 * work with, we take it. If, however, the lock is not at the
1503 * required level, we go thru the full cycle. One way this could
1504 * happen is if a process requesting an upconvert to PR is
1505 * closely followed by another requesting upconvert to an EX.
1506 * If the process requesting EX lands here, we want it to
1507 * continue attempting to upconvert and let the process
1508 * requesting PR take the lock.
1509 * If multiple processes request upconvert to PR, the first one
1510 * here will take the lock. The others will have to go thru the
1511 * OCFS2_LOCK_BLOCKED check to ensure that there is no pending
1512 * downconvert request.
1514 if (level
<= lockres
->l_level
)
1515 goto update_holders
;
1518 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
&&
1519 !ocfs2_may_continue_on_blocked_lock(lockres
, level
)) {
1520 /* is the lock is currently blocked on behalf of
1522 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BLOCKED
, 0);
1527 if (level
> lockres
->l_level
) {
1528 if (noqueue_attempted
> 0) {
1532 if (lkm_flags
& DLM_LKF_NOQUEUE
)
1533 noqueue_attempted
= 1;
1535 if (lockres
->l_action
!= OCFS2_AST_INVALID
)
1536 mlog(ML_ERROR
, "lockres %s has action %u pending\n",
1537 lockres
->l_name
, lockres
->l_action
);
1539 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1540 lockres
->l_action
= OCFS2_AST_ATTACH
;
1541 lkm_flags
&= ~DLM_LKF_CONVERT
;
1543 lockres
->l_action
= OCFS2_AST_CONVERT
;
1544 lkm_flags
|= DLM_LKF_CONVERT
;
1547 lockres
->l_requested
= level
;
1548 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
1549 gen
= lockres_set_pending(lockres
);
1550 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1552 BUG_ON(level
== DLM_LOCK_IV
);
1553 BUG_ON(level
== DLM_LOCK_NL
);
1555 mlog(ML_BASTS
, "lockres %s, convert from %d to %d\n",
1556 lockres
->l_name
, lockres
->l_level
, level
);
1558 /* call dlm_lock to upgrade lock now */
1559 ret
= ocfs2_dlm_lock(osb
->cconn
,
1564 OCFS2_LOCK_ID_MAX_LEN
- 1);
1565 lockres_clear_pending(lockres
, gen
, osb
);
1567 if (!(lkm_flags
& DLM_LKF_NOQUEUE
) ||
1569 ocfs2_log_dlm_error("ocfs2_dlm_lock",
1572 ocfs2_recover_from_dlm_error(lockres
, 1);
1577 mlog(0, "lock %s, successful return from ocfs2_dlm_lock\n",
1580 /* At this point we've gone inside the dlm and need to
1581 * complete our work regardless. */
1584 /* wait for busy to clear and carry on */
1589 /* Ok, if we get here then we're good to go. */
1590 ocfs2_inc_holders(lockres
, level
);
1594 lockres_clear_flags(lockres
, OCFS2_LOCK_UPCONVERT_FINISHING
);
1596 /* ocfs2_unblock_lock reques on seeing OCFS2_LOCK_UPCONVERT_FINISHING */
1597 kick_dc
= (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
);
1599 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1601 ocfs2_wake_downconvert_thread(osb
);
1604 * This is helping work around a lock inversion between the page lock
1605 * and dlm locks. One path holds the page lock while calling aops
1606 * which block acquiring dlm locks. The voting thread holds dlm
1607 * locks while acquiring page locks while down converting data locks.
1608 * This block is helping an aop path notice the inversion and back
1609 * off to unlock its page lock before trying the dlm lock again.
1611 if (wait
&& arg_flags
& OCFS2_LOCK_NONBLOCK
&&
1612 mw
.mw_mask
& (OCFS2_LOCK_BUSY
|OCFS2_LOCK_BLOCKED
)) {
1614 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1615 if (__lockres_remove_mask_waiter(lockres
, &mw
)) {
1617 lockres_or_flags(lockres
,
1618 OCFS2_LOCK_NONBLOCK_FINISHED
);
1619 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1622 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1627 ret
= ocfs2_wait_for_mask(&mw
);
1632 ocfs2_update_lock_stats(lockres
, level
, &mw
, ret
);
1634 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1635 if (!ret
&& lockres
->l_lockdep_map
.key
!= NULL
) {
1636 if (level
== DLM_LOCK_PR
)
1637 rwsem_acquire_read(&lockres
->l_lockdep_map
, l_subclass
,
1638 !!(arg_flags
& OCFS2_META_LOCK_NOQUEUE
),
1641 rwsem_acquire(&lockres
->l_lockdep_map
, l_subclass
,
1642 !!(arg_flags
& OCFS2_META_LOCK_NOQUEUE
),
1649 static inline int ocfs2_cluster_lock(struct ocfs2_super
*osb
,
1650 struct ocfs2_lock_res
*lockres
,
1655 return __ocfs2_cluster_lock(osb
, lockres
, level
, lkm_flags
, arg_flags
,
1660 static void __ocfs2_cluster_unlock(struct ocfs2_super
*osb
,
1661 struct ocfs2_lock_res
*lockres
,
1663 unsigned long caller_ip
)
1665 unsigned long flags
;
1667 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1668 ocfs2_dec_holders(lockres
, level
);
1669 ocfs2_downconvert_on_unlock(osb
, lockres
);
1670 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1671 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1672 if (lockres
->l_lockdep_map
.key
!= NULL
)
1673 rwsem_release(&lockres
->l_lockdep_map
, 1, caller_ip
);
1677 static int ocfs2_create_new_lock(struct ocfs2_super
*osb
,
1678 struct ocfs2_lock_res
*lockres
,
1682 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1683 unsigned long flags
;
1684 u32 lkm_flags
= local
? DLM_LKF_LOCAL
: 0;
1686 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1687 BUG_ON(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
);
1688 lockres_or_flags(lockres
, OCFS2_LOCK_LOCAL
);
1689 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1691 return ocfs2_lock_create(osb
, lockres
, level
, lkm_flags
);
1694 /* Grants us an EX lock on the data and metadata resources, skipping
1695 * the normal cluster directory lookup. Use this ONLY on newly created
1696 * inodes which other nodes can't possibly see, and which haven't been
1697 * hashed in the inode hash yet. This can give us a good performance
1698 * increase as it'll skip the network broadcast normally associated
1699 * with creating a new lock resource. */
1700 int ocfs2_create_new_inode_locks(struct inode
*inode
)
1703 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1705 BUG_ON(!ocfs2_inode_is_new(inode
));
1707 mlog(0, "Inode %llu\n", (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1709 /* NOTE: That we don't increment any of the holder counts, nor
1710 * do we add anything to a journal handle. Since this is
1711 * supposed to be a new inode which the cluster doesn't know
1712 * about yet, there is no need to. As far as the LVB handling
1713 * is concerned, this is basically like acquiring an EX lock
1714 * on a resource which has an invalid one -- we'll set it
1715 * valid when we release the EX. */
1717 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_rw_lockres
, 1, 1);
1724 * We don't want to use DLM_LKF_LOCAL on a meta data lock as they
1725 * don't use a generation in their lock names.
1727 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_inode_lockres
, 1, 0);
1733 ret
= ocfs2_create_new_lock(osb
, &OCFS2_I(inode
)->ip_open_lockres
, 0, 0);
1741 int ocfs2_rw_lock(struct inode
*inode
, int write
)
1744 struct ocfs2_lock_res
*lockres
;
1745 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1747 mlog(0, "inode %llu take %s RW lock\n",
1748 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1749 write
? "EXMODE" : "PRMODE");
1751 if (ocfs2_mount_local(osb
))
1754 lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1756 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1758 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
1765 int ocfs2_try_rw_lock(struct inode
*inode
, int write
)
1768 struct ocfs2_lock_res
*lockres
;
1769 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1771 mlog(0, "inode %llu try to take %s RW lock\n",
1772 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1773 write
? "EXMODE" : "PRMODE");
1775 if (ocfs2_mount_local(osb
))
1778 lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1780 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1782 status
= ocfs2_cluster_lock(osb
, lockres
, level
, DLM_LKF_NOQUEUE
, 0);
1786 void ocfs2_rw_unlock(struct inode
*inode
, int write
)
1788 int level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1789 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_rw_lockres
;
1790 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1792 mlog(0, "inode %llu drop %s RW lock\n",
1793 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1794 write
? "EXMODE" : "PRMODE");
1796 if (!ocfs2_mount_local(osb
))
1797 ocfs2_cluster_unlock(osb
, lockres
, level
);
1801 * ocfs2_open_lock always get PR mode lock.
1803 int ocfs2_open_lock(struct inode
*inode
)
1806 struct ocfs2_lock_res
*lockres
;
1807 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1809 mlog(0, "inode %llu take PRMODE open lock\n",
1810 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1812 if (ocfs2_is_hard_readonly(osb
) || ocfs2_mount_local(osb
))
1815 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1817 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_PR
, 0, 0);
1825 int ocfs2_try_open_lock(struct inode
*inode
, int write
)
1827 int status
= 0, level
;
1828 struct ocfs2_lock_res
*lockres
;
1829 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1831 mlog(0, "inode %llu try to take %s open lock\n",
1832 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
1833 write
? "EXMODE" : "PRMODE");
1835 if (ocfs2_is_hard_readonly(osb
)) {
1841 if (ocfs2_mount_local(osb
))
1844 lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1846 level
= write
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1849 * The file system may already holding a PRMODE/EXMODE open lock.
1850 * Since we pass DLM_LKF_NOQUEUE, the request won't block waiting on
1851 * other nodes and the -EAGAIN will indicate to the caller that
1852 * this inode is still in use.
1854 status
= ocfs2_cluster_lock(osb
, lockres
, level
, DLM_LKF_NOQUEUE
, 0);
1861 * ocfs2_open_unlock unlock PR and EX mode open locks.
1863 void ocfs2_open_unlock(struct inode
*inode
)
1865 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_open_lockres
;
1866 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
1868 mlog(0, "inode %llu drop open lock\n",
1869 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
1871 if (ocfs2_mount_local(osb
))
1874 if(lockres
->l_ro_holders
)
1875 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_PR
);
1876 if(lockres
->l_ex_holders
)
1877 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
1883 static int ocfs2_flock_handle_signal(struct ocfs2_lock_res
*lockres
,
1887 struct ocfs2_super
*osb
= ocfs2_get_lockres_osb(lockres
);
1888 unsigned long flags
;
1889 struct ocfs2_mask_waiter mw
;
1891 ocfs2_init_mask_waiter(&mw
);
1894 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1895 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
1896 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
1898 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1899 ret
= ocfs2_cancel_convert(osb
, lockres
);
1906 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1907 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1909 ocfs2_wait_for_mask(&mw
);
1915 * We may still have gotten the lock, in which case there's no
1916 * point to restarting the syscall.
1918 if (lockres
->l_level
== level
)
1921 mlog(0, "Cancel returning %d. flags: 0x%lx, level: %d, act: %d\n", ret
,
1922 lockres
->l_flags
, lockres
->l_level
, lockres
->l_action
);
1924 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1931 * ocfs2_file_lock() and ocfs2_file_unlock() map to a single pair of
1932 * flock() calls. The locking approach this requires is sufficiently
1933 * different from all other cluster lock types that we implement a
1934 * separate path to the "low-level" dlm calls. In particular:
1936 * - No optimization of lock levels is done - we take at exactly
1937 * what's been requested.
1939 * - No lock caching is employed. We immediately downconvert to
1940 * no-lock at unlock time. This also means flock locks never go on
1941 * the blocking list).
1943 * - Since userspace can trivially deadlock itself with flock, we make
1944 * sure to allow cancellation of a misbehaving applications flock()
1947 * - Access to any flock lockres doesn't require concurrency, so we
1948 * can simplify the code by requiring the caller to guarantee
1949 * serialization of dlmglue flock calls.
1951 int ocfs2_file_lock(struct file
*file
, int ex
, int trylock
)
1953 int ret
, level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
1954 unsigned int lkm_flags
= trylock
? DLM_LKF_NOQUEUE
: 0;
1955 unsigned long flags
;
1956 struct ocfs2_file_private
*fp
= file
->private_data
;
1957 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
1958 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
1959 struct ocfs2_mask_waiter mw
;
1961 ocfs2_init_mask_waiter(&mw
);
1963 if ((lockres
->l_flags
& OCFS2_LOCK_BUSY
) ||
1964 (lockres
->l_level
> DLM_LOCK_NL
)) {
1966 "File lock \"%s\" has busy or locked state: flags: 0x%lx, "
1967 "level: %u\n", lockres
->l_name
, lockres
->l_flags
,
1972 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1973 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
1974 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
1975 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
1978 * Get the lock at NLMODE to start - that way we
1979 * can cancel the upconvert request if need be.
1981 ret
= ocfs2_lock_create(osb
, lockres
, DLM_LOCK_NL
, 0);
1987 ret
= ocfs2_wait_for_mask(&mw
);
1992 spin_lock_irqsave(&lockres
->l_lock
, flags
);
1995 lockres
->l_action
= OCFS2_AST_CONVERT
;
1996 lkm_flags
|= DLM_LKF_CONVERT
;
1997 lockres
->l_requested
= level
;
1998 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
2000 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
2001 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2003 ret
= ocfs2_dlm_lock(osb
->cconn
, level
, &lockres
->l_lksb
, lkm_flags
,
2004 lockres
->l_name
, OCFS2_LOCK_ID_MAX_LEN
- 1);
2006 if (!trylock
|| (ret
!= -EAGAIN
)) {
2007 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
2011 ocfs2_recover_from_dlm_error(lockres
, 1);
2012 lockres_remove_mask_waiter(lockres
, &mw
);
2016 ret
= ocfs2_wait_for_mask_interruptible(&mw
, lockres
);
2017 if (ret
== -ERESTARTSYS
) {
2019 * Userspace can cause deadlock itself with
2020 * flock(). Current behavior locally is to allow the
2021 * deadlock, but abort the system call if a signal is
2022 * received. We follow this example, otherwise a
2023 * poorly written program could sit in kernel until
2026 * Handling this is a bit more complicated for Ocfs2
2027 * though. We can't exit this function with an
2028 * outstanding lock request, so a cancel convert is
2029 * required. We intentionally overwrite 'ret' - if the
2030 * cancel fails and the lock was granted, it's easier
2031 * to just bubble success back up to the user.
2033 ret
= ocfs2_flock_handle_signal(lockres
, level
);
2034 } else if (!ret
&& (level
> lockres
->l_level
)) {
2035 /* Trylock failed asynchronously */
2042 mlog(0, "Lock: \"%s\" ex: %d, trylock: %d, returns: %d\n",
2043 lockres
->l_name
, ex
, trylock
, ret
);
2047 void ocfs2_file_unlock(struct file
*file
)
2051 unsigned long flags
;
2052 struct ocfs2_file_private
*fp
= file
->private_data
;
2053 struct ocfs2_lock_res
*lockres
= &fp
->fp_flock
;
2054 struct ocfs2_super
*osb
= OCFS2_SB(file
->f_mapping
->host
->i_sb
);
2055 struct ocfs2_mask_waiter mw
;
2057 ocfs2_init_mask_waiter(&mw
);
2059 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
))
2062 if (lockres
->l_level
== DLM_LOCK_NL
)
2065 mlog(0, "Unlock: \"%s\" flags: 0x%lx, level: %d, act: %d\n",
2066 lockres
->l_name
, lockres
->l_flags
, lockres
->l_level
,
2069 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2071 * Fake a blocking ast for the downconvert code.
2073 lockres_or_flags(lockres
, OCFS2_LOCK_BLOCKED
);
2074 lockres
->l_blocking
= DLM_LOCK_EX
;
2076 gen
= ocfs2_prepare_downconvert(lockres
, DLM_LOCK_NL
);
2077 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_BUSY
, 0);
2078 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2080 ret
= ocfs2_downconvert_lock(osb
, lockres
, DLM_LOCK_NL
, 0, gen
);
2086 ret
= ocfs2_wait_for_mask(&mw
);
2091 static void ocfs2_downconvert_on_unlock(struct ocfs2_super
*osb
,
2092 struct ocfs2_lock_res
*lockres
)
2096 /* If we know that another node is waiting on our lock, kick
2097 * the downconvert thread * pre-emptively when we reach a release
2099 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
) {
2100 switch(lockres
->l_blocking
) {
2102 if (!lockres
->l_ex_holders
&& !lockres
->l_ro_holders
)
2106 if (!lockres
->l_ex_holders
)
2115 ocfs2_wake_downconvert_thread(osb
);
2118 #define OCFS2_SEC_BITS 34
2119 #define OCFS2_SEC_SHIFT (64 - 34)
2120 #define OCFS2_NSEC_MASK ((1ULL << OCFS2_SEC_SHIFT) - 1)
2122 /* LVB only has room for 64 bits of time here so we pack it for
2124 static u64
ocfs2_pack_timespec(struct timespec
*spec
)
2127 u64 sec
= spec
->tv_sec
;
2128 u32 nsec
= spec
->tv_nsec
;
2130 res
= (sec
<< OCFS2_SEC_SHIFT
) | (nsec
& OCFS2_NSEC_MASK
);
2135 /* Call this with the lockres locked. I am reasonably sure we don't
2136 * need ip_lock in this function as anyone who would be changing those
2137 * values is supposed to be blocked in ocfs2_inode_lock right now. */
2138 static void __ocfs2_stuff_meta_lvb(struct inode
*inode
)
2140 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
2141 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
2142 struct ocfs2_meta_lvb
*lvb
;
2144 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2147 * Invalidate the LVB of a deleted inode - this way other
2148 * nodes are forced to go to disk and discover the new inode
2151 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
2152 lvb
->lvb_version
= 0;
2156 lvb
->lvb_version
= OCFS2_LVB_VERSION
;
2157 lvb
->lvb_isize
= cpu_to_be64(i_size_read(inode
));
2158 lvb
->lvb_iclusters
= cpu_to_be32(oi
->ip_clusters
);
2159 lvb
->lvb_iuid
= cpu_to_be32(i_uid_read(inode
));
2160 lvb
->lvb_igid
= cpu_to_be32(i_gid_read(inode
));
2161 lvb
->lvb_imode
= cpu_to_be16(inode
->i_mode
);
2162 lvb
->lvb_inlink
= cpu_to_be16(inode
->i_nlink
);
2163 lvb
->lvb_iatime_packed
=
2164 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_atime
));
2165 lvb
->lvb_ictime_packed
=
2166 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_ctime
));
2167 lvb
->lvb_imtime_packed
=
2168 cpu_to_be64(ocfs2_pack_timespec(&inode
->i_mtime
));
2169 lvb
->lvb_iattr
= cpu_to_be32(oi
->ip_attr
);
2170 lvb
->lvb_idynfeatures
= cpu_to_be16(oi
->ip_dyn_features
);
2171 lvb
->lvb_igeneration
= cpu_to_be32(inode
->i_generation
);
2174 mlog_meta_lvb(0, lockres
);
2177 static void ocfs2_unpack_timespec(struct timespec
*spec
,
2180 spec
->tv_sec
= packed_time
>> OCFS2_SEC_SHIFT
;
2181 spec
->tv_nsec
= packed_time
& OCFS2_NSEC_MASK
;
2184 static void ocfs2_refresh_inode_from_lvb(struct inode
*inode
)
2186 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
2187 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
2188 struct ocfs2_meta_lvb
*lvb
;
2190 mlog_meta_lvb(0, lockres
);
2192 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2194 /* We're safe here without the lockres lock... */
2195 spin_lock(&oi
->ip_lock
);
2196 oi
->ip_clusters
= be32_to_cpu(lvb
->lvb_iclusters
);
2197 i_size_write(inode
, be64_to_cpu(lvb
->lvb_isize
));
2199 oi
->ip_attr
= be32_to_cpu(lvb
->lvb_iattr
);
2200 oi
->ip_dyn_features
= be16_to_cpu(lvb
->lvb_idynfeatures
);
2201 ocfs2_set_inode_flags(inode
);
2203 /* fast-symlinks are a special case */
2204 if (S_ISLNK(inode
->i_mode
) && !oi
->ip_clusters
)
2205 inode
->i_blocks
= 0;
2207 inode
->i_blocks
= ocfs2_inode_sector_count(inode
);
2209 i_uid_write(inode
, be32_to_cpu(lvb
->lvb_iuid
));
2210 i_gid_write(inode
, be32_to_cpu(lvb
->lvb_igid
));
2211 inode
->i_mode
= be16_to_cpu(lvb
->lvb_imode
);
2212 set_nlink(inode
, be16_to_cpu(lvb
->lvb_inlink
));
2213 ocfs2_unpack_timespec(&inode
->i_atime
,
2214 be64_to_cpu(lvb
->lvb_iatime_packed
));
2215 ocfs2_unpack_timespec(&inode
->i_mtime
,
2216 be64_to_cpu(lvb
->lvb_imtime_packed
));
2217 ocfs2_unpack_timespec(&inode
->i_ctime
,
2218 be64_to_cpu(lvb
->lvb_ictime_packed
));
2219 spin_unlock(&oi
->ip_lock
);
2222 static inline int ocfs2_meta_lvb_is_trustable(struct inode
*inode
,
2223 struct ocfs2_lock_res
*lockres
)
2225 struct ocfs2_meta_lvb
*lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2227 if (ocfs2_dlm_lvb_valid(&lockres
->l_lksb
)
2228 && lvb
->lvb_version
== OCFS2_LVB_VERSION
2229 && be32_to_cpu(lvb
->lvb_igeneration
) == inode
->i_generation
)
2234 /* Determine whether a lock resource needs to be refreshed, and
2235 * arbitrate who gets to refresh it.
2237 * 0 means no refresh needed.
2239 * > 0 means you need to refresh this and you MUST call
2240 * ocfs2_complete_lock_res_refresh afterwards. */
2241 static int ocfs2_should_refresh_lock_res(struct ocfs2_lock_res
*lockres
)
2243 unsigned long flags
;
2247 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2248 if (!(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
)) {
2249 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2253 if (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
) {
2254 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2256 ocfs2_wait_on_refreshing_lock(lockres
);
2260 /* Ok, I'll be the one to refresh this lock. */
2261 lockres_or_flags(lockres
, OCFS2_LOCK_REFRESHING
);
2262 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2266 mlog(0, "status %d\n", status
);
2270 /* If status is non zero, I'll mark it as not being in refresh
2271 * anymroe, but i won't clear the needs refresh flag. */
2272 static inline void ocfs2_complete_lock_res_refresh(struct ocfs2_lock_res
*lockres
,
2275 unsigned long flags
;
2277 spin_lock_irqsave(&lockres
->l_lock
, flags
);
2278 lockres_clear_flags(lockres
, OCFS2_LOCK_REFRESHING
);
2280 lockres_clear_flags(lockres
, OCFS2_LOCK_NEEDS_REFRESH
);
2281 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
2283 wake_up(&lockres
->l_event
);
2286 /* may or may not return a bh if it went to disk. */
2287 static int ocfs2_inode_lock_update(struct inode
*inode
,
2288 struct buffer_head
**bh
)
2291 struct ocfs2_inode_info
*oi
= OCFS2_I(inode
);
2292 struct ocfs2_lock_res
*lockres
= &oi
->ip_inode_lockres
;
2293 struct ocfs2_dinode
*fe
;
2294 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2296 if (ocfs2_mount_local(osb
))
2299 spin_lock(&oi
->ip_lock
);
2300 if (oi
->ip_flags
& OCFS2_INODE_DELETED
) {
2301 mlog(0, "Orphaned inode %llu was deleted while we "
2302 "were waiting on a lock. ip_flags = 0x%x\n",
2303 (unsigned long long)oi
->ip_blkno
, oi
->ip_flags
);
2304 spin_unlock(&oi
->ip_lock
);
2308 spin_unlock(&oi
->ip_lock
);
2310 if (!ocfs2_should_refresh_lock_res(lockres
))
2313 /* This will discard any caching information we might have had
2314 * for the inode metadata. */
2315 ocfs2_metadata_cache_purge(INODE_CACHE(inode
));
2317 ocfs2_extent_map_trunc(inode
, 0);
2319 if (ocfs2_meta_lvb_is_trustable(inode
, lockres
)) {
2320 mlog(0, "Trusting LVB on inode %llu\n",
2321 (unsigned long long)oi
->ip_blkno
);
2322 ocfs2_refresh_inode_from_lvb(inode
);
2324 /* Boo, we have to go to disk. */
2325 /* read bh, cast, ocfs2_refresh_inode */
2326 status
= ocfs2_read_inode_block(inode
, bh
);
2331 fe
= (struct ocfs2_dinode
*) (*bh
)->b_data
;
2333 /* This is a good chance to make sure we're not
2334 * locking an invalid object. ocfs2_read_inode_block()
2335 * already checked that the inode block is sane.
2337 * We bug on a stale inode here because we checked
2338 * above whether it was wiped from disk. The wiping
2339 * node provides a guarantee that we receive that
2340 * message and can mark the inode before dropping any
2341 * locks associated with it. */
2342 mlog_bug_on_msg(inode
->i_generation
!=
2343 le32_to_cpu(fe
->i_generation
),
2344 "Invalid dinode %llu disk generation: %u "
2345 "inode->i_generation: %u\n",
2346 (unsigned long long)oi
->ip_blkno
,
2347 le32_to_cpu(fe
->i_generation
),
2348 inode
->i_generation
);
2349 mlog_bug_on_msg(le64_to_cpu(fe
->i_dtime
) ||
2350 !(fe
->i_flags
& cpu_to_le32(OCFS2_VALID_FL
)),
2351 "Stale dinode %llu dtime: %llu flags: 0x%x\n",
2352 (unsigned long long)oi
->ip_blkno
,
2353 (unsigned long long)le64_to_cpu(fe
->i_dtime
),
2354 le32_to_cpu(fe
->i_flags
));
2356 ocfs2_refresh_inode(inode
, fe
);
2357 ocfs2_track_lock_refresh(lockres
);
2362 ocfs2_complete_lock_res_refresh(lockres
, status
);
2367 static int ocfs2_assign_bh(struct inode
*inode
,
2368 struct buffer_head
**ret_bh
,
2369 struct buffer_head
*passed_bh
)
2374 /* Ok, the update went to disk for us, use the
2376 *ret_bh
= passed_bh
;
2382 status
= ocfs2_read_inode_block(inode
, ret_bh
);
2390 * returns < 0 error if the callback will never be called, otherwise
2391 * the result of the lock will be communicated via the callback.
2393 int ocfs2_inode_lock_full_nested(struct inode
*inode
,
2394 struct buffer_head
**ret_bh
,
2399 int status
, level
, acquired
;
2401 struct ocfs2_lock_res
*lockres
= NULL
;
2402 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2403 struct buffer_head
*local_bh
= NULL
;
2405 mlog(0, "inode %llu, take %s META lock\n",
2406 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2407 ex
? "EXMODE" : "PRMODE");
2411 /* We'll allow faking a readonly metadata lock for
2413 if (ocfs2_is_hard_readonly(osb
)) {
2419 if ((arg_flags
& OCFS2_META_LOCK_GETBH
) ||
2420 ocfs2_mount_local(osb
))
2423 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
2424 ocfs2_wait_for_recovery(osb
);
2426 lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2427 level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2429 if (arg_flags
& OCFS2_META_LOCK_NOQUEUE
)
2430 dlm_flags
|= DLM_LKF_NOQUEUE
;
2432 status
= __ocfs2_cluster_lock(osb
, lockres
, level
, dlm_flags
,
2433 arg_flags
, subclass
, _RET_IP_
);
2435 if (status
!= -EAGAIN
)
2440 /* Notify the error cleanup path to drop the cluster lock. */
2443 /* We wait twice because a node may have died while we were in
2444 * the lower dlm layers. The second time though, we've
2445 * committed to owning this lock so we don't allow signals to
2446 * abort the operation. */
2447 if (!(arg_flags
& OCFS2_META_LOCK_RECOVERY
))
2448 ocfs2_wait_for_recovery(osb
);
2452 * We only see this flag if we're being called from
2453 * ocfs2_read_locked_inode(). It means we're locking an inode
2454 * which hasn't been populated yet, so clear the refresh flag
2455 * and let the caller handle it.
2457 if (inode
->i_state
& I_NEW
) {
2460 ocfs2_complete_lock_res_refresh(lockres
, 0);
2464 /* This is fun. The caller may want a bh back, or it may
2465 * not. ocfs2_inode_lock_update definitely wants one in, but
2466 * may or may not read one, depending on what's in the
2467 * LVB. The result of all of this is that we've *only* gone to
2468 * disk if we have to, so the complexity is worthwhile. */
2469 status
= ocfs2_inode_lock_update(inode
, &local_bh
);
2471 if (status
!= -ENOENT
)
2477 status
= ocfs2_assign_bh(inode
, ret_bh
, local_bh
);
2486 if (ret_bh
&& (*ret_bh
)) {
2491 ocfs2_inode_unlock(inode
, ex
);
2501 * This is working around a lock inversion between tasks acquiring DLM
2502 * locks while holding a page lock and the downconvert thread which
2503 * blocks dlm lock acquiry while acquiring page locks.
2505 * ** These _with_page variantes are only intended to be called from aop
2506 * methods that hold page locks and return a very specific *positive* error
2507 * code that aop methods pass up to the VFS -- test for errors with != 0. **
2509 * The DLM is called such that it returns -EAGAIN if it would have
2510 * blocked waiting for the downconvert thread. In that case we unlock
2511 * our page so the downconvert thread can make progress. Once we've
2512 * done this we have to return AOP_TRUNCATED_PAGE so the aop method
2513 * that called us can bubble that back up into the VFS who will then
2514 * immediately retry the aop call.
2516 int ocfs2_inode_lock_with_page(struct inode
*inode
,
2517 struct buffer_head
**ret_bh
,
2523 ret
= ocfs2_inode_lock_full(inode
, ret_bh
, ex
, OCFS2_LOCK_NONBLOCK
);
2524 if (ret
== -EAGAIN
) {
2527 * If we can't get inode lock immediately, we should not return
2528 * directly here, since this will lead to a softlockup problem.
2529 * The method is to get a blocking lock and immediately unlock
2530 * before returning, this can avoid CPU resource waste due to
2531 * lots of retries, and benefits fairness in getting lock.
2533 if (ocfs2_inode_lock(inode
, ret_bh
, ex
) == 0)
2534 ocfs2_inode_unlock(inode
, ex
);
2535 ret
= AOP_TRUNCATED_PAGE
;
2541 int ocfs2_inode_lock_atime(struct inode
*inode
,
2542 struct vfsmount
*vfsmnt
,
2543 int *level
, int wait
)
2548 ret
= ocfs2_inode_lock(inode
, NULL
, 0);
2550 ret
= ocfs2_try_inode_lock(inode
, NULL
, 0);
2559 * If we should update atime, we will get EX lock,
2560 * otherwise we just get PR lock.
2562 if (ocfs2_should_update_atime(inode
, vfsmnt
)) {
2563 struct buffer_head
*bh
= NULL
;
2565 ocfs2_inode_unlock(inode
, 0);
2567 ret
= ocfs2_inode_lock(inode
, &bh
, 1);
2569 ret
= ocfs2_try_inode_lock(inode
, &bh
, 1);
2577 if (ocfs2_should_update_atime(inode
, vfsmnt
))
2578 ocfs2_update_inode_atime(inode
, bh
);
2587 void ocfs2_inode_unlock(struct inode
*inode
,
2590 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2591 struct ocfs2_lock_res
*lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2592 struct ocfs2_super
*osb
= OCFS2_SB(inode
->i_sb
);
2594 mlog(0, "inode %llu drop %s META lock\n",
2595 (unsigned long long)OCFS2_I(inode
)->ip_blkno
,
2596 ex
? "EXMODE" : "PRMODE");
2598 if (!ocfs2_is_hard_readonly(osb
) &&
2599 !ocfs2_mount_local(osb
))
2600 ocfs2_cluster_unlock(osb
, lockres
, level
);
2604 * This _tracker variantes are introduced to deal with the recursive cluster
2605 * locking issue. The idea is to keep track of a lock holder on the stack of
2606 * the current process. If there's a lock holder on the stack, we know the
2607 * task context is already protected by cluster locking. Currently, they're
2608 * used in some VFS entry routines.
2610 * return < 0 on error, return == 0 if there's no lock holder on the stack
2611 * before this call, return == 1 if this call would be a recursive locking.
2612 * return == -1 if this lock attempt will cause an upgrade which is forbidden.
2614 * When taking lock levels into account,we face some different situations.
2616 * 1. no lock is held
2617 * In this case, just lock the inode as requested and return 0
2619 * 2. We are holding a lock
2620 * For this situation, things diverges into several cases
2622 * wanted holding what to do
2623 * ex ex see 2.1 below
2624 * ex pr see 2.2 below
2625 * pr ex see 2.1 below
2626 * pr pr see 2.1 below
2628 * 2.1 lock level that is been held is compatible
2629 * with the wanted level, so no lock action will be tacken.
2631 * 2.2 Otherwise, an upgrade is needed, but it is forbidden.
2633 * Reason why upgrade within a process is forbidden is that
2634 * lock upgrade may cause dead lock. The following illustrates
2637 * thread on node1 thread on node2
2638 * ocfs2_inode_lock_tracker(ex=0)
2640 * <====== ocfs2_inode_lock_tracker(ex=1)
2642 * ocfs2_inode_lock_tracker(ex=1)
2644 int ocfs2_inode_lock_tracker(struct inode
*inode
,
2645 struct buffer_head
**ret_bh
,
2647 struct ocfs2_lock_holder
*oh
)
2650 struct ocfs2_lock_res
*lockres
;
2651 struct ocfs2_lock_holder
*tmp_oh
;
2652 struct pid
*pid
= task_pid(current
);
2655 lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2656 tmp_oh
= ocfs2_pid_holder(lockres
, pid
);
2660 * This corresponds to the case 1.
2661 * We haven't got any lock before.
2663 status
= ocfs2_inode_lock_full(inode
, ret_bh
, ex
, 0);
2665 if (status
!= -ENOENT
)
2671 ocfs2_add_holder(lockres
, oh
);
2675 if (unlikely(ex
&& !tmp_oh
->oh_ex
)) {
2677 * case 2.2 upgrade may cause dead lock, forbid it.
2679 mlog(ML_ERROR
, "Recursive locking is not permitted to "
2680 "upgrade to EX level from PR level.\n");
2686 * case 2.1 OCFS2_META_LOCK_GETBH flag make ocfs2_inode_lock_full.
2687 * ignore the lock level and just update it.
2690 status
= ocfs2_inode_lock_full(inode
, ret_bh
, ex
,
2691 OCFS2_META_LOCK_GETBH
);
2693 if (status
!= -ENOENT
)
2698 return tmp_oh
? 1 : 0;
2701 void ocfs2_inode_unlock_tracker(struct inode
*inode
,
2703 struct ocfs2_lock_holder
*oh
,
2706 struct ocfs2_lock_res
*lockres
;
2708 lockres
= &OCFS2_I(inode
)->ip_inode_lockres
;
2709 /* had_lock means that the currect process already takes the cluster
2711 * If had_lock is 1, we have nothing to do here.
2712 * If had_lock is 0, we will release the lock.
2715 ocfs2_inode_unlock(inode
, oh
->oh_ex
);
2716 ocfs2_remove_holder(lockres
, oh
);
2720 int ocfs2_orphan_scan_lock(struct ocfs2_super
*osb
, u32
*seqno
)
2722 struct ocfs2_lock_res
*lockres
;
2723 struct ocfs2_orphan_scan_lvb
*lvb
;
2726 if (ocfs2_is_hard_readonly(osb
))
2729 if (ocfs2_mount_local(osb
))
2732 lockres
= &osb
->osb_orphan_scan
.os_lockres
;
2733 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_EX
, 0, 0);
2737 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2738 if (ocfs2_dlm_lvb_valid(&lockres
->l_lksb
) &&
2739 lvb
->lvb_version
== OCFS2_ORPHAN_LVB_VERSION
)
2740 *seqno
= be32_to_cpu(lvb
->lvb_os_seqno
);
2742 *seqno
= osb
->osb_orphan_scan
.os_seqno
+ 1;
2747 void ocfs2_orphan_scan_unlock(struct ocfs2_super
*osb
, u32 seqno
)
2749 struct ocfs2_lock_res
*lockres
;
2750 struct ocfs2_orphan_scan_lvb
*lvb
;
2752 if (!ocfs2_is_hard_readonly(osb
) && !ocfs2_mount_local(osb
)) {
2753 lockres
= &osb
->osb_orphan_scan
.os_lockres
;
2754 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2755 lvb
->lvb_version
= OCFS2_ORPHAN_LVB_VERSION
;
2756 lvb
->lvb_os_seqno
= cpu_to_be32(seqno
);
2757 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
2761 int ocfs2_super_lock(struct ocfs2_super
*osb
,
2765 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2766 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2768 if (ocfs2_is_hard_readonly(osb
))
2771 if (ocfs2_mount_local(osb
))
2774 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
2780 /* The super block lock path is really in the best position to
2781 * know when resources covered by the lock need to be
2782 * refreshed, so we do it here. Of course, making sense of
2783 * everything is up to the caller :) */
2784 status
= ocfs2_should_refresh_lock_res(lockres
);
2786 status
= ocfs2_refresh_slot_info(osb
);
2788 ocfs2_complete_lock_res_refresh(lockres
, status
);
2791 ocfs2_cluster_unlock(osb
, lockres
, level
);
2794 ocfs2_track_lock_refresh(lockres
);
2800 void ocfs2_super_unlock(struct ocfs2_super
*osb
,
2803 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2804 struct ocfs2_lock_res
*lockres
= &osb
->osb_super_lockres
;
2806 if (!ocfs2_mount_local(osb
))
2807 ocfs2_cluster_unlock(osb
, lockres
, level
);
2810 int ocfs2_rename_lock(struct ocfs2_super
*osb
)
2813 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2815 if (ocfs2_is_hard_readonly(osb
))
2818 if (ocfs2_mount_local(osb
))
2821 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_EX
, 0, 0);
2828 void ocfs2_rename_unlock(struct ocfs2_super
*osb
)
2830 struct ocfs2_lock_res
*lockres
= &osb
->osb_rename_lockres
;
2832 if (!ocfs2_mount_local(osb
))
2833 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
2836 int ocfs2_nfs_sync_lock(struct ocfs2_super
*osb
, int ex
)
2839 struct ocfs2_lock_res
*lockres
= &osb
->osb_nfs_sync_lockres
;
2841 if (ocfs2_is_hard_readonly(osb
))
2844 if (ocfs2_mount_local(osb
))
2847 status
= ocfs2_cluster_lock(osb
, lockres
, ex
? LKM_EXMODE
: LKM_PRMODE
,
2850 mlog(ML_ERROR
, "lock on nfs sync lock failed %d\n", status
);
2855 void ocfs2_nfs_sync_unlock(struct ocfs2_super
*osb
, int ex
)
2857 struct ocfs2_lock_res
*lockres
= &osb
->osb_nfs_sync_lockres
;
2859 if (!ocfs2_mount_local(osb
))
2860 ocfs2_cluster_unlock(osb
, lockres
,
2861 ex
? LKM_EXMODE
: LKM_PRMODE
);
2864 int ocfs2_trim_fs_lock(struct ocfs2_super
*osb
,
2865 struct ocfs2_trim_fs_info
*info
, int trylock
)
2868 struct ocfs2_trim_fs_lvb
*lvb
;
2869 struct ocfs2_lock_res
*lockres
= &osb
->osb_trim_fs_lockres
;
2874 if (ocfs2_is_hard_readonly(osb
))
2877 if (ocfs2_mount_local(osb
))
2880 status
= ocfs2_cluster_lock(osb
, lockres
, DLM_LOCK_EX
,
2881 trylock
? DLM_LKF_NOQUEUE
: 0, 0);
2883 if (status
!= -EAGAIN
)
2889 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2890 if (ocfs2_dlm_lvb_valid(&lockres
->l_lksb
) &&
2891 lvb
->lvb_version
== OCFS2_TRIMFS_LVB_VERSION
) {
2893 info
->tf_success
= lvb
->lvb_success
;
2894 info
->tf_nodenum
= be32_to_cpu(lvb
->lvb_nodenum
);
2895 info
->tf_start
= be64_to_cpu(lvb
->lvb_start
);
2896 info
->tf_len
= be64_to_cpu(lvb
->lvb_len
);
2897 info
->tf_minlen
= be64_to_cpu(lvb
->lvb_minlen
);
2898 info
->tf_trimlen
= be64_to_cpu(lvb
->lvb_trimlen
);
2905 void ocfs2_trim_fs_unlock(struct ocfs2_super
*osb
,
2906 struct ocfs2_trim_fs_info
*info
)
2908 struct ocfs2_trim_fs_lvb
*lvb
;
2909 struct ocfs2_lock_res
*lockres
= &osb
->osb_trim_fs_lockres
;
2911 if (ocfs2_mount_local(osb
))
2915 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
2916 lvb
->lvb_version
= OCFS2_TRIMFS_LVB_VERSION
;
2917 lvb
->lvb_success
= info
->tf_success
;
2918 lvb
->lvb_nodenum
= cpu_to_be32(info
->tf_nodenum
);
2919 lvb
->lvb_start
= cpu_to_be64(info
->tf_start
);
2920 lvb
->lvb_len
= cpu_to_be64(info
->tf_len
);
2921 lvb
->lvb_minlen
= cpu_to_be64(info
->tf_minlen
);
2922 lvb
->lvb_trimlen
= cpu_to_be64(info
->tf_trimlen
);
2925 ocfs2_cluster_unlock(osb
, lockres
, DLM_LOCK_EX
);
2928 int ocfs2_dentry_lock(struct dentry
*dentry
, int ex
)
2931 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2932 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2933 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2937 if (ocfs2_is_hard_readonly(osb
)) {
2943 if (ocfs2_mount_local(osb
))
2946 ret
= ocfs2_cluster_lock(osb
, &dl
->dl_lockres
, level
, 0, 0);
2953 void ocfs2_dentry_unlock(struct dentry
*dentry
, int ex
)
2955 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
2956 struct ocfs2_dentry_lock
*dl
= dentry
->d_fsdata
;
2957 struct ocfs2_super
*osb
= OCFS2_SB(dentry
->d_sb
);
2959 if (!ocfs2_is_hard_readonly(osb
) && !ocfs2_mount_local(osb
))
2960 ocfs2_cluster_unlock(osb
, &dl
->dl_lockres
, level
);
2963 /* Reference counting of the dlm debug structure. We want this because
2964 * open references on the debug inodes can live on after a mount, so
2965 * we can't rely on the ocfs2_super to always exist. */
2966 static void ocfs2_dlm_debug_free(struct kref
*kref
)
2968 struct ocfs2_dlm_debug
*dlm_debug
;
2970 dlm_debug
= container_of(kref
, struct ocfs2_dlm_debug
, d_refcnt
);
2975 void ocfs2_put_dlm_debug(struct ocfs2_dlm_debug
*dlm_debug
)
2978 kref_put(&dlm_debug
->d_refcnt
, ocfs2_dlm_debug_free
);
2981 static void ocfs2_get_dlm_debug(struct ocfs2_dlm_debug
*debug
)
2983 kref_get(&debug
->d_refcnt
);
2986 struct ocfs2_dlm_debug
*ocfs2_new_dlm_debug(void)
2988 struct ocfs2_dlm_debug
*dlm_debug
;
2990 dlm_debug
= kmalloc(sizeof(struct ocfs2_dlm_debug
), GFP_KERNEL
);
2992 mlog_errno(-ENOMEM
);
2996 kref_init(&dlm_debug
->d_refcnt
);
2997 INIT_LIST_HEAD(&dlm_debug
->d_lockres_tracking
);
2998 dlm_debug
->d_locking_state
= NULL
;
3003 /* Access to this is arbitrated for us via seq_file->sem. */
3004 struct ocfs2_dlm_seq_priv
{
3005 struct ocfs2_dlm_debug
*p_dlm_debug
;
3006 struct ocfs2_lock_res p_iter_res
;
3007 struct ocfs2_lock_res p_tmp_res
;
3010 static struct ocfs2_lock_res
*ocfs2_dlm_next_res(struct ocfs2_lock_res
*start
,
3011 struct ocfs2_dlm_seq_priv
*priv
)
3013 struct ocfs2_lock_res
*iter
, *ret
= NULL
;
3014 struct ocfs2_dlm_debug
*dlm_debug
= priv
->p_dlm_debug
;
3016 assert_spin_locked(&ocfs2_dlm_tracking_lock
);
3018 list_for_each_entry(iter
, &start
->l_debug_list
, l_debug_list
) {
3019 /* discover the head of the list */
3020 if (&iter
->l_debug_list
== &dlm_debug
->d_lockres_tracking
) {
3021 mlog(0, "End of list found, %p\n", ret
);
3025 /* We track our "dummy" iteration lockres' by a NULL
3027 if (iter
->l_ops
!= NULL
) {
3036 static void *ocfs2_dlm_seq_start(struct seq_file
*m
, loff_t
*pos
)
3038 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
3039 struct ocfs2_lock_res
*iter
;
3041 spin_lock(&ocfs2_dlm_tracking_lock
);
3042 iter
= ocfs2_dlm_next_res(&priv
->p_iter_res
, priv
);
3044 /* Since lockres' have the lifetime of their container
3045 * (which can be inodes, ocfs2_supers, etc) we want to
3046 * copy this out to a temporary lockres while still
3047 * under the spinlock. Obviously after this we can't
3048 * trust any pointers on the copy returned, but that's
3049 * ok as the information we want isn't typically held
3051 priv
->p_tmp_res
= *iter
;
3052 iter
= &priv
->p_tmp_res
;
3054 spin_unlock(&ocfs2_dlm_tracking_lock
);
3059 static void ocfs2_dlm_seq_stop(struct seq_file
*m
, void *v
)
3063 static void *ocfs2_dlm_seq_next(struct seq_file
*m
, void *v
, loff_t
*pos
)
3065 struct ocfs2_dlm_seq_priv
*priv
= m
->private;
3066 struct ocfs2_lock_res
*iter
= v
;
3067 struct ocfs2_lock_res
*dummy
= &priv
->p_iter_res
;
3069 spin_lock(&ocfs2_dlm_tracking_lock
);
3070 iter
= ocfs2_dlm_next_res(iter
, priv
);
3071 list_del_init(&dummy
->l_debug_list
);
3073 list_add(&dummy
->l_debug_list
, &iter
->l_debug_list
);
3074 priv
->p_tmp_res
= *iter
;
3075 iter
= &priv
->p_tmp_res
;
3077 spin_unlock(&ocfs2_dlm_tracking_lock
);
3083 * Version is used by debugfs.ocfs2 to determine the format being used
3086 * - Lock stats printed
3088 * - Max time in lock stats is in usecs (instead of nsecs)
3090 #define OCFS2_DLM_DEBUG_STR_VERSION 3
3091 static int ocfs2_dlm_seq_show(struct seq_file
*m
, void *v
)
3095 struct ocfs2_lock_res
*lockres
= v
;
3100 seq_printf(m
, "0x%x\t", OCFS2_DLM_DEBUG_STR_VERSION
);
3102 if (lockres
->l_type
== OCFS2_LOCK_TYPE_DENTRY
)
3103 seq_printf(m
, "%.*s%08x\t", OCFS2_DENTRY_LOCK_INO_START
- 1,
3105 (unsigned int)ocfs2_get_dentry_lock_ino(lockres
));
3107 seq_printf(m
, "%.*s\t", OCFS2_LOCK_ID_MAX_LEN
, lockres
->l_name
);
3109 seq_printf(m
, "%d\t"
3120 lockres
->l_unlock_action
,
3121 lockres
->l_ro_holders
,
3122 lockres
->l_ex_holders
,
3123 lockres
->l_requested
,
3124 lockres
->l_blocking
);
3126 /* Dump the raw LVB */
3127 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
3128 for(i
= 0; i
< DLM_LVB_LEN
; i
++)
3129 seq_printf(m
, "0x%x\t", lvb
[i
]);
3131 #ifdef CONFIG_OCFS2_FS_STATS
3132 # define lock_num_prmode(_l) ((_l)->l_lock_prmode.ls_gets)
3133 # define lock_num_exmode(_l) ((_l)->l_lock_exmode.ls_gets)
3134 # define lock_num_prmode_failed(_l) ((_l)->l_lock_prmode.ls_fail)
3135 # define lock_num_exmode_failed(_l) ((_l)->l_lock_exmode.ls_fail)
3136 # define lock_total_prmode(_l) ((_l)->l_lock_prmode.ls_total)
3137 # define lock_total_exmode(_l) ((_l)->l_lock_exmode.ls_total)
3138 # define lock_max_prmode(_l) ((_l)->l_lock_prmode.ls_max)
3139 # define lock_max_exmode(_l) ((_l)->l_lock_exmode.ls_max)
3140 # define lock_refresh(_l) ((_l)->l_lock_refresh)
3142 # define lock_num_prmode(_l) (0)
3143 # define lock_num_exmode(_l) (0)
3144 # define lock_num_prmode_failed(_l) (0)
3145 # define lock_num_exmode_failed(_l) (0)
3146 # define lock_total_prmode(_l) (0ULL)
3147 # define lock_total_exmode(_l) (0ULL)
3148 # define lock_max_prmode(_l) (0)
3149 # define lock_max_exmode(_l) (0)
3150 # define lock_refresh(_l) (0)
3152 /* The following seq_print was added in version 2 of this output */
3153 seq_printf(m
, "%u\t"
3162 lock_num_prmode(lockres
),
3163 lock_num_exmode(lockres
),
3164 lock_num_prmode_failed(lockres
),
3165 lock_num_exmode_failed(lockres
),
3166 lock_total_prmode(lockres
),
3167 lock_total_exmode(lockres
),
3168 lock_max_prmode(lockres
),
3169 lock_max_exmode(lockres
),
3170 lock_refresh(lockres
));
3173 seq_printf(m
, "\n");
3177 static const struct seq_operations ocfs2_dlm_seq_ops
= {
3178 .start
= ocfs2_dlm_seq_start
,
3179 .stop
= ocfs2_dlm_seq_stop
,
3180 .next
= ocfs2_dlm_seq_next
,
3181 .show
= ocfs2_dlm_seq_show
,
3184 static int ocfs2_dlm_debug_release(struct inode
*inode
, struct file
*file
)
3186 struct seq_file
*seq
= file
->private_data
;
3187 struct ocfs2_dlm_seq_priv
*priv
= seq
->private;
3188 struct ocfs2_lock_res
*res
= &priv
->p_iter_res
;
3190 ocfs2_remove_lockres_tracking(res
);
3191 ocfs2_put_dlm_debug(priv
->p_dlm_debug
);
3192 return seq_release_private(inode
, file
);
3195 static int ocfs2_dlm_debug_open(struct inode
*inode
, struct file
*file
)
3197 struct ocfs2_dlm_seq_priv
*priv
;
3198 struct ocfs2_super
*osb
;
3200 priv
= __seq_open_private(file
, &ocfs2_dlm_seq_ops
, sizeof(*priv
));
3202 mlog_errno(-ENOMEM
);
3206 osb
= inode
->i_private
;
3207 ocfs2_get_dlm_debug(osb
->osb_dlm_debug
);
3208 priv
->p_dlm_debug
= osb
->osb_dlm_debug
;
3209 INIT_LIST_HEAD(&priv
->p_iter_res
.l_debug_list
);
3211 ocfs2_add_lockres_tracking(&priv
->p_iter_res
,
3217 static const struct file_operations ocfs2_dlm_debug_fops
= {
3218 .open
= ocfs2_dlm_debug_open
,
3219 .release
= ocfs2_dlm_debug_release
,
3221 .llseek
= seq_lseek
,
3224 static int ocfs2_dlm_init_debug(struct ocfs2_super
*osb
)
3227 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
3229 dlm_debug
->d_locking_state
= debugfs_create_file("locking_state",
3231 osb
->osb_debug_root
,
3233 &ocfs2_dlm_debug_fops
);
3234 if (!dlm_debug
->d_locking_state
) {
3237 "Unable to create locking state debugfs file.\n");
3241 ocfs2_get_dlm_debug(dlm_debug
);
3246 static void ocfs2_dlm_shutdown_debug(struct ocfs2_super
*osb
)
3248 struct ocfs2_dlm_debug
*dlm_debug
= osb
->osb_dlm_debug
;
3251 debugfs_remove(dlm_debug
->d_locking_state
);
3252 ocfs2_put_dlm_debug(dlm_debug
);
3256 int ocfs2_dlm_init(struct ocfs2_super
*osb
)
3259 struct ocfs2_cluster_connection
*conn
= NULL
;
3261 if (ocfs2_mount_local(osb
)) {
3266 status
= ocfs2_dlm_init_debug(osb
);
3272 /* launch downconvert thread */
3273 osb
->dc_task
= kthread_run(ocfs2_downconvert_thread
, osb
, "ocfs2dc-%s",
3275 if (IS_ERR(osb
->dc_task
)) {
3276 status
= PTR_ERR(osb
->dc_task
);
3277 osb
->dc_task
= NULL
;
3282 /* for now, uuid == domain */
3283 status
= ocfs2_cluster_connect(osb
->osb_cluster_stack
,
3284 osb
->osb_cluster_name
,
3285 strlen(osb
->osb_cluster_name
),
3287 strlen(osb
->uuid_str
),
3288 &lproto
, ocfs2_do_node_down
, osb
,
3295 status
= ocfs2_cluster_this_node(conn
, &osb
->node_num
);
3299 "could not find this host's node number\n");
3300 ocfs2_cluster_disconnect(conn
, 0);
3305 ocfs2_super_lock_res_init(&osb
->osb_super_lockres
, osb
);
3306 ocfs2_rename_lock_res_init(&osb
->osb_rename_lockres
, osb
);
3307 ocfs2_nfs_sync_lock_res_init(&osb
->osb_nfs_sync_lockres
, osb
);
3308 ocfs2_orphan_scan_lock_res_init(&osb
->osb_orphan_scan
.os_lockres
, osb
);
3313 ocfs2_dlm_shutdown_debug(osb
);
3315 kthread_stop(osb
->dc_task
);
3321 void ocfs2_dlm_shutdown(struct ocfs2_super
*osb
,
3324 ocfs2_drop_osb_locks(osb
);
3327 * Now that we have dropped all locks and ocfs2_dismount_volume()
3328 * has disabled recovery, the DLM won't be talking to us. It's
3329 * safe to tear things down before disconnecting the cluster.
3333 kthread_stop(osb
->dc_task
);
3334 osb
->dc_task
= NULL
;
3337 ocfs2_lock_res_free(&osb
->osb_super_lockres
);
3338 ocfs2_lock_res_free(&osb
->osb_rename_lockres
);
3339 ocfs2_lock_res_free(&osb
->osb_nfs_sync_lockres
);
3340 ocfs2_lock_res_free(&osb
->osb_orphan_scan
.os_lockres
);
3342 ocfs2_cluster_disconnect(osb
->cconn
, hangup_pending
);
3345 ocfs2_dlm_shutdown_debug(osb
);
3348 static int ocfs2_drop_lock(struct ocfs2_super
*osb
,
3349 struct ocfs2_lock_res
*lockres
)
3352 unsigned long flags
;
3355 /* We didn't get anywhere near actually using this lockres. */
3356 if (!(lockres
->l_flags
& OCFS2_LOCK_INITIALIZED
))
3359 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
3360 lkm_flags
|= DLM_LKF_VALBLK
;
3362 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3364 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_FREEING
),
3365 "lockres %s, flags 0x%lx\n",
3366 lockres
->l_name
, lockres
->l_flags
);
3368 while (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
3369 mlog(0, "waiting on busy lock \"%s\": flags = %lx, action = "
3370 "%u, unlock_action = %u\n",
3371 lockres
->l_name
, lockres
->l_flags
, lockres
->l_action
,
3372 lockres
->l_unlock_action
);
3374 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3376 /* XXX: Today we just wait on any busy
3377 * locks... Perhaps we need to cancel converts in the
3379 ocfs2_wait_on_busy_lock(lockres
);
3381 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3384 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
3385 if (lockres
->l_flags
& OCFS2_LOCK_ATTACHED
&&
3386 lockres
->l_level
== DLM_LOCK_EX
&&
3387 !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
3388 lockres
->l_ops
->set_lvb(lockres
);
3391 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
)
3392 mlog(ML_ERROR
, "destroying busy lock: \"%s\"\n",
3394 if (lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)
3395 mlog(0, "destroying blocked lock: \"%s\"\n", lockres
->l_name
);
3397 if (!(lockres
->l_flags
& OCFS2_LOCK_ATTACHED
)) {
3398 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3402 lockres_clear_flags(lockres
, OCFS2_LOCK_ATTACHED
);
3404 /* make sure we never get here while waiting for an ast to
3406 BUG_ON(lockres
->l_action
!= OCFS2_AST_INVALID
);
3408 /* is this necessary? */
3409 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
3410 lockres
->l_unlock_action
= OCFS2_UNLOCK_DROP_LOCK
;
3411 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3413 mlog(0, "lock %s\n", lockres
->l_name
);
3415 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
, lkm_flags
);
3417 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
3418 mlog(ML_ERROR
, "lockres flags: %lu\n", lockres
->l_flags
);
3419 ocfs2_dlm_dump_lksb(&lockres
->l_lksb
);
3422 mlog(0, "lock %s, successful return from ocfs2_dlm_unlock\n",
3425 ocfs2_wait_on_busy_lock(lockres
);
3430 static void ocfs2_process_blocked_lock(struct ocfs2_super
*osb
,
3431 struct ocfs2_lock_res
*lockres
);
3433 /* Mark the lockres as being dropped. It will no longer be
3434 * queued if blocking, but we still may have to wait on it
3435 * being dequeued from the downconvert thread before we can consider
3438 * You can *not* attempt to call cluster_lock on this lockres anymore. */
3439 void ocfs2_mark_lockres_freeing(struct ocfs2_super
*osb
,
3440 struct ocfs2_lock_res
*lockres
)
3443 struct ocfs2_mask_waiter mw
;
3444 unsigned long flags
, flags2
;
3446 ocfs2_init_mask_waiter(&mw
);
3448 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3449 lockres
->l_flags
|= OCFS2_LOCK_FREEING
;
3450 if (lockres
->l_flags
& OCFS2_LOCK_QUEUED
&& current
== osb
->dc_task
) {
3452 * We know the downconvert is queued but not in progress
3453 * because we are the downconvert thread and processing
3454 * different lock. So we can just remove the lock from the
3455 * queue. This is not only an optimization but also a way
3456 * to avoid the following deadlock:
3457 * ocfs2_dentry_post_unlock()
3458 * ocfs2_dentry_lock_put()
3459 * ocfs2_drop_dentry_lock()
3461 * ocfs2_evict_inode()
3462 * ocfs2_clear_inode()
3463 * ocfs2_mark_lockres_freeing()
3464 * ... blocks waiting for OCFS2_LOCK_QUEUED
3465 * since we are the downconvert thread which
3466 * should clear the flag.
3468 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3469 spin_lock_irqsave(&osb
->dc_task_lock
, flags2
);
3470 list_del_init(&lockres
->l_blocked_list
);
3471 osb
->blocked_lock_count
--;
3472 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags2
);
3474 * Warn if we recurse into another post_unlock call. Strictly
3475 * speaking it isn't a problem but we need to be careful if
3476 * that happens (stack overflow, deadlocks, ...) so warn if
3477 * ocfs2 grows a path for which this can happen.
3479 WARN_ON_ONCE(lockres
->l_ops
->post_unlock
);
3480 /* Since the lock is freeing we don't do much in the fn below */
3481 ocfs2_process_blocked_lock(osb
, lockres
);
3484 while (lockres
->l_flags
& OCFS2_LOCK_QUEUED
) {
3485 lockres_add_mask_waiter(lockres
, &mw
, OCFS2_LOCK_QUEUED
, 0);
3486 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3488 mlog(0, "Waiting on lockres %s\n", lockres
->l_name
);
3490 status
= ocfs2_wait_for_mask(&mw
);
3494 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3496 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3499 void ocfs2_simple_drop_lockres(struct ocfs2_super
*osb
,
3500 struct ocfs2_lock_res
*lockres
)
3504 ocfs2_mark_lockres_freeing(osb
, lockres
);
3505 ret
= ocfs2_drop_lock(osb
, lockres
);
3510 static void ocfs2_drop_osb_locks(struct ocfs2_super
*osb
)
3512 ocfs2_simple_drop_lockres(osb
, &osb
->osb_super_lockres
);
3513 ocfs2_simple_drop_lockres(osb
, &osb
->osb_rename_lockres
);
3514 ocfs2_simple_drop_lockres(osb
, &osb
->osb_nfs_sync_lockres
);
3515 ocfs2_simple_drop_lockres(osb
, &osb
->osb_orphan_scan
.os_lockres
);
3518 int ocfs2_drop_inode_locks(struct inode
*inode
)
3522 /* No need to call ocfs2_mark_lockres_freeing here -
3523 * ocfs2_clear_inode has done it for us. */
3525 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
3526 &OCFS2_I(inode
)->ip_open_lockres
);
3532 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
3533 &OCFS2_I(inode
)->ip_inode_lockres
);
3536 if (err
< 0 && !status
)
3539 err
= ocfs2_drop_lock(OCFS2_SB(inode
->i_sb
),
3540 &OCFS2_I(inode
)->ip_rw_lockres
);
3543 if (err
< 0 && !status
)
3549 static unsigned int ocfs2_prepare_downconvert(struct ocfs2_lock_res
*lockres
,
3552 assert_spin_locked(&lockres
->l_lock
);
3554 BUG_ON(lockres
->l_blocking
<= DLM_LOCK_NL
);
3556 if (lockres
->l_level
<= new_level
) {
3557 mlog(ML_ERROR
, "lockres %s, lvl %d <= %d, blcklst %d, mask %d, "
3558 "type %d, flags 0x%lx, hold %d %d, act %d %d, req %d, "
3559 "block %d, pgen %d\n", lockres
->l_name
, lockres
->l_level
,
3560 new_level
, list_empty(&lockres
->l_blocked_list
),
3561 list_empty(&lockres
->l_mask_waiters
), lockres
->l_type
,
3562 lockres
->l_flags
, lockres
->l_ro_holders
,
3563 lockres
->l_ex_holders
, lockres
->l_action
,
3564 lockres
->l_unlock_action
, lockres
->l_requested
,
3565 lockres
->l_blocking
, lockres
->l_pending_gen
);
3569 mlog(ML_BASTS
, "lockres %s, level %d => %d, blocking %d\n",
3570 lockres
->l_name
, lockres
->l_level
, new_level
, lockres
->l_blocking
);
3572 lockres
->l_action
= OCFS2_AST_DOWNCONVERT
;
3573 lockres
->l_requested
= new_level
;
3574 lockres_or_flags(lockres
, OCFS2_LOCK_BUSY
);
3575 return lockres_set_pending(lockres
);
3578 static int ocfs2_downconvert_lock(struct ocfs2_super
*osb
,
3579 struct ocfs2_lock_res
*lockres
,
3582 unsigned int generation
)
3585 u32 dlm_flags
= DLM_LKF_CONVERT
;
3587 mlog(ML_BASTS
, "lockres %s, level %d => %d\n", lockres
->l_name
,
3588 lockres
->l_level
, new_level
);
3591 * On DLM_LKF_VALBLK, fsdlm behaves differently with o2cb. It always
3592 * expects DLM_LKF_VALBLK being set if the LKB has LVB, so that
3593 * we can recover correctly from node failure. Otherwise, we may get
3594 * invalid LVB in LKB, but without DLM_SBF_VALNOTVALID being set.
3596 if (!ocfs2_is_o2cb_active() &&
3597 lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
)
3601 dlm_flags
|= DLM_LKF_VALBLK
;
3603 ret
= ocfs2_dlm_lock(osb
->cconn
,
3608 OCFS2_LOCK_ID_MAX_LEN
- 1);
3609 lockres_clear_pending(lockres
, generation
, osb
);
3611 ocfs2_log_dlm_error("ocfs2_dlm_lock", ret
, lockres
);
3612 ocfs2_recover_from_dlm_error(lockres
, 1);
3621 /* returns 1 when the caller should unlock and call ocfs2_dlm_unlock */
3622 static int ocfs2_prepare_cancel_convert(struct ocfs2_super
*osb
,
3623 struct ocfs2_lock_res
*lockres
)
3625 assert_spin_locked(&lockres
->l_lock
);
3627 if (lockres
->l_unlock_action
== OCFS2_UNLOCK_CANCEL_CONVERT
) {
3628 /* If we're already trying to cancel a lock conversion
3629 * then just drop the spinlock and allow the caller to
3630 * requeue this lock. */
3631 mlog(ML_BASTS
, "lockres %s, skip convert\n", lockres
->l_name
);
3635 /* were we in a convert when we got the bast fire? */
3636 BUG_ON(lockres
->l_action
!= OCFS2_AST_CONVERT
&&
3637 lockres
->l_action
!= OCFS2_AST_DOWNCONVERT
);
3638 /* set things up for the unlockast to know to just
3639 * clear out the ast_action and unset busy, etc. */
3640 lockres
->l_unlock_action
= OCFS2_UNLOCK_CANCEL_CONVERT
;
3642 mlog_bug_on_msg(!(lockres
->l_flags
& OCFS2_LOCK_BUSY
),
3643 "lock %s, invalid flags: 0x%lx\n",
3644 lockres
->l_name
, lockres
->l_flags
);
3646 mlog(ML_BASTS
, "lockres %s\n", lockres
->l_name
);
3651 static int ocfs2_cancel_convert(struct ocfs2_super
*osb
,
3652 struct ocfs2_lock_res
*lockres
)
3656 ret
= ocfs2_dlm_unlock(osb
->cconn
, &lockres
->l_lksb
,
3659 ocfs2_log_dlm_error("ocfs2_dlm_unlock", ret
, lockres
);
3660 ocfs2_recover_from_dlm_error(lockres
, 0);
3663 mlog(ML_BASTS
, "lockres %s\n", lockres
->l_name
);
3668 static int ocfs2_unblock_lock(struct ocfs2_super
*osb
,
3669 struct ocfs2_lock_res
*lockres
,
3670 struct ocfs2_unblock_ctl
*ctl
)
3672 unsigned long flags
;
3680 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3684 * Is it still blocking? If not, we have no more work to do.
3686 if (!(lockres
->l_flags
& OCFS2_LOCK_BLOCKED
)) {
3687 BUG_ON(lockres
->l_blocking
!= DLM_LOCK_NL
);
3688 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3693 if (lockres
->l_flags
& OCFS2_LOCK_BUSY
) {
3695 * This is a *big* race. The OCFS2_LOCK_PENDING flag
3696 * exists entirely for one reason - another thread has set
3697 * OCFS2_LOCK_BUSY, but has *NOT* yet called dlm_lock().
3699 * If we do ocfs2_cancel_convert() before the other thread
3700 * calls dlm_lock(), our cancel will do nothing. We will
3701 * get no ast, and we will have no way of knowing the
3702 * cancel failed. Meanwhile, the other thread will call
3703 * into dlm_lock() and wait...forever.
3705 * Why forever? Because another node has asked for the
3706 * lock first; that's why we're here in unblock_lock().
3708 * The solution is OCFS2_LOCK_PENDING. When PENDING is
3709 * set, we just requeue the unblock. Only when the other
3710 * thread has called dlm_lock() and cleared PENDING will
3711 * we then cancel their request.
3713 * All callers of dlm_lock() must set OCFS2_DLM_PENDING
3714 * at the same time they set OCFS2_DLM_BUSY. They must
3715 * clear OCFS2_DLM_PENDING after dlm_lock() returns.
3717 if (lockres
->l_flags
& OCFS2_LOCK_PENDING
) {
3718 mlog(ML_BASTS
, "lockres %s, ReQ: Pending\n",
3724 ret
= ocfs2_prepare_cancel_convert(osb
, lockres
);
3725 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3727 ret
= ocfs2_cancel_convert(osb
, lockres
);
3735 * This prevents livelocks. OCFS2_LOCK_UPCONVERT_FINISHING flag is
3736 * set when the ast is received for an upconvert just before the
3737 * OCFS2_LOCK_BUSY flag is cleared. Now if the fs received a bast
3738 * on the heels of the ast, we want to delay the downconvert just
3739 * enough to allow the up requestor to do its task. Because this
3740 * lock is in the blocked queue, the lock will be downconverted
3741 * as soon as the requestor is done with the lock.
3743 if (lockres
->l_flags
& OCFS2_LOCK_UPCONVERT_FINISHING
)
3747 * How can we block and yet be at NL? We were trying to upconvert
3748 * from NL and got canceled. The code comes back here, and now
3749 * we notice and clear BLOCKING.
3751 if (lockres
->l_level
== DLM_LOCK_NL
) {
3752 BUG_ON(lockres
->l_ex_holders
|| lockres
->l_ro_holders
);
3753 mlog(ML_BASTS
, "lockres %s, Aborting dc\n", lockres
->l_name
);
3754 lockres
->l_blocking
= DLM_LOCK_NL
;
3755 lockres_clear_flags(lockres
, OCFS2_LOCK_BLOCKED
);
3756 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3760 /* if we're blocking an exclusive and we have *any* holders,
3762 if ((lockres
->l_blocking
== DLM_LOCK_EX
)
3763 && (lockres
->l_ex_holders
|| lockres
->l_ro_holders
)) {
3764 mlog(ML_BASTS
, "lockres %s, ReQ: EX/PR Holders %u,%u\n",
3765 lockres
->l_name
, lockres
->l_ex_holders
,
3766 lockres
->l_ro_holders
);
3770 /* If it's a PR we're blocking, then only
3771 * requeue if we've got any EX holders */
3772 if (lockres
->l_blocking
== DLM_LOCK_PR
&&
3773 lockres
->l_ex_holders
) {
3774 mlog(ML_BASTS
, "lockres %s, ReQ: EX Holders %u\n",
3775 lockres
->l_name
, lockres
->l_ex_holders
);
3780 * Can we get a lock in this state if the holder counts are
3781 * zero? The meta data unblock code used to check this.
3783 if ((lockres
->l_ops
->flags
& LOCK_TYPE_REQUIRES_REFRESH
)
3784 && (lockres
->l_flags
& OCFS2_LOCK_REFRESHING
)) {
3785 mlog(ML_BASTS
, "lockres %s, ReQ: Lock Refreshing\n",
3790 new_level
= ocfs2_highest_compat_lock_level(lockres
->l_blocking
);
3792 if (lockres
->l_ops
->check_downconvert
3793 && !lockres
->l_ops
->check_downconvert(lockres
, new_level
)) {
3794 mlog(ML_BASTS
, "lockres %s, ReQ: Checkpointing\n",
3799 /* If we get here, then we know that there are no more
3800 * incompatible holders (and anyone asking for an incompatible
3801 * lock is blocked). We can now downconvert the lock */
3802 if (!lockres
->l_ops
->downconvert_worker
)
3805 /* Some lockres types want to do a bit of work before
3806 * downconverting a lock. Allow that here. The worker function
3807 * may sleep, so we save off a copy of what we're blocking as
3808 * it may change while we're not holding the spin lock. */
3809 blocking
= lockres
->l_blocking
;
3810 level
= lockres
->l_level
;
3811 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3813 ctl
->unblock_action
= lockres
->l_ops
->downconvert_worker(lockres
, blocking
);
3815 if (ctl
->unblock_action
== UNBLOCK_STOP_POST
) {
3816 mlog(ML_BASTS
, "lockres %s, UNBLOCK_STOP_POST\n",
3821 spin_lock_irqsave(&lockres
->l_lock
, flags
);
3822 if ((blocking
!= lockres
->l_blocking
) || (level
!= lockres
->l_level
)) {
3823 /* If this changed underneath us, then we can't drop
3825 mlog(ML_BASTS
, "lockres %s, block=%d:%d, level=%d:%d, "
3826 "Recheck\n", lockres
->l_name
, blocking
,
3827 lockres
->l_blocking
, level
, lockres
->l_level
);
3834 if (lockres
->l_ops
->flags
& LOCK_TYPE_USES_LVB
) {
3835 if (lockres
->l_level
== DLM_LOCK_EX
)
3839 * We only set the lvb if the lock has been fully
3840 * refreshed - otherwise we risk setting stale
3841 * data. Otherwise, there's no need to actually clear
3842 * out the lvb here as it's value is still valid.
3844 if (set_lvb
&& !(lockres
->l_flags
& OCFS2_LOCK_NEEDS_REFRESH
))
3845 lockres
->l_ops
->set_lvb(lockres
);
3848 gen
= ocfs2_prepare_downconvert(lockres
, new_level
);
3849 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3850 ret
= ocfs2_downconvert_lock(osb
, lockres
, new_level
, set_lvb
,
3859 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
3865 static int ocfs2_data_convert_worker(struct ocfs2_lock_res
*lockres
,
3868 struct inode
*inode
;
3869 struct address_space
*mapping
;
3870 struct ocfs2_inode_info
*oi
;
3872 inode
= ocfs2_lock_res_inode(lockres
);
3873 mapping
= inode
->i_mapping
;
3875 if (S_ISDIR(inode
->i_mode
)) {
3876 oi
= OCFS2_I(inode
);
3877 oi
->ip_dir_lock_gen
++;
3878 mlog(0, "generation: %u\n", oi
->ip_dir_lock_gen
);
3882 if (!S_ISREG(inode
->i_mode
))
3886 * We need this before the filemap_fdatawrite() so that it can
3887 * transfer the dirty bit from the PTE to the
3888 * page. Unfortunately this means that even for EX->PR
3889 * downconverts, we'll lose our mappings and have to build
3892 unmap_mapping_range(mapping
, 0, 0, 0);
3894 if (filemap_fdatawrite(mapping
)) {
3895 mlog(ML_ERROR
, "Could not sync inode %llu for downconvert!",
3896 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
3898 sync_mapping_buffers(mapping
);
3899 if (blocking
== DLM_LOCK_EX
) {
3900 truncate_inode_pages(mapping
, 0);
3902 /* We only need to wait on the I/O if we're not also
3903 * truncating pages because truncate_inode_pages waits
3904 * for us above. We don't truncate pages if we're
3905 * blocking anything < EXMODE because we want to keep
3906 * them around in that case. */
3907 filemap_fdatawait(mapping
);
3910 forget_all_cached_acls(inode
);
3913 return UNBLOCK_CONTINUE
;
3916 static int ocfs2_ci_checkpointed(struct ocfs2_caching_info
*ci
,
3917 struct ocfs2_lock_res
*lockres
,
3920 int checkpointed
= ocfs2_ci_fully_checkpointed(ci
);
3922 BUG_ON(new_level
!= DLM_LOCK_NL
&& new_level
!= DLM_LOCK_PR
);
3923 BUG_ON(lockres
->l_level
!= DLM_LOCK_EX
&& !checkpointed
);
3928 ocfs2_start_checkpoint(OCFS2_SB(ocfs2_metadata_cache_get_super(ci
)));
3932 static int ocfs2_check_meta_downconvert(struct ocfs2_lock_res
*lockres
,
3935 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3937 return ocfs2_ci_checkpointed(INODE_CACHE(inode
), lockres
, new_level
);
3940 static void ocfs2_set_meta_lvb(struct ocfs2_lock_res
*lockres
)
3942 struct inode
*inode
= ocfs2_lock_res_inode(lockres
);
3944 __ocfs2_stuff_meta_lvb(inode
);
3948 * Does the final reference drop on our dentry lock. Right now this
3949 * happens in the downconvert thread, but we could choose to simplify the
3950 * dlmglue API and push these off to the ocfs2_wq in the future.
3952 static void ocfs2_dentry_post_unlock(struct ocfs2_super
*osb
,
3953 struct ocfs2_lock_res
*lockres
)
3955 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3956 ocfs2_dentry_lock_put(osb
, dl
);
3960 * d_delete() matching dentries before the lock downconvert.
3962 * At this point, any process waiting to destroy the
3963 * dentry_lock due to last ref count is stopped by the
3964 * OCFS2_LOCK_QUEUED flag.
3966 * We have two potential problems
3968 * 1) If we do the last reference drop on our dentry_lock (via dput)
3969 * we'll wind up in ocfs2_release_dentry_lock(), waiting on
3970 * the downconvert to finish. Instead we take an elevated
3971 * reference and push the drop until after we've completed our
3972 * unblock processing.
3974 * 2) There might be another process with a final reference,
3975 * waiting on us to finish processing. If this is the case, we
3976 * detect it and exit out - there's no more dentries anyway.
3978 static int ocfs2_dentry_convert_worker(struct ocfs2_lock_res
*lockres
,
3981 struct ocfs2_dentry_lock
*dl
= ocfs2_lock_res_dl(lockres
);
3982 struct ocfs2_inode_info
*oi
= OCFS2_I(dl
->dl_inode
);
3983 struct dentry
*dentry
;
3984 unsigned long flags
;
3988 * This node is blocking another node from getting a read
3989 * lock. This happens when we've renamed within a
3990 * directory. We've forced the other nodes to d_delete(), but
3991 * we never actually dropped our lock because it's still
3992 * valid. The downconvert code will retain a PR for this node,
3993 * so there's no further work to do.
3995 if (blocking
== DLM_LOCK_PR
)
3996 return UNBLOCK_CONTINUE
;
3999 * Mark this inode as potentially orphaned. The code in
4000 * ocfs2_delete_inode() will figure out whether it actually
4001 * needs to be freed or not.
4003 spin_lock(&oi
->ip_lock
);
4004 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
4005 spin_unlock(&oi
->ip_lock
);
4008 * Yuck. We need to make sure however that the check of
4009 * OCFS2_LOCK_FREEING and the extra reference are atomic with
4010 * respect to a reference decrement or the setting of that
4013 spin_lock_irqsave(&lockres
->l_lock
, flags
);
4014 spin_lock(&dentry_attach_lock
);
4015 if (!(lockres
->l_flags
& OCFS2_LOCK_FREEING
)
4020 spin_unlock(&dentry_attach_lock
);
4021 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
4023 mlog(0, "extra_ref = %d\n", extra_ref
);
4026 * We have a process waiting on us in ocfs2_dentry_iput(),
4027 * which means we can't have any more outstanding
4028 * aliases. There's no need to do any more work.
4031 return UNBLOCK_CONTINUE
;
4033 spin_lock(&dentry_attach_lock
);
4035 dentry
= ocfs2_find_local_alias(dl
->dl_inode
,
4036 dl
->dl_parent_blkno
, 1);
4039 spin_unlock(&dentry_attach_lock
);
4041 if (S_ISDIR(dl
->dl_inode
->i_mode
))
4042 shrink_dcache_parent(dentry
);
4044 mlog(0, "d_delete(%pd);\n", dentry
);
4047 * The following dcache calls may do an
4048 * iput(). Normally we don't want that from the
4049 * downconverting thread, but in this case it's ok
4050 * because the requesting node already has an
4051 * exclusive lock on the inode, so it can't be queued
4052 * for a downconvert.
4057 spin_lock(&dentry_attach_lock
);
4059 spin_unlock(&dentry_attach_lock
);
4062 * If we are the last holder of this dentry lock, there is no
4063 * reason to downconvert so skip straight to the unlock.
4065 if (dl
->dl_count
== 1)
4066 return UNBLOCK_STOP_POST
;
4068 return UNBLOCK_CONTINUE_POST
;
4071 static int ocfs2_check_refcount_downconvert(struct ocfs2_lock_res
*lockres
,
4074 struct ocfs2_refcount_tree
*tree
=
4075 ocfs2_lock_res_refcount_tree(lockres
);
4077 return ocfs2_ci_checkpointed(&tree
->rf_ci
, lockres
, new_level
);
4080 static int ocfs2_refcount_convert_worker(struct ocfs2_lock_res
*lockres
,
4083 struct ocfs2_refcount_tree
*tree
=
4084 ocfs2_lock_res_refcount_tree(lockres
);
4086 ocfs2_metadata_cache_purge(&tree
->rf_ci
);
4088 return UNBLOCK_CONTINUE
;
4091 static void ocfs2_set_qinfo_lvb(struct ocfs2_lock_res
*lockres
)
4093 struct ocfs2_qinfo_lvb
*lvb
;
4094 struct ocfs2_mem_dqinfo
*oinfo
= ocfs2_lock_res_qinfo(lockres
);
4095 struct mem_dqinfo
*info
= sb_dqinfo(oinfo
->dqi_gi
.dqi_sb
,
4096 oinfo
->dqi_gi
.dqi_type
);
4098 lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
4099 lvb
->lvb_version
= OCFS2_QINFO_LVB_VERSION
;
4100 lvb
->lvb_bgrace
= cpu_to_be32(info
->dqi_bgrace
);
4101 lvb
->lvb_igrace
= cpu_to_be32(info
->dqi_igrace
);
4102 lvb
->lvb_syncms
= cpu_to_be32(oinfo
->dqi_syncms
);
4103 lvb
->lvb_blocks
= cpu_to_be32(oinfo
->dqi_gi
.dqi_blocks
);
4104 lvb
->lvb_free_blk
= cpu_to_be32(oinfo
->dqi_gi
.dqi_free_blk
);
4105 lvb
->lvb_free_entry
= cpu_to_be32(oinfo
->dqi_gi
.dqi_free_entry
);
4108 void ocfs2_qinfo_unlock(struct ocfs2_mem_dqinfo
*oinfo
, int ex
)
4110 struct ocfs2_lock_res
*lockres
= &oinfo
->dqi_gqlock
;
4111 struct ocfs2_super
*osb
= OCFS2_SB(oinfo
->dqi_gi
.dqi_sb
);
4112 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
4114 if (!ocfs2_is_hard_readonly(osb
) && !ocfs2_mount_local(osb
))
4115 ocfs2_cluster_unlock(osb
, lockres
, level
);
4118 static int ocfs2_refresh_qinfo(struct ocfs2_mem_dqinfo
*oinfo
)
4120 struct mem_dqinfo
*info
= sb_dqinfo(oinfo
->dqi_gi
.dqi_sb
,
4121 oinfo
->dqi_gi
.dqi_type
);
4122 struct ocfs2_lock_res
*lockres
= &oinfo
->dqi_gqlock
;
4123 struct ocfs2_qinfo_lvb
*lvb
= ocfs2_dlm_lvb(&lockres
->l_lksb
);
4124 struct buffer_head
*bh
= NULL
;
4125 struct ocfs2_global_disk_dqinfo
*gdinfo
;
4128 if (ocfs2_dlm_lvb_valid(&lockres
->l_lksb
) &&
4129 lvb
->lvb_version
== OCFS2_QINFO_LVB_VERSION
) {
4130 info
->dqi_bgrace
= be32_to_cpu(lvb
->lvb_bgrace
);
4131 info
->dqi_igrace
= be32_to_cpu(lvb
->lvb_igrace
);
4132 oinfo
->dqi_syncms
= be32_to_cpu(lvb
->lvb_syncms
);
4133 oinfo
->dqi_gi
.dqi_blocks
= be32_to_cpu(lvb
->lvb_blocks
);
4134 oinfo
->dqi_gi
.dqi_free_blk
= be32_to_cpu(lvb
->lvb_free_blk
);
4135 oinfo
->dqi_gi
.dqi_free_entry
=
4136 be32_to_cpu(lvb
->lvb_free_entry
);
4138 status
= ocfs2_read_quota_phys_block(oinfo
->dqi_gqinode
,
4139 oinfo
->dqi_giblk
, &bh
);
4144 gdinfo
= (struct ocfs2_global_disk_dqinfo
*)
4145 (bh
->b_data
+ OCFS2_GLOBAL_INFO_OFF
);
4146 info
->dqi_bgrace
= le32_to_cpu(gdinfo
->dqi_bgrace
);
4147 info
->dqi_igrace
= le32_to_cpu(gdinfo
->dqi_igrace
);
4148 oinfo
->dqi_syncms
= le32_to_cpu(gdinfo
->dqi_syncms
);
4149 oinfo
->dqi_gi
.dqi_blocks
= le32_to_cpu(gdinfo
->dqi_blocks
);
4150 oinfo
->dqi_gi
.dqi_free_blk
= le32_to_cpu(gdinfo
->dqi_free_blk
);
4151 oinfo
->dqi_gi
.dqi_free_entry
=
4152 le32_to_cpu(gdinfo
->dqi_free_entry
);
4154 ocfs2_track_lock_refresh(lockres
);
4161 /* Lock quota info, this function expects at least shared lock on the quota file
4162 * so that we can safely refresh quota info from disk. */
4163 int ocfs2_qinfo_lock(struct ocfs2_mem_dqinfo
*oinfo
, int ex
)
4165 struct ocfs2_lock_res
*lockres
= &oinfo
->dqi_gqlock
;
4166 struct ocfs2_super
*osb
= OCFS2_SB(oinfo
->dqi_gi
.dqi_sb
);
4167 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
4170 /* On RO devices, locking really isn't needed... */
4171 if (ocfs2_is_hard_readonly(osb
)) {
4176 if (ocfs2_mount_local(osb
))
4179 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
4184 if (!ocfs2_should_refresh_lock_res(lockres
))
4186 /* OK, we have the lock but we need to refresh the quota info */
4187 status
= ocfs2_refresh_qinfo(oinfo
);
4189 ocfs2_qinfo_unlock(oinfo
, ex
);
4190 ocfs2_complete_lock_res_refresh(lockres
, status
);
4195 int ocfs2_refcount_lock(struct ocfs2_refcount_tree
*ref_tree
, int ex
)
4198 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
4199 struct ocfs2_lock_res
*lockres
= &ref_tree
->rf_lockres
;
4200 struct ocfs2_super
*osb
= lockres
->l_priv
;
4203 if (ocfs2_is_hard_readonly(osb
))
4206 if (ocfs2_mount_local(osb
))
4209 status
= ocfs2_cluster_lock(osb
, lockres
, level
, 0, 0);
4216 void ocfs2_refcount_unlock(struct ocfs2_refcount_tree
*ref_tree
, int ex
)
4218 int level
= ex
? DLM_LOCK_EX
: DLM_LOCK_PR
;
4219 struct ocfs2_lock_res
*lockres
= &ref_tree
->rf_lockres
;
4220 struct ocfs2_super
*osb
= lockres
->l_priv
;
4222 if (!ocfs2_mount_local(osb
))
4223 ocfs2_cluster_unlock(osb
, lockres
, level
);
4226 static void ocfs2_process_blocked_lock(struct ocfs2_super
*osb
,
4227 struct ocfs2_lock_res
*lockres
)
4230 struct ocfs2_unblock_ctl ctl
= {0, 0,};
4231 unsigned long flags
;
4233 /* Our reference to the lockres in this function can be
4234 * considered valid until we remove the OCFS2_LOCK_QUEUED
4238 BUG_ON(!lockres
->l_ops
);
4240 mlog(ML_BASTS
, "lockres %s blocked\n", lockres
->l_name
);
4242 /* Detect whether a lock has been marked as going away while
4243 * the downconvert thread was processing other things. A lock can
4244 * still be marked with OCFS2_LOCK_FREEING after this check,
4245 * but short circuiting here will still save us some
4247 spin_lock_irqsave(&lockres
->l_lock
, flags
);
4248 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
)
4250 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
4252 status
= ocfs2_unblock_lock(osb
, lockres
, &ctl
);
4256 spin_lock_irqsave(&lockres
->l_lock
, flags
);
4258 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
|| !ctl
.requeue
) {
4259 lockres_clear_flags(lockres
, OCFS2_LOCK_QUEUED
);
4261 ocfs2_schedule_blocked_lock(osb
, lockres
);
4263 mlog(ML_BASTS
, "lockres %s, requeue = %s.\n", lockres
->l_name
,
4264 ctl
.requeue
? "yes" : "no");
4265 spin_unlock_irqrestore(&lockres
->l_lock
, flags
);
4267 if (ctl
.unblock_action
!= UNBLOCK_CONTINUE
4268 && lockres
->l_ops
->post_unlock
)
4269 lockres
->l_ops
->post_unlock(osb
, lockres
);
4272 static void ocfs2_schedule_blocked_lock(struct ocfs2_super
*osb
,
4273 struct ocfs2_lock_res
*lockres
)
4275 unsigned long flags
;
4277 assert_spin_locked(&lockres
->l_lock
);
4279 if (lockres
->l_flags
& OCFS2_LOCK_FREEING
) {
4280 /* Do not schedule a lock for downconvert when it's on
4281 * the way to destruction - any nodes wanting access
4282 * to the resource will get it soon. */
4283 mlog(ML_BASTS
, "lockres %s won't be scheduled: flags 0x%lx\n",
4284 lockres
->l_name
, lockres
->l_flags
);
4288 lockres_or_flags(lockres
, OCFS2_LOCK_QUEUED
);
4290 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4291 if (list_empty(&lockres
->l_blocked_list
)) {
4292 list_add_tail(&lockres
->l_blocked_list
,
4293 &osb
->blocked_lock_list
);
4294 osb
->blocked_lock_count
++;
4296 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4299 static void ocfs2_downconvert_thread_do_work(struct ocfs2_super
*osb
)
4301 unsigned long processed
;
4302 unsigned long flags
;
4303 struct ocfs2_lock_res
*lockres
;
4305 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4306 /* grab this early so we know to try again if a state change and
4307 * wake happens part-way through our work */
4308 osb
->dc_work_sequence
= osb
->dc_wake_sequence
;
4310 processed
= osb
->blocked_lock_count
;
4312 * blocked lock processing in this loop might call iput which can
4313 * remove items off osb->blocked_lock_list. Downconvert up to
4314 * 'processed' number of locks, but stop short if we had some
4315 * removed in ocfs2_mark_lockres_freeing when downconverting.
4317 while (processed
&& !list_empty(&osb
->blocked_lock_list
)) {
4318 lockres
= list_entry(osb
->blocked_lock_list
.next
,
4319 struct ocfs2_lock_res
, l_blocked_list
);
4320 list_del_init(&lockres
->l_blocked_list
);
4321 osb
->blocked_lock_count
--;
4322 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4327 ocfs2_process_blocked_lock(osb
, lockres
);
4329 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4331 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4334 static int ocfs2_downconvert_thread_lists_empty(struct ocfs2_super
*osb
)
4337 unsigned long flags
;
4339 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4340 if (list_empty(&osb
->blocked_lock_list
))
4343 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4347 static int ocfs2_downconvert_thread_should_wake(struct ocfs2_super
*osb
)
4349 int should_wake
= 0;
4350 unsigned long flags
;
4352 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4353 if (osb
->dc_work_sequence
!= osb
->dc_wake_sequence
)
4355 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4360 static int ocfs2_downconvert_thread(void *arg
)
4363 struct ocfs2_super
*osb
= arg
;
4365 /* only quit once we've been asked to stop and there is no more
4367 while (!(kthread_should_stop() &&
4368 ocfs2_downconvert_thread_lists_empty(osb
))) {
4370 wait_event_interruptible(osb
->dc_event
,
4371 ocfs2_downconvert_thread_should_wake(osb
) ||
4372 kthread_should_stop());
4374 mlog(0, "downconvert_thread: awoken\n");
4376 ocfs2_downconvert_thread_do_work(osb
);
4379 osb
->dc_task
= NULL
;
4383 void ocfs2_wake_downconvert_thread(struct ocfs2_super
*osb
)
4385 unsigned long flags
;
4387 spin_lock_irqsave(&osb
->dc_task_lock
, flags
);
4388 /* make sure the voting thread gets a swipe at whatever changes
4389 * the caller may have made to the voting state */
4390 osb
->dc_wake_sequence
++;
4391 spin_unlock_irqrestore(&osb
->dc_task_lock
, flags
);
4392 wake_up(&osb
->dc_event
);