1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Defines functions of journalling api
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.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/kthread.h>
32 #define MLOG_MASK_PREFIX ML_JOURNAL
33 #include <cluster/masklog.h>
40 #include "extent_map.h"
41 #include "heartbeat.h"
44 #include "localalloc.h"
49 #include "buffer_head_io.h"
51 DEFINE_SPINLOCK(trans_inc_lock
);
53 static int ocfs2_force_read_journal(struct inode
*inode
);
54 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
56 static int __ocfs2_recovery_thread(void *arg
);
57 static int ocfs2_commit_cache(struct ocfs2_super
*osb
);
58 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
);
59 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
60 int dirty
, int replayed
);
61 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
63 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
65 static int ocfs2_commit_thread(void *arg
);
69 * The recovery_list is a simple linked list of node numbers to recover.
70 * It is protected by the recovery_lock.
73 struct ocfs2_recovery_map
{
75 unsigned int *rm_entries
;
78 int ocfs2_recovery_init(struct ocfs2_super
*osb
)
80 struct ocfs2_recovery_map
*rm
;
82 mutex_init(&osb
->recovery_lock
);
83 osb
->disable_recovery
= 0;
84 osb
->recovery_thread_task
= NULL
;
85 init_waitqueue_head(&osb
->recovery_event
);
87 rm
= kzalloc(sizeof(struct ocfs2_recovery_map
) +
88 osb
->max_slots
* sizeof(unsigned int),
95 rm
->rm_entries
= (unsigned int *)((char *)rm
+
96 sizeof(struct ocfs2_recovery_map
));
97 osb
->recovery_map
= rm
;
102 /* we can't grab the goofy sem lock from inside wait_event, so we use
103 * memory barriers to make sure that we'll see the null task before
105 static int ocfs2_recovery_thread_running(struct ocfs2_super
*osb
)
108 return osb
->recovery_thread_task
!= NULL
;
111 void ocfs2_recovery_exit(struct ocfs2_super
*osb
)
113 struct ocfs2_recovery_map
*rm
;
115 /* disable any new recovery threads and wait for any currently
116 * running ones to exit. Do this before setting the vol_state. */
117 mutex_lock(&osb
->recovery_lock
);
118 osb
->disable_recovery
= 1;
119 mutex_unlock(&osb
->recovery_lock
);
120 wait_event(osb
->recovery_event
, !ocfs2_recovery_thread_running(osb
));
122 /* At this point, we know that no more recovery threads can be
123 * launched, so wait for any recovery completion work to
125 flush_workqueue(ocfs2_wq
);
128 * Now that recovery is shut down, and the osb is about to be
129 * freed, the osb_lock is not taken here.
131 rm
= osb
->recovery_map
;
132 /* XXX: Should we bug if there are dirty entries? */
137 static int __ocfs2_recovery_map_test(struct ocfs2_super
*osb
,
138 unsigned int node_num
)
141 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
143 assert_spin_locked(&osb
->osb_lock
);
145 for (i
= 0; i
< rm
->rm_used
; i
++) {
146 if (rm
->rm_entries
[i
] == node_num
)
153 /* Behaves like test-and-set. Returns the previous value */
154 static int ocfs2_recovery_map_set(struct ocfs2_super
*osb
,
155 unsigned int node_num
)
157 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
159 spin_lock(&osb
->osb_lock
);
160 if (__ocfs2_recovery_map_test(osb
, node_num
)) {
161 spin_unlock(&osb
->osb_lock
);
165 /* XXX: Can this be exploited? Not from o2dlm... */
166 BUG_ON(rm
->rm_used
>= osb
->max_slots
);
168 rm
->rm_entries
[rm
->rm_used
] = node_num
;
170 spin_unlock(&osb
->osb_lock
);
175 static void ocfs2_recovery_map_clear(struct ocfs2_super
*osb
,
176 unsigned int node_num
)
179 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
181 spin_lock(&osb
->osb_lock
);
183 for (i
= 0; i
< rm
->rm_used
; i
++) {
184 if (rm
->rm_entries
[i
] == node_num
)
188 if (i
< rm
->rm_used
) {
189 /* XXX: be careful with the pointer math */
190 memmove(&(rm
->rm_entries
[i
]), &(rm
->rm_entries
[i
+ 1]),
191 (rm
->rm_used
- i
- 1) * sizeof(unsigned int));
195 spin_unlock(&osb
->osb_lock
);
198 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
201 unsigned int flushed
;
202 unsigned long old_id
;
203 struct ocfs2_journal
*journal
= NULL
;
207 journal
= osb
->journal
;
209 /* Flush all pending commits and checkpoint the journal. */
210 down_write(&journal
->j_trans_barrier
);
212 if (atomic_read(&journal
->j_num_trans
) == 0) {
213 up_write(&journal
->j_trans_barrier
);
214 mlog(0, "No transactions for me to flush!\n");
218 jbd2_journal_lock_updates(journal
->j_journal
);
219 status
= jbd2_journal_flush(journal
->j_journal
);
220 jbd2_journal_unlock_updates(journal
->j_journal
);
222 up_write(&journal
->j_trans_barrier
);
227 old_id
= ocfs2_inc_trans_id(journal
);
229 flushed
= atomic_read(&journal
->j_num_trans
);
230 atomic_set(&journal
->j_num_trans
, 0);
231 up_write(&journal
->j_trans_barrier
);
233 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
234 journal
->j_trans_id
, flushed
);
236 ocfs2_wake_downconvert_thread(osb
);
237 wake_up(&journal
->j_checkpointed
);
243 /* pass it NULL and it will allocate a new handle object for you. If
244 * you pass it a handle however, it may still return error, in which
245 * case it has free'd the passed handle for you. */
246 handle_t
*ocfs2_start_trans(struct ocfs2_super
*osb
, int max_buffs
)
248 journal_t
*journal
= osb
->journal
->j_journal
;
251 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
253 if (ocfs2_is_hard_readonly(osb
))
254 return ERR_PTR(-EROFS
);
256 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
257 BUG_ON(max_buffs
<= 0);
259 /* JBD might support this, but our journalling code doesn't yet. */
260 if (journal_current_handle()) {
261 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
265 down_read(&osb
->journal
->j_trans_barrier
);
267 handle
= jbd2_journal_start(journal
, max_buffs
);
268 if (IS_ERR(handle
)) {
269 up_read(&osb
->journal
->j_trans_barrier
);
271 mlog_errno(PTR_ERR(handle
));
273 if (is_journal_aborted(journal
)) {
274 ocfs2_abort(osb
->sb
, "Detected aborted journal");
275 handle
= ERR_PTR(-EROFS
);
278 if (!ocfs2_mount_local(osb
))
279 atomic_inc(&(osb
->journal
->j_num_trans
));
285 int ocfs2_commit_trans(struct ocfs2_super
*osb
,
289 struct ocfs2_journal
*journal
= osb
->journal
;
293 ret
= jbd2_journal_stop(handle
);
297 up_read(&journal
->j_trans_barrier
);
303 * 'nblocks' is what you want to add to the current
304 * transaction. extend_trans will either extend the current handle by
305 * nblocks, or commit it and start a new one with nblocks credits.
307 * This might call jbd2_journal_restart() which will commit dirty buffers
308 * and then restart the transaction. Before calling
309 * ocfs2_extend_trans(), any changed blocks should have been
310 * dirtied. After calling it, all blocks which need to be changed must
311 * go through another set of journal_access/journal_dirty calls.
313 * WARNING: This will not release any semaphores or disk locks taken
314 * during the transaction, so make sure they were taken *before*
315 * start_trans or we'll have ordering deadlocks.
317 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
318 * good because transaction ids haven't yet been recorded on the
319 * cluster locks associated with this handle.
321 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
330 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
332 #ifdef CONFIG_OCFS2_DEBUG_FS
335 status
= jbd2_journal_extend(handle
, nblocks
);
344 "jbd2_journal_extend failed, trying "
345 "jbd2_journal_restart\n");
346 status
= jbd2_journal_restart(handle
, nblocks
);
360 int ocfs2_journal_access(handle_t
*handle
,
362 struct buffer_head
*bh
,
371 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
372 (unsigned long long)bh
->b_blocknr
, type
,
373 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
374 "OCFS2_JOURNAL_ACCESS_CREATE" :
375 "OCFS2_JOURNAL_ACCESS_WRITE",
378 /* we can safely remove this assertion after testing. */
379 if (!buffer_uptodate(bh
)) {
380 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
381 mlog(ML_ERROR
, "b_blocknr=%llu\n",
382 (unsigned long long)bh
->b_blocknr
);
386 /* Set the current transaction information on the inode so
387 * that the locking code knows whether it can drop it's locks
388 * on this inode or not. We're protected from the commit
389 * thread updating the current transaction id until
390 * ocfs2_commit_trans() because ocfs2_start_trans() took
391 * j_trans_barrier for us. */
392 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
394 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
396 case OCFS2_JOURNAL_ACCESS_CREATE
:
397 case OCFS2_JOURNAL_ACCESS_WRITE
:
398 status
= jbd2_journal_get_write_access(handle
, bh
);
401 case OCFS2_JOURNAL_ACCESS_UNDO
:
402 status
= jbd2_journal_get_undo_access(handle
, bh
);
407 mlog(ML_ERROR
, "Uknown access type!\n");
409 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
412 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
419 int ocfs2_journal_dirty(handle_t
*handle
,
420 struct buffer_head
*bh
)
424 mlog_entry("(bh->b_blocknr=%llu)\n",
425 (unsigned long long)bh
->b_blocknr
);
427 status
= jbd2_journal_dirty_metadata(handle
, bh
);
429 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
430 "(bh->b_blocknr=%llu)\n",
431 (unsigned long long)bh
->b_blocknr
);
437 #ifdef CONFIG_OCFS2_COMPAT_JBD
438 int ocfs2_journal_dirty_data(handle_t
*handle
,
439 struct buffer_head
*bh
)
441 int err
= journal_dirty_data(handle
, bh
);
444 /* TODO: When we can handle it, abort the handle and go RO on
451 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
453 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
455 journal_t
*journal
= osb
->journal
->j_journal
;
456 unsigned long commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
458 if (osb
->osb_commit_interval
)
459 commit_interval
= osb
->osb_commit_interval
;
461 spin_lock(&journal
->j_state_lock
);
462 journal
->j_commit_interval
= commit_interval
;
463 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
464 journal
->j_flags
|= JBD2_BARRIER
;
466 journal
->j_flags
&= ~JBD2_BARRIER
;
467 spin_unlock(&journal
->j_state_lock
);
470 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
473 struct inode
*inode
= NULL
; /* the journal inode */
474 journal_t
*j_journal
= NULL
;
475 struct ocfs2_dinode
*di
= NULL
;
476 struct buffer_head
*bh
= NULL
;
477 struct ocfs2_super
*osb
;
484 osb
= journal
->j_osb
;
486 /* already have the inode for our journal */
487 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
494 if (is_bad_inode(inode
)) {
495 mlog(ML_ERROR
, "access error (bad inode)\n");
502 SET_INODE_JOURNAL(inode
);
503 OCFS2_I(inode
)->ip_open_count
++;
505 /* Skip recovery waits here - journal inode metadata never
506 * changes in a live cluster so it can be considered an
507 * exception to the rule. */
508 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
510 if (status
!= -ERESTARTSYS
)
511 mlog(ML_ERROR
, "Could not get lock on journal!\n");
516 di
= (struct ocfs2_dinode
*)bh
->b_data
;
518 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
519 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
525 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
526 mlog(0, "inode->i_blocks = %llu\n",
527 (unsigned long long)inode
->i_blocks
);
528 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
530 /* call the kernels journal init function now */
531 j_journal
= jbd2_journal_init_inode(inode
);
532 if (j_journal
== NULL
) {
533 mlog(ML_ERROR
, "Linux journal layer error\n");
538 mlog(0, "Returned from jbd2_journal_init_inode\n");
539 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
541 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
542 OCFS2_JOURNAL_DIRTY_FL
);
544 journal
->j_journal
= j_journal
;
545 journal
->j_inode
= inode
;
548 ocfs2_set_journal_params(osb
);
550 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
556 ocfs2_inode_unlock(inode
, 1);
560 OCFS2_I(inode
)->ip_open_count
--;
569 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode
*di
)
571 le32_add_cpu(&(di
->id1
.journal1
.ij_recovery_generation
), 1);
574 static u32
ocfs2_get_recovery_generation(struct ocfs2_dinode
*di
)
576 return le32_to_cpu(di
->id1
.journal1
.ij_recovery_generation
);
579 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
580 int dirty
, int replayed
)
584 struct ocfs2_journal
*journal
= osb
->journal
;
585 struct buffer_head
*bh
= journal
->j_bh
;
586 struct ocfs2_dinode
*fe
;
590 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
591 if (!OCFS2_IS_VALID_DINODE(fe
)) {
592 /* This is called from startup/shutdown which will
593 * handle the errors in a specific manner, so no need
594 * to call ocfs2_error() here. */
595 mlog(ML_ERROR
, "Journal dinode %llu has invalid "
597 (unsigned long long)le64_to_cpu(fe
->i_blkno
), 7,
603 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
605 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
607 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
608 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
611 ocfs2_bump_recovery_generation(fe
);
613 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
623 * If the journal has been kmalloc'd it needs to be freed after this
626 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
628 struct ocfs2_journal
*journal
= NULL
;
630 struct inode
*inode
= NULL
;
631 int num_running_trans
= 0;
637 journal
= osb
->journal
;
641 inode
= journal
->j_inode
;
643 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
646 /* need to inc inode use count - jbd2_journal_destroy will iput. */
650 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
651 if (num_running_trans
> 0)
652 mlog(0, "Shutting down journal: must wait on %d "
653 "running transactions!\n",
656 /* Do a commit_cache here. It will flush our journal, *and*
657 * release any locks that are still held.
658 * set the SHUTDOWN flag and release the trans lock.
659 * the commit thread will take the trans lock for us below. */
660 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
662 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
663 * drop the trans_lock (which we want to hold until we
664 * completely destroy the journal. */
665 if (osb
->commit_task
) {
666 /* Wait for the commit thread */
667 mlog(0, "Waiting for ocfs2commit to exit....\n");
668 kthread_stop(osb
->commit_task
);
669 osb
->commit_task
= NULL
;
672 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
674 if (ocfs2_mount_local(osb
)) {
675 jbd2_journal_lock_updates(journal
->j_journal
);
676 status
= jbd2_journal_flush(journal
->j_journal
);
677 jbd2_journal_unlock_updates(journal
->j_journal
);
684 * Do not toggle if flush was unsuccessful otherwise
685 * will leave dirty metadata in a "clean" journal
687 status
= ocfs2_journal_toggle_dirty(osb
, 0, 0);
692 /* Shutdown the kernel journal system */
693 jbd2_journal_destroy(journal
->j_journal
);
695 OCFS2_I(inode
)->ip_open_count
--;
697 /* unlock our journal */
698 ocfs2_inode_unlock(inode
, 1);
700 brelse(journal
->j_bh
);
701 journal
->j_bh
= NULL
;
703 journal
->j_state
= OCFS2_JOURNAL_FREE
;
705 // up_write(&journal->j_trans_barrier);
712 static void ocfs2_clear_journal_error(struct super_block
*sb
,
718 olderr
= jbd2_journal_errno(journal
);
720 mlog(ML_ERROR
, "File system error %d recorded in "
721 "journal %u.\n", olderr
, slot
);
722 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
725 jbd2_journal_ack_err(journal
);
726 jbd2_journal_clear_err(journal
);
730 int ocfs2_journal_load(struct ocfs2_journal
*journal
, int local
, int replayed
)
733 struct ocfs2_super
*osb
;
739 osb
= journal
->j_osb
;
741 status
= jbd2_journal_load(journal
->j_journal
);
743 mlog(ML_ERROR
, "Failed to load journal!\n");
747 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
749 status
= ocfs2_journal_toggle_dirty(osb
, 1, replayed
);
755 /* Launch the commit thread */
757 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
,
759 if (IS_ERR(osb
->commit_task
)) {
760 status
= PTR_ERR(osb
->commit_task
);
761 osb
->commit_task
= NULL
;
762 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, "
767 osb
->commit_task
= NULL
;
775 /* 'full' flag tells us whether we clear out all blocks or if we just
776 * mark the journal clean */
777 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
785 status
= jbd2_journal_wipe(journal
->j_journal
, full
);
791 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0, 0);
800 static int ocfs2_recovery_completed(struct ocfs2_super
*osb
)
803 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
805 spin_lock(&osb
->osb_lock
);
806 empty
= (rm
->rm_used
== 0);
807 spin_unlock(&osb
->osb_lock
);
812 void ocfs2_wait_for_recovery(struct ocfs2_super
*osb
)
814 wait_event(osb
->recovery_event
, ocfs2_recovery_completed(osb
));
818 * JBD Might read a cached version of another nodes journal file. We
819 * don't want this as this file changes often and we get no
820 * notification on those changes. The only way to be sure that we've
821 * got the most up to date version of those blocks then is to force
822 * read them off disk. Just searching through the buffer cache won't
823 * work as there may be pages backing this file which are still marked
824 * up to date. We know things can't change on this file underneath us
825 * as we have the lock by now :)
827 static int ocfs2_force_read_journal(struct inode
*inode
)
831 u64 v_blkno
, p_blkno
, p_blocks
, num_blocks
;
832 #define CONCURRENT_JOURNAL_FILL 32ULL
833 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
837 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
839 num_blocks
= ocfs2_blocks_for_bytes(inode
->i_sb
, inode
->i_size
);
841 while (v_blkno
< num_blocks
) {
842 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
843 &p_blkno
, &p_blocks
, NULL
);
849 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
850 p_blocks
= CONCURRENT_JOURNAL_FILL
;
852 /* We are reading journal data which should not
853 * be put in the uptodate cache */
854 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
855 p_blkno
, p_blocks
, bhs
, 0,
862 for(i
= 0; i
< p_blocks
; i
++) {
871 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
878 struct ocfs2_la_recovery_item
{
879 struct list_head lri_list
;
881 struct ocfs2_dinode
*lri_la_dinode
;
882 struct ocfs2_dinode
*lri_tl_dinode
;
885 /* Does the second half of the recovery process. By this point, the
886 * node is marked clean and can actually be considered recovered,
887 * hence it's no longer in the recovery map, but there's still some
888 * cleanup we can do which shouldn't happen within the recovery thread
889 * as locking in that context becomes very difficult if we are to take
890 * recovering nodes into account.
892 * NOTE: This function can and will sleep on recovery of other nodes
893 * during cluster locking, just like any other ocfs2 process.
895 void ocfs2_complete_recovery(struct work_struct
*work
)
898 struct ocfs2_journal
*journal
=
899 container_of(work
, struct ocfs2_journal
, j_recovery_work
);
900 struct ocfs2_super
*osb
= journal
->j_osb
;
901 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
902 struct ocfs2_la_recovery_item
*item
, *n
;
903 LIST_HEAD(tmp_la_list
);
907 mlog(0, "completing recovery from keventd\n");
909 spin_lock(&journal
->j_lock
);
910 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
911 spin_unlock(&journal
->j_lock
);
913 list_for_each_entry_safe(item
, n
, &tmp_la_list
, lri_list
) {
914 list_del_init(&item
->lri_list
);
916 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
918 la_dinode
= item
->lri_la_dinode
;
920 mlog(0, "Clean up local alloc %llu\n",
921 (unsigned long long)le64_to_cpu(la_dinode
->i_blkno
));
923 ret
= ocfs2_complete_local_alloc_recovery(osb
,
931 tl_dinode
= item
->lri_tl_dinode
;
933 mlog(0, "Clean up truncate log %llu\n",
934 (unsigned long long)le64_to_cpu(tl_dinode
->i_blkno
));
936 ret
= ocfs2_complete_truncate_log_recovery(osb
,
944 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
951 mlog(0, "Recovery completion\n");
955 /* NOTE: This function always eats your references to la_dinode and
956 * tl_dinode, either manually on error, or by passing them to
957 * ocfs2_complete_recovery */
958 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
960 struct ocfs2_dinode
*la_dinode
,
961 struct ocfs2_dinode
*tl_dinode
)
963 struct ocfs2_la_recovery_item
*item
;
965 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
967 /* Though we wish to avoid it, we are in fact safe in
968 * skipping local alloc cleanup as fsck.ocfs2 is more
969 * than capable of reclaiming unused space. */
980 INIT_LIST_HEAD(&item
->lri_list
);
981 item
->lri_la_dinode
= la_dinode
;
982 item
->lri_slot
= slot_num
;
983 item
->lri_tl_dinode
= tl_dinode
;
985 spin_lock(&journal
->j_lock
);
986 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
987 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
988 spin_unlock(&journal
->j_lock
);
991 /* Called by the mount code to queue recovery the last part of
992 * recovery for it's own slot. */
993 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
995 struct ocfs2_journal
*journal
= osb
->journal
;
998 /* No need to queue up our truncate_log as regular
999 * cleanup will catch that. */
1000 ocfs2_queue_recovery_completion(journal
,
1002 osb
->local_alloc_copy
,
1004 ocfs2_schedule_truncate_log_flush(osb
, 0);
1006 osb
->local_alloc_copy
= NULL
;
1011 static int __ocfs2_recovery_thread(void *arg
)
1013 int status
, node_num
;
1014 struct ocfs2_super
*osb
= arg
;
1015 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
1019 status
= ocfs2_wait_on_mount(osb
);
1025 status
= ocfs2_super_lock(osb
, 1);
1031 spin_lock(&osb
->osb_lock
);
1032 while (rm
->rm_used
) {
1033 /* It's always safe to remove entry zero, as we won't
1034 * clear it until ocfs2_recover_node() has succeeded. */
1035 node_num
= rm
->rm_entries
[0];
1036 spin_unlock(&osb
->osb_lock
);
1038 status
= ocfs2_recover_node(osb
, node_num
);
1040 ocfs2_recovery_map_clear(osb
, node_num
);
1043 "Error %d recovering node %d on device (%u,%u)!\n",
1045 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1046 mlog(ML_ERROR
, "Volume requires unmount.\n");
1049 spin_lock(&osb
->osb_lock
);
1051 spin_unlock(&osb
->osb_lock
);
1052 mlog(0, "All nodes recovered\n");
1054 /* Refresh all journal recovery generations from disk */
1055 status
= ocfs2_check_journals_nolocks(osb
);
1056 status
= (status
== -EROFS
) ? 0 : status
;
1060 ocfs2_super_unlock(osb
, 1);
1062 /* We always run recovery on our own orphan dir - the dead
1063 * node(s) may have disallowd a previos inode delete. Re-processing
1064 * is therefore required. */
1065 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
1069 mutex_lock(&osb
->recovery_lock
);
1070 if (!status
&& !ocfs2_recovery_completed(osb
)) {
1071 mutex_unlock(&osb
->recovery_lock
);
1075 osb
->recovery_thread_task
= NULL
;
1076 mb(); /* sync with ocfs2_recovery_thread_running */
1077 wake_up(&osb
->recovery_event
);
1079 mutex_unlock(&osb
->recovery_lock
);
1082 /* no one is callint kthread_stop() for us so the kthread() api
1083 * requires that we call do_exit(). And it isn't exported, but
1084 * complete_and_exit() seems to be a minimal wrapper around it. */
1085 complete_and_exit(NULL
, status
);
1089 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
1091 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1092 node_num
, osb
->node_num
);
1094 mutex_lock(&osb
->recovery_lock
);
1095 if (osb
->disable_recovery
)
1098 /* People waiting on recovery will wait on
1099 * the recovery map to empty. */
1100 if (ocfs2_recovery_map_set(osb
, node_num
))
1101 mlog(0, "node %d already in recovery map.\n", node_num
);
1103 mlog(0, "starting recovery thread...\n");
1105 if (osb
->recovery_thread_task
)
1108 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
1110 if (IS_ERR(osb
->recovery_thread_task
)) {
1111 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
1112 osb
->recovery_thread_task
= NULL
;
1116 mutex_unlock(&osb
->recovery_lock
);
1117 wake_up(&osb
->recovery_event
);
1122 static int ocfs2_read_journal_inode(struct ocfs2_super
*osb
,
1124 struct buffer_head
**bh
,
1125 struct inode
**ret_inode
)
1127 int status
= -EACCES
;
1128 struct inode
*inode
= NULL
;
1130 BUG_ON(slot_num
>= osb
->max_slots
);
1132 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1134 if (!inode
|| is_bad_inode(inode
)) {
1138 SET_INODE_JOURNAL(inode
);
1140 status
= ocfs2_read_block(osb
, OCFS2_I(inode
)->ip_blkno
, bh
, 0, inode
);
1150 if (status
|| !ret_inode
)
1158 /* Does the actual journal replay and marks the journal inode as
1159 * clean. Will only replay if the journal inode is marked dirty. */
1160 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
1167 struct inode
*inode
= NULL
;
1168 struct ocfs2_dinode
*fe
;
1169 journal_t
*journal
= NULL
;
1170 struct buffer_head
*bh
= NULL
;
1173 status
= ocfs2_read_journal_inode(osb
, slot_num
, &bh
, &inode
);
1179 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
1180 slot_reco_gen
= ocfs2_get_recovery_generation(fe
);
1185 * As the fs recovery is asynchronous, there is a small chance that
1186 * another node mounted (and recovered) the slot before the recovery
1187 * thread could get the lock. To handle that, we dirty read the journal
1188 * inode for that slot to get the recovery generation. If it is
1189 * different than what we expected, the slot has been recovered.
1190 * If not, it needs recovery.
1192 if (osb
->slot_recovery_generations
[slot_num
] != slot_reco_gen
) {
1193 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num
,
1194 osb
->slot_recovery_generations
[slot_num
], slot_reco_gen
);
1195 osb
->slot_recovery_generations
[slot_num
] = slot_reco_gen
;
1200 /* Continue with recovery as the journal has not yet been recovered */
1202 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
1204 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status
);
1205 if (status
!= -ERESTARTSYS
)
1206 mlog(ML_ERROR
, "Could not lock journal!\n");
1211 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
1213 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1214 slot_reco_gen
= ocfs2_get_recovery_generation(fe
);
1216 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
1217 mlog(0, "No recovery required for node %d\n", node_num
);
1218 /* Refresh recovery generation for the slot */
1219 osb
->slot_recovery_generations
[slot_num
] = slot_reco_gen
;
1223 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
1225 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1227 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
1229 status
= ocfs2_force_read_journal(inode
);
1235 mlog(0, "calling journal_init_inode\n");
1236 journal
= jbd2_journal_init_inode(inode
);
1237 if (journal
== NULL
) {
1238 mlog(ML_ERROR
, "Linux journal layer error\n");
1243 status
= jbd2_journal_load(journal
);
1248 jbd2_journal_destroy(journal
);
1252 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1254 /* wipe the journal */
1255 mlog(0, "flushing the journal.\n");
1256 jbd2_journal_lock_updates(journal
);
1257 status
= jbd2_journal_flush(journal
);
1258 jbd2_journal_unlock_updates(journal
);
1262 /* This will mark the node clean */
1263 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1264 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1265 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1267 /* Increment recovery generation to indicate successful recovery */
1268 ocfs2_bump_recovery_generation(fe
);
1269 osb
->slot_recovery_generations
[slot_num
] =
1270 ocfs2_get_recovery_generation(fe
);
1272 status
= ocfs2_write_block(osb
, bh
, inode
);
1279 jbd2_journal_destroy(journal
);
1282 /* drop the lock on this nodes journal */
1284 ocfs2_inode_unlock(inode
, 1);
1297 * Do the most important parts of node recovery:
1298 * - Replay it's journal
1299 * - Stamp a clean local allocator file
1300 * - Stamp a clean truncate log
1301 * - Mark the node clean
1303 * If this function completes without error, a node in OCFS2 can be
1304 * said to have been safely recovered. As a result, failure during the
1305 * second part of a nodes recovery process (local alloc recovery) is
1306 * far less concerning.
1308 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1313 struct ocfs2_dinode
*la_copy
= NULL
;
1314 struct ocfs2_dinode
*tl_copy
= NULL
;
1316 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1317 node_num
, osb
->node_num
);
1319 mlog(0, "checking node %d\n", node_num
);
1321 /* Should not ever be called to recover ourselves -- in that
1322 * case we should've called ocfs2_journal_load instead. */
1323 BUG_ON(osb
->node_num
== node_num
);
1325 slot_num
= ocfs2_node_num_to_slot(osb
, node_num
);
1326 if (slot_num
== -ENOENT
) {
1328 mlog(0, "no slot for this node, so no recovery required.\n");
1332 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1334 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1336 if (status
== -EBUSY
) {
1337 mlog(0, "Skipping recovery for slot %u (node %u) "
1338 "as another node has recovered it\n", slot_num
,
1347 /* Stamp a clean local alloc file AFTER recovering the journal... */
1348 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1354 /* An error from begin_truncate_log_recovery is not
1355 * serious enough to warrant halting the rest of
1357 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1361 /* Likewise, this would be a strange but ultimately not so
1362 * harmful place to get an error... */
1363 status
= ocfs2_clear_slot(osb
, slot_num
);
1367 /* This will kfree the memory pointed to by la_copy and tl_copy */
1368 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1378 /* Test node liveness by trylocking his journal. If we get the lock,
1379 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1380 * still alive (we couldn't get the lock) and < 0 on error. */
1381 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1385 struct inode
*inode
= NULL
;
1387 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1389 if (inode
== NULL
) {
1390 mlog(ML_ERROR
, "access error\n");
1394 if (is_bad_inode(inode
)) {
1395 mlog(ML_ERROR
, "access error (bad inode)\n");
1401 SET_INODE_JOURNAL(inode
);
1403 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1404 status
= ocfs2_inode_lock_full(inode
, NULL
, 1, flags
);
1406 if (status
!= -EAGAIN
)
1411 ocfs2_inode_unlock(inode
, 1);
1419 /* Call this underneath ocfs2_super_lock. It also assumes that the
1420 * slot info struct has been updated from disk. */
1421 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1423 unsigned int node_num
;
1426 struct buffer_head
*bh
= NULL
;
1427 struct ocfs2_dinode
*di
;
1429 /* This is called with the super block cluster lock, so we
1430 * know that the slot map can't change underneath us. */
1432 for (i
= 0; i
< osb
->max_slots
; i
++) {
1433 /* Read journal inode to get the recovery generation */
1434 status
= ocfs2_read_journal_inode(osb
, i
, &bh
, NULL
);
1439 di
= (struct ocfs2_dinode
*)bh
->b_data
;
1440 gen
= ocfs2_get_recovery_generation(di
);
1444 spin_lock(&osb
->osb_lock
);
1445 osb
->slot_recovery_generations
[i
] = gen
;
1447 mlog(0, "Slot %u recovery generation is %u\n", i
,
1448 osb
->slot_recovery_generations
[i
]);
1450 if (i
== osb
->slot_num
) {
1451 spin_unlock(&osb
->osb_lock
);
1455 status
= ocfs2_slot_to_node_num_locked(osb
, i
, &node_num
);
1456 if (status
== -ENOENT
) {
1457 spin_unlock(&osb
->osb_lock
);
1461 if (__ocfs2_recovery_map_test(osb
, node_num
)) {
1462 spin_unlock(&osb
->osb_lock
);
1465 spin_unlock(&osb
->osb_lock
);
1467 /* Ok, we have a slot occupied by another node which
1468 * is not in the recovery map. We trylock his journal
1469 * file here to test if he's alive. */
1470 status
= ocfs2_trylock_journal(osb
, i
);
1472 /* Since we're called from mount, we know that
1473 * the recovery thread can't race us on
1474 * setting / checking the recovery bits. */
1475 ocfs2_recovery_thread(osb
, node_num
);
1476 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1488 struct ocfs2_orphan_filldir_priv
{
1490 struct ocfs2_super
*osb
;
1493 static int ocfs2_orphan_filldir(void *priv
, const char *name
, int name_len
,
1494 loff_t pos
, u64 ino
, unsigned type
)
1496 struct ocfs2_orphan_filldir_priv
*p
= priv
;
1499 if (name_len
== 1 && !strncmp(".", name
, 1))
1501 if (name_len
== 2 && !strncmp("..", name
, 2))
1504 /* Skip bad inodes so that recovery can continue */
1505 iter
= ocfs2_iget(p
->osb
, ino
,
1506 OCFS2_FI_FLAG_ORPHAN_RECOVERY
, 0);
1510 mlog(0, "queue orphan %llu\n",
1511 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
1512 /* No locking is required for the next_orphan queue as there
1513 * is only ever a single process doing orphan recovery. */
1514 OCFS2_I(iter
)->ip_next_orphan
= p
->head
;
1520 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
1522 struct inode
**head
)
1525 struct inode
*orphan_dir_inode
= NULL
;
1526 struct ocfs2_orphan_filldir_priv priv
;
1532 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1533 ORPHAN_DIR_SYSTEM_INODE
,
1535 if (!orphan_dir_inode
) {
1541 mutex_lock(&orphan_dir_inode
->i_mutex
);
1542 status
= ocfs2_inode_lock(orphan_dir_inode
, NULL
, 0);
1548 status
= ocfs2_dir_foreach(orphan_dir_inode
, &pos
, &priv
,
1549 ocfs2_orphan_filldir
);
1558 ocfs2_inode_unlock(orphan_dir_inode
, 0);
1560 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1561 iput(orphan_dir_inode
);
1565 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
1570 spin_lock(&osb
->osb_lock
);
1571 ret
= !osb
->osb_orphan_wipes
[slot
];
1572 spin_unlock(&osb
->osb_lock
);
1576 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
1579 spin_lock(&osb
->osb_lock
);
1580 /* Mark ourselves such that new processes in delete_inode()
1581 * know to quit early. */
1582 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1583 while (osb
->osb_orphan_wipes
[slot
]) {
1584 /* If any processes are already in the middle of an
1585 * orphan wipe on this dir, then we need to wait for
1587 spin_unlock(&osb
->osb_lock
);
1588 wait_event_interruptible(osb
->osb_wipe_event
,
1589 ocfs2_orphan_recovery_can_continue(osb
, slot
));
1590 spin_lock(&osb
->osb_lock
);
1592 spin_unlock(&osb
->osb_lock
);
1595 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
1598 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1602 * Orphan recovery. Each mounted node has it's own orphan dir which we
1603 * must run during recovery. Our strategy here is to build a list of
1604 * the inodes in the orphan dir and iget/iput them. The VFS does
1605 * (most) of the rest of the work.
1607 * Orphan recovery can happen at any time, not just mount so we have a
1608 * couple of extra considerations.
1610 * - We grab as many inodes as we can under the orphan dir lock -
1611 * doing iget() outside the orphan dir risks getting a reference on
1613 * - We must be sure not to deadlock with other processes on the
1614 * system wanting to run delete_inode(). This can happen when they go
1615 * to lock the orphan dir and the orphan recovery process attempts to
1616 * iget() inside the orphan dir lock. This can be avoided by
1617 * advertising our state to ocfs2_delete_inode().
1619 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1623 struct inode
*inode
= NULL
;
1625 struct ocfs2_inode_info
*oi
;
1627 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1629 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
1630 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
1631 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
1633 /* Error here should be noted, but we want to continue with as
1634 * many queued inodes as we've got. */
1639 oi
= OCFS2_I(inode
);
1640 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
1642 iter
= oi
->ip_next_orphan
;
1644 spin_lock(&oi
->ip_lock
);
1645 /* The remote delete code may have set these on the
1646 * assumption that the other node would wipe them
1647 * successfully. If they are still in the node's
1648 * orphan dir, we need to reset that state. */
1649 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1651 /* Set the proper information to get us going into
1652 * ocfs2_delete_inode. */
1653 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1654 spin_unlock(&oi
->ip_lock
);
1664 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1666 /* This check is good because ocfs2 will wait on our recovery
1667 * thread before changing it to something other than MOUNTED
1669 wait_event(osb
->osb_mount_event
,
1670 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1671 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1673 /* If there's an error on mount, then we may never get to the
1674 * MOUNTED flag, but this is set right before
1675 * dismount_volume() so we can trust it. */
1676 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1677 mlog(0, "mount error, exiting!\n");
1684 static int ocfs2_commit_thread(void *arg
)
1687 struct ocfs2_super
*osb
= arg
;
1688 struct ocfs2_journal
*journal
= osb
->journal
;
1690 /* we can trust j_num_trans here because _should_stop() is only set in
1691 * shutdown and nobody other than ourselves should be able to start
1692 * transactions. committing on shutdown might take a few iterations
1693 * as final transactions put deleted inodes on the list */
1694 while (!(kthread_should_stop() &&
1695 atomic_read(&journal
->j_num_trans
) == 0)) {
1697 wait_event_interruptible(osb
->checkpoint_event
,
1698 atomic_read(&journal
->j_num_trans
)
1699 || kthread_should_stop());
1701 status
= ocfs2_commit_cache(osb
);
1705 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1707 "commit_thread: %u transactions pending on "
1709 atomic_read(&journal
->j_num_trans
));
1716 /* Reads all the journal inodes without taking any cluster locks. Used
1717 * for hard readonly access to determine whether any journal requires
1718 * recovery. Also used to refresh the recovery generation numbers after
1719 * a journal has been recovered by another node.
1721 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1725 struct buffer_head
*di_bh
= NULL
;
1726 struct ocfs2_dinode
*di
;
1727 int journal_dirty
= 0;
1729 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1730 ret
= ocfs2_read_journal_inode(osb
, slot
, &di_bh
, NULL
);
1736 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1738 osb
->slot_recovery_generations
[slot
] =
1739 ocfs2_get_recovery_generation(di
);
1741 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1742 OCFS2_JOURNAL_DIRTY_FL
)