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>
39 #include "extent_map.h"
40 #include "heartbeat.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
52 spinlock_t trans_inc_lock
= SPIN_LOCK_UNLOCKED
;
54 static int ocfs2_force_read_journal(struct inode
*inode
);
55 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
57 static int __ocfs2_recovery_thread(void *arg
);
58 static int ocfs2_commit_cache(struct ocfs2_super
*osb
);
59 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
);
60 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal
*journal
,
61 struct ocfs2_journal_handle
*handle
);
62 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle
*handle
);
63 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
65 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
67 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
69 static int ocfs2_commit_thread(void *arg
);
71 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
76 struct ocfs2_journal
*journal
= NULL
;
80 journal
= osb
->journal
;
82 /* Flush all pending commits and checkpoint the journal. */
83 down_write(&journal
->j_trans_barrier
);
85 if (atomic_read(&journal
->j_num_trans
) == 0) {
86 up_write(&journal
->j_trans_barrier
);
87 mlog(0, "No transactions for me to flush!\n");
91 journal_lock_updates(journal
->j_journal
);
92 status
= journal_flush(journal
->j_journal
);
93 journal_unlock_updates(journal
->j_journal
);
95 up_write(&journal
->j_trans_barrier
);
100 old_id
= ocfs2_inc_trans_id(journal
);
102 flushed
= atomic_read(&journal
->j_num_trans
);
103 atomic_set(&journal
->j_num_trans
, 0);
104 up_write(&journal
->j_trans_barrier
);
106 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
107 journal
->j_trans_id
, flushed
);
109 ocfs2_kick_vote_thread(osb
);
110 wake_up(&journal
->j_checkpointed
);
116 struct ocfs2_journal_handle
*ocfs2_alloc_handle(struct ocfs2_super
*osb
)
118 struct ocfs2_journal_handle
*retval
= NULL
;
120 retval
= kcalloc(1, sizeof(*retval
), GFP_KERNEL
);
122 mlog(ML_ERROR
, "Failed to allocate memory for journal "
127 retval
->max_buffs
= 0;
128 retval
->num_locks
= 0;
129 retval
->k_handle
= NULL
;
131 INIT_LIST_HEAD(&retval
->locks
);
132 INIT_LIST_HEAD(&retval
->inode_list
);
133 retval
->journal
= osb
->journal
;
138 /* pass it NULL and it will allocate a new handle object for you. If
139 * you pass it a handle however, it may still return error, in which
140 * case it has free'd the passed handle for you. */
141 struct ocfs2_journal_handle
*ocfs2_start_trans(struct ocfs2_super
*osb
,
142 struct ocfs2_journal_handle
*handle
,
146 journal_t
*journal
= osb
->journal
->j_journal
;
148 mlog_entry("(max_buffs = %d)\n", max_buffs
);
150 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
152 if (ocfs2_is_hard_readonly(osb
)) {
157 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
158 BUG_ON(max_buffs
<= 0);
160 /* JBD might support this, but our journalling code doesn't yet. */
161 if (journal_current_handle()) {
162 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
167 handle
= ocfs2_alloc_handle(osb
);
170 mlog(ML_ERROR
, "Failed to allocate memory for journal "
175 handle
->max_buffs
= max_buffs
;
177 down_read(&osb
->journal
->j_trans_barrier
);
179 /* actually start the transaction now */
180 handle
->k_handle
= journal_start(journal
, max_buffs
);
181 if (IS_ERR(handle
->k_handle
)) {
182 up_read(&osb
->journal
->j_trans_barrier
);
184 ret
= PTR_ERR(handle
->k_handle
);
185 handle
->k_handle
= NULL
;
188 if (is_journal_aborted(journal
)) {
189 ocfs2_abort(osb
->sb
, "Detected aborted journal");
195 atomic_inc(&(osb
->journal
->j_num_trans
));
196 handle
->flags
|= OCFS2_HANDLE_STARTED
;
198 mlog_exit_ptr(handle
);
203 ocfs2_commit_unstarted_handle(handle
); /* will kfree handle */
209 void ocfs2_handle_add_inode(struct ocfs2_journal_handle
*handle
,
215 atomic_inc(&inode
->i_count
);
217 /* we're obviously changing it... */
218 mutex_lock(&inode
->i_mutex
);
221 BUG_ON(OCFS2_I(inode
)->ip_handle
);
222 BUG_ON(!list_empty(&OCFS2_I(inode
)->ip_handle_list
));
224 OCFS2_I(inode
)->ip_handle
= handle
;
225 list_del(&(OCFS2_I(inode
)->ip_handle_list
));
226 list_add_tail(&(OCFS2_I(inode
)->ip_handle_list
), &(handle
->inode_list
));
229 static void ocfs2_handle_unlock_inodes(struct ocfs2_journal_handle
*handle
)
231 struct list_head
*p
, *n
;
233 struct ocfs2_inode_info
*oi
;
235 list_for_each_safe(p
, n
, &handle
->inode_list
) {
236 oi
= list_entry(p
, struct ocfs2_inode_info
,
238 inode
= &oi
->vfs_inode
;
240 OCFS2_I(inode
)->ip_handle
= NULL
;
241 list_del_init(&OCFS2_I(inode
)->ip_handle_list
);
243 mutex_unlock(&inode
->i_mutex
);
248 /* This is trivial so we do it out of the main commit
249 * paths. Beware, it can be called from start_trans too! */
250 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle
*handle
)
254 BUG_ON(handle
->flags
& OCFS2_HANDLE_STARTED
);
256 ocfs2_handle_unlock_inodes(handle
);
257 /* You are allowed to add journal locks before the transaction
259 ocfs2_handle_cleanup_locks(handle
->journal
, handle
);
266 void ocfs2_commit_trans(struct ocfs2_journal_handle
*handle
)
268 handle_t
*jbd_handle
;
270 struct ocfs2_journal
*journal
= handle
->journal
;
276 if (!(handle
->flags
& OCFS2_HANDLE_STARTED
)) {
277 ocfs2_commit_unstarted_handle(handle
);
282 /* release inode semaphores we took during this transaction */
283 ocfs2_handle_unlock_inodes(handle
);
285 /* ocfs2_extend_trans may have had to call journal_restart
286 * which will always commit the transaction, but may return
287 * error for any number of reasons. If this is the case, we
288 * clear k_handle as it's not valid any more. */
289 if (handle
->k_handle
) {
290 jbd_handle
= handle
->k_handle
;
292 if (handle
->flags
& OCFS2_HANDLE_SYNC
)
293 jbd_handle
->h_sync
= 1;
295 jbd_handle
->h_sync
= 0;
297 /* actually stop the transaction. if we've set h_sync,
298 * it'll have been committed when we return */
299 retval
= journal_stop(jbd_handle
);
302 mlog(ML_ERROR
, "Could not commit transaction\n");
306 handle
->k_handle
= NULL
; /* it's been free'd in journal_stop */
309 ocfs2_handle_cleanup_locks(journal
, handle
);
311 up_read(&journal
->j_trans_barrier
);
318 * 'nblocks' is what you want to add to the current
319 * transaction. extend_trans will either extend the current handle by
320 * nblocks, or commit it and start a new one with nblocks credits.
322 * WARNING: This will not release any semaphores or disk locks taken
323 * during the transaction, so make sure they were taken *before*
324 * start_trans or we'll have ordering deadlocks.
326 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
327 * good because transaction ids haven't yet been recorded on the
328 * cluster locks associated with this handle.
330 int ocfs2_extend_trans(struct ocfs2_journal_handle
*handle
,
336 BUG_ON(!(handle
->flags
& OCFS2_HANDLE_STARTED
));
341 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
343 status
= journal_extend(handle
->k_handle
, nblocks
);
350 mlog(0, "journal_extend failed, trying journal_restart\n");
351 status
= journal_restart(handle
->k_handle
, nblocks
);
353 handle
->k_handle
= NULL
;
357 handle
->max_buffs
= nblocks
;
359 handle
->max_buffs
+= nblocks
;
368 int ocfs2_journal_access(struct ocfs2_journal_handle
*handle
,
370 struct buffer_head
*bh
,
378 BUG_ON(!(handle
->flags
& OCFS2_HANDLE_STARTED
));
380 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %hu\n",
381 (unsigned long long)bh
->b_blocknr
, type
,
382 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
383 "OCFS2_JOURNAL_ACCESS_CREATE" :
384 "OCFS2_JOURNAL_ACCESS_WRITE",
387 /* we can safely remove this assertion after testing. */
388 if (!buffer_uptodate(bh
)) {
389 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
390 mlog(ML_ERROR
, "b_blocknr=%llu\n",
391 (unsigned long long)bh
->b_blocknr
);
395 /* Set the current transaction information on the inode so
396 * that the locking code knows whether it can drop it's locks
397 * on this inode or not. We're protected from the commit
398 * thread updating the current transaction id until
399 * ocfs2_commit_trans() because ocfs2_start_trans() took
400 * j_trans_barrier for us. */
401 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
403 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
405 case OCFS2_JOURNAL_ACCESS_CREATE
:
406 case OCFS2_JOURNAL_ACCESS_WRITE
:
407 status
= journal_get_write_access(handle
->k_handle
, bh
);
410 case OCFS2_JOURNAL_ACCESS_UNDO
:
411 status
= journal_get_undo_access(handle
->k_handle
, bh
);
416 mlog(ML_ERROR
, "Uknown access type!\n");
418 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
421 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
428 int ocfs2_journal_dirty(struct ocfs2_journal_handle
*handle
,
429 struct buffer_head
*bh
)
433 BUG_ON(!(handle
->flags
& OCFS2_HANDLE_STARTED
));
435 mlog_entry("(bh->b_blocknr=%llu)\n",
436 (unsigned long long)bh
->b_blocknr
);
438 status
= journal_dirty_metadata(handle
->k_handle
, bh
);
440 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
441 "(bh->b_blocknr=%llu)\n",
442 (unsigned long long)bh
->b_blocknr
);
448 int ocfs2_journal_dirty_data(handle_t
*handle
,
449 struct buffer_head
*bh
)
451 int err
= journal_dirty_data(handle
, bh
);
454 /* TODO: When we can handle it, abort the handle and go RO on
460 /* We always assume you're adding a metadata lock at level 'ex' */
461 int ocfs2_handle_add_lock(struct ocfs2_journal_handle
*handle
,
465 struct ocfs2_journal_lock
*lock
;
469 lock
= kmem_cache_alloc(ocfs2_lock_cache
, GFP_NOFS
);
478 lock
->jl_inode
= inode
;
480 list_add_tail(&(lock
->jl_lock_list
), &(handle
->locks
));
489 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal
*journal
,
490 struct ocfs2_journal_handle
*handle
)
492 struct list_head
*p
, *n
;
493 struct ocfs2_journal_lock
*lock
;
496 list_for_each_safe(p
, n
, &(handle
->locks
)) {
497 lock
= list_entry(p
, struct ocfs2_journal_lock
,
499 list_del(&lock
->jl_lock_list
);
502 inode
= lock
->jl_inode
;
503 ocfs2_meta_unlock(inode
, 1);
504 if (atomic_read(&inode
->i_count
) == 1)
506 "Inode %"MLFu64
", I'm doing a last iput for!",
507 OCFS2_I(inode
)->ip_blkno
);
509 kmem_cache_free(ocfs2_lock_cache
, lock
);
513 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
515 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
517 journal_t
*journal
= osb
->journal
->j_journal
;
519 spin_lock(&journal
->j_state_lock
);
520 journal
->j_commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
521 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
522 journal
->j_flags
|= JFS_BARRIER
;
524 journal
->j_flags
&= ~JFS_BARRIER
;
525 spin_unlock(&journal
->j_state_lock
);
528 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
531 struct inode
*inode
= NULL
; /* the journal inode */
532 journal_t
*j_journal
= NULL
;
533 struct ocfs2_dinode
*di
= NULL
;
534 struct buffer_head
*bh
= NULL
;
535 struct ocfs2_super
*osb
;
542 osb
= journal
->j_osb
;
544 /* already have the inode for our journal */
545 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
552 if (is_bad_inode(inode
)) {
553 mlog(ML_ERROR
, "access error (bad inode)\n");
560 SET_INODE_JOURNAL(inode
);
561 OCFS2_I(inode
)->ip_open_count
++;
563 status
= ocfs2_meta_lock(inode
, NULL
, &bh
, 1);
565 if (status
!= -ERESTARTSYS
)
566 mlog(ML_ERROR
, "Could not get lock on journal!\n");
571 di
= (struct ocfs2_dinode
*)bh
->b_data
;
573 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
574 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
580 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
581 mlog(0, "inode->i_blocks = %lu\n", inode
->i_blocks
);
582 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
584 /* call the kernels journal init function now */
585 j_journal
= journal_init_inode(inode
);
586 if (j_journal
== NULL
) {
587 mlog(ML_ERROR
, "Linux journal layer error\n");
592 mlog(0, "Returned from journal_init_inode\n");
593 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
595 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
596 OCFS2_JOURNAL_DIRTY_FL
);
598 journal
->j_journal
= j_journal
;
599 journal
->j_inode
= inode
;
602 ocfs2_set_journal_params(osb
);
604 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
610 ocfs2_meta_unlock(inode
, 1);
614 OCFS2_I(inode
)->ip_open_count
--;
623 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
628 struct ocfs2_journal
*journal
= osb
->journal
;
629 struct buffer_head
*bh
= journal
->j_bh
;
630 struct ocfs2_dinode
*fe
;
634 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
635 if (!OCFS2_IS_VALID_DINODE(fe
)) {
636 /* This is called from startup/shutdown which will
637 * handle the errors in a specific manner, so no need
638 * to call ocfs2_error() here. */
639 mlog(ML_ERROR
, "Journal dinode %"MLFu64
" has invalid "
640 "signature: %.*s", fe
->i_blkno
, 7, fe
->i_signature
);
645 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
647 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
649 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
650 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
652 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
662 * If the journal has been kmalloc'd it needs to be freed after this
665 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
667 struct ocfs2_journal
*journal
= NULL
;
669 struct inode
*inode
= NULL
;
670 int num_running_trans
= 0;
676 journal
= osb
->journal
;
680 inode
= journal
->j_inode
;
682 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
685 /* need to inc inode use count as journal_destroy will iput. */
689 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
690 if (num_running_trans
> 0)
691 mlog(0, "Shutting down journal: must wait on %d "
692 "running transactions!\n",
695 /* Do a commit_cache here. It will flush our journal, *and*
696 * release any locks that are still held.
697 * set the SHUTDOWN flag and release the trans lock.
698 * the commit thread will take the trans lock for us below. */
699 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
701 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
702 * drop the trans_lock (which we want to hold until we
703 * completely destroy the journal. */
704 if (osb
->commit_task
) {
705 /* Wait for the commit thread */
706 mlog(0, "Waiting for ocfs2commit to exit....\n");
707 kthread_stop(osb
->commit_task
);
708 osb
->commit_task
= NULL
;
711 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
713 status
= ocfs2_journal_toggle_dirty(osb
, 0);
717 /* Shutdown the kernel journal system */
718 journal_destroy(journal
->j_journal
);
720 OCFS2_I(inode
)->ip_open_count
--;
722 /* unlock our journal */
723 ocfs2_meta_unlock(inode
, 1);
725 brelse(journal
->j_bh
);
726 journal
->j_bh
= NULL
;
728 journal
->j_state
= OCFS2_JOURNAL_FREE
;
730 // up_write(&journal->j_trans_barrier);
737 static void ocfs2_clear_journal_error(struct super_block
*sb
,
743 olderr
= journal_errno(journal
);
745 mlog(ML_ERROR
, "File system error %d recorded in "
746 "journal %u.\n", olderr
, slot
);
747 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
750 journal_ack_err(journal
);
751 journal_clear_err(journal
);
755 int ocfs2_journal_load(struct ocfs2_journal
*journal
)
758 struct ocfs2_super
*osb
;
765 osb
= journal
->j_osb
;
767 status
= journal_load(journal
->j_journal
);
769 mlog(ML_ERROR
, "Failed to load journal!\n");
773 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
775 status
= ocfs2_journal_toggle_dirty(osb
, 1);
781 /* Launch the commit thread */
782 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
, "ocfs2cmt-%d",
784 if (IS_ERR(osb
->commit_task
)) {
785 status
= PTR_ERR(osb
->commit_task
);
786 osb
->commit_task
= NULL
;
787 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, error=%d",
798 /* 'full' flag tells us whether we clear out all blocks or if we just
799 * mark the journal clean */
800 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
808 status
= journal_wipe(journal
->j_journal
, full
);
814 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0);
824 * JBD Might read a cached version of another nodes journal file. We
825 * don't want this as this file changes often and we get no
826 * notification on those changes. The only way to be sure that we've
827 * got the most up to date version of those blocks then is to force
828 * read them off disk. Just searching through the buffer cache won't
829 * work as there may be pages backing this file which are still marked
830 * up to date. We know things can't change on this file underneath us
831 * as we have the lock by now :)
833 static int ocfs2_force_read_journal(struct inode
*inode
)
837 u64 v_blkno
, p_blkno
;
838 #define CONCURRENT_JOURNAL_FILL 32
839 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
843 BUG_ON(inode
->i_blocks
!=
844 ocfs2_align_bytes_to_sectors(i_size_read(inode
)));
846 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
848 mlog(0, "Force reading %lu blocks\n",
849 (inode
->i_blocks
>> (inode
->i_sb
->s_blocksize_bits
- 9)));
853 (inode
->i_blocks
>> (inode
->i_sb
->s_blocksize_bits
- 9))) {
855 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
863 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
864 p_blocks
= CONCURRENT_JOURNAL_FILL
;
866 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
867 p_blkno
, p_blocks
, bhs
, 0,
874 for(i
= 0; i
< p_blocks
; i
++) {
883 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
890 struct ocfs2_la_recovery_item
{
891 struct list_head lri_list
;
893 struct ocfs2_dinode
*lri_la_dinode
;
894 struct ocfs2_dinode
*lri_tl_dinode
;
897 /* Does the second half of the recovery process. By this point, the
898 * node is marked clean and can actually be considered recovered,
899 * hence it's no longer in the recovery map, but there's still some
900 * cleanup we can do which shouldn't happen within the recovery thread
901 * as locking in that context becomes very difficult if we are to take
902 * recovering nodes into account.
904 * NOTE: This function can and will sleep on recovery of other nodes
905 * during cluster locking, just like any other ocfs2 process.
907 void ocfs2_complete_recovery(void *data
)
910 struct ocfs2_super
*osb
= data
;
911 struct ocfs2_journal
*journal
= osb
->journal
;
912 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
913 struct ocfs2_la_recovery_item
*item
;
914 struct list_head
*p
, *n
;
915 LIST_HEAD(tmp_la_list
);
919 mlog(0, "completing recovery from keventd\n");
921 spin_lock(&journal
->j_lock
);
922 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
923 spin_unlock(&journal
->j_lock
);
925 list_for_each_safe(p
, n
, &tmp_la_list
) {
926 item
= list_entry(p
, struct ocfs2_la_recovery_item
, lri_list
);
927 list_del_init(&item
->lri_list
);
929 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
931 la_dinode
= item
->lri_la_dinode
;
933 mlog(0, "Clean up local alloc %"MLFu64
"\n",
936 ret
= ocfs2_complete_local_alloc_recovery(osb
,
944 tl_dinode
= item
->lri_tl_dinode
;
946 mlog(0, "Clean up truncate log %"MLFu64
"\n",
949 ret
= ocfs2_complete_truncate_log_recovery(osb
,
957 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
964 mlog(0, "Recovery completion\n");
968 /* NOTE: This function always eats your references to la_dinode and
969 * tl_dinode, either manually on error, or by passing them to
970 * ocfs2_complete_recovery */
971 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
973 struct ocfs2_dinode
*la_dinode
,
974 struct ocfs2_dinode
*tl_dinode
)
976 struct ocfs2_la_recovery_item
*item
;
978 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_KERNEL
);
980 /* Though we wish to avoid it, we are in fact safe in
981 * skipping local alloc cleanup as fsck.ocfs2 is more
982 * than capable of reclaiming unused space. */
993 INIT_LIST_HEAD(&item
->lri_list
);
994 item
->lri_la_dinode
= la_dinode
;
995 item
->lri_slot
= slot_num
;
996 item
->lri_tl_dinode
= tl_dinode
;
998 spin_lock(&journal
->j_lock
);
999 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
1000 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
1001 spin_unlock(&journal
->j_lock
);
1004 /* Called by the mount code to queue recovery the last part of
1005 * recovery for it's own slot. */
1006 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
1008 struct ocfs2_journal
*journal
= osb
->journal
;
1011 /* No need to queue up our truncate_log as regular
1012 * cleanup will catch that. */
1013 ocfs2_queue_recovery_completion(journal
,
1015 osb
->local_alloc_copy
,
1017 ocfs2_schedule_truncate_log_flush(osb
, 0);
1019 osb
->local_alloc_copy
= NULL
;
1024 static int __ocfs2_recovery_thread(void *arg
)
1026 int status
, node_num
;
1027 struct ocfs2_super
*osb
= arg
;
1031 status
= ocfs2_wait_on_mount(osb
);
1037 status
= ocfs2_super_lock(osb
, 1);
1043 while(!ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
1044 node_num
= ocfs2_node_map_first_set_bit(osb
,
1045 &osb
->recovery_map
);
1046 if (node_num
== O2NM_INVALID_NODE_NUM
) {
1047 mlog(0, "Out of nodes to recover.\n");
1051 status
= ocfs2_recover_node(osb
, node_num
);
1054 "Error %d recovering node %d on device (%u,%u)!\n",
1056 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1057 mlog(ML_ERROR
, "Volume requires unmount.\n");
1061 ocfs2_recovery_map_clear(osb
, node_num
);
1063 ocfs2_super_unlock(osb
, 1);
1065 /* We always run recovery on our own orphan dir - the dead
1066 * node(s) may have voted "no" on an inode delete earlier. A
1067 * revote is therefore required. */
1068 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
1072 mutex_lock(&osb
->recovery_lock
);
1074 !ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
1075 mutex_unlock(&osb
->recovery_lock
);
1079 osb
->recovery_thread_task
= NULL
;
1080 mb(); /* sync with ocfs2_recovery_thread_running */
1081 wake_up(&osb
->recovery_event
);
1083 mutex_unlock(&osb
->recovery_lock
);
1086 /* no one is callint kthread_stop() for us so the kthread() api
1087 * requires that we call do_exit(). And it isn't exported, but
1088 * complete_and_exit() seems to be a minimal wrapper around it. */
1089 complete_and_exit(NULL
, status
);
1093 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
1095 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1096 node_num
, osb
->node_num
);
1098 mutex_lock(&osb
->recovery_lock
);
1099 if (osb
->disable_recovery
)
1102 /* People waiting on recovery will wait on
1103 * the recovery map to empty. */
1104 if (!ocfs2_recovery_map_set(osb
, node_num
))
1105 mlog(0, "node %d already be in recovery.\n", node_num
);
1107 mlog(0, "starting recovery thread...\n");
1109 if (osb
->recovery_thread_task
)
1112 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
1113 "ocfs2rec-%d", osb
->osb_id
);
1114 if (IS_ERR(osb
->recovery_thread_task
)) {
1115 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
1116 osb
->recovery_thread_task
= NULL
;
1120 mutex_unlock(&osb
->recovery_lock
);
1121 wake_up(&osb
->recovery_event
);
1126 /* Does the actual journal replay and marks the journal inode as
1127 * clean. Will only replay if the journal inode is marked dirty. */
1128 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
1135 struct inode
*inode
= NULL
;
1136 struct ocfs2_dinode
*fe
;
1137 journal_t
*journal
= NULL
;
1138 struct buffer_head
*bh
= NULL
;
1140 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1142 if (inode
== NULL
) {
1147 if (is_bad_inode(inode
)) {
1154 SET_INODE_JOURNAL(inode
);
1156 status
= ocfs2_meta_lock_full(inode
, NULL
, &bh
, 1,
1157 OCFS2_META_LOCK_RECOVERY
);
1159 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status
);
1160 if (status
!= -ERESTARTSYS
)
1161 mlog(ML_ERROR
, "Could not lock journal!\n");
1166 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
1168 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1170 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
1171 mlog(0, "No recovery required for node %d\n", node_num
);
1175 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
1177 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1179 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
1181 status
= ocfs2_force_read_journal(inode
);
1187 mlog(0, "calling journal_init_inode\n");
1188 journal
= journal_init_inode(inode
);
1189 if (journal
== NULL
) {
1190 mlog(ML_ERROR
, "Linux journal layer error\n");
1195 status
= journal_load(journal
);
1200 journal_destroy(journal
);
1204 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1206 /* wipe the journal */
1207 mlog(0, "flushing the journal.\n");
1208 journal_lock_updates(journal
);
1209 status
= journal_flush(journal
);
1210 journal_unlock_updates(journal
);
1214 /* This will mark the node clean */
1215 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1216 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1217 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1219 status
= ocfs2_write_block(osb
, bh
, inode
);
1226 journal_destroy(journal
);
1229 /* drop the lock on this nodes journal */
1231 ocfs2_meta_unlock(inode
, 1);
1244 * Do the most important parts of node recovery:
1245 * - Replay it's journal
1246 * - Stamp a clean local allocator file
1247 * - Stamp a clean truncate log
1248 * - Mark the node clean
1250 * If this function completes without error, a node in OCFS2 can be
1251 * said to have been safely recovered. As a result, failure during the
1252 * second part of a nodes recovery process (local alloc recovery) is
1253 * far less concerning.
1255 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1260 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1261 struct ocfs2_dinode
*la_copy
= NULL
;
1262 struct ocfs2_dinode
*tl_copy
= NULL
;
1264 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1265 node_num
, osb
->node_num
);
1267 mlog(0, "checking node %d\n", node_num
);
1269 /* Should not ever be called to recover ourselves -- in that
1270 * case we should've called ocfs2_journal_load instead. */
1271 BUG_ON(osb
->node_num
== node_num
);
1273 slot_num
= ocfs2_node_num_to_slot(si
, node_num
);
1274 if (slot_num
== OCFS2_INVALID_SLOT
) {
1276 mlog(0, "no slot for this node, so no recovery required.\n");
1280 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1282 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1288 /* Stamp a clean local alloc file AFTER recovering the journal... */
1289 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1295 /* An error from begin_truncate_log_recovery is not
1296 * serious enough to warrant halting the rest of
1298 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1302 /* Likewise, this would be a strange but ultimately not so
1303 * harmful place to get an error... */
1304 ocfs2_clear_slot(si
, slot_num
);
1305 status
= ocfs2_update_disk_slots(osb
, si
);
1309 /* This will kfree the memory pointed to by la_copy and tl_copy */
1310 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1320 /* Test node liveness by trylocking his journal. If we get the lock,
1321 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1322 * still alive (we couldn't get the lock) and < 0 on error. */
1323 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1327 struct inode
*inode
= NULL
;
1329 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1331 if (inode
== NULL
) {
1332 mlog(ML_ERROR
, "access error\n");
1336 if (is_bad_inode(inode
)) {
1337 mlog(ML_ERROR
, "access error (bad inode)\n");
1343 SET_INODE_JOURNAL(inode
);
1345 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1346 status
= ocfs2_meta_lock_full(inode
, NULL
, NULL
, 1, flags
);
1348 if (status
!= -EAGAIN
)
1353 ocfs2_meta_unlock(inode
, 1);
1361 /* Call this underneath ocfs2_super_lock. It also assumes that the
1362 * slot info struct has been updated from disk. */
1363 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1365 int status
, i
, node_num
;
1366 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1368 /* This is called with the super block cluster lock, so we
1369 * know that the slot map can't change underneath us. */
1371 spin_lock(&si
->si_lock
);
1372 for(i
= 0; i
< si
->si_num_slots
; i
++) {
1373 if (i
== osb
->slot_num
)
1375 if (ocfs2_is_empty_slot(si
, i
))
1378 node_num
= si
->si_global_node_nums
[i
];
1379 if (ocfs2_node_map_test_bit(osb
, &osb
->recovery_map
, node_num
))
1381 spin_unlock(&si
->si_lock
);
1383 /* Ok, we have a slot occupied by another node which
1384 * is not in the recovery map. We trylock his journal
1385 * file here to test if he's alive. */
1386 status
= ocfs2_trylock_journal(osb
, i
);
1388 /* Since we're called from mount, we know that
1389 * the recovery thread can't race us on
1390 * setting / checking the recovery bits. */
1391 ocfs2_recovery_thread(osb
, node_num
);
1392 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1397 spin_lock(&si
->si_lock
);
1399 spin_unlock(&si
->si_lock
);
1407 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1411 int have_disk_lock
= 0;
1412 struct inode
*inode
= NULL
;
1414 struct inode
*orphan_dir_inode
= NULL
;
1415 unsigned long offset
, blk
, local
;
1416 struct buffer_head
*bh
= NULL
;
1417 struct ocfs2_dir_entry
*de
;
1418 struct super_block
*sb
= osb
->sb
;
1419 struct ocfs2_inode_info
*oi
;
1421 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1423 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1424 ORPHAN_DIR_SYSTEM_INODE
,
1426 if (!orphan_dir_inode
) {
1432 mutex_lock(&orphan_dir_inode
->i_mutex
);
1433 status
= ocfs2_meta_lock(orphan_dir_inode
, NULL
, NULL
, 0);
1435 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1443 while(offset
< i_size_read(orphan_dir_inode
)) {
1444 blk
= offset
>> sb
->s_blocksize_bits
;
1446 bh
= ocfs2_bread(orphan_dir_inode
, blk
, &status
, 0);
1450 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1458 while(offset
< i_size_read(orphan_dir_inode
)
1459 && local
< sb
->s_blocksize
) {
1460 de
= (struct ocfs2_dir_entry
*) (bh
->b_data
+ local
);
1462 if (!ocfs2_check_dir_entry(orphan_dir_inode
,
1464 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1471 local
+= le16_to_cpu(de
->rec_len
);
1472 offset
+= le16_to_cpu(de
->rec_len
);
1474 /* I guess we silently fail on no inode? */
1475 if (!le64_to_cpu(de
->inode
))
1477 if (de
->file_type
> OCFS2_FT_MAX
) {
1479 "block %llu contains invalid de: "
1480 "inode = %"MLFu64
", rec_len = %u, "
1481 "name_len = %u, file_type = %u, "
1483 (unsigned long long)bh
->b_blocknr
,
1484 le64_to_cpu(de
->inode
),
1485 le16_to_cpu(de
->rec_len
),
1492 if (de
->name_len
== 1 && !strncmp(".", de
->name
, 1))
1494 if (de
->name_len
== 2 && !strncmp("..", de
->name
, 2))
1497 iter
= ocfs2_iget(osb
, le64_to_cpu(de
->inode
));
1501 mlog(0, "queue orphan %"MLFu64
"\n",
1502 OCFS2_I(iter
)->ip_blkno
);
1503 OCFS2_I(iter
)->ip_next_orphan
= inode
;
1508 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1510 ocfs2_meta_unlock(orphan_dir_inode
, 0);
1513 iput(orphan_dir_inode
);
1514 orphan_dir_inode
= NULL
;
1517 oi
= OCFS2_I(inode
);
1518 mlog(0, "iput orphan %"MLFu64
"\n", oi
->ip_blkno
);
1520 iter
= oi
->ip_next_orphan
;
1522 spin_lock(&oi
->ip_lock
);
1523 /* Delete voting may have set these on the assumption
1524 * that the other node would wipe them successfully.
1525 * If they are still in the node's orphan dir, we need
1526 * to reset that state. */
1527 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1529 /* Set the proper information to get us going into
1530 * ocfs2_delete_inode. */
1531 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1532 oi
->ip_orphaned_slot
= slot
;
1533 spin_unlock(&oi
->ip_lock
);
1542 ocfs2_meta_unlock(orphan_dir_inode
, 0);
1544 if (orphan_dir_inode
)
1545 iput(orphan_dir_inode
);
1550 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1552 /* This check is good because ocfs2 will wait on our recovery
1553 * thread before changing it to something other than MOUNTED
1555 wait_event(osb
->osb_mount_event
,
1556 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1557 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1559 /* If there's an error on mount, then we may never get to the
1560 * MOUNTED flag, but this is set right before
1561 * dismount_volume() so we can trust it. */
1562 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1563 mlog(0, "mount error, exiting!\n");
1570 static int ocfs2_commit_thread(void *arg
)
1573 struct ocfs2_super
*osb
= arg
;
1574 struct ocfs2_journal
*journal
= osb
->journal
;
1576 /* we can trust j_num_trans here because _should_stop() is only set in
1577 * shutdown and nobody other than ourselves should be able to start
1578 * transactions. committing on shutdown might take a few iterations
1579 * as final transactions put deleted inodes on the list */
1580 while (!(kthread_should_stop() &&
1581 atomic_read(&journal
->j_num_trans
) == 0)) {
1583 wait_event_interruptible_timeout(osb
->checkpoint_event
,
1584 atomic_read(&journal
->j_num_trans
)
1585 || kthread_should_stop(),
1586 OCFS2_CHECKPOINT_INTERVAL
);
1588 status
= ocfs2_commit_cache(osb
);
1592 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1594 "commit_thread: %u transactions pending on "
1596 atomic_read(&journal
->j_num_trans
));
1603 /* Look for a dirty journal without taking any cluster locks. Used for
1604 * hard readonly access to determine whether the file system journals
1605 * require recovery. */
1606 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1610 struct buffer_head
*di_bh
;
1611 struct ocfs2_dinode
*di
;
1612 struct inode
*journal
= NULL
;
1614 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1615 journal
= ocfs2_get_system_file_inode(osb
,
1616 JOURNAL_SYSTEM_INODE
,
1618 if (!journal
|| is_bad_inode(journal
)) {
1625 ret
= ocfs2_read_block(osb
, OCFS2_I(journal
)->ip_blkno
, &di_bh
,
1632 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1634 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1635 OCFS2_JOURNAL_DIRTY_FL
)