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 DEFINE_SPINLOCK(trans_inc_lock
);
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 int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
62 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
64 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
66 static int ocfs2_commit_thread(void *arg
);
68 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
73 struct ocfs2_journal
*journal
= NULL
;
77 journal
= osb
->journal
;
79 /* Flush all pending commits and checkpoint the journal. */
80 down_write(&journal
->j_trans_barrier
);
82 if (atomic_read(&journal
->j_num_trans
) == 0) {
83 up_write(&journal
->j_trans_barrier
);
84 mlog(0, "No transactions for me to flush!\n");
88 journal_lock_updates(journal
->j_journal
);
89 status
= journal_flush(journal
->j_journal
);
90 journal_unlock_updates(journal
->j_journal
);
92 up_write(&journal
->j_trans_barrier
);
97 old_id
= ocfs2_inc_trans_id(journal
);
99 flushed
= atomic_read(&journal
->j_num_trans
);
100 atomic_set(&journal
->j_num_trans
, 0);
101 up_write(&journal
->j_trans_barrier
);
103 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
104 journal
->j_trans_id
, flushed
);
106 ocfs2_kick_vote_thread(osb
);
107 wake_up(&journal
->j_checkpointed
);
113 /* pass it NULL and it will allocate a new handle object for you. If
114 * you pass it a handle however, it may still return error, in which
115 * case it has free'd the passed handle for you. */
116 handle_t
*ocfs2_start_trans(struct ocfs2_super
*osb
, int max_buffs
)
118 journal_t
*journal
= osb
->journal
->j_journal
;
121 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
123 if (ocfs2_is_hard_readonly(osb
))
124 return ERR_PTR(-EROFS
);
126 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
127 BUG_ON(max_buffs
<= 0);
129 /* JBD might support this, but our journalling code doesn't yet. */
130 if (journal_current_handle()) {
131 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
135 down_read(&osb
->journal
->j_trans_barrier
);
137 handle
= journal_start(journal
, max_buffs
);
138 if (IS_ERR(handle
)) {
139 up_read(&osb
->journal
->j_trans_barrier
);
141 mlog_errno(PTR_ERR(handle
));
143 if (is_journal_aborted(journal
)) {
144 ocfs2_abort(osb
->sb
, "Detected aborted journal");
145 handle
= ERR_PTR(-EROFS
);
148 atomic_inc(&(osb
->journal
->j_num_trans
));
153 int ocfs2_commit_trans(struct ocfs2_super
*osb
,
157 struct ocfs2_journal
*journal
= osb
->journal
;
161 ret
= journal_stop(handle
);
165 up_read(&journal
->j_trans_barrier
);
171 * 'nblocks' is what you want to add to the current
172 * transaction. extend_trans will either extend the current handle by
173 * nblocks, or commit it and start a new one with nblocks credits.
175 * WARNING: This will not release any semaphores or disk locks taken
176 * during the transaction, so make sure they were taken *before*
177 * start_trans or we'll have ordering deadlocks.
179 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
180 * good because transaction ids haven't yet been recorded on the
181 * cluster locks associated with this handle.
183 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
192 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
194 status
= journal_extend(handle
, nblocks
);
201 mlog(0, "journal_extend failed, trying journal_restart\n");
202 status
= journal_restart(handle
, nblocks
);
216 int ocfs2_journal_access(handle_t
*handle
,
218 struct buffer_head
*bh
,
227 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
228 (unsigned long long)bh
->b_blocknr
, type
,
229 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
230 "OCFS2_JOURNAL_ACCESS_CREATE" :
231 "OCFS2_JOURNAL_ACCESS_WRITE",
234 /* we can safely remove this assertion after testing. */
235 if (!buffer_uptodate(bh
)) {
236 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
237 mlog(ML_ERROR
, "b_blocknr=%llu\n",
238 (unsigned long long)bh
->b_blocknr
);
242 /* Set the current transaction information on the inode so
243 * that the locking code knows whether it can drop it's locks
244 * on this inode or not. We're protected from the commit
245 * thread updating the current transaction id until
246 * ocfs2_commit_trans() because ocfs2_start_trans() took
247 * j_trans_barrier for us. */
248 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
250 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
252 case OCFS2_JOURNAL_ACCESS_CREATE
:
253 case OCFS2_JOURNAL_ACCESS_WRITE
:
254 status
= journal_get_write_access(handle
, bh
);
257 case OCFS2_JOURNAL_ACCESS_UNDO
:
258 status
= journal_get_undo_access(handle
, bh
);
263 mlog(ML_ERROR
, "Uknown access type!\n");
265 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
268 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
275 int ocfs2_journal_dirty(handle_t
*handle
,
276 struct buffer_head
*bh
)
280 mlog_entry("(bh->b_blocknr=%llu)\n",
281 (unsigned long long)bh
->b_blocknr
);
283 status
= journal_dirty_metadata(handle
, bh
);
285 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
286 "(bh->b_blocknr=%llu)\n",
287 (unsigned long long)bh
->b_blocknr
);
293 int ocfs2_journal_dirty_data(handle_t
*handle
,
294 struct buffer_head
*bh
)
296 int err
= journal_dirty_data(handle
, bh
);
299 /* TODO: When we can handle it, abort the handle and go RO on
305 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
307 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
309 journal_t
*journal
= osb
->journal
->j_journal
;
311 spin_lock(&journal
->j_state_lock
);
312 journal
->j_commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
313 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
314 journal
->j_flags
|= JFS_BARRIER
;
316 journal
->j_flags
&= ~JFS_BARRIER
;
317 spin_unlock(&journal
->j_state_lock
);
320 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
323 struct inode
*inode
= NULL
; /* the journal inode */
324 journal_t
*j_journal
= NULL
;
325 struct ocfs2_dinode
*di
= NULL
;
326 struct buffer_head
*bh
= NULL
;
327 struct ocfs2_super
*osb
;
334 osb
= journal
->j_osb
;
336 /* already have the inode for our journal */
337 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
344 if (is_bad_inode(inode
)) {
345 mlog(ML_ERROR
, "access error (bad inode)\n");
352 SET_INODE_JOURNAL(inode
);
353 OCFS2_I(inode
)->ip_open_count
++;
355 /* Skip recovery waits here - journal inode metadata never
356 * changes in a live cluster so it can be considered an
357 * exception to the rule. */
358 status
= ocfs2_meta_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
360 if (status
!= -ERESTARTSYS
)
361 mlog(ML_ERROR
, "Could not get lock on journal!\n");
366 di
= (struct ocfs2_dinode
*)bh
->b_data
;
368 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
369 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
375 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
376 mlog(0, "inode->i_blocks = %llu\n",
377 (unsigned long long)inode
->i_blocks
);
378 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
380 /* call the kernels journal init function now */
381 j_journal
= journal_init_inode(inode
);
382 if (j_journal
== NULL
) {
383 mlog(ML_ERROR
, "Linux journal layer error\n");
388 mlog(0, "Returned from journal_init_inode\n");
389 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
391 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
392 OCFS2_JOURNAL_DIRTY_FL
);
394 journal
->j_journal
= j_journal
;
395 journal
->j_inode
= inode
;
398 ocfs2_set_journal_params(osb
);
400 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
406 ocfs2_meta_unlock(inode
, 1);
410 OCFS2_I(inode
)->ip_open_count
--;
419 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
424 struct ocfs2_journal
*journal
= osb
->journal
;
425 struct buffer_head
*bh
= journal
->j_bh
;
426 struct ocfs2_dinode
*fe
;
430 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
431 if (!OCFS2_IS_VALID_DINODE(fe
)) {
432 /* This is called from startup/shutdown which will
433 * handle the errors in a specific manner, so no need
434 * to call ocfs2_error() here. */
435 mlog(ML_ERROR
, "Journal dinode %llu has invalid "
436 "signature: %.*s", (unsigned long long)fe
->i_blkno
, 7,
442 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
444 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
446 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
447 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
449 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
459 * If the journal has been kmalloc'd it needs to be freed after this
462 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
464 struct ocfs2_journal
*journal
= NULL
;
466 struct inode
*inode
= NULL
;
467 int num_running_trans
= 0;
473 journal
= osb
->journal
;
477 inode
= journal
->j_inode
;
479 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
482 /* need to inc inode use count as journal_destroy will iput. */
486 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
487 if (num_running_trans
> 0)
488 mlog(0, "Shutting down journal: must wait on %d "
489 "running transactions!\n",
492 /* Do a commit_cache here. It will flush our journal, *and*
493 * release any locks that are still held.
494 * set the SHUTDOWN flag and release the trans lock.
495 * the commit thread will take the trans lock for us below. */
496 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
498 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
499 * drop the trans_lock (which we want to hold until we
500 * completely destroy the journal. */
501 if (osb
->commit_task
) {
502 /* Wait for the commit thread */
503 mlog(0, "Waiting for ocfs2commit to exit....\n");
504 kthread_stop(osb
->commit_task
);
505 osb
->commit_task
= NULL
;
508 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
510 status
= ocfs2_journal_toggle_dirty(osb
, 0);
514 /* Shutdown the kernel journal system */
515 journal_destroy(journal
->j_journal
);
517 OCFS2_I(inode
)->ip_open_count
--;
519 /* unlock our journal */
520 ocfs2_meta_unlock(inode
, 1);
522 brelse(journal
->j_bh
);
523 journal
->j_bh
= NULL
;
525 journal
->j_state
= OCFS2_JOURNAL_FREE
;
527 // up_write(&journal->j_trans_barrier);
534 static void ocfs2_clear_journal_error(struct super_block
*sb
,
540 olderr
= journal_errno(journal
);
542 mlog(ML_ERROR
, "File system error %d recorded in "
543 "journal %u.\n", olderr
, slot
);
544 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
547 journal_ack_err(journal
);
548 journal_clear_err(journal
);
552 int ocfs2_journal_load(struct ocfs2_journal
*journal
)
555 struct ocfs2_super
*osb
;
562 osb
= journal
->j_osb
;
564 status
= journal_load(journal
->j_journal
);
566 mlog(ML_ERROR
, "Failed to load journal!\n");
570 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
572 status
= ocfs2_journal_toggle_dirty(osb
, 1);
578 /* Launch the commit thread */
579 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
, "ocfs2cmt");
580 if (IS_ERR(osb
->commit_task
)) {
581 status
= PTR_ERR(osb
->commit_task
);
582 osb
->commit_task
= NULL
;
583 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, error=%d",
594 /* 'full' flag tells us whether we clear out all blocks or if we just
595 * mark the journal clean */
596 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
604 status
= journal_wipe(journal
->j_journal
, full
);
610 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0);
620 * JBD Might read a cached version of another nodes journal file. We
621 * don't want this as this file changes often and we get no
622 * notification on those changes. The only way to be sure that we've
623 * got the most up to date version of those blocks then is to force
624 * read them off disk. Just searching through the buffer cache won't
625 * work as there may be pages backing this file which are still marked
626 * up to date. We know things can't change on this file underneath us
627 * as we have the lock by now :)
629 static int ocfs2_force_read_journal(struct inode
*inode
)
633 u64 v_blkno
, p_blkno
;
634 #define CONCURRENT_JOURNAL_FILL 32
635 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
639 BUG_ON(inode
->i_blocks
!=
640 ocfs2_align_bytes_to_sectors(i_size_read(inode
)));
642 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
644 mlog(0, "Force reading %llu blocks\n",
645 (unsigned long long)(inode
->i_blocks
>>
646 (inode
->i_sb
->s_blocksize_bits
- 9)));
650 (inode
->i_blocks
>> (inode
->i_sb
->s_blocksize_bits
- 9))) {
652 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
660 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
661 p_blocks
= CONCURRENT_JOURNAL_FILL
;
663 /* We are reading journal data which should not
664 * be put in the uptodate cache */
665 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
666 p_blkno
, p_blocks
, bhs
, 0,
673 for(i
= 0; i
< p_blocks
; i
++) {
682 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
689 struct ocfs2_la_recovery_item
{
690 struct list_head lri_list
;
692 struct ocfs2_dinode
*lri_la_dinode
;
693 struct ocfs2_dinode
*lri_tl_dinode
;
696 /* Does the second half of the recovery process. By this point, the
697 * node is marked clean and can actually be considered recovered,
698 * hence it's no longer in the recovery map, but there's still some
699 * cleanup we can do which shouldn't happen within the recovery thread
700 * as locking in that context becomes very difficult if we are to take
701 * recovering nodes into account.
703 * NOTE: This function can and will sleep on recovery of other nodes
704 * during cluster locking, just like any other ocfs2 process.
706 void ocfs2_complete_recovery(struct work_struct
*work
)
709 struct ocfs2_journal
*journal
=
710 container_of(work
, struct ocfs2_journal
, j_recovery_work
);
711 struct ocfs2_super
*osb
= journal
->j_osb
;
712 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
713 struct ocfs2_la_recovery_item
*item
;
714 struct list_head
*p
, *n
;
715 LIST_HEAD(tmp_la_list
);
719 mlog(0, "completing recovery from keventd\n");
721 spin_lock(&journal
->j_lock
);
722 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
723 spin_unlock(&journal
->j_lock
);
725 list_for_each_safe(p
, n
, &tmp_la_list
) {
726 item
= list_entry(p
, struct ocfs2_la_recovery_item
, lri_list
);
727 list_del_init(&item
->lri_list
);
729 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
731 la_dinode
= item
->lri_la_dinode
;
733 mlog(0, "Clean up local alloc %llu\n",
734 (unsigned long long)la_dinode
->i_blkno
);
736 ret
= ocfs2_complete_local_alloc_recovery(osb
,
744 tl_dinode
= item
->lri_tl_dinode
;
746 mlog(0, "Clean up truncate log %llu\n",
747 (unsigned long long)tl_dinode
->i_blkno
);
749 ret
= ocfs2_complete_truncate_log_recovery(osb
,
757 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
764 mlog(0, "Recovery completion\n");
768 /* NOTE: This function always eats your references to la_dinode and
769 * tl_dinode, either manually on error, or by passing them to
770 * ocfs2_complete_recovery */
771 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
773 struct ocfs2_dinode
*la_dinode
,
774 struct ocfs2_dinode
*tl_dinode
)
776 struct ocfs2_la_recovery_item
*item
;
778 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
780 /* Though we wish to avoid it, we are in fact safe in
781 * skipping local alloc cleanup as fsck.ocfs2 is more
782 * than capable of reclaiming unused space. */
793 INIT_LIST_HEAD(&item
->lri_list
);
794 item
->lri_la_dinode
= la_dinode
;
795 item
->lri_slot
= slot_num
;
796 item
->lri_tl_dinode
= tl_dinode
;
798 spin_lock(&journal
->j_lock
);
799 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
800 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
801 spin_unlock(&journal
->j_lock
);
804 /* Called by the mount code to queue recovery the last part of
805 * recovery for it's own slot. */
806 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
808 struct ocfs2_journal
*journal
= osb
->journal
;
811 /* No need to queue up our truncate_log as regular
812 * cleanup will catch that. */
813 ocfs2_queue_recovery_completion(journal
,
815 osb
->local_alloc_copy
,
817 ocfs2_schedule_truncate_log_flush(osb
, 0);
819 osb
->local_alloc_copy
= NULL
;
824 static int __ocfs2_recovery_thread(void *arg
)
826 int status
, node_num
;
827 struct ocfs2_super
*osb
= arg
;
831 status
= ocfs2_wait_on_mount(osb
);
837 status
= ocfs2_super_lock(osb
, 1);
843 while(!ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
844 node_num
= ocfs2_node_map_first_set_bit(osb
,
846 if (node_num
== O2NM_INVALID_NODE_NUM
) {
847 mlog(0, "Out of nodes to recover.\n");
851 status
= ocfs2_recover_node(osb
, node_num
);
854 "Error %d recovering node %d on device (%u,%u)!\n",
856 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
857 mlog(ML_ERROR
, "Volume requires unmount.\n");
861 ocfs2_recovery_map_clear(osb
, node_num
);
863 ocfs2_super_unlock(osb
, 1);
865 /* We always run recovery on our own orphan dir - the dead
866 * node(s) may have voted "no" on an inode delete earlier. A
867 * revote is therefore required. */
868 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
872 mutex_lock(&osb
->recovery_lock
);
874 !ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
875 mutex_unlock(&osb
->recovery_lock
);
879 osb
->recovery_thread_task
= NULL
;
880 mb(); /* sync with ocfs2_recovery_thread_running */
881 wake_up(&osb
->recovery_event
);
883 mutex_unlock(&osb
->recovery_lock
);
886 /* no one is callint kthread_stop() for us so the kthread() api
887 * requires that we call do_exit(). And it isn't exported, but
888 * complete_and_exit() seems to be a minimal wrapper around it. */
889 complete_and_exit(NULL
, status
);
893 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
895 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
896 node_num
, osb
->node_num
);
898 mutex_lock(&osb
->recovery_lock
);
899 if (osb
->disable_recovery
)
902 /* People waiting on recovery will wait on
903 * the recovery map to empty. */
904 if (!ocfs2_recovery_map_set(osb
, node_num
))
905 mlog(0, "node %d already be in recovery.\n", node_num
);
907 mlog(0, "starting recovery thread...\n");
909 if (osb
->recovery_thread_task
)
912 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
914 if (IS_ERR(osb
->recovery_thread_task
)) {
915 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
916 osb
->recovery_thread_task
= NULL
;
920 mutex_unlock(&osb
->recovery_lock
);
921 wake_up(&osb
->recovery_event
);
926 /* Does the actual journal replay and marks the journal inode as
927 * clean. Will only replay if the journal inode is marked dirty. */
928 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
935 struct inode
*inode
= NULL
;
936 struct ocfs2_dinode
*fe
;
937 journal_t
*journal
= NULL
;
938 struct buffer_head
*bh
= NULL
;
940 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
947 if (is_bad_inode(inode
)) {
954 SET_INODE_JOURNAL(inode
);
956 status
= ocfs2_meta_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
958 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status
);
959 if (status
!= -ERESTARTSYS
)
960 mlog(ML_ERROR
, "Could not lock journal!\n");
965 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
967 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
969 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
970 mlog(0, "No recovery required for node %d\n", node_num
);
974 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
976 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
978 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
980 status
= ocfs2_force_read_journal(inode
);
986 mlog(0, "calling journal_init_inode\n");
987 journal
= journal_init_inode(inode
);
988 if (journal
== NULL
) {
989 mlog(ML_ERROR
, "Linux journal layer error\n");
994 status
= journal_load(journal
);
999 journal_destroy(journal
);
1003 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1005 /* wipe the journal */
1006 mlog(0, "flushing the journal.\n");
1007 journal_lock_updates(journal
);
1008 status
= journal_flush(journal
);
1009 journal_unlock_updates(journal
);
1013 /* This will mark the node clean */
1014 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1015 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1016 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1018 status
= ocfs2_write_block(osb
, bh
, inode
);
1025 journal_destroy(journal
);
1028 /* drop the lock on this nodes journal */
1030 ocfs2_meta_unlock(inode
, 1);
1043 * Do the most important parts of node recovery:
1044 * - Replay it's journal
1045 * - Stamp a clean local allocator file
1046 * - Stamp a clean truncate log
1047 * - Mark the node clean
1049 * If this function completes without error, a node in OCFS2 can be
1050 * said to have been safely recovered. As a result, failure during the
1051 * second part of a nodes recovery process (local alloc recovery) is
1052 * far less concerning.
1054 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1059 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1060 struct ocfs2_dinode
*la_copy
= NULL
;
1061 struct ocfs2_dinode
*tl_copy
= NULL
;
1063 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1064 node_num
, osb
->node_num
);
1066 mlog(0, "checking node %d\n", node_num
);
1068 /* Should not ever be called to recover ourselves -- in that
1069 * case we should've called ocfs2_journal_load instead. */
1070 BUG_ON(osb
->node_num
== node_num
);
1072 slot_num
= ocfs2_node_num_to_slot(si
, node_num
);
1073 if (slot_num
== OCFS2_INVALID_SLOT
) {
1075 mlog(0, "no slot for this node, so no recovery required.\n");
1079 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1081 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1087 /* Stamp a clean local alloc file AFTER recovering the journal... */
1088 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1094 /* An error from begin_truncate_log_recovery is not
1095 * serious enough to warrant halting the rest of
1097 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1101 /* Likewise, this would be a strange but ultimately not so
1102 * harmful place to get an error... */
1103 ocfs2_clear_slot(si
, slot_num
);
1104 status
= ocfs2_update_disk_slots(osb
, si
);
1108 /* This will kfree the memory pointed to by la_copy and tl_copy */
1109 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1119 /* Test node liveness by trylocking his journal. If we get the lock,
1120 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1121 * still alive (we couldn't get the lock) and < 0 on error. */
1122 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1126 struct inode
*inode
= NULL
;
1128 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1130 if (inode
== NULL
) {
1131 mlog(ML_ERROR
, "access error\n");
1135 if (is_bad_inode(inode
)) {
1136 mlog(ML_ERROR
, "access error (bad inode)\n");
1142 SET_INODE_JOURNAL(inode
);
1144 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1145 status
= ocfs2_meta_lock_full(inode
, NULL
, 1, flags
);
1147 if (status
!= -EAGAIN
)
1152 ocfs2_meta_unlock(inode
, 1);
1160 /* Call this underneath ocfs2_super_lock. It also assumes that the
1161 * slot info struct has been updated from disk. */
1162 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1164 int status
, i
, node_num
;
1165 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1167 /* This is called with the super block cluster lock, so we
1168 * know that the slot map can't change underneath us. */
1170 spin_lock(&si
->si_lock
);
1171 for(i
= 0; i
< si
->si_num_slots
; i
++) {
1172 if (i
== osb
->slot_num
)
1174 if (ocfs2_is_empty_slot(si
, i
))
1177 node_num
= si
->si_global_node_nums
[i
];
1178 if (ocfs2_node_map_test_bit(osb
, &osb
->recovery_map
, node_num
))
1180 spin_unlock(&si
->si_lock
);
1182 /* Ok, we have a slot occupied by another node which
1183 * is not in the recovery map. We trylock his journal
1184 * file here to test if he's alive. */
1185 status
= ocfs2_trylock_journal(osb
, i
);
1187 /* Since we're called from mount, we know that
1188 * the recovery thread can't race us on
1189 * setting / checking the recovery bits. */
1190 ocfs2_recovery_thread(osb
, node_num
);
1191 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1196 spin_lock(&si
->si_lock
);
1198 spin_unlock(&si
->si_lock
);
1206 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
1208 struct inode
**head
)
1211 struct inode
*orphan_dir_inode
= NULL
;
1213 unsigned long offset
, blk
, local
;
1214 struct buffer_head
*bh
= NULL
;
1215 struct ocfs2_dir_entry
*de
;
1216 struct super_block
*sb
= osb
->sb
;
1218 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1219 ORPHAN_DIR_SYSTEM_INODE
,
1221 if (!orphan_dir_inode
) {
1227 mutex_lock(&orphan_dir_inode
->i_mutex
);
1228 status
= ocfs2_meta_lock(orphan_dir_inode
, NULL
, 0);
1236 while(offset
< i_size_read(orphan_dir_inode
)) {
1237 blk
= offset
>> sb
->s_blocksize_bits
;
1239 bh
= ocfs2_bread(orphan_dir_inode
, blk
, &status
, 0);
1250 while(offset
< i_size_read(orphan_dir_inode
)
1251 && local
< sb
->s_blocksize
) {
1252 de
= (struct ocfs2_dir_entry
*) (bh
->b_data
+ local
);
1254 if (!ocfs2_check_dir_entry(orphan_dir_inode
,
1262 local
+= le16_to_cpu(de
->rec_len
);
1263 offset
+= le16_to_cpu(de
->rec_len
);
1265 /* I guess we silently fail on no inode? */
1266 if (!le64_to_cpu(de
->inode
))
1268 if (de
->file_type
> OCFS2_FT_MAX
) {
1270 "block %llu contains invalid de: "
1271 "inode = %llu, rec_len = %u, "
1272 "name_len = %u, file_type = %u, "
1274 (unsigned long long)bh
->b_blocknr
,
1275 (unsigned long long)le64_to_cpu(de
->inode
),
1276 le16_to_cpu(de
->rec_len
),
1283 if (de
->name_len
== 1 && !strncmp(".", de
->name
, 1))
1285 if (de
->name_len
== 2 && !strncmp("..", de
->name
, 2))
1288 iter
= ocfs2_iget(osb
, le64_to_cpu(de
->inode
),
1289 OCFS2_FI_FLAG_NOLOCK
);
1293 mlog(0, "queue orphan %llu\n",
1294 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
1295 /* No locking is required for the next_orphan
1296 * queue as there is only ever a single
1297 * process doing orphan recovery. */
1298 OCFS2_I(iter
)->ip_next_orphan
= *head
;
1305 ocfs2_meta_unlock(orphan_dir_inode
, 0);
1307 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1308 iput(orphan_dir_inode
);
1312 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
1317 spin_lock(&osb
->osb_lock
);
1318 ret
= !osb
->osb_orphan_wipes
[slot
];
1319 spin_unlock(&osb
->osb_lock
);
1323 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
1326 spin_lock(&osb
->osb_lock
);
1327 /* Mark ourselves such that new processes in delete_inode()
1328 * know to quit early. */
1329 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1330 while (osb
->osb_orphan_wipes
[slot
]) {
1331 /* If any processes are already in the middle of an
1332 * orphan wipe on this dir, then we need to wait for
1334 spin_unlock(&osb
->osb_lock
);
1335 wait_event_interruptible(osb
->osb_wipe_event
,
1336 ocfs2_orphan_recovery_can_continue(osb
, slot
));
1337 spin_lock(&osb
->osb_lock
);
1339 spin_unlock(&osb
->osb_lock
);
1342 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
1345 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1349 * Orphan recovery. Each mounted node has it's own orphan dir which we
1350 * must run during recovery. Our strategy here is to build a list of
1351 * the inodes in the orphan dir and iget/iput them. The VFS does
1352 * (most) of the rest of the work.
1354 * Orphan recovery can happen at any time, not just mount so we have a
1355 * couple of extra considerations.
1357 * - We grab as many inodes as we can under the orphan dir lock -
1358 * doing iget() outside the orphan dir risks getting a reference on
1360 * - We must be sure not to deadlock with other processes on the
1361 * system wanting to run delete_inode(). This can happen when they go
1362 * to lock the orphan dir and the orphan recovery process attempts to
1363 * iget() inside the orphan dir lock. This can be avoided by
1364 * advertising our state to ocfs2_delete_inode().
1366 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1370 struct inode
*inode
= NULL
;
1372 struct ocfs2_inode_info
*oi
;
1374 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1376 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
1377 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
1378 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
1380 /* Error here should be noted, but we want to continue with as
1381 * many queued inodes as we've got. */
1386 oi
= OCFS2_I(inode
);
1387 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
1389 iter
= oi
->ip_next_orphan
;
1391 spin_lock(&oi
->ip_lock
);
1392 /* Delete voting may have set these on the assumption
1393 * that the other node would wipe them successfully.
1394 * If they are still in the node's orphan dir, we need
1395 * to reset that state. */
1396 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1398 /* Set the proper information to get us going into
1399 * ocfs2_delete_inode. */
1400 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1401 oi
->ip_orphaned_slot
= slot
;
1402 spin_unlock(&oi
->ip_lock
);
1412 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1414 /* This check is good because ocfs2 will wait on our recovery
1415 * thread before changing it to something other than MOUNTED
1417 wait_event(osb
->osb_mount_event
,
1418 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1419 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1421 /* If there's an error on mount, then we may never get to the
1422 * MOUNTED flag, but this is set right before
1423 * dismount_volume() so we can trust it. */
1424 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1425 mlog(0, "mount error, exiting!\n");
1432 static int ocfs2_commit_thread(void *arg
)
1435 struct ocfs2_super
*osb
= arg
;
1436 struct ocfs2_journal
*journal
= osb
->journal
;
1438 /* we can trust j_num_trans here because _should_stop() is only set in
1439 * shutdown and nobody other than ourselves should be able to start
1440 * transactions. committing on shutdown might take a few iterations
1441 * as final transactions put deleted inodes on the list */
1442 while (!(kthread_should_stop() &&
1443 atomic_read(&journal
->j_num_trans
) == 0)) {
1445 wait_event_interruptible(osb
->checkpoint_event
,
1446 atomic_read(&journal
->j_num_trans
)
1447 || kthread_should_stop());
1449 status
= ocfs2_commit_cache(osb
);
1453 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1455 "commit_thread: %u transactions pending on "
1457 atomic_read(&journal
->j_num_trans
));
1464 /* Look for a dirty journal without taking any cluster locks. Used for
1465 * hard readonly access to determine whether the file system journals
1466 * require recovery. */
1467 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1471 struct buffer_head
*di_bh
;
1472 struct ocfs2_dinode
*di
;
1473 struct inode
*journal
= NULL
;
1475 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1476 journal
= ocfs2_get_system_file_inode(osb
,
1477 JOURNAL_SYSTEM_INODE
,
1479 if (!journal
|| is_bad_inode(journal
)) {
1486 ret
= ocfs2_read_block(osb
, OCFS2_I(journal
)->ip_blkno
, &di_bh
,
1493 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1495 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1496 OCFS2_JOURNAL_DIRTY_FL
)