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
,
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
);
67 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
72 struct ocfs2_journal
*journal
= NULL
;
76 journal
= osb
->journal
;
78 /* Flush all pending commits and checkpoint the journal. */
79 down_write(&journal
->j_trans_barrier
);
81 if (atomic_read(&journal
->j_num_trans
) == 0) {
82 up_write(&journal
->j_trans_barrier
);
83 mlog(0, "No transactions for me to flush!\n");
87 journal_lock_updates(journal
->j_journal
);
88 status
= journal_flush(journal
->j_journal
);
89 journal_unlock_updates(journal
->j_journal
);
91 up_write(&journal
->j_trans_barrier
);
96 old_id
= ocfs2_inc_trans_id(journal
);
98 flushed
= atomic_read(&journal
->j_num_trans
);
99 atomic_set(&journal
->j_num_trans
, 0);
100 up_write(&journal
->j_trans_barrier
);
102 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
103 journal
->j_trans_id
, flushed
);
105 ocfs2_wake_downconvert_thread(osb
);
106 wake_up(&journal
->j_checkpointed
);
112 /* pass it NULL and it will allocate a new handle object for you. If
113 * you pass it a handle however, it may still return error, in which
114 * case it has free'd the passed handle for you. */
115 handle_t
*ocfs2_start_trans(struct ocfs2_super
*osb
, int max_buffs
)
117 journal_t
*journal
= osb
->journal
->j_journal
;
120 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
122 if (ocfs2_is_hard_readonly(osb
))
123 return ERR_PTR(-EROFS
);
125 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
126 BUG_ON(max_buffs
<= 0);
128 /* JBD might support this, but our journalling code doesn't yet. */
129 if (journal_current_handle()) {
130 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
134 down_read(&osb
->journal
->j_trans_barrier
);
136 handle
= journal_start(journal
, max_buffs
);
137 if (IS_ERR(handle
)) {
138 up_read(&osb
->journal
->j_trans_barrier
);
140 mlog_errno(PTR_ERR(handle
));
142 if (is_journal_aborted(journal
)) {
143 ocfs2_abort(osb
->sb
, "Detected aborted journal");
144 handle
= ERR_PTR(-EROFS
);
147 if (!ocfs2_mount_local(osb
))
148 atomic_inc(&(osb
->journal
->j_num_trans
));
154 int ocfs2_commit_trans(struct ocfs2_super
*osb
,
158 struct ocfs2_journal
*journal
= osb
->journal
;
162 ret
= journal_stop(handle
);
166 up_read(&journal
->j_trans_barrier
);
172 * 'nblocks' is what you want to add to the current
173 * transaction. extend_trans will either extend the current handle by
174 * nblocks, or commit it and start a new one with nblocks credits.
176 * This might call journal_restart() which will commit dirty buffers
177 * and then restart the transaction. Before calling
178 * ocfs2_extend_trans(), any changed blocks should have been
179 * dirtied. After calling it, all blocks which need to be changed must
180 * go through another set of journal_access/journal_dirty calls.
182 * WARNING: This will not release any semaphores or disk locks taken
183 * during the transaction, so make sure they were taken *before*
184 * start_trans or we'll have ordering deadlocks.
186 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
187 * good because transaction ids haven't yet been recorded on the
188 * cluster locks associated with this handle.
190 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
199 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
201 #ifdef OCFS2_DEBUG_FS
204 status
= journal_extend(handle
, nblocks
);
212 mlog(0, "journal_extend failed, trying journal_restart\n");
213 status
= journal_restart(handle
, nblocks
);
227 int ocfs2_journal_access(handle_t
*handle
,
229 struct buffer_head
*bh
,
238 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
239 (unsigned long long)bh
->b_blocknr
, type
,
240 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
241 "OCFS2_JOURNAL_ACCESS_CREATE" :
242 "OCFS2_JOURNAL_ACCESS_WRITE",
245 /* we can safely remove this assertion after testing. */
246 if (!buffer_uptodate(bh
)) {
247 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
248 mlog(ML_ERROR
, "b_blocknr=%llu\n",
249 (unsigned long long)bh
->b_blocknr
);
253 /* Set the current transaction information on the inode so
254 * that the locking code knows whether it can drop it's locks
255 * on this inode or not. We're protected from the commit
256 * thread updating the current transaction id until
257 * ocfs2_commit_trans() because ocfs2_start_trans() took
258 * j_trans_barrier for us. */
259 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
261 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
263 case OCFS2_JOURNAL_ACCESS_CREATE
:
264 case OCFS2_JOURNAL_ACCESS_WRITE
:
265 status
= journal_get_write_access(handle
, bh
);
268 case OCFS2_JOURNAL_ACCESS_UNDO
:
269 status
= journal_get_undo_access(handle
, bh
);
274 mlog(ML_ERROR
, "Uknown access type!\n");
276 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
279 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
286 int ocfs2_journal_dirty(handle_t
*handle
,
287 struct buffer_head
*bh
)
291 mlog_entry("(bh->b_blocknr=%llu)\n",
292 (unsigned long long)bh
->b_blocknr
);
294 status
= journal_dirty_metadata(handle
, bh
);
296 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
297 "(bh->b_blocknr=%llu)\n",
298 (unsigned long long)bh
->b_blocknr
);
304 int ocfs2_journal_dirty_data(handle_t
*handle
,
305 struct buffer_head
*bh
)
307 int err
= journal_dirty_data(handle
, bh
);
310 /* TODO: When we can handle it, abort the handle and go RO on
316 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
318 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
320 journal_t
*journal
= osb
->journal
->j_journal
;
322 spin_lock(&journal
->j_state_lock
);
323 journal
->j_commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
324 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
325 journal
->j_flags
|= JFS_BARRIER
;
327 journal
->j_flags
&= ~JFS_BARRIER
;
328 spin_unlock(&journal
->j_state_lock
);
331 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
334 struct inode
*inode
= NULL
; /* the journal inode */
335 journal_t
*j_journal
= NULL
;
336 struct ocfs2_dinode
*di
= NULL
;
337 struct buffer_head
*bh
= NULL
;
338 struct ocfs2_super
*osb
;
345 osb
= journal
->j_osb
;
347 /* already have the inode for our journal */
348 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
355 if (is_bad_inode(inode
)) {
356 mlog(ML_ERROR
, "access error (bad inode)\n");
363 SET_INODE_JOURNAL(inode
);
364 OCFS2_I(inode
)->ip_open_count
++;
366 /* Skip recovery waits here - journal inode metadata never
367 * changes in a live cluster so it can be considered an
368 * exception to the rule. */
369 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
371 if (status
!= -ERESTARTSYS
)
372 mlog(ML_ERROR
, "Could not get lock on journal!\n");
377 di
= (struct ocfs2_dinode
*)bh
->b_data
;
379 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
380 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
386 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
387 mlog(0, "inode->i_blocks = %llu\n",
388 (unsigned long long)inode
->i_blocks
);
389 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
391 /* call the kernels journal init function now */
392 j_journal
= journal_init_inode(inode
);
393 if (j_journal
== NULL
) {
394 mlog(ML_ERROR
, "Linux journal layer error\n");
399 mlog(0, "Returned from journal_init_inode\n");
400 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
402 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
403 OCFS2_JOURNAL_DIRTY_FL
);
405 journal
->j_journal
= j_journal
;
406 journal
->j_inode
= inode
;
409 ocfs2_set_journal_params(osb
);
411 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
417 ocfs2_inode_unlock(inode
, 1);
421 OCFS2_I(inode
)->ip_open_count
--;
430 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
435 struct ocfs2_journal
*journal
= osb
->journal
;
436 struct buffer_head
*bh
= journal
->j_bh
;
437 struct ocfs2_dinode
*fe
;
441 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
442 if (!OCFS2_IS_VALID_DINODE(fe
)) {
443 /* This is called from startup/shutdown which will
444 * handle the errors in a specific manner, so no need
445 * to call ocfs2_error() here. */
446 mlog(ML_ERROR
, "Journal dinode %llu has invalid "
448 (unsigned long long)le64_to_cpu(fe
->i_blkno
), 7,
454 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
456 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
458 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
459 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
461 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
471 * If the journal has been kmalloc'd it needs to be freed after this
474 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
476 struct ocfs2_journal
*journal
= NULL
;
478 struct inode
*inode
= NULL
;
479 int num_running_trans
= 0;
485 journal
= osb
->journal
;
489 inode
= journal
->j_inode
;
491 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
494 /* need to inc inode use count as journal_destroy will iput. */
498 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
499 if (num_running_trans
> 0)
500 mlog(0, "Shutting down journal: must wait on %d "
501 "running transactions!\n",
504 /* Do a commit_cache here. It will flush our journal, *and*
505 * release any locks that are still held.
506 * set the SHUTDOWN flag and release the trans lock.
507 * the commit thread will take the trans lock for us below. */
508 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
510 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
511 * drop the trans_lock (which we want to hold until we
512 * completely destroy the journal. */
513 if (osb
->commit_task
) {
514 /* Wait for the commit thread */
515 mlog(0, "Waiting for ocfs2commit to exit....\n");
516 kthread_stop(osb
->commit_task
);
517 osb
->commit_task
= NULL
;
520 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
522 if (ocfs2_mount_local(osb
)) {
523 journal_lock_updates(journal
->j_journal
);
524 status
= journal_flush(journal
->j_journal
);
525 journal_unlock_updates(journal
->j_journal
);
532 * Do not toggle if flush was unsuccessful otherwise
533 * will leave dirty metadata in a "clean" journal
535 status
= ocfs2_journal_toggle_dirty(osb
, 0);
540 /* Shutdown the kernel journal system */
541 journal_destroy(journal
->j_journal
);
543 OCFS2_I(inode
)->ip_open_count
--;
545 /* unlock our journal */
546 ocfs2_inode_unlock(inode
, 1);
548 brelse(journal
->j_bh
);
549 journal
->j_bh
= NULL
;
551 journal
->j_state
= OCFS2_JOURNAL_FREE
;
553 // up_write(&journal->j_trans_barrier);
560 static void ocfs2_clear_journal_error(struct super_block
*sb
,
566 olderr
= journal_errno(journal
);
568 mlog(ML_ERROR
, "File system error %d recorded in "
569 "journal %u.\n", olderr
, slot
);
570 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
573 journal_ack_err(journal
);
574 journal_clear_err(journal
);
578 int ocfs2_journal_load(struct ocfs2_journal
*journal
, int local
)
581 struct ocfs2_super
*osb
;
588 osb
= journal
->j_osb
;
590 status
= journal_load(journal
->j_journal
);
592 mlog(ML_ERROR
, "Failed to load journal!\n");
596 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
598 status
= ocfs2_journal_toggle_dirty(osb
, 1);
604 /* Launch the commit thread */
606 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
,
608 if (IS_ERR(osb
->commit_task
)) {
609 status
= PTR_ERR(osb
->commit_task
);
610 osb
->commit_task
= NULL
;
611 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, "
616 osb
->commit_task
= NULL
;
624 /* 'full' flag tells us whether we clear out all blocks or if we just
625 * mark the journal clean */
626 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
634 status
= journal_wipe(journal
->j_journal
, full
);
640 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0);
650 * JBD Might read a cached version of another nodes journal file. We
651 * don't want this as this file changes often and we get no
652 * notification on those changes. The only way to be sure that we've
653 * got the most up to date version of those blocks then is to force
654 * read them off disk. Just searching through the buffer cache won't
655 * work as there may be pages backing this file which are still marked
656 * up to date. We know things can't change on this file underneath us
657 * as we have the lock by now :)
659 static int ocfs2_force_read_journal(struct inode
*inode
)
663 u64 v_blkno
, p_blkno
, p_blocks
, num_blocks
;
664 #define CONCURRENT_JOURNAL_FILL 32ULL
665 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
669 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
671 num_blocks
= ocfs2_blocks_for_bytes(inode
->i_sb
, inode
->i_size
);
673 while (v_blkno
< num_blocks
) {
674 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
675 &p_blkno
, &p_blocks
, NULL
);
681 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
682 p_blocks
= CONCURRENT_JOURNAL_FILL
;
684 /* We are reading journal data which should not
685 * be put in the uptodate cache */
686 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
687 p_blkno
, p_blocks
, bhs
, 0,
694 for(i
= 0; i
< p_blocks
; i
++) {
703 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
710 struct ocfs2_la_recovery_item
{
711 struct list_head lri_list
;
713 struct ocfs2_dinode
*lri_la_dinode
;
714 struct ocfs2_dinode
*lri_tl_dinode
;
717 /* Does the second half of the recovery process. By this point, the
718 * node is marked clean and can actually be considered recovered,
719 * hence it's no longer in the recovery map, but there's still some
720 * cleanup we can do which shouldn't happen within the recovery thread
721 * as locking in that context becomes very difficult if we are to take
722 * recovering nodes into account.
724 * NOTE: This function can and will sleep on recovery of other nodes
725 * during cluster locking, just like any other ocfs2 process.
727 void ocfs2_complete_recovery(struct work_struct
*work
)
730 struct ocfs2_journal
*journal
=
731 container_of(work
, struct ocfs2_journal
, j_recovery_work
);
732 struct ocfs2_super
*osb
= journal
->j_osb
;
733 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
734 struct ocfs2_la_recovery_item
*item
, *n
;
735 LIST_HEAD(tmp_la_list
);
739 mlog(0, "completing recovery from keventd\n");
741 spin_lock(&journal
->j_lock
);
742 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
743 spin_unlock(&journal
->j_lock
);
745 list_for_each_entry_safe(item
, n
, &tmp_la_list
, lri_list
) {
746 list_del_init(&item
->lri_list
);
748 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
750 la_dinode
= item
->lri_la_dinode
;
752 mlog(0, "Clean up local alloc %llu\n",
753 (unsigned long long)le64_to_cpu(la_dinode
->i_blkno
));
755 ret
= ocfs2_complete_local_alloc_recovery(osb
,
763 tl_dinode
= item
->lri_tl_dinode
;
765 mlog(0, "Clean up truncate log %llu\n",
766 (unsigned long long)le64_to_cpu(tl_dinode
->i_blkno
));
768 ret
= ocfs2_complete_truncate_log_recovery(osb
,
776 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
783 mlog(0, "Recovery completion\n");
787 /* NOTE: This function always eats your references to la_dinode and
788 * tl_dinode, either manually on error, or by passing them to
789 * ocfs2_complete_recovery */
790 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
792 struct ocfs2_dinode
*la_dinode
,
793 struct ocfs2_dinode
*tl_dinode
)
795 struct ocfs2_la_recovery_item
*item
;
797 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
799 /* Though we wish to avoid it, we are in fact safe in
800 * skipping local alloc cleanup as fsck.ocfs2 is more
801 * than capable of reclaiming unused space. */
812 INIT_LIST_HEAD(&item
->lri_list
);
813 item
->lri_la_dinode
= la_dinode
;
814 item
->lri_slot
= slot_num
;
815 item
->lri_tl_dinode
= tl_dinode
;
817 spin_lock(&journal
->j_lock
);
818 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
819 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
820 spin_unlock(&journal
->j_lock
);
823 /* Called by the mount code to queue recovery the last part of
824 * recovery for it's own slot. */
825 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
827 struct ocfs2_journal
*journal
= osb
->journal
;
830 /* No need to queue up our truncate_log as regular
831 * cleanup will catch that. */
832 ocfs2_queue_recovery_completion(journal
,
834 osb
->local_alloc_copy
,
836 ocfs2_schedule_truncate_log_flush(osb
, 0);
838 osb
->local_alloc_copy
= NULL
;
843 static int __ocfs2_recovery_thread(void *arg
)
845 int status
, node_num
;
846 struct ocfs2_super
*osb
= arg
;
850 status
= ocfs2_wait_on_mount(osb
);
856 status
= ocfs2_super_lock(osb
, 1);
862 while(!ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
863 node_num
= ocfs2_node_map_first_set_bit(osb
,
865 if (node_num
== O2NM_INVALID_NODE_NUM
) {
866 mlog(0, "Out of nodes to recover.\n");
870 status
= ocfs2_recover_node(osb
, node_num
);
873 "Error %d recovering node %d on device (%u,%u)!\n",
875 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
876 mlog(ML_ERROR
, "Volume requires unmount.\n");
880 ocfs2_recovery_map_clear(osb
, node_num
);
882 ocfs2_super_unlock(osb
, 1);
884 /* We always run recovery on our own orphan dir - the dead
885 * node(s) may have disallowd a previos inode delete. Re-processing
886 * is therefore required. */
887 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
891 mutex_lock(&osb
->recovery_lock
);
893 !ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
894 mutex_unlock(&osb
->recovery_lock
);
898 osb
->recovery_thread_task
= NULL
;
899 mb(); /* sync with ocfs2_recovery_thread_running */
900 wake_up(&osb
->recovery_event
);
902 mutex_unlock(&osb
->recovery_lock
);
905 /* no one is callint kthread_stop() for us so the kthread() api
906 * requires that we call do_exit(). And it isn't exported, but
907 * complete_and_exit() seems to be a minimal wrapper around it. */
908 complete_and_exit(NULL
, status
);
912 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
914 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
915 node_num
, osb
->node_num
);
917 mutex_lock(&osb
->recovery_lock
);
918 if (osb
->disable_recovery
)
921 /* People waiting on recovery will wait on
922 * the recovery map to empty. */
923 if (!ocfs2_recovery_map_set(osb
, node_num
))
924 mlog(0, "node %d already be in recovery.\n", node_num
);
926 mlog(0, "starting recovery thread...\n");
928 if (osb
->recovery_thread_task
)
931 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
933 if (IS_ERR(osb
->recovery_thread_task
)) {
934 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
935 osb
->recovery_thread_task
= NULL
;
939 mutex_unlock(&osb
->recovery_lock
);
940 wake_up(&osb
->recovery_event
);
945 /* Does the actual journal replay and marks the journal inode as
946 * clean. Will only replay if the journal inode is marked dirty. */
947 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
954 struct inode
*inode
= NULL
;
955 struct ocfs2_dinode
*fe
;
956 journal_t
*journal
= NULL
;
957 struct buffer_head
*bh
= NULL
;
959 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
966 if (is_bad_inode(inode
)) {
973 SET_INODE_JOURNAL(inode
);
975 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
977 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status
);
978 if (status
!= -ERESTARTSYS
)
979 mlog(ML_ERROR
, "Could not lock journal!\n");
984 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
986 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
988 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
989 mlog(0, "No recovery required for node %d\n", node_num
);
993 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
995 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
997 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
999 status
= ocfs2_force_read_journal(inode
);
1005 mlog(0, "calling journal_init_inode\n");
1006 journal
= journal_init_inode(inode
);
1007 if (journal
== NULL
) {
1008 mlog(ML_ERROR
, "Linux journal layer error\n");
1013 status
= journal_load(journal
);
1018 journal_destroy(journal
);
1022 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1024 /* wipe the journal */
1025 mlog(0, "flushing the journal.\n");
1026 journal_lock_updates(journal
);
1027 status
= journal_flush(journal
);
1028 journal_unlock_updates(journal
);
1032 /* This will mark the node clean */
1033 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1034 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1035 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1037 status
= ocfs2_write_block(osb
, bh
, inode
);
1044 journal_destroy(journal
);
1047 /* drop the lock on this nodes journal */
1049 ocfs2_inode_unlock(inode
, 1);
1062 * Do the most important parts of node recovery:
1063 * - Replay it's journal
1064 * - Stamp a clean local allocator file
1065 * - Stamp a clean truncate log
1066 * - Mark the node clean
1068 * If this function completes without error, a node in OCFS2 can be
1069 * said to have been safely recovered. As a result, failure during the
1070 * second part of a nodes recovery process (local alloc recovery) is
1071 * far less concerning.
1073 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1078 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1079 struct ocfs2_dinode
*la_copy
= NULL
;
1080 struct ocfs2_dinode
*tl_copy
= NULL
;
1082 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1083 node_num
, osb
->node_num
);
1085 mlog(0, "checking node %d\n", node_num
);
1087 /* Should not ever be called to recover ourselves -- in that
1088 * case we should've called ocfs2_journal_load instead. */
1089 BUG_ON(osb
->node_num
== node_num
);
1091 slot_num
= ocfs2_node_num_to_slot(si
, node_num
);
1092 if (slot_num
== OCFS2_INVALID_SLOT
) {
1094 mlog(0, "no slot for this node, so no recovery required.\n");
1098 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1100 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1106 /* Stamp a clean local alloc file AFTER recovering the journal... */
1107 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1113 /* An error from begin_truncate_log_recovery is not
1114 * serious enough to warrant halting the rest of
1116 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1120 /* Likewise, this would be a strange but ultimately not so
1121 * harmful place to get an error... */
1122 ocfs2_clear_slot(si
, slot_num
);
1123 status
= ocfs2_update_disk_slots(osb
, si
);
1127 /* This will kfree the memory pointed to by la_copy and tl_copy */
1128 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1138 /* Test node liveness by trylocking his journal. If we get the lock,
1139 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1140 * still alive (we couldn't get the lock) and < 0 on error. */
1141 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1145 struct inode
*inode
= NULL
;
1147 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1149 if (inode
== NULL
) {
1150 mlog(ML_ERROR
, "access error\n");
1154 if (is_bad_inode(inode
)) {
1155 mlog(ML_ERROR
, "access error (bad inode)\n");
1161 SET_INODE_JOURNAL(inode
);
1163 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1164 status
= ocfs2_inode_lock_full(inode
, NULL
, 1, flags
);
1166 if (status
!= -EAGAIN
)
1171 ocfs2_inode_unlock(inode
, 1);
1179 /* Call this underneath ocfs2_super_lock. It also assumes that the
1180 * slot info struct has been updated from disk. */
1181 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1183 int status
, i
, node_num
;
1184 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1186 /* This is called with the super block cluster lock, so we
1187 * know that the slot map can't change underneath us. */
1189 spin_lock(&si
->si_lock
);
1190 for(i
= 0; i
< si
->si_num_slots
; i
++) {
1191 if (i
== osb
->slot_num
)
1193 if (ocfs2_is_empty_slot(si
, i
))
1196 node_num
= si
->si_global_node_nums
[i
];
1197 if (ocfs2_node_map_test_bit(osb
, &osb
->recovery_map
, node_num
))
1199 spin_unlock(&si
->si_lock
);
1201 /* Ok, we have a slot occupied by another node which
1202 * is not in the recovery map. We trylock his journal
1203 * file here to test if he's alive. */
1204 status
= ocfs2_trylock_journal(osb
, i
);
1206 /* Since we're called from mount, we know that
1207 * the recovery thread can't race us on
1208 * setting / checking the recovery bits. */
1209 ocfs2_recovery_thread(osb
, node_num
);
1210 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1215 spin_lock(&si
->si_lock
);
1217 spin_unlock(&si
->si_lock
);
1225 struct ocfs2_orphan_filldir_priv
{
1227 struct ocfs2_super
*osb
;
1230 static int ocfs2_orphan_filldir(void *priv
, const char *name
, int name_len
,
1231 loff_t pos
, u64 ino
, unsigned type
)
1233 struct ocfs2_orphan_filldir_priv
*p
= priv
;
1236 if (name_len
== 1 && !strncmp(".", name
, 1))
1238 if (name_len
== 2 && !strncmp("..", name
, 2))
1241 /* Skip bad inodes so that recovery can continue */
1242 iter
= ocfs2_iget(p
->osb
, ino
,
1243 OCFS2_FI_FLAG_ORPHAN_RECOVERY
);
1247 mlog(0, "queue orphan %llu\n",
1248 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
1249 /* No locking is required for the next_orphan queue as there
1250 * is only ever a single process doing orphan recovery. */
1251 OCFS2_I(iter
)->ip_next_orphan
= p
->head
;
1257 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
1259 struct inode
**head
)
1262 struct inode
*orphan_dir_inode
= NULL
;
1263 struct ocfs2_orphan_filldir_priv priv
;
1269 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1270 ORPHAN_DIR_SYSTEM_INODE
,
1272 if (!orphan_dir_inode
) {
1278 mutex_lock(&orphan_dir_inode
->i_mutex
);
1279 status
= ocfs2_inode_lock(orphan_dir_inode
, NULL
, 0);
1285 status
= ocfs2_dir_foreach(orphan_dir_inode
, &pos
, &priv
,
1286 ocfs2_orphan_filldir
);
1295 ocfs2_inode_unlock(orphan_dir_inode
, 0);
1297 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1298 iput(orphan_dir_inode
);
1302 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
1307 spin_lock(&osb
->osb_lock
);
1308 ret
= !osb
->osb_orphan_wipes
[slot
];
1309 spin_unlock(&osb
->osb_lock
);
1313 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
1316 spin_lock(&osb
->osb_lock
);
1317 /* Mark ourselves such that new processes in delete_inode()
1318 * know to quit early. */
1319 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1320 while (osb
->osb_orphan_wipes
[slot
]) {
1321 /* If any processes are already in the middle of an
1322 * orphan wipe on this dir, then we need to wait for
1324 spin_unlock(&osb
->osb_lock
);
1325 wait_event_interruptible(osb
->osb_wipe_event
,
1326 ocfs2_orphan_recovery_can_continue(osb
, slot
));
1327 spin_lock(&osb
->osb_lock
);
1329 spin_unlock(&osb
->osb_lock
);
1332 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
1335 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1339 * Orphan recovery. Each mounted node has it's own orphan dir which we
1340 * must run during recovery. Our strategy here is to build a list of
1341 * the inodes in the orphan dir and iget/iput them. The VFS does
1342 * (most) of the rest of the work.
1344 * Orphan recovery can happen at any time, not just mount so we have a
1345 * couple of extra considerations.
1347 * - We grab as many inodes as we can under the orphan dir lock -
1348 * doing iget() outside the orphan dir risks getting a reference on
1350 * - We must be sure not to deadlock with other processes on the
1351 * system wanting to run delete_inode(). This can happen when they go
1352 * to lock the orphan dir and the orphan recovery process attempts to
1353 * iget() inside the orphan dir lock. This can be avoided by
1354 * advertising our state to ocfs2_delete_inode().
1356 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1360 struct inode
*inode
= NULL
;
1362 struct ocfs2_inode_info
*oi
;
1364 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1366 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
1367 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
1368 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
1370 /* Error here should be noted, but we want to continue with as
1371 * many queued inodes as we've got. */
1376 oi
= OCFS2_I(inode
);
1377 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
1379 iter
= oi
->ip_next_orphan
;
1381 spin_lock(&oi
->ip_lock
);
1382 /* The remote delete code may have set these on the
1383 * assumption that the other node would wipe them
1384 * successfully. If they are still in the node's
1385 * orphan dir, we need to reset that state. */
1386 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1388 /* Set the proper information to get us going into
1389 * ocfs2_delete_inode. */
1390 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1391 spin_unlock(&oi
->ip_lock
);
1401 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1403 /* This check is good because ocfs2 will wait on our recovery
1404 * thread before changing it to something other than MOUNTED
1406 wait_event(osb
->osb_mount_event
,
1407 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1408 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1410 /* If there's an error on mount, then we may never get to the
1411 * MOUNTED flag, but this is set right before
1412 * dismount_volume() so we can trust it. */
1413 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1414 mlog(0, "mount error, exiting!\n");
1421 static int ocfs2_commit_thread(void *arg
)
1424 struct ocfs2_super
*osb
= arg
;
1425 struct ocfs2_journal
*journal
= osb
->journal
;
1427 /* we can trust j_num_trans here because _should_stop() is only set in
1428 * shutdown and nobody other than ourselves should be able to start
1429 * transactions. committing on shutdown might take a few iterations
1430 * as final transactions put deleted inodes on the list */
1431 while (!(kthread_should_stop() &&
1432 atomic_read(&journal
->j_num_trans
) == 0)) {
1434 wait_event_interruptible(osb
->checkpoint_event
,
1435 atomic_read(&journal
->j_num_trans
)
1436 || kthread_should_stop());
1438 status
= ocfs2_commit_cache(osb
);
1442 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1444 "commit_thread: %u transactions pending on "
1446 atomic_read(&journal
->j_num_trans
));
1453 /* Look for a dirty journal without taking any cluster locks. Used for
1454 * hard readonly access to determine whether the file system journals
1455 * require recovery. */
1456 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1460 struct buffer_head
*di_bh
;
1461 struct ocfs2_dinode
*di
;
1462 struct inode
*journal
= NULL
;
1464 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1465 journal
= ocfs2_get_system_file_inode(osb
,
1466 JOURNAL_SYSTEM_INODE
,
1468 if (!journal
|| is_bad_inode(journal
)) {
1475 ret
= ocfs2_read_block(osb
, OCFS2_I(journal
)->ip_blkno
, &di_bh
,
1482 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1484 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1485 OCFS2_JOURNAL_DIRTY_FL
)