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 struct ocfs2_journal_handle
*ocfs2_alloc_handle(struct ocfs2_super
*osb
)
115 struct ocfs2_journal_handle
*retval
= NULL
;
117 retval
= kcalloc(1, sizeof(*retval
), GFP_NOFS
);
119 mlog(ML_ERROR
, "Failed to allocate memory for journal "
123 retval
->k_handle
= NULL
;
125 retval
->journal
= osb
->journal
;
130 /* pass it NULL and it will allocate a new handle object for you. If
131 * you pass it a handle however, it may still return error, in which
132 * case it has free'd the passed handle for you. */
133 struct ocfs2_journal_handle
*ocfs2_start_trans(struct ocfs2_super
*osb
,
134 struct ocfs2_journal_handle
*handle
,
138 journal_t
*journal
= osb
->journal
->j_journal
;
140 mlog_entry("(max_buffs = %d)\n", max_buffs
);
142 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
144 if (ocfs2_is_hard_readonly(osb
)) {
149 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
150 BUG_ON(max_buffs
<= 0);
152 /* JBD might support this, but our journalling code doesn't yet. */
153 if (journal_current_handle()) {
154 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
159 handle
= ocfs2_alloc_handle(osb
);
162 mlog(ML_ERROR
, "Failed to allocate memory for journal "
167 down_read(&osb
->journal
->j_trans_barrier
);
169 /* actually start the transaction now */
170 handle
->k_handle
= journal_start(journal
, max_buffs
);
171 if (IS_ERR(handle
->k_handle
)) {
172 up_read(&osb
->journal
->j_trans_barrier
);
174 ret
= PTR_ERR(handle
->k_handle
);
175 handle
->k_handle
= NULL
;
178 if (is_journal_aborted(journal
)) {
179 ocfs2_abort(osb
->sb
, "Detected aborted journal");
185 atomic_inc(&(osb
->journal
->j_num_trans
));
187 mlog_exit_ptr(handle
);
198 void ocfs2_commit_trans(struct ocfs2_journal_handle
*handle
)
200 handle_t
*jbd_handle
;
202 struct ocfs2_journal
*journal
= handle
->journal
;
208 if (!handle
->k_handle
) {
214 /* ocfs2_extend_trans may have had to call journal_restart
215 * which will always commit the transaction, but may return
216 * error for any number of reasons. If this is the case, we
217 * clear k_handle as it's not valid any more. */
218 if (handle
->k_handle
) {
219 jbd_handle
= handle
->k_handle
;
221 /* actually stop the transaction. if we've set h_sync,
222 * it'll have been committed when we return */
223 retval
= journal_stop(jbd_handle
);
226 mlog(ML_ERROR
, "Could not commit transaction\n");
230 handle
->k_handle
= NULL
; /* it's been free'd in journal_stop */
233 up_read(&journal
->j_trans_barrier
);
240 * 'nblocks' is what you want to add to the current
241 * transaction. extend_trans will either extend the current handle by
242 * nblocks, or commit it and start a new one with nblocks credits.
244 * WARNING: This will not release any semaphores or disk locks taken
245 * during the transaction, so make sure they were taken *before*
246 * start_trans or we'll have ordering deadlocks.
248 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
249 * good because transaction ids haven't yet been recorded on the
250 * cluster locks associated with this handle.
252 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
261 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
263 status
= journal_extend(handle
, nblocks
);
270 mlog(0, "journal_extend failed, trying journal_restart\n");
271 status
= journal_restart(handle
, nblocks
);
285 int ocfs2_journal_access(struct ocfs2_journal_handle
*handle
,
287 struct buffer_head
*bh
,
296 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
297 (unsigned long long)bh
->b_blocknr
, type
,
298 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
299 "OCFS2_JOURNAL_ACCESS_CREATE" :
300 "OCFS2_JOURNAL_ACCESS_WRITE",
303 /* we can safely remove this assertion after testing. */
304 if (!buffer_uptodate(bh
)) {
305 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
306 mlog(ML_ERROR
, "b_blocknr=%llu\n",
307 (unsigned long long)bh
->b_blocknr
);
311 /* Set the current transaction information on the inode so
312 * that the locking code knows whether it can drop it's locks
313 * on this inode or not. We're protected from the commit
314 * thread updating the current transaction id until
315 * ocfs2_commit_trans() because ocfs2_start_trans() took
316 * j_trans_barrier for us. */
317 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
319 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
321 case OCFS2_JOURNAL_ACCESS_CREATE
:
322 case OCFS2_JOURNAL_ACCESS_WRITE
:
323 status
= journal_get_write_access(handle
->k_handle
, bh
);
326 case OCFS2_JOURNAL_ACCESS_UNDO
:
327 status
= journal_get_undo_access(handle
->k_handle
, bh
);
332 mlog(ML_ERROR
, "Uknown access type!\n");
334 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
337 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
344 int ocfs2_journal_dirty(struct ocfs2_journal_handle
*handle
,
345 struct buffer_head
*bh
)
349 mlog_entry("(bh->b_blocknr=%llu)\n",
350 (unsigned long long)bh
->b_blocknr
);
352 status
= journal_dirty_metadata(handle
->k_handle
, bh
);
354 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
355 "(bh->b_blocknr=%llu)\n",
356 (unsigned long long)bh
->b_blocknr
);
362 int ocfs2_journal_dirty_data(handle_t
*handle
,
363 struct buffer_head
*bh
)
365 int err
= journal_dirty_data(handle
, bh
);
368 /* TODO: When we can handle it, abort the handle and go RO on
374 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
376 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
378 journal_t
*journal
= osb
->journal
->j_journal
;
380 spin_lock(&journal
->j_state_lock
);
381 journal
->j_commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
382 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
383 journal
->j_flags
|= JFS_BARRIER
;
385 journal
->j_flags
&= ~JFS_BARRIER
;
386 spin_unlock(&journal
->j_state_lock
);
389 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
392 struct inode
*inode
= NULL
; /* the journal inode */
393 journal_t
*j_journal
= NULL
;
394 struct ocfs2_dinode
*di
= NULL
;
395 struct buffer_head
*bh
= NULL
;
396 struct ocfs2_super
*osb
;
403 osb
= journal
->j_osb
;
405 /* already have the inode for our journal */
406 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
413 if (is_bad_inode(inode
)) {
414 mlog(ML_ERROR
, "access error (bad inode)\n");
421 SET_INODE_JOURNAL(inode
);
422 OCFS2_I(inode
)->ip_open_count
++;
424 /* Skip recovery waits here - journal inode metadata never
425 * changes in a live cluster so it can be considered an
426 * exception to the rule. */
427 status
= ocfs2_meta_lock_full(inode
, NULL
, &bh
, 1,
428 OCFS2_META_LOCK_RECOVERY
);
430 if (status
!= -ERESTARTSYS
)
431 mlog(ML_ERROR
, "Could not get lock on journal!\n");
436 di
= (struct ocfs2_dinode
*)bh
->b_data
;
438 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
439 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
445 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
446 mlog(0, "inode->i_blocks = %llu\n",
447 (unsigned long long)inode
->i_blocks
);
448 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
450 /* call the kernels journal init function now */
451 j_journal
= journal_init_inode(inode
);
452 if (j_journal
== NULL
) {
453 mlog(ML_ERROR
, "Linux journal layer error\n");
458 mlog(0, "Returned from journal_init_inode\n");
459 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
461 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
462 OCFS2_JOURNAL_DIRTY_FL
);
464 journal
->j_journal
= j_journal
;
465 journal
->j_inode
= inode
;
468 ocfs2_set_journal_params(osb
);
470 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
476 ocfs2_meta_unlock(inode
, 1);
480 OCFS2_I(inode
)->ip_open_count
--;
489 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
494 struct ocfs2_journal
*journal
= osb
->journal
;
495 struct buffer_head
*bh
= journal
->j_bh
;
496 struct ocfs2_dinode
*fe
;
500 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
501 if (!OCFS2_IS_VALID_DINODE(fe
)) {
502 /* This is called from startup/shutdown which will
503 * handle the errors in a specific manner, so no need
504 * to call ocfs2_error() here. */
505 mlog(ML_ERROR
, "Journal dinode %llu has invalid "
506 "signature: %.*s", (unsigned long long)fe
->i_blkno
, 7,
512 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
514 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
516 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
517 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
519 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
529 * If the journal has been kmalloc'd it needs to be freed after this
532 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
534 struct ocfs2_journal
*journal
= NULL
;
536 struct inode
*inode
= NULL
;
537 int num_running_trans
= 0;
543 journal
= osb
->journal
;
547 inode
= journal
->j_inode
;
549 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
552 /* need to inc inode use count as journal_destroy will iput. */
556 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
557 if (num_running_trans
> 0)
558 mlog(0, "Shutting down journal: must wait on %d "
559 "running transactions!\n",
562 /* Do a commit_cache here. It will flush our journal, *and*
563 * release any locks that are still held.
564 * set the SHUTDOWN flag and release the trans lock.
565 * the commit thread will take the trans lock for us below. */
566 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
568 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
569 * drop the trans_lock (which we want to hold until we
570 * completely destroy the journal. */
571 if (osb
->commit_task
) {
572 /* Wait for the commit thread */
573 mlog(0, "Waiting for ocfs2commit to exit....\n");
574 kthread_stop(osb
->commit_task
);
575 osb
->commit_task
= NULL
;
578 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
580 status
= ocfs2_journal_toggle_dirty(osb
, 0);
584 /* Shutdown the kernel journal system */
585 journal_destroy(journal
->j_journal
);
587 OCFS2_I(inode
)->ip_open_count
--;
589 /* unlock our journal */
590 ocfs2_meta_unlock(inode
, 1);
592 brelse(journal
->j_bh
);
593 journal
->j_bh
= NULL
;
595 journal
->j_state
= OCFS2_JOURNAL_FREE
;
597 // up_write(&journal->j_trans_barrier);
604 static void ocfs2_clear_journal_error(struct super_block
*sb
,
610 olderr
= journal_errno(journal
);
612 mlog(ML_ERROR
, "File system error %d recorded in "
613 "journal %u.\n", olderr
, slot
);
614 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
617 journal_ack_err(journal
);
618 journal_clear_err(journal
);
622 int ocfs2_journal_load(struct ocfs2_journal
*journal
)
625 struct ocfs2_super
*osb
;
632 osb
= journal
->j_osb
;
634 status
= journal_load(journal
->j_journal
);
636 mlog(ML_ERROR
, "Failed to load journal!\n");
640 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
642 status
= ocfs2_journal_toggle_dirty(osb
, 1);
648 /* Launch the commit thread */
649 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
, "ocfs2cmt");
650 if (IS_ERR(osb
->commit_task
)) {
651 status
= PTR_ERR(osb
->commit_task
);
652 osb
->commit_task
= NULL
;
653 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, error=%d",
664 /* 'full' flag tells us whether we clear out all blocks or if we just
665 * mark the journal clean */
666 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
674 status
= journal_wipe(journal
->j_journal
, full
);
680 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0);
690 * JBD Might read a cached version of another nodes journal file. We
691 * don't want this as this file changes often and we get no
692 * notification on those changes. The only way to be sure that we've
693 * got the most up to date version of those blocks then is to force
694 * read them off disk. Just searching through the buffer cache won't
695 * work as there may be pages backing this file which are still marked
696 * up to date. We know things can't change on this file underneath us
697 * as we have the lock by now :)
699 static int ocfs2_force_read_journal(struct inode
*inode
)
703 u64 v_blkno
, p_blkno
;
704 #define CONCURRENT_JOURNAL_FILL 32
705 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
709 BUG_ON(inode
->i_blocks
!=
710 ocfs2_align_bytes_to_sectors(i_size_read(inode
)));
712 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
714 mlog(0, "Force reading %llu blocks\n",
715 (unsigned long long)(inode
->i_blocks
>>
716 (inode
->i_sb
->s_blocksize_bits
- 9)));
720 (inode
->i_blocks
>> (inode
->i_sb
->s_blocksize_bits
- 9))) {
722 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
730 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
731 p_blocks
= CONCURRENT_JOURNAL_FILL
;
733 /* We are reading journal data which should not
734 * be put in the uptodate cache */
735 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
736 p_blkno
, p_blocks
, bhs
, 0,
743 for(i
= 0; i
< p_blocks
; i
++) {
752 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
759 struct ocfs2_la_recovery_item
{
760 struct list_head lri_list
;
762 struct ocfs2_dinode
*lri_la_dinode
;
763 struct ocfs2_dinode
*lri_tl_dinode
;
766 /* Does the second half of the recovery process. By this point, the
767 * node is marked clean and can actually be considered recovered,
768 * hence it's no longer in the recovery map, but there's still some
769 * cleanup we can do which shouldn't happen within the recovery thread
770 * as locking in that context becomes very difficult if we are to take
771 * recovering nodes into account.
773 * NOTE: This function can and will sleep on recovery of other nodes
774 * during cluster locking, just like any other ocfs2 process.
776 void ocfs2_complete_recovery(void *data
)
779 struct ocfs2_super
*osb
= data
;
780 struct ocfs2_journal
*journal
= osb
->journal
;
781 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
782 struct ocfs2_la_recovery_item
*item
;
783 struct list_head
*p
, *n
;
784 LIST_HEAD(tmp_la_list
);
788 mlog(0, "completing recovery from keventd\n");
790 spin_lock(&journal
->j_lock
);
791 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
792 spin_unlock(&journal
->j_lock
);
794 list_for_each_safe(p
, n
, &tmp_la_list
) {
795 item
= list_entry(p
, struct ocfs2_la_recovery_item
, lri_list
);
796 list_del_init(&item
->lri_list
);
798 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
800 la_dinode
= item
->lri_la_dinode
;
802 mlog(0, "Clean up local alloc %llu\n",
803 (unsigned long long)la_dinode
->i_blkno
);
805 ret
= ocfs2_complete_local_alloc_recovery(osb
,
813 tl_dinode
= item
->lri_tl_dinode
;
815 mlog(0, "Clean up truncate log %llu\n",
816 (unsigned long long)tl_dinode
->i_blkno
);
818 ret
= ocfs2_complete_truncate_log_recovery(osb
,
826 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
833 mlog(0, "Recovery completion\n");
837 /* NOTE: This function always eats your references to la_dinode and
838 * tl_dinode, either manually on error, or by passing them to
839 * ocfs2_complete_recovery */
840 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
842 struct ocfs2_dinode
*la_dinode
,
843 struct ocfs2_dinode
*tl_dinode
)
845 struct ocfs2_la_recovery_item
*item
;
847 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
849 /* Though we wish to avoid it, we are in fact safe in
850 * skipping local alloc cleanup as fsck.ocfs2 is more
851 * than capable of reclaiming unused space. */
862 INIT_LIST_HEAD(&item
->lri_list
);
863 item
->lri_la_dinode
= la_dinode
;
864 item
->lri_slot
= slot_num
;
865 item
->lri_tl_dinode
= tl_dinode
;
867 spin_lock(&journal
->j_lock
);
868 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
869 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
870 spin_unlock(&journal
->j_lock
);
873 /* Called by the mount code to queue recovery the last part of
874 * recovery for it's own slot. */
875 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
877 struct ocfs2_journal
*journal
= osb
->journal
;
880 /* No need to queue up our truncate_log as regular
881 * cleanup will catch that. */
882 ocfs2_queue_recovery_completion(journal
,
884 osb
->local_alloc_copy
,
886 ocfs2_schedule_truncate_log_flush(osb
, 0);
888 osb
->local_alloc_copy
= NULL
;
893 static int __ocfs2_recovery_thread(void *arg
)
895 int status
, node_num
;
896 struct ocfs2_super
*osb
= arg
;
900 status
= ocfs2_wait_on_mount(osb
);
906 status
= ocfs2_super_lock(osb
, 1);
912 while(!ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
913 node_num
= ocfs2_node_map_first_set_bit(osb
,
915 if (node_num
== O2NM_INVALID_NODE_NUM
) {
916 mlog(0, "Out of nodes to recover.\n");
920 status
= ocfs2_recover_node(osb
, node_num
);
923 "Error %d recovering node %d on device (%u,%u)!\n",
925 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
926 mlog(ML_ERROR
, "Volume requires unmount.\n");
930 ocfs2_recovery_map_clear(osb
, node_num
);
932 ocfs2_super_unlock(osb
, 1);
934 /* We always run recovery on our own orphan dir - the dead
935 * node(s) may have voted "no" on an inode delete earlier. A
936 * revote is therefore required. */
937 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
941 mutex_lock(&osb
->recovery_lock
);
943 !ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
944 mutex_unlock(&osb
->recovery_lock
);
948 osb
->recovery_thread_task
= NULL
;
949 mb(); /* sync with ocfs2_recovery_thread_running */
950 wake_up(&osb
->recovery_event
);
952 mutex_unlock(&osb
->recovery_lock
);
955 /* no one is callint kthread_stop() for us so the kthread() api
956 * requires that we call do_exit(). And it isn't exported, but
957 * complete_and_exit() seems to be a minimal wrapper around it. */
958 complete_and_exit(NULL
, status
);
962 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
964 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
965 node_num
, osb
->node_num
);
967 mutex_lock(&osb
->recovery_lock
);
968 if (osb
->disable_recovery
)
971 /* People waiting on recovery will wait on
972 * the recovery map to empty. */
973 if (!ocfs2_recovery_map_set(osb
, node_num
))
974 mlog(0, "node %d already be in recovery.\n", node_num
);
976 mlog(0, "starting recovery thread...\n");
978 if (osb
->recovery_thread_task
)
981 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
983 if (IS_ERR(osb
->recovery_thread_task
)) {
984 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
985 osb
->recovery_thread_task
= NULL
;
989 mutex_unlock(&osb
->recovery_lock
);
990 wake_up(&osb
->recovery_event
);
995 /* Does the actual journal replay and marks the journal inode as
996 * clean. Will only replay if the journal inode is marked dirty. */
997 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
1004 struct inode
*inode
= NULL
;
1005 struct ocfs2_dinode
*fe
;
1006 journal_t
*journal
= NULL
;
1007 struct buffer_head
*bh
= NULL
;
1009 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1011 if (inode
== NULL
) {
1016 if (is_bad_inode(inode
)) {
1023 SET_INODE_JOURNAL(inode
);
1025 status
= ocfs2_meta_lock_full(inode
, NULL
, &bh
, 1,
1026 OCFS2_META_LOCK_RECOVERY
);
1028 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status
);
1029 if (status
!= -ERESTARTSYS
)
1030 mlog(ML_ERROR
, "Could not lock journal!\n");
1035 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
1037 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1039 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
1040 mlog(0, "No recovery required for node %d\n", node_num
);
1044 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
1046 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1048 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
1050 status
= ocfs2_force_read_journal(inode
);
1056 mlog(0, "calling journal_init_inode\n");
1057 journal
= journal_init_inode(inode
);
1058 if (journal
== NULL
) {
1059 mlog(ML_ERROR
, "Linux journal layer error\n");
1064 status
= journal_load(journal
);
1069 journal_destroy(journal
);
1073 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1075 /* wipe the journal */
1076 mlog(0, "flushing the journal.\n");
1077 journal_lock_updates(journal
);
1078 status
= journal_flush(journal
);
1079 journal_unlock_updates(journal
);
1083 /* This will mark the node clean */
1084 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1085 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1086 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1088 status
= ocfs2_write_block(osb
, bh
, inode
);
1095 journal_destroy(journal
);
1098 /* drop the lock on this nodes journal */
1100 ocfs2_meta_unlock(inode
, 1);
1113 * Do the most important parts of node recovery:
1114 * - Replay it's journal
1115 * - Stamp a clean local allocator file
1116 * - Stamp a clean truncate log
1117 * - Mark the node clean
1119 * If this function completes without error, a node in OCFS2 can be
1120 * said to have been safely recovered. As a result, failure during the
1121 * second part of a nodes recovery process (local alloc recovery) is
1122 * far less concerning.
1124 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1129 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1130 struct ocfs2_dinode
*la_copy
= NULL
;
1131 struct ocfs2_dinode
*tl_copy
= NULL
;
1133 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1134 node_num
, osb
->node_num
);
1136 mlog(0, "checking node %d\n", node_num
);
1138 /* Should not ever be called to recover ourselves -- in that
1139 * case we should've called ocfs2_journal_load instead. */
1140 BUG_ON(osb
->node_num
== node_num
);
1142 slot_num
= ocfs2_node_num_to_slot(si
, node_num
);
1143 if (slot_num
== OCFS2_INVALID_SLOT
) {
1145 mlog(0, "no slot for this node, so no recovery required.\n");
1149 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1151 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1157 /* Stamp a clean local alloc file AFTER recovering the journal... */
1158 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1164 /* An error from begin_truncate_log_recovery is not
1165 * serious enough to warrant halting the rest of
1167 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1171 /* Likewise, this would be a strange but ultimately not so
1172 * harmful place to get an error... */
1173 ocfs2_clear_slot(si
, slot_num
);
1174 status
= ocfs2_update_disk_slots(osb
, si
);
1178 /* This will kfree the memory pointed to by la_copy and tl_copy */
1179 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1189 /* Test node liveness by trylocking his journal. If we get the lock,
1190 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1191 * still alive (we couldn't get the lock) and < 0 on error. */
1192 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1196 struct inode
*inode
= NULL
;
1198 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1200 if (inode
== NULL
) {
1201 mlog(ML_ERROR
, "access error\n");
1205 if (is_bad_inode(inode
)) {
1206 mlog(ML_ERROR
, "access error (bad inode)\n");
1212 SET_INODE_JOURNAL(inode
);
1214 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1215 status
= ocfs2_meta_lock_full(inode
, NULL
, NULL
, 1, flags
);
1217 if (status
!= -EAGAIN
)
1222 ocfs2_meta_unlock(inode
, 1);
1230 /* Call this underneath ocfs2_super_lock. It also assumes that the
1231 * slot info struct has been updated from disk. */
1232 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1234 int status
, i
, node_num
;
1235 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1237 /* This is called with the super block cluster lock, so we
1238 * know that the slot map can't change underneath us. */
1240 spin_lock(&si
->si_lock
);
1241 for(i
= 0; i
< si
->si_num_slots
; i
++) {
1242 if (i
== osb
->slot_num
)
1244 if (ocfs2_is_empty_slot(si
, i
))
1247 node_num
= si
->si_global_node_nums
[i
];
1248 if (ocfs2_node_map_test_bit(osb
, &osb
->recovery_map
, node_num
))
1250 spin_unlock(&si
->si_lock
);
1252 /* Ok, we have a slot occupied by another node which
1253 * is not in the recovery map. We trylock his journal
1254 * file here to test if he's alive. */
1255 status
= ocfs2_trylock_journal(osb
, i
);
1257 /* Since we're called from mount, we know that
1258 * the recovery thread can't race us on
1259 * setting / checking the recovery bits. */
1260 ocfs2_recovery_thread(osb
, node_num
);
1261 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1266 spin_lock(&si
->si_lock
);
1268 spin_unlock(&si
->si_lock
);
1276 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
1278 struct inode
**head
)
1281 struct inode
*orphan_dir_inode
= NULL
;
1283 unsigned long offset
, blk
, local
;
1284 struct buffer_head
*bh
= NULL
;
1285 struct ocfs2_dir_entry
*de
;
1286 struct super_block
*sb
= osb
->sb
;
1288 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1289 ORPHAN_DIR_SYSTEM_INODE
,
1291 if (!orphan_dir_inode
) {
1297 mutex_lock(&orphan_dir_inode
->i_mutex
);
1298 status
= ocfs2_meta_lock(orphan_dir_inode
, NULL
, NULL
, 0);
1306 while(offset
< i_size_read(orphan_dir_inode
)) {
1307 blk
= offset
>> sb
->s_blocksize_bits
;
1309 bh
= ocfs2_bread(orphan_dir_inode
, blk
, &status
, 0);
1320 while(offset
< i_size_read(orphan_dir_inode
)
1321 && local
< sb
->s_blocksize
) {
1322 de
= (struct ocfs2_dir_entry
*) (bh
->b_data
+ local
);
1324 if (!ocfs2_check_dir_entry(orphan_dir_inode
,
1332 local
+= le16_to_cpu(de
->rec_len
);
1333 offset
+= le16_to_cpu(de
->rec_len
);
1335 /* I guess we silently fail on no inode? */
1336 if (!le64_to_cpu(de
->inode
))
1338 if (de
->file_type
> OCFS2_FT_MAX
) {
1340 "block %llu contains invalid de: "
1341 "inode = %llu, rec_len = %u, "
1342 "name_len = %u, file_type = %u, "
1344 (unsigned long long)bh
->b_blocknr
,
1345 (unsigned long long)le64_to_cpu(de
->inode
),
1346 le16_to_cpu(de
->rec_len
),
1353 if (de
->name_len
== 1 && !strncmp(".", de
->name
, 1))
1355 if (de
->name_len
== 2 && !strncmp("..", de
->name
, 2))
1358 iter
= ocfs2_iget(osb
, le64_to_cpu(de
->inode
),
1359 OCFS2_FI_FLAG_NOLOCK
);
1363 mlog(0, "queue orphan %llu\n",
1364 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
1365 /* No locking is required for the next_orphan
1366 * queue as there is only ever a single
1367 * process doing orphan recovery. */
1368 OCFS2_I(iter
)->ip_next_orphan
= *head
;
1375 ocfs2_meta_unlock(orphan_dir_inode
, 0);
1377 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1378 iput(orphan_dir_inode
);
1382 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
1387 spin_lock(&osb
->osb_lock
);
1388 ret
= !osb
->osb_orphan_wipes
[slot
];
1389 spin_unlock(&osb
->osb_lock
);
1393 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
1396 spin_lock(&osb
->osb_lock
);
1397 /* Mark ourselves such that new processes in delete_inode()
1398 * know to quit early. */
1399 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1400 while (osb
->osb_orphan_wipes
[slot
]) {
1401 /* If any processes are already in the middle of an
1402 * orphan wipe on this dir, then we need to wait for
1404 spin_unlock(&osb
->osb_lock
);
1405 wait_event_interruptible(osb
->osb_wipe_event
,
1406 ocfs2_orphan_recovery_can_continue(osb
, slot
));
1407 spin_lock(&osb
->osb_lock
);
1409 spin_unlock(&osb
->osb_lock
);
1412 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
1415 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1419 * Orphan recovery. Each mounted node has it's own orphan dir which we
1420 * must run during recovery. Our strategy here is to build a list of
1421 * the inodes in the orphan dir and iget/iput them. The VFS does
1422 * (most) of the rest of the work.
1424 * Orphan recovery can happen at any time, not just mount so we have a
1425 * couple of extra considerations.
1427 * - We grab as many inodes as we can under the orphan dir lock -
1428 * doing iget() outside the orphan dir risks getting a reference on
1430 * - We must be sure not to deadlock with other processes on the
1431 * system wanting to run delete_inode(). This can happen when they go
1432 * to lock the orphan dir and the orphan recovery process attempts to
1433 * iget() inside the orphan dir lock. This can be avoided by
1434 * advertising our state to ocfs2_delete_inode().
1436 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1440 struct inode
*inode
= NULL
;
1442 struct ocfs2_inode_info
*oi
;
1444 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1446 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
1447 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
1448 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
1450 /* Error here should be noted, but we want to continue with as
1451 * many queued inodes as we've got. */
1456 oi
= OCFS2_I(inode
);
1457 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
1459 iter
= oi
->ip_next_orphan
;
1461 spin_lock(&oi
->ip_lock
);
1462 /* Delete voting may have set these on the assumption
1463 * that the other node would wipe them successfully.
1464 * If they are still in the node's orphan dir, we need
1465 * to reset that state. */
1466 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1468 /* Set the proper information to get us going into
1469 * ocfs2_delete_inode. */
1470 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1471 oi
->ip_orphaned_slot
= slot
;
1472 spin_unlock(&oi
->ip_lock
);
1482 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1484 /* This check is good because ocfs2 will wait on our recovery
1485 * thread before changing it to something other than MOUNTED
1487 wait_event(osb
->osb_mount_event
,
1488 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1489 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1491 /* If there's an error on mount, then we may never get to the
1492 * MOUNTED flag, but this is set right before
1493 * dismount_volume() so we can trust it. */
1494 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1495 mlog(0, "mount error, exiting!\n");
1502 static int ocfs2_commit_thread(void *arg
)
1505 struct ocfs2_super
*osb
= arg
;
1506 struct ocfs2_journal
*journal
= osb
->journal
;
1508 /* we can trust j_num_trans here because _should_stop() is only set in
1509 * shutdown and nobody other than ourselves should be able to start
1510 * transactions. committing on shutdown might take a few iterations
1511 * as final transactions put deleted inodes on the list */
1512 while (!(kthread_should_stop() &&
1513 atomic_read(&journal
->j_num_trans
) == 0)) {
1515 wait_event_interruptible(osb
->checkpoint_event
,
1516 atomic_read(&journal
->j_num_trans
)
1517 || kthread_should_stop());
1519 status
= ocfs2_commit_cache(osb
);
1523 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1525 "commit_thread: %u transactions pending on "
1527 atomic_read(&journal
->j_num_trans
));
1534 /* Look for a dirty journal without taking any cluster locks. Used for
1535 * hard readonly access to determine whether the file system journals
1536 * require recovery. */
1537 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1541 struct buffer_head
*di_bh
;
1542 struct ocfs2_dinode
*di
;
1543 struct inode
*journal
= NULL
;
1545 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1546 journal
= ocfs2_get_system_file_inode(osb
,
1547 JOURNAL_SYSTEM_INODE
,
1549 if (!journal
|| is_bad_inode(journal
)) {
1556 ret
= ocfs2_read_block(osb
, OCFS2_I(journal
)->ip_blkno
, &di_bh
,
1563 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1565 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1566 OCFS2_JOURNAL_DIRTY_FL
)