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 void ocfs2_handle_cleanup_locks(struct ocfs2_journal
*journal
,
61 struct ocfs2_journal_handle
*handle
);
62 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle
*handle
);
63 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
65 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
67 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
69 static int ocfs2_commit_thread(void *arg
);
71 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
76 struct ocfs2_journal
*journal
= NULL
;
80 journal
= osb
->journal
;
82 /* Flush all pending commits and checkpoint the journal. */
83 down_write(&journal
->j_trans_barrier
);
85 if (atomic_read(&journal
->j_num_trans
) == 0) {
86 up_write(&journal
->j_trans_barrier
);
87 mlog(0, "No transactions for me to flush!\n");
91 journal_lock_updates(journal
->j_journal
);
92 status
= journal_flush(journal
->j_journal
);
93 journal_unlock_updates(journal
->j_journal
);
95 up_write(&journal
->j_trans_barrier
);
100 old_id
= ocfs2_inc_trans_id(journal
);
102 flushed
= atomic_read(&journal
->j_num_trans
);
103 atomic_set(&journal
->j_num_trans
, 0);
104 up_write(&journal
->j_trans_barrier
);
106 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
107 journal
->j_trans_id
, flushed
);
109 ocfs2_kick_vote_thread(osb
);
110 wake_up(&journal
->j_checkpointed
);
116 struct ocfs2_journal_handle
*ocfs2_alloc_handle(struct ocfs2_super
*osb
)
118 struct ocfs2_journal_handle
*retval
= NULL
;
120 retval
= kcalloc(1, sizeof(*retval
), GFP_NOFS
);
122 mlog(ML_ERROR
, "Failed to allocate memory for journal "
127 retval
->num_locks
= 0;
128 retval
->k_handle
= NULL
;
130 INIT_LIST_HEAD(&retval
->locks
);
131 retval
->journal
= osb
->journal
;
136 /* pass it NULL and it will allocate a new handle object for you. If
137 * you pass it a handle however, it may still return error, in which
138 * case it has free'd the passed handle for you. */
139 struct ocfs2_journal_handle
*ocfs2_start_trans(struct ocfs2_super
*osb
,
140 struct ocfs2_journal_handle
*handle
,
144 journal_t
*journal
= osb
->journal
->j_journal
;
146 mlog_entry("(max_buffs = %d)\n", max_buffs
);
148 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
150 if (ocfs2_is_hard_readonly(osb
)) {
155 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
156 BUG_ON(max_buffs
<= 0);
158 /* JBD might support this, but our journalling code doesn't yet. */
159 if (journal_current_handle()) {
160 mlog(ML_ERROR
, "Recursive transaction attempted!\n");
165 handle
= ocfs2_alloc_handle(osb
);
168 mlog(ML_ERROR
, "Failed to allocate memory for journal "
173 down_read(&osb
->journal
->j_trans_barrier
);
175 /* actually start the transaction now */
176 handle
->k_handle
= journal_start(journal
, max_buffs
);
177 if (IS_ERR(handle
->k_handle
)) {
178 up_read(&osb
->journal
->j_trans_barrier
);
180 ret
= PTR_ERR(handle
->k_handle
);
181 handle
->k_handle
= NULL
;
184 if (is_journal_aborted(journal
)) {
185 ocfs2_abort(osb
->sb
, "Detected aborted journal");
191 atomic_inc(&(osb
->journal
->j_num_trans
));
193 mlog_exit_ptr(handle
);
198 ocfs2_commit_unstarted_handle(handle
); /* will kfree handle */
204 /* This is trivial so we do it out of the main commit
205 * paths. Beware, it can be called from start_trans too! */
206 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle
*handle
)
210 /* You are allowed to add journal locks before the transaction
212 ocfs2_handle_cleanup_locks(handle
->journal
, handle
);
219 void ocfs2_commit_trans(struct ocfs2_journal_handle
*handle
)
221 handle_t
*jbd_handle
;
223 struct ocfs2_journal
*journal
= handle
->journal
;
229 if (!handle
->k_handle
) {
230 ocfs2_commit_unstarted_handle(handle
);
235 /* ocfs2_extend_trans may have had to call journal_restart
236 * which will always commit the transaction, but may return
237 * error for any number of reasons. If this is the case, we
238 * clear k_handle as it's not valid any more. */
239 if (handle
->k_handle
) {
240 jbd_handle
= handle
->k_handle
;
242 /* actually stop the transaction. if we've set h_sync,
243 * it'll have been committed when we return */
244 retval
= journal_stop(jbd_handle
);
247 mlog(ML_ERROR
, "Could not commit transaction\n");
251 handle
->k_handle
= NULL
; /* it's been free'd in journal_stop */
254 ocfs2_handle_cleanup_locks(journal
, handle
);
256 up_read(&journal
->j_trans_barrier
);
263 * 'nblocks' is what you want to add to the current
264 * transaction. extend_trans will either extend the current handle by
265 * nblocks, or commit it and start a new one with nblocks credits.
267 * WARNING: This will not release any semaphores or disk locks taken
268 * during the transaction, so make sure they were taken *before*
269 * start_trans or we'll have ordering deadlocks.
271 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
272 * good because transaction ids haven't yet been recorded on the
273 * cluster locks associated with this handle.
275 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
284 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
286 status
= journal_extend(handle
, nblocks
);
293 mlog(0, "journal_extend failed, trying journal_restart\n");
294 status
= journal_restart(handle
, nblocks
);
308 int ocfs2_journal_access(struct ocfs2_journal_handle
*handle
,
310 struct buffer_head
*bh
,
319 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
320 (unsigned long long)bh
->b_blocknr
, type
,
321 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
322 "OCFS2_JOURNAL_ACCESS_CREATE" :
323 "OCFS2_JOURNAL_ACCESS_WRITE",
326 /* we can safely remove this assertion after testing. */
327 if (!buffer_uptodate(bh
)) {
328 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
329 mlog(ML_ERROR
, "b_blocknr=%llu\n",
330 (unsigned long long)bh
->b_blocknr
);
334 /* Set the current transaction information on the inode so
335 * that the locking code knows whether it can drop it's locks
336 * on this inode or not. We're protected from the commit
337 * thread updating the current transaction id until
338 * ocfs2_commit_trans() because ocfs2_start_trans() took
339 * j_trans_barrier for us. */
340 ocfs2_set_inode_lock_trans(OCFS2_SB(inode
->i_sb
)->journal
, inode
);
342 mutex_lock(&OCFS2_I(inode
)->ip_io_mutex
);
344 case OCFS2_JOURNAL_ACCESS_CREATE
:
345 case OCFS2_JOURNAL_ACCESS_WRITE
:
346 status
= journal_get_write_access(handle
->k_handle
, bh
);
349 case OCFS2_JOURNAL_ACCESS_UNDO
:
350 status
= journal_get_undo_access(handle
->k_handle
, bh
);
355 mlog(ML_ERROR
, "Uknown access type!\n");
357 mutex_unlock(&OCFS2_I(inode
)->ip_io_mutex
);
360 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
367 int ocfs2_journal_dirty(struct ocfs2_journal_handle
*handle
,
368 struct buffer_head
*bh
)
372 mlog_entry("(bh->b_blocknr=%llu)\n",
373 (unsigned long long)bh
->b_blocknr
);
375 status
= journal_dirty_metadata(handle
->k_handle
, bh
);
377 mlog(ML_ERROR
, "Could not dirty metadata buffer. "
378 "(bh->b_blocknr=%llu)\n",
379 (unsigned long long)bh
->b_blocknr
);
385 int ocfs2_journal_dirty_data(handle_t
*handle
,
386 struct buffer_head
*bh
)
388 int err
= journal_dirty_data(handle
, bh
);
391 /* TODO: When we can handle it, abort the handle and go RO on
397 /* We always assume you're adding a metadata lock at level 'ex' */
398 int ocfs2_handle_add_lock(struct ocfs2_journal_handle
*handle
,
402 struct ocfs2_journal_lock
*lock
;
406 lock
= kmem_cache_alloc(ocfs2_lock_cache
, GFP_NOFS
);
415 lock
->jl_inode
= inode
;
417 list_add_tail(&(lock
->jl_lock_list
), &(handle
->locks
));
426 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal
*journal
,
427 struct ocfs2_journal_handle
*handle
)
429 struct list_head
*p
, *n
;
430 struct ocfs2_journal_lock
*lock
;
433 list_for_each_safe(p
, n
, &(handle
->locks
)) {
434 lock
= list_entry(p
, struct ocfs2_journal_lock
,
436 list_del(&lock
->jl_lock_list
);
439 inode
= lock
->jl_inode
;
440 ocfs2_meta_unlock(inode
, 1);
441 if (atomic_read(&inode
->i_count
) == 1)
443 "Inode %llu, I'm doing a last iput for!",
444 (unsigned long long)OCFS2_I(inode
)->ip_blkno
);
446 kmem_cache_free(ocfs2_lock_cache
, lock
);
450 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
452 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
454 journal_t
*journal
= osb
->journal
->j_journal
;
456 spin_lock(&journal
->j_state_lock
);
457 journal
->j_commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
458 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
459 journal
->j_flags
|= JFS_BARRIER
;
461 journal
->j_flags
&= ~JFS_BARRIER
;
462 spin_unlock(&journal
->j_state_lock
);
465 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
468 struct inode
*inode
= NULL
; /* the journal inode */
469 journal_t
*j_journal
= NULL
;
470 struct ocfs2_dinode
*di
= NULL
;
471 struct buffer_head
*bh
= NULL
;
472 struct ocfs2_super
*osb
;
479 osb
= journal
->j_osb
;
481 /* already have the inode for our journal */
482 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
489 if (is_bad_inode(inode
)) {
490 mlog(ML_ERROR
, "access error (bad inode)\n");
497 SET_INODE_JOURNAL(inode
);
498 OCFS2_I(inode
)->ip_open_count
++;
500 /* Skip recovery waits here - journal inode metadata never
501 * changes in a live cluster so it can be considered an
502 * exception to the rule. */
503 status
= ocfs2_meta_lock_full(inode
, NULL
, &bh
, 1,
504 OCFS2_META_LOCK_RECOVERY
);
506 if (status
!= -ERESTARTSYS
)
507 mlog(ML_ERROR
, "Could not get lock on journal!\n");
512 di
= (struct ocfs2_dinode
*)bh
->b_data
;
514 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
515 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
521 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
522 mlog(0, "inode->i_blocks = %llu\n",
523 (unsigned long long)inode
->i_blocks
);
524 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
526 /* call the kernels journal init function now */
527 j_journal
= journal_init_inode(inode
);
528 if (j_journal
== NULL
) {
529 mlog(ML_ERROR
, "Linux journal layer error\n");
534 mlog(0, "Returned from journal_init_inode\n");
535 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
537 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
538 OCFS2_JOURNAL_DIRTY_FL
);
540 journal
->j_journal
= j_journal
;
541 journal
->j_inode
= inode
;
544 ocfs2_set_journal_params(osb
);
546 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
552 ocfs2_meta_unlock(inode
, 1);
556 OCFS2_I(inode
)->ip_open_count
--;
565 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
570 struct ocfs2_journal
*journal
= osb
->journal
;
571 struct buffer_head
*bh
= journal
->j_bh
;
572 struct ocfs2_dinode
*fe
;
576 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
577 if (!OCFS2_IS_VALID_DINODE(fe
)) {
578 /* This is called from startup/shutdown which will
579 * handle the errors in a specific manner, so no need
580 * to call ocfs2_error() here. */
581 mlog(ML_ERROR
, "Journal dinode %llu has invalid "
582 "signature: %.*s", (unsigned long long)fe
->i_blkno
, 7,
588 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
590 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
592 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
593 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
595 status
= ocfs2_write_block(osb
, bh
, journal
->j_inode
);
605 * If the journal has been kmalloc'd it needs to be freed after this
608 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
610 struct ocfs2_journal
*journal
= NULL
;
612 struct inode
*inode
= NULL
;
613 int num_running_trans
= 0;
619 journal
= osb
->journal
;
623 inode
= journal
->j_inode
;
625 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
628 /* need to inc inode use count as journal_destroy will iput. */
632 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
633 if (num_running_trans
> 0)
634 mlog(0, "Shutting down journal: must wait on %d "
635 "running transactions!\n",
638 /* Do a commit_cache here. It will flush our journal, *and*
639 * release any locks that are still held.
640 * set the SHUTDOWN flag and release the trans lock.
641 * the commit thread will take the trans lock for us below. */
642 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
644 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
645 * drop the trans_lock (which we want to hold until we
646 * completely destroy the journal. */
647 if (osb
->commit_task
) {
648 /* Wait for the commit thread */
649 mlog(0, "Waiting for ocfs2commit to exit....\n");
650 kthread_stop(osb
->commit_task
);
651 osb
->commit_task
= NULL
;
654 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
656 status
= ocfs2_journal_toggle_dirty(osb
, 0);
660 /* Shutdown the kernel journal system */
661 journal_destroy(journal
->j_journal
);
663 OCFS2_I(inode
)->ip_open_count
--;
665 /* unlock our journal */
666 ocfs2_meta_unlock(inode
, 1);
668 brelse(journal
->j_bh
);
669 journal
->j_bh
= NULL
;
671 journal
->j_state
= OCFS2_JOURNAL_FREE
;
673 // up_write(&journal->j_trans_barrier);
680 static void ocfs2_clear_journal_error(struct super_block
*sb
,
686 olderr
= journal_errno(journal
);
688 mlog(ML_ERROR
, "File system error %d recorded in "
689 "journal %u.\n", olderr
, slot
);
690 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
693 journal_ack_err(journal
);
694 journal_clear_err(journal
);
698 int ocfs2_journal_load(struct ocfs2_journal
*journal
)
701 struct ocfs2_super
*osb
;
708 osb
= journal
->j_osb
;
710 status
= journal_load(journal
->j_journal
);
712 mlog(ML_ERROR
, "Failed to load journal!\n");
716 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
718 status
= ocfs2_journal_toggle_dirty(osb
, 1);
724 /* Launch the commit thread */
725 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
, "ocfs2cmt");
726 if (IS_ERR(osb
->commit_task
)) {
727 status
= PTR_ERR(osb
->commit_task
);
728 osb
->commit_task
= NULL
;
729 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, error=%d",
740 /* 'full' flag tells us whether we clear out all blocks or if we just
741 * mark the journal clean */
742 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
750 status
= journal_wipe(journal
->j_journal
, full
);
756 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0);
766 * JBD Might read a cached version of another nodes journal file. We
767 * don't want this as this file changes often and we get no
768 * notification on those changes. The only way to be sure that we've
769 * got the most up to date version of those blocks then is to force
770 * read them off disk. Just searching through the buffer cache won't
771 * work as there may be pages backing this file which are still marked
772 * up to date. We know things can't change on this file underneath us
773 * as we have the lock by now :)
775 static int ocfs2_force_read_journal(struct inode
*inode
)
779 u64 v_blkno
, p_blkno
;
780 #define CONCURRENT_JOURNAL_FILL 32
781 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
785 BUG_ON(inode
->i_blocks
!=
786 ocfs2_align_bytes_to_sectors(i_size_read(inode
)));
788 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
790 mlog(0, "Force reading %llu blocks\n",
791 (unsigned long long)(inode
->i_blocks
>>
792 (inode
->i_sb
->s_blocksize_bits
- 9)));
796 (inode
->i_blocks
>> (inode
->i_sb
->s_blocksize_bits
- 9))) {
798 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
806 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
807 p_blocks
= CONCURRENT_JOURNAL_FILL
;
809 /* We are reading journal data which should not
810 * be put in the uptodate cache */
811 status
= ocfs2_read_blocks(OCFS2_SB(inode
->i_sb
),
812 p_blkno
, p_blocks
, bhs
, 0,
819 for(i
= 0; i
< p_blocks
; i
++) {
828 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
835 struct ocfs2_la_recovery_item
{
836 struct list_head lri_list
;
838 struct ocfs2_dinode
*lri_la_dinode
;
839 struct ocfs2_dinode
*lri_tl_dinode
;
842 /* Does the second half of the recovery process. By this point, the
843 * node is marked clean and can actually be considered recovered,
844 * hence it's no longer in the recovery map, but there's still some
845 * cleanup we can do which shouldn't happen within the recovery thread
846 * as locking in that context becomes very difficult if we are to take
847 * recovering nodes into account.
849 * NOTE: This function can and will sleep on recovery of other nodes
850 * during cluster locking, just like any other ocfs2 process.
852 void ocfs2_complete_recovery(void *data
)
855 struct ocfs2_super
*osb
= data
;
856 struct ocfs2_journal
*journal
= osb
->journal
;
857 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
858 struct ocfs2_la_recovery_item
*item
;
859 struct list_head
*p
, *n
;
860 LIST_HEAD(tmp_la_list
);
864 mlog(0, "completing recovery from keventd\n");
866 spin_lock(&journal
->j_lock
);
867 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
868 spin_unlock(&journal
->j_lock
);
870 list_for_each_safe(p
, n
, &tmp_la_list
) {
871 item
= list_entry(p
, struct ocfs2_la_recovery_item
, lri_list
);
872 list_del_init(&item
->lri_list
);
874 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
876 la_dinode
= item
->lri_la_dinode
;
878 mlog(0, "Clean up local alloc %llu\n",
879 (unsigned long long)la_dinode
->i_blkno
);
881 ret
= ocfs2_complete_local_alloc_recovery(osb
,
889 tl_dinode
= item
->lri_tl_dinode
;
891 mlog(0, "Clean up truncate log %llu\n",
892 (unsigned long long)tl_dinode
->i_blkno
);
894 ret
= ocfs2_complete_truncate_log_recovery(osb
,
902 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
909 mlog(0, "Recovery completion\n");
913 /* NOTE: This function always eats your references to la_dinode and
914 * tl_dinode, either manually on error, or by passing them to
915 * ocfs2_complete_recovery */
916 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
918 struct ocfs2_dinode
*la_dinode
,
919 struct ocfs2_dinode
*tl_dinode
)
921 struct ocfs2_la_recovery_item
*item
;
923 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
925 /* Though we wish to avoid it, we are in fact safe in
926 * skipping local alloc cleanup as fsck.ocfs2 is more
927 * than capable of reclaiming unused space. */
938 INIT_LIST_HEAD(&item
->lri_list
);
939 item
->lri_la_dinode
= la_dinode
;
940 item
->lri_slot
= slot_num
;
941 item
->lri_tl_dinode
= tl_dinode
;
943 spin_lock(&journal
->j_lock
);
944 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
945 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
946 spin_unlock(&journal
->j_lock
);
949 /* Called by the mount code to queue recovery the last part of
950 * recovery for it's own slot. */
951 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
953 struct ocfs2_journal
*journal
= osb
->journal
;
956 /* No need to queue up our truncate_log as regular
957 * cleanup will catch that. */
958 ocfs2_queue_recovery_completion(journal
,
960 osb
->local_alloc_copy
,
962 ocfs2_schedule_truncate_log_flush(osb
, 0);
964 osb
->local_alloc_copy
= NULL
;
969 static int __ocfs2_recovery_thread(void *arg
)
971 int status
, node_num
;
972 struct ocfs2_super
*osb
= arg
;
976 status
= ocfs2_wait_on_mount(osb
);
982 status
= ocfs2_super_lock(osb
, 1);
988 while(!ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
989 node_num
= ocfs2_node_map_first_set_bit(osb
,
991 if (node_num
== O2NM_INVALID_NODE_NUM
) {
992 mlog(0, "Out of nodes to recover.\n");
996 status
= ocfs2_recover_node(osb
, node_num
);
999 "Error %d recovering node %d on device (%u,%u)!\n",
1001 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1002 mlog(ML_ERROR
, "Volume requires unmount.\n");
1006 ocfs2_recovery_map_clear(osb
, node_num
);
1008 ocfs2_super_unlock(osb
, 1);
1010 /* We always run recovery on our own orphan dir - the dead
1011 * node(s) may have voted "no" on an inode delete earlier. A
1012 * revote is therefore required. */
1013 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
1017 mutex_lock(&osb
->recovery_lock
);
1019 !ocfs2_node_map_is_empty(osb
, &osb
->recovery_map
)) {
1020 mutex_unlock(&osb
->recovery_lock
);
1024 osb
->recovery_thread_task
= NULL
;
1025 mb(); /* sync with ocfs2_recovery_thread_running */
1026 wake_up(&osb
->recovery_event
);
1028 mutex_unlock(&osb
->recovery_lock
);
1031 /* no one is callint kthread_stop() for us so the kthread() api
1032 * requires that we call do_exit(). And it isn't exported, but
1033 * complete_and_exit() seems to be a minimal wrapper around it. */
1034 complete_and_exit(NULL
, status
);
1038 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
1040 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1041 node_num
, osb
->node_num
);
1043 mutex_lock(&osb
->recovery_lock
);
1044 if (osb
->disable_recovery
)
1047 /* People waiting on recovery will wait on
1048 * the recovery map to empty. */
1049 if (!ocfs2_recovery_map_set(osb
, node_num
))
1050 mlog(0, "node %d already be in recovery.\n", node_num
);
1052 mlog(0, "starting recovery thread...\n");
1054 if (osb
->recovery_thread_task
)
1057 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
1059 if (IS_ERR(osb
->recovery_thread_task
)) {
1060 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
1061 osb
->recovery_thread_task
= NULL
;
1065 mutex_unlock(&osb
->recovery_lock
);
1066 wake_up(&osb
->recovery_event
);
1071 /* Does the actual journal replay and marks the journal inode as
1072 * clean. Will only replay if the journal inode is marked dirty. */
1073 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
1080 struct inode
*inode
= NULL
;
1081 struct ocfs2_dinode
*fe
;
1082 journal_t
*journal
= NULL
;
1083 struct buffer_head
*bh
= NULL
;
1085 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1087 if (inode
== NULL
) {
1092 if (is_bad_inode(inode
)) {
1099 SET_INODE_JOURNAL(inode
);
1101 status
= ocfs2_meta_lock_full(inode
, NULL
, &bh
, 1,
1102 OCFS2_META_LOCK_RECOVERY
);
1104 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status
);
1105 if (status
!= -ERESTARTSYS
)
1106 mlog(ML_ERROR
, "Could not lock journal!\n");
1111 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
1113 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1115 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
1116 mlog(0, "No recovery required for node %d\n", node_num
);
1120 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
1122 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1124 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
1126 status
= ocfs2_force_read_journal(inode
);
1132 mlog(0, "calling journal_init_inode\n");
1133 journal
= journal_init_inode(inode
);
1134 if (journal
== NULL
) {
1135 mlog(ML_ERROR
, "Linux journal layer error\n");
1140 status
= journal_load(journal
);
1145 journal_destroy(journal
);
1149 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1151 /* wipe the journal */
1152 mlog(0, "flushing the journal.\n");
1153 journal_lock_updates(journal
);
1154 status
= journal_flush(journal
);
1155 journal_unlock_updates(journal
);
1159 /* This will mark the node clean */
1160 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1161 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1162 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1164 status
= ocfs2_write_block(osb
, bh
, inode
);
1171 journal_destroy(journal
);
1174 /* drop the lock on this nodes journal */
1176 ocfs2_meta_unlock(inode
, 1);
1189 * Do the most important parts of node recovery:
1190 * - Replay it's journal
1191 * - Stamp a clean local allocator file
1192 * - Stamp a clean truncate log
1193 * - Mark the node clean
1195 * If this function completes without error, a node in OCFS2 can be
1196 * said to have been safely recovered. As a result, failure during the
1197 * second part of a nodes recovery process (local alloc recovery) is
1198 * far less concerning.
1200 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1205 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1206 struct ocfs2_dinode
*la_copy
= NULL
;
1207 struct ocfs2_dinode
*tl_copy
= NULL
;
1209 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1210 node_num
, osb
->node_num
);
1212 mlog(0, "checking node %d\n", node_num
);
1214 /* Should not ever be called to recover ourselves -- in that
1215 * case we should've called ocfs2_journal_load instead. */
1216 BUG_ON(osb
->node_num
== node_num
);
1218 slot_num
= ocfs2_node_num_to_slot(si
, node_num
);
1219 if (slot_num
== OCFS2_INVALID_SLOT
) {
1221 mlog(0, "no slot for this node, so no recovery required.\n");
1225 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1227 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1233 /* Stamp a clean local alloc file AFTER recovering the journal... */
1234 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1240 /* An error from begin_truncate_log_recovery is not
1241 * serious enough to warrant halting the rest of
1243 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1247 /* Likewise, this would be a strange but ultimately not so
1248 * harmful place to get an error... */
1249 ocfs2_clear_slot(si
, slot_num
);
1250 status
= ocfs2_update_disk_slots(osb
, si
);
1254 /* This will kfree the memory pointed to by la_copy and tl_copy */
1255 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1265 /* Test node liveness by trylocking his journal. If we get the lock,
1266 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1267 * still alive (we couldn't get the lock) and < 0 on error. */
1268 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1272 struct inode
*inode
= NULL
;
1274 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1276 if (inode
== NULL
) {
1277 mlog(ML_ERROR
, "access error\n");
1281 if (is_bad_inode(inode
)) {
1282 mlog(ML_ERROR
, "access error (bad inode)\n");
1288 SET_INODE_JOURNAL(inode
);
1290 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1291 status
= ocfs2_meta_lock_full(inode
, NULL
, NULL
, 1, flags
);
1293 if (status
!= -EAGAIN
)
1298 ocfs2_meta_unlock(inode
, 1);
1306 /* Call this underneath ocfs2_super_lock. It also assumes that the
1307 * slot info struct has been updated from disk. */
1308 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1310 int status
, i
, node_num
;
1311 struct ocfs2_slot_info
*si
= osb
->slot_info
;
1313 /* This is called with the super block cluster lock, so we
1314 * know that the slot map can't change underneath us. */
1316 spin_lock(&si
->si_lock
);
1317 for(i
= 0; i
< si
->si_num_slots
; i
++) {
1318 if (i
== osb
->slot_num
)
1320 if (ocfs2_is_empty_slot(si
, i
))
1323 node_num
= si
->si_global_node_nums
[i
];
1324 if (ocfs2_node_map_test_bit(osb
, &osb
->recovery_map
, node_num
))
1326 spin_unlock(&si
->si_lock
);
1328 /* Ok, we have a slot occupied by another node which
1329 * is not in the recovery map. We trylock his journal
1330 * file here to test if he's alive. */
1331 status
= ocfs2_trylock_journal(osb
, i
);
1333 /* Since we're called from mount, we know that
1334 * the recovery thread can't race us on
1335 * setting / checking the recovery bits. */
1336 ocfs2_recovery_thread(osb
, node_num
);
1337 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1342 spin_lock(&si
->si_lock
);
1344 spin_unlock(&si
->si_lock
);
1352 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
1354 struct inode
**head
)
1357 struct inode
*orphan_dir_inode
= NULL
;
1359 unsigned long offset
, blk
, local
;
1360 struct buffer_head
*bh
= NULL
;
1361 struct ocfs2_dir_entry
*de
;
1362 struct super_block
*sb
= osb
->sb
;
1364 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
1365 ORPHAN_DIR_SYSTEM_INODE
,
1367 if (!orphan_dir_inode
) {
1373 mutex_lock(&orphan_dir_inode
->i_mutex
);
1374 status
= ocfs2_meta_lock(orphan_dir_inode
, NULL
, NULL
, 0);
1382 while(offset
< i_size_read(orphan_dir_inode
)) {
1383 blk
= offset
>> sb
->s_blocksize_bits
;
1385 bh
= ocfs2_bread(orphan_dir_inode
, blk
, &status
, 0);
1396 while(offset
< i_size_read(orphan_dir_inode
)
1397 && local
< sb
->s_blocksize
) {
1398 de
= (struct ocfs2_dir_entry
*) (bh
->b_data
+ local
);
1400 if (!ocfs2_check_dir_entry(orphan_dir_inode
,
1408 local
+= le16_to_cpu(de
->rec_len
);
1409 offset
+= le16_to_cpu(de
->rec_len
);
1411 /* I guess we silently fail on no inode? */
1412 if (!le64_to_cpu(de
->inode
))
1414 if (de
->file_type
> OCFS2_FT_MAX
) {
1416 "block %llu contains invalid de: "
1417 "inode = %llu, rec_len = %u, "
1418 "name_len = %u, file_type = %u, "
1420 (unsigned long long)bh
->b_blocknr
,
1421 (unsigned long long)le64_to_cpu(de
->inode
),
1422 le16_to_cpu(de
->rec_len
),
1429 if (de
->name_len
== 1 && !strncmp(".", de
->name
, 1))
1431 if (de
->name_len
== 2 && !strncmp("..", de
->name
, 2))
1434 iter
= ocfs2_iget(osb
, le64_to_cpu(de
->inode
),
1435 OCFS2_FI_FLAG_NOLOCK
);
1439 mlog(0, "queue orphan %llu\n",
1440 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
1441 /* No locking is required for the next_orphan
1442 * queue as there is only ever a single
1443 * process doing orphan recovery. */
1444 OCFS2_I(iter
)->ip_next_orphan
= *head
;
1451 ocfs2_meta_unlock(orphan_dir_inode
, 0);
1453 mutex_unlock(&orphan_dir_inode
->i_mutex
);
1454 iput(orphan_dir_inode
);
1458 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
1463 spin_lock(&osb
->osb_lock
);
1464 ret
= !osb
->osb_orphan_wipes
[slot
];
1465 spin_unlock(&osb
->osb_lock
);
1469 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
1472 spin_lock(&osb
->osb_lock
);
1473 /* Mark ourselves such that new processes in delete_inode()
1474 * know to quit early. */
1475 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1476 while (osb
->osb_orphan_wipes
[slot
]) {
1477 /* If any processes are already in the middle of an
1478 * orphan wipe on this dir, then we need to wait for
1480 spin_unlock(&osb
->osb_lock
);
1481 wait_event_interruptible(osb
->osb_wipe_event
,
1482 ocfs2_orphan_recovery_can_continue(osb
, slot
));
1483 spin_lock(&osb
->osb_lock
);
1485 spin_unlock(&osb
->osb_lock
);
1488 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
1491 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
1495 * Orphan recovery. Each mounted node has it's own orphan dir which we
1496 * must run during recovery. Our strategy here is to build a list of
1497 * the inodes in the orphan dir and iget/iput them. The VFS does
1498 * (most) of the rest of the work.
1500 * Orphan recovery can happen at any time, not just mount so we have a
1501 * couple of extra considerations.
1503 * - We grab as many inodes as we can under the orphan dir lock -
1504 * doing iget() outside the orphan dir risks getting a reference on
1506 * - We must be sure not to deadlock with other processes on the
1507 * system wanting to run delete_inode(). This can happen when they go
1508 * to lock the orphan dir and the orphan recovery process attempts to
1509 * iget() inside the orphan dir lock. This can be avoided by
1510 * advertising our state to ocfs2_delete_inode().
1512 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
1516 struct inode
*inode
= NULL
;
1518 struct ocfs2_inode_info
*oi
;
1520 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
1522 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
1523 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
1524 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
1526 /* Error here should be noted, but we want to continue with as
1527 * many queued inodes as we've got. */
1532 oi
= OCFS2_I(inode
);
1533 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
1535 iter
= oi
->ip_next_orphan
;
1537 spin_lock(&oi
->ip_lock
);
1538 /* Delete voting may have set these on the assumption
1539 * that the other node would wipe them successfully.
1540 * If they are still in the node's orphan dir, we need
1541 * to reset that state. */
1542 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
1544 /* Set the proper information to get us going into
1545 * ocfs2_delete_inode. */
1546 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
1547 oi
->ip_orphaned_slot
= slot
;
1548 spin_unlock(&oi
->ip_lock
);
1558 static int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
1560 /* This check is good because ocfs2 will wait on our recovery
1561 * thread before changing it to something other than MOUNTED
1563 wait_event(osb
->osb_mount_event
,
1564 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
||
1565 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
1567 /* If there's an error on mount, then we may never get to the
1568 * MOUNTED flag, but this is set right before
1569 * dismount_volume() so we can trust it. */
1570 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
1571 mlog(0, "mount error, exiting!\n");
1578 static int ocfs2_commit_thread(void *arg
)
1581 struct ocfs2_super
*osb
= arg
;
1582 struct ocfs2_journal
*journal
= osb
->journal
;
1584 /* we can trust j_num_trans here because _should_stop() is only set in
1585 * shutdown and nobody other than ourselves should be able to start
1586 * transactions. committing on shutdown might take a few iterations
1587 * as final transactions put deleted inodes on the list */
1588 while (!(kthread_should_stop() &&
1589 atomic_read(&journal
->j_num_trans
) == 0)) {
1591 wait_event_interruptible(osb
->checkpoint_event
,
1592 atomic_read(&journal
->j_num_trans
)
1593 || kthread_should_stop());
1595 status
= ocfs2_commit_cache(osb
);
1599 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
1601 "commit_thread: %u transactions pending on "
1603 atomic_read(&journal
->j_num_trans
));
1610 /* Look for a dirty journal without taking any cluster locks. Used for
1611 * hard readonly access to determine whether the file system journals
1612 * require recovery. */
1613 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
1617 struct buffer_head
*di_bh
;
1618 struct ocfs2_dinode
*di
;
1619 struct inode
*journal
= NULL
;
1621 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
1622 journal
= ocfs2_get_system_file_inode(osb
,
1623 JOURNAL_SYSTEM_INODE
,
1625 if (!journal
|| is_bad_inode(journal
)) {
1632 ret
= ocfs2_read_block(osb
, OCFS2_I(journal
)->ip_blkno
, &di_bh
,
1639 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
1641 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
1642 OCFS2_JOURNAL_DIRTY_FL
)