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>
31 #include <linux/time.h>
32 #include <linux/random.h>
34 #define MLOG_MASK_PREFIX ML_JOURNAL
35 #include <cluster/masklog.h>
40 #include "blockcheck.h"
43 #include "extent_map.h"
44 #include "heartbeat.h"
47 #include "localalloc.h"
54 #include "buffer_head_io.h"
56 DEFINE_SPINLOCK(trans_inc_lock
);
58 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
60 static int ocfs2_force_read_journal(struct inode
*inode
);
61 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
62 int node_num
, int slot_num
);
63 static int __ocfs2_recovery_thread(void *arg
);
64 static int ocfs2_commit_cache(struct ocfs2_super
*osb
);
65 static int __ocfs2_wait_on_mount(struct ocfs2_super
*osb
, int quota
);
66 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
67 int dirty
, int replayed
);
68 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
70 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
72 static int ocfs2_commit_thread(void *arg
);
73 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
75 struct ocfs2_dinode
*la_dinode
,
76 struct ocfs2_dinode
*tl_dinode
,
77 struct ocfs2_quota_recovery
*qrec
);
79 static inline int ocfs2_wait_on_mount(struct ocfs2_super
*osb
)
81 return __ocfs2_wait_on_mount(osb
, 0);
84 static inline int ocfs2_wait_on_quotas(struct ocfs2_super
*osb
)
86 return __ocfs2_wait_on_mount(osb
, 1);
90 * This replay_map is to track online/offline slots, so we could recover
91 * offline slots during recovery and mount
94 enum ocfs2_replay_state
{
95 REPLAY_UNNEEDED
= 0, /* Replay is not needed, so ignore this map */
96 REPLAY_NEEDED
, /* Replay slots marked in rm_replay_slots */
97 REPLAY_DONE
/* Replay was already queued */
100 struct ocfs2_replay_map
{
101 unsigned int rm_slots
;
102 enum ocfs2_replay_state rm_state
;
103 unsigned char rm_replay_slots
[0];
106 void ocfs2_replay_map_set_state(struct ocfs2_super
*osb
, int state
)
108 if (!osb
->replay_map
)
111 /* If we've already queued the replay, we don't have any more to do */
112 if (osb
->replay_map
->rm_state
== REPLAY_DONE
)
115 osb
->replay_map
->rm_state
= state
;
118 int ocfs2_compute_replay_slots(struct ocfs2_super
*osb
)
120 struct ocfs2_replay_map
*replay_map
;
123 /* If replay map is already set, we don't do it again */
127 replay_map
= kzalloc(sizeof(struct ocfs2_replay_map
) +
128 (osb
->max_slots
* sizeof(char)), GFP_KERNEL
);
135 spin_lock(&osb
->osb_lock
);
137 replay_map
->rm_slots
= osb
->max_slots
;
138 replay_map
->rm_state
= REPLAY_UNNEEDED
;
140 /* set rm_replay_slots for offline slot(s) */
141 for (i
= 0; i
< replay_map
->rm_slots
; i
++) {
142 if (ocfs2_slot_to_node_num_locked(osb
, i
, &node_num
) == -ENOENT
)
143 replay_map
->rm_replay_slots
[i
] = 1;
146 osb
->replay_map
= replay_map
;
147 spin_unlock(&osb
->osb_lock
);
151 void ocfs2_queue_replay_slots(struct ocfs2_super
*osb
)
153 struct ocfs2_replay_map
*replay_map
= osb
->replay_map
;
159 if (replay_map
->rm_state
!= REPLAY_NEEDED
)
162 for (i
= 0; i
< replay_map
->rm_slots
; i
++)
163 if (replay_map
->rm_replay_slots
[i
])
164 ocfs2_queue_recovery_completion(osb
->journal
, i
, NULL
,
166 replay_map
->rm_state
= REPLAY_DONE
;
169 void ocfs2_free_replay_slots(struct ocfs2_super
*osb
)
171 struct ocfs2_replay_map
*replay_map
= osb
->replay_map
;
173 if (!osb
->replay_map
)
177 osb
->replay_map
= NULL
;
180 int ocfs2_recovery_init(struct ocfs2_super
*osb
)
182 struct ocfs2_recovery_map
*rm
;
184 mutex_init(&osb
->recovery_lock
);
185 osb
->disable_recovery
= 0;
186 osb
->recovery_thread_task
= NULL
;
187 init_waitqueue_head(&osb
->recovery_event
);
189 rm
= kzalloc(sizeof(struct ocfs2_recovery_map
) +
190 osb
->max_slots
* sizeof(unsigned int),
197 rm
->rm_entries
= (unsigned int *)((char *)rm
+
198 sizeof(struct ocfs2_recovery_map
));
199 osb
->recovery_map
= rm
;
204 /* we can't grab the goofy sem lock from inside wait_event, so we use
205 * memory barriers to make sure that we'll see the null task before
207 static int ocfs2_recovery_thread_running(struct ocfs2_super
*osb
)
210 return osb
->recovery_thread_task
!= NULL
;
213 void ocfs2_recovery_exit(struct ocfs2_super
*osb
)
215 struct ocfs2_recovery_map
*rm
;
217 /* disable any new recovery threads and wait for any currently
218 * running ones to exit. Do this before setting the vol_state. */
219 mutex_lock(&osb
->recovery_lock
);
220 osb
->disable_recovery
= 1;
221 mutex_unlock(&osb
->recovery_lock
);
222 wait_event(osb
->recovery_event
, !ocfs2_recovery_thread_running(osb
));
224 /* At this point, we know that no more recovery threads can be
225 * launched, so wait for any recovery completion work to
227 flush_workqueue(ocfs2_wq
);
230 * Now that recovery is shut down, and the osb is about to be
231 * freed, the osb_lock is not taken here.
233 rm
= osb
->recovery_map
;
238 static int __ocfs2_recovery_map_test(struct ocfs2_super
*osb
,
239 unsigned int node_num
)
242 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
244 assert_spin_locked(&osb
->osb_lock
);
246 for (i
= 0; i
< rm
->rm_used
; i
++) {
247 if (rm
->rm_entries
[i
] == node_num
)
254 /* Behaves like test-and-set. Returns the previous value */
255 static int ocfs2_recovery_map_set(struct ocfs2_super
*osb
,
256 unsigned int node_num
)
258 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
260 spin_lock(&osb
->osb_lock
);
261 if (__ocfs2_recovery_map_test(osb
, node_num
)) {
262 spin_unlock(&osb
->osb_lock
);
266 BUG_ON(rm
->rm_used
>= osb
->max_slots
);
268 rm
->rm_entries
[rm
->rm_used
] = node_num
;
270 spin_unlock(&osb
->osb_lock
);
275 static void ocfs2_recovery_map_clear(struct ocfs2_super
*osb
,
276 unsigned int node_num
)
279 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
281 spin_lock(&osb
->osb_lock
);
283 for (i
= 0; i
< rm
->rm_used
; i
++) {
284 if (rm
->rm_entries
[i
] == node_num
)
288 if (i
< rm
->rm_used
) {
289 memmove(&(rm
->rm_entries
[i
]), &(rm
->rm_entries
[i
+ 1]),
290 (rm
->rm_used
- i
- 1) * sizeof(unsigned int));
294 spin_unlock(&osb
->osb_lock
);
297 static int ocfs2_commit_cache(struct ocfs2_super
*osb
)
300 unsigned int flushed
;
301 unsigned long old_id
;
302 struct ocfs2_journal
*journal
= NULL
;
306 journal
= osb
->journal
;
308 /* Flush all pending commits and checkpoint the journal. */
309 down_write(&journal
->j_trans_barrier
);
311 if (atomic_read(&journal
->j_num_trans
) == 0) {
312 up_write(&journal
->j_trans_barrier
);
313 mlog(0, "No transactions for me to flush!\n");
317 jbd2_journal_lock_updates(journal
->j_journal
);
318 status
= jbd2_journal_flush(journal
->j_journal
);
319 jbd2_journal_unlock_updates(journal
->j_journal
);
321 up_write(&journal
->j_trans_barrier
);
326 old_id
= ocfs2_inc_trans_id(journal
);
328 flushed
= atomic_read(&journal
->j_num_trans
);
329 atomic_set(&journal
->j_num_trans
, 0);
330 up_write(&journal
->j_trans_barrier
);
332 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
333 journal
->j_trans_id
, flushed
);
335 ocfs2_wake_downconvert_thread(osb
);
336 wake_up(&journal
->j_checkpointed
);
342 /* pass it NULL and it will allocate a new handle object for you. If
343 * you pass it a handle however, it may still return error, in which
344 * case it has free'd the passed handle for you. */
345 handle_t
*ocfs2_start_trans(struct ocfs2_super
*osb
, int max_buffs
)
347 journal_t
*journal
= osb
->journal
->j_journal
;
350 BUG_ON(!osb
|| !osb
->journal
->j_journal
);
352 if (ocfs2_is_hard_readonly(osb
))
353 return ERR_PTR(-EROFS
);
355 BUG_ON(osb
->journal
->j_state
== OCFS2_JOURNAL_FREE
);
356 BUG_ON(max_buffs
<= 0);
358 /* Nested transaction? Just return the handle... */
359 if (journal_current_handle())
360 return jbd2_journal_start(journal
, max_buffs
);
362 down_read(&osb
->journal
->j_trans_barrier
);
364 handle
= jbd2_journal_start(journal
, max_buffs
);
365 if (IS_ERR(handle
)) {
366 up_read(&osb
->journal
->j_trans_barrier
);
368 mlog_errno(PTR_ERR(handle
));
370 if (is_journal_aborted(journal
)) {
371 ocfs2_abort(osb
->sb
, "Detected aborted journal");
372 handle
= ERR_PTR(-EROFS
);
375 if (!ocfs2_mount_local(osb
))
376 atomic_inc(&(osb
->journal
->j_num_trans
));
382 int ocfs2_commit_trans(struct ocfs2_super
*osb
,
386 struct ocfs2_journal
*journal
= osb
->journal
;
390 nested
= handle
->h_ref
> 1;
391 ret
= jbd2_journal_stop(handle
);
396 up_read(&journal
->j_trans_barrier
);
402 * 'nblocks' is what you want to add to the current transaction.
404 * This might call jbd2_journal_restart() which will commit dirty buffers
405 * and then restart the transaction. Before calling
406 * ocfs2_extend_trans(), any changed blocks should have been
407 * dirtied. After calling it, all blocks which need to be changed must
408 * go through another set of journal_access/journal_dirty calls.
410 * WARNING: This will not release any semaphores or disk locks taken
411 * during the transaction, so make sure they were taken *before*
412 * start_trans or we'll have ordering deadlocks.
414 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
415 * good because transaction ids haven't yet been recorded on the
416 * cluster locks associated with this handle.
418 int ocfs2_extend_trans(handle_t
*handle
, int nblocks
)
420 int status
, old_nblocks
;
428 old_nblocks
= handle
->h_buffer_credits
;
431 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks
);
433 #ifdef CONFIG_OCFS2_DEBUG_FS
436 status
= jbd2_journal_extend(handle
, nblocks
);
445 "jbd2_journal_extend failed, trying "
446 "jbd2_journal_restart\n");
447 status
= jbd2_journal_restart(handle
,
448 old_nblocks
+ nblocks
);
462 struct ocfs2_triggers
{
463 struct jbd2_buffer_trigger_type ot_triggers
;
467 static inline struct ocfs2_triggers
*to_ocfs2_trigger(struct jbd2_buffer_trigger_type
*triggers
)
469 return container_of(triggers
, struct ocfs2_triggers
, ot_triggers
);
472 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type
*triggers
,
473 struct buffer_head
*bh
,
474 void *data
, size_t size
)
476 struct ocfs2_triggers
*ot
= to_ocfs2_trigger(triggers
);
479 * We aren't guaranteed to have the superblock here, so we
480 * must unconditionally compute the ecc data.
481 * __ocfs2_journal_access() will only set the triggers if
482 * metaecc is enabled.
484 ocfs2_block_check_compute(data
, size
, data
+ ot
->ot_offset
);
488 * Quota blocks have their own trigger because the struct ocfs2_block_check
489 * offset depends on the blocksize.
491 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type
*triggers
,
492 struct buffer_head
*bh
,
493 void *data
, size_t size
)
495 struct ocfs2_disk_dqtrailer
*dqt
=
496 ocfs2_block_dqtrailer(size
, data
);
499 * We aren't guaranteed to have the superblock here, so we
500 * must unconditionally compute the ecc data.
501 * __ocfs2_journal_access() will only set the triggers if
502 * metaecc is enabled.
504 ocfs2_block_check_compute(data
, size
, &dqt
->dq_check
);
508 * Directory blocks also have their own trigger because the
509 * struct ocfs2_block_check offset depends on the blocksize.
511 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type
*triggers
,
512 struct buffer_head
*bh
,
513 void *data
, size_t size
)
515 struct ocfs2_dir_block_trailer
*trailer
=
516 ocfs2_dir_trailer_from_size(size
, data
);
519 * We aren't guaranteed to have the superblock here, so we
520 * must unconditionally compute the ecc data.
521 * __ocfs2_journal_access() will only set the triggers if
522 * metaecc is enabled.
524 ocfs2_block_check_compute(data
, size
, &trailer
->db_check
);
527 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type
*triggers
,
528 struct buffer_head
*bh
)
531 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
532 "bh->b_blocknr = %llu\n",
534 (unsigned long long)bh
->b_blocknr
);
536 /* We aren't guaranteed to have the superblock here - but if we
537 * don't, it'll just crash. */
538 ocfs2_error(bh
->b_assoc_map
->host
->i_sb
,
539 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
542 static struct ocfs2_triggers di_triggers
= {
544 .t_frozen
= ocfs2_frozen_trigger
,
545 .t_abort
= ocfs2_abort_trigger
,
547 .ot_offset
= offsetof(struct ocfs2_dinode
, i_check
),
550 static struct ocfs2_triggers eb_triggers
= {
552 .t_frozen
= ocfs2_frozen_trigger
,
553 .t_abort
= ocfs2_abort_trigger
,
555 .ot_offset
= offsetof(struct ocfs2_extent_block
, h_check
),
558 static struct ocfs2_triggers rb_triggers
= {
560 .t_frozen
= ocfs2_frozen_trigger
,
561 .t_abort
= ocfs2_abort_trigger
,
563 .ot_offset
= offsetof(struct ocfs2_refcount_block
, rf_check
),
566 static struct ocfs2_triggers gd_triggers
= {
568 .t_frozen
= ocfs2_frozen_trigger
,
569 .t_abort
= ocfs2_abort_trigger
,
571 .ot_offset
= offsetof(struct ocfs2_group_desc
, bg_check
),
574 static struct ocfs2_triggers db_triggers
= {
576 .t_frozen
= ocfs2_db_frozen_trigger
,
577 .t_abort
= ocfs2_abort_trigger
,
581 static struct ocfs2_triggers xb_triggers
= {
583 .t_frozen
= ocfs2_frozen_trigger
,
584 .t_abort
= ocfs2_abort_trigger
,
586 .ot_offset
= offsetof(struct ocfs2_xattr_block
, xb_check
),
589 static struct ocfs2_triggers dq_triggers
= {
591 .t_frozen
= ocfs2_dq_frozen_trigger
,
592 .t_abort
= ocfs2_abort_trigger
,
596 static struct ocfs2_triggers dr_triggers
= {
598 .t_frozen
= ocfs2_frozen_trigger
,
599 .t_abort
= ocfs2_abort_trigger
,
601 .ot_offset
= offsetof(struct ocfs2_dx_root_block
, dr_check
),
604 static struct ocfs2_triggers dl_triggers
= {
606 .t_frozen
= ocfs2_frozen_trigger
,
607 .t_abort
= ocfs2_abort_trigger
,
609 .ot_offset
= offsetof(struct ocfs2_dx_leaf
, dl_check
),
612 static int __ocfs2_journal_access(handle_t
*handle
,
613 struct ocfs2_caching_info
*ci
,
614 struct buffer_head
*bh
,
615 struct ocfs2_triggers
*triggers
,
619 struct ocfs2_super
*osb
=
620 OCFS2_SB(ocfs2_metadata_cache_get_super(ci
));
622 BUG_ON(!ci
|| !ci
->ci_ops
);
626 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
627 (unsigned long long)bh
->b_blocknr
, type
,
628 (type
== OCFS2_JOURNAL_ACCESS_CREATE
) ?
629 "OCFS2_JOURNAL_ACCESS_CREATE" :
630 "OCFS2_JOURNAL_ACCESS_WRITE",
633 /* we can safely remove this assertion after testing. */
634 if (!buffer_uptodate(bh
)) {
635 mlog(ML_ERROR
, "giving me a buffer that's not uptodate!\n");
636 mlog(ML_ERROR
, "b_blocknr=%llu\n",
637 (unsigned long long)bh
->b_blocknr
);
641 /* Set the current transaction information on the ci so
642 * that the locking code knows whether it can drop it's locks
643 * on this ci or not. We're protected from the commit
644 * thread updating the current transaction id until
645 * ocfs2_commit_trans() because ocfs2_start_trans() took
646 * j_trans_barrier for us. */
647 ocfs2_set_ci_lock_trans(osb
->journal
, ci
);
649 ocfs2_metadata_cache_io_lock(ci
);
651 case OCFS2_JOURNAL_ACCESS_CREATE
:
652 case OCFS2_JOURNAL_ACCESS_WRITE
:
653 status
= jbd2_journal_get_write_access(handle
, bh
);
656 case OCFS2_JOURNAL_ACCESS_UNDO
:
657 status
= jbd2_journal_get_undo_access(handle
, bh
);
662 mlog(ML_ERROR
, "Unknown access type!\n");
664 if (!status
&& ocfs2_meta_ecc(osb
) && triggers
)
665 jbd2_journal_set_triggers(bh
, &triggers
->ot_triggers
);
666 ocfs2_metadata_cache_io_unlock(ci
);
669 mlog(ML_ERROR
, "Error %d getting %d access to buffer!\n",
676 int ocfs2_journal_access_di(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
677 struct buffer_head
*bh
, int type
)
679 return __ocfs2_journal_access(handle
, ci
, bh
, &di_triggers
, type
);
682 int ocfs2_journal_access_eb(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
683 struct buffer_head
*bh
, int type
)
685 return __ocfs2_journal_access(handle
, ci
, bh
, &eb_triggers
, type
);
688 int ocfs2_journal_access_rb(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
689 struct buffer_head
*bh
, int type
)
691 return __ocfs2_journal_access(handle
, ci
, bh
, &rb_triggers
,
695 int ocfs2_journal_access_gd(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
696 struct buffer_head
*bh
, int type
)
698 return __ocfs2_journal_access(handle
, ci
, bh
, &gd_triggers
, type
);
701 int ocfs2_journal_access_db(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
702 struct buffer_head
*bh
, int type
)
704 return __ocfs2_journal_access(handle
, ci
, bh
, &db_triggers
, type
);
707 int ocfs2_journal_access_xb(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
708 struct buffer_head
*bh
, int type
)
710 return __ocfs2_journal_access(handle
, ci
, bh
, &xb_triggers
, type
);
713 int ocfs2_journal_access_dq(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
714 struct buffer_head
*bh
, int type
)
716 return __ocfs2_journal_access(handle
, ci
, bh
, &dq_triggers
, type
);
719 int ocfs2_journal_access_dr(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
720 struct buffer_head
*bh
, int type
)
722 return __ocfs2_journal_access(handle
, ci
, bh
, &dr_triggers
, type
);
725 int ocfs2_journal_access_dl(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
726 struct buffer_head
*bh
, int type
)
728 return __ocfs2_journal_access(handle
, ci
, bh
, &dl_triggers
, type
);
731 int ocfs2_journal_access(handle_t
*handle
, struct ocfs2_caching_info
*ci
,
732 struct buffer_head
*bh
, int type
)
734 return __ocfs2_journal_access(handle
, ci
, bh
, NULL
, type
);
737 void ocfs2_journal_dirty(handle_t
*handle
, struct buffer_head
*bh
)
741 mlog_entry("(bh->b_blocknr=%llu)\n",
742 (unsigned long long)bh
->b_blocknr
);
744 status
= jbd2_journal_dirty_metadata(handle
, bh
);
750 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
752 void ocfs2_set_journal_params(struct ocfs2_super
*osb
)
754 journal_t
*journal
= osb
->journal
->j_journal
;
755 unsigned long commit_interval
= OCFS2_DEFAULT_COMMIT_INTERVAL
;
757 if (osb
->osb_commit_interval
)
758 commit_interval
= osb
->osb_commit_interval
;
760 write_lock(&journal
->j_state_lock
);
761 journal
->j_commit_interval
= commit_interval
;
762 if (osb
->s_mount_opt
& OCFS2_MOUNT_BARRIER
)
763 journal
->j_flags
|= JBD2_BARRIER
;
765 journal
->j_flags
&= ~JBD2_BARRIER
;
766 write_unlock(&journal
->j_state_lock
);
769 int ocfs2_journal_init(struct ocfs2_journal
*journal
, int *dirty
)
772 struct inode
*inode
= NULL
; /* the journal inode */
773 journal_t
*j_journal
= NULL
;
774 struct ocfs2_dinode
*di
= NULL
;
775 struct buffer_head
*bh
= NULL
;
776 struct ocfs2_super
*osb
;
783 osb
= journal
->j_osb
;
785 /* already have the inode for our journal */
786 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
793 if (is_bad_inode(inode
)) {
794 mlog(ML_ERROR
, "access error (bad inode)\n");
801 SET_INODE_JOURNAL(inode
);
802 OCFS2_I(inode
)->ip_open_count
++;
804 /* Skip recovery waits here - journal inode metadata never
805 * changes in a live cluster so it can be considered an
806 * exception to the rule. */
807 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
809 if (status
!= -ERESTARTSYS
)
810 mlog(ML_ERROR
, "Could not get lock on journal!\n");
815 di
= (struct ocfs2_dinode
*)bh
->b_data
;
817 if (inode
->i_size
< OCFS2_MIN_JOURNAL_SIZE
) {
818 mlog(ML_ERROR
, "Journal file size (%lld) is too small!\n",
824 mlog(0, "inode->i_size = %lld\n", inode
->i_size
);
825 mlog(0, "inode->i_blocks = %llu\n",
826 (unsigned long long)inode
->i_blocks
);
827 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode
)->ip_clusters
);
829 /* call the kernels journal init function now */
830 j_journal
= jbd2_journal_init_inode(inode
);
831 if (j_journal
== NULL
) {
832 mlog(ML_ERROR
, "Linux journal layer error\n");
837 mlog(0, "Returned from jbd2_journal_init_inode\n");
838 mlog(0, "j_journal->j_maxlen = %u\n", j_journal
->j_maxlen
);
840 *dirty
= (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
841 OCFS2_JOURNAL_DIRTY_FL
);
843 journal
->j_journal
= j_journal
;
844 journal
->j_inode
= inode
;
847 ocfs2_set_journal_params(osb
);
849 journal
->j_state
= OCFS2_JOURNAL_LOADED
;
855 ocfs2_inode_unlock(inode
, 1);
858 OCFS2_I(inode
)->ip_open_count
--;
867 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode
*di
)
869 le32_add_cpu(&(di
->id1
.journal1
.ij_recovery_generation
), 1);
872 static u32
ocfs2_get_recovery_generation(struct ocfs2_dinode
*di
)
874 return le32_to_cpu(di
->id1
.journal1
.ij_recovery_generation
);
877 static int ocfs2_journal_toggle_dirty(struct ocfs2_super
*osb
,
878 int dirty
, int replayed
)
882 struct ocfs2_journal
*journal
= osb
->journal
;
883 struct buffer_head
*bh
= journal
->j_bh
;
884 struct ocfs2_dinode
*fe
;
888 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
890 /* The journal bh on the osb always comes from ocfs2_journal_init()
891 * and was validated there inside ocfs2_inode_lock_full(). It's a
892 * code bug if we mess it up. */
893 BUG_ON(!OCFS2_IS_VALID_DINODE(fe
));
895 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
897 flags
|= OCFS2_JOURNAL_DIRTY_FL
;
899 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
900 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
903 ocfs2_bump_recovery_generation(fe
);
905 ocfs2_compute_meta_ecc(osb
->sb
, bh
->b_data
, &fe
->i_check
);
906 status
= ocfs2_write_block(osb
, bh
, INODE_CACHE(journal
->j_inode
));
915 * If the journal has been kmalloc'd it needs to be freed after this
918 void ocfs2_journal_shutdown(struct ocfs2_super
*osb
)
920 struct ocfs2_journal
*journal
= NULL
;
922 struct inode
*inode
= NULL
;
923 int num_running_trans
= 0;
929 journal
= osb
->journal
;
933 inode
= journal
->j_inode
;
935 if (journal
->j_state
!= OCFS2_JOURNAL_LOADED
)
938 /* need to inc inode use count - jbd2_journal_destroy will iput. */
942 num_running_trans
= atomic_read(&(osb
->journal
->j_num_trans
));
943 if (num_running_trans
> 0)
944 mlog(0, "Shutting down journal: must wait on %d "
945 "running transactions!\n",
948 /* Do a commit_cache here. It will flush our journal, *and*
949 * release any locks that are still held.
950 * set the SHUTDOWN flag and release the trans lock.
951 * the commit thread will take the trans lock for us below. */
952 journal
->j_state
= OCFS2_JOURNAL_IN_SHUTDOWN
;
954 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
955 * drop the trans_lock (which we want to hold until we
956 * completely destroy the journal. */
957 if (osb
->commit_task
) {
958 /* Wait for the commit thread */
959 mlog(0, "Waiting for ocfs2commit to exit....\n");
960 kthread_stop(osb
->commit_task
);
961 osb
->commit_task
= NULL
;
964 BUG_ON(atomic_read(&(osb
->journal
->j_num_trans
)) != 0);
966 if (ocfs2_mount_local(osb
)) {
967 jbd2_journal_lock_updates(journal
->j_journal
);
968 status
= jbd2_journal_flush(journal
->j_journal
);
969 jbd2_journal_unlock_updates(journal
->j_journal
);
976 * Do not toggle if flush was unsuccessful otherwise
977 * will leave dirty metadata in a "clean" journal
979 status
= ocfs2_journal_toggle_dirty(osb
, 0, 0);
984 /* Shutdown the kernel journal system */
985 jbd2_journal_destroy(journal
->j_journal
);
986 journal
->j_journal
= NULL
;
988 OCFS2_I(inode
)->ip_open_count
--;
990 /* unlock our journal */
991 ocfs2_inode_unlock(inode
, 1);
993 brelse(journal
->j_bh
);
994 journal
->j_bh
= NULL
;
996 journal
->j_state
= OCFS2_JOURNAL_FREE
;
998 // up_write(&journal->j_trans_barrier);
1005 static void ocfs2_clear_journal_error(struct super_block
*sb
,
1011 olderr
= jbd2_journal_errno(journal
);
1013 mlog(ML_ERROR
, "File system error %d recorded in "
1014 "journal %u.\n", olderr
, slot
);
1015 mlog(ML_ERROR
, "File system on device %s needs checking.\n",
1018 jbd2_journal_ack_err(journal
);
1019 jbd2_journal_clear_err(journal
);
1023 int ocfs2_journal_load(struct ocfs2_journal
*journal
, int local
, int replayed
)
1026 struct ocfs2_super
*osb
;
1032 osb
= journal
->j_osb
;
1034 status
= jbd2_journal_load(journal
->j_journal
);
1036 mlog(ML_ERROR
, "Failed to load journal!\n");
1040 ocfs2_clear_journal_error(osb
->sb
, journal
->j_journal
, osb
->slot_num
);
1042 status
= ocfs2_journal_toggle_dirty(osb
, 1, replayed
);
1048 /* Launch the commit thread */
1050 osb
->commit_task
= kthread_run(ocfs2_commit_thread
, osb
,
1052 if (IS_ERR(osb
->commit_task
)) {
1053 status
= PTR_ERR(osb
->commit_task
);
1054 osb
->commit_task
= NULL
;
1055 mlog(ML_ERROR
, "unable to launch ocfs2commit thread, "
1056 "error=%d", status
);
1060 osb
->commit_task
= NULL
;
1068 /* 'full' flag tells us whether we clear out all blocks or if we just
1069 * mark the journal clean */
1070 int ocfs2_journal_wipe(struct ocfs2_journal
*journal
, int full
)
1078 status
= jbd2_journal_wipe(journal
->j_journal
, full
);
1084 status
= ocfs2_journal_toggle_dirty(journal
->j_osb
, 0, 0);
1093 static int ocfs2_recovery_completed(struct ocfs2_super
*osb
)
1096 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
1098 spin_lock(&osb
->osb_lock
);
1099 empty
= (rm
->rm_used
== 0);
1100 spin_unlock(&osb
->osb_lock
);
1105 void ocfs2_wait_for_recovery(struct ocfs2_super
*osb
)
1107 wait_event(osb
->recovery_event
, ocfs2_recovery_completed(osb
));
1111 * JBD Might read a cached version of another nodes journal file. We
1112 * don't want this as this file changes often and we get no
1113 * notification on those changes. The only way to be sure that we've
1114 * got the most up to date version of those blocks then is to force
1115 * read them off disk. Just searching through the buffer cache won't
1116 * work as there may be pages backing this file which are still marked
1117 * up to date. We know things can't change on this file underneath us
1118 * as we have the lock by now :)
1120 static int ocfs2_force_read_journal(struct inode
*inode
)
1124 u64 v_blkno
, p_blkno
, p_blocks
, num_blocks
;
1125 #define CONCURRENT_JOURNAL_FILL 32ULL
1126 struct buffer_head
*bhs
[CONCURRENT_JOURNAL_FILL
];
1130 memset(bhs
, 0, sizeof(struct buffer_head
*) * CONCURRENT_JOURNAL_FILL
);
1132 num_blocks
= ocfs2_blocks_for_bytes(inode
->i_sb
, inode
->i_size
);
1134 while (v_blkno
< num_blocks
) {
1135 status
= ocfs2_extent_map_get_blocks(inode
, v_blkno
,
1136 &p_blkno
, &p_blocks
, NULL
);
1142 if (p_blocks
> CONCURRENT_JOURNAL_FILL
)
1143 p_blocks
= CONCURRENT_JOURNAL_FILL
;
1145 /* We are reading journal data which should not
1146 * be put in the uptodate cache */
1147 status
= ocfs2_read_blocks_sync(OCFS2_SB(inode
->i_sb
),
1148 p_blkno
, p_blocks
, bhs
);
1154 for(i
= 0; i
< p_blocks
; i
++) {
1159 v_blkno
+= p_blocks
;
1163 for(i
= 0; i
< CONCURRENT_JOURNAL_FILL
; i
++)
1169 struct ocfs2_la_recovery_item
{
1170 struct list_head lri_list
;
1172 struct ocfs2_dinode
*lri_la_dinode
;
1173 struct ocfs2_dinode
*lri_tl_dinode
;
1174 struct ocfs2_quota_recovery
*lri_qrec
;
1177 /* Does the second half of the recovery process. By this point, the
1178 * node is marked clean and can actually be considered recovered,
1179 * hence it's no longer in the recovery map, but there's still some
1180 * cleanup we can do which shouldn't happen within the recovery thread
1181 * as locking in that context becomes very difficult if we are to take
1182 * recovering nodes into account.
1184 * NOTE: This function can and will sleep on recovery of other nodes
1185 * during cluster locking, just like any other ocfs2 process.
1187 void ocfs2_complete_recovery(struct work_struct
*work
)
1190 struct ocfs2_journal
*journal
=
1191 container_of(work
, struct ocfs2_journal
, j_recovery_work
);
1192 struct ocfs2_super
*osb
= journal
->j_osb
;
1193 struct ocfs2_dinode
*la_dinode
, *tl_dinode
;
1194 struct ocfs2_la_recovery_item
*item
, *n
;
1195 struct ocfs2_quota_recovery
*qrec
;
1196 LIST_HEAD(tmp_la_list
);
1200 mlog(0, "completing recovery from keventd\n");
1202 spin_lock(&journal
->j_lock
);
1203 list_splice_init(&journal
->j_la_cleanups
, &tmp_la_list
);
1204 spin_unlock(&journal
->j_lock
);
1206 list_for_each_entry_safe(item
, n
, &tmp_la_list
, lri_list
) {
1207 list_del_init(&item
->lri_list
);
1209 mlog(0, "Complete recovery for slot %d\n", item
->lri_slot
);
1211 ocfs2_wait_on_quotas(osb
);
1213 la_dinode
= item
->lri_la_dinode
;
1215 mlog(0, "Clean up local alloc %llu\n",
1216 (unsigned long long)le64_to_cpu(la_dinode
->i_blkno
));
1218 ret
= ocfs2_complete_local_alloc_recovery(osb
,
1226 tl_dinode
= item
->lri_tl_dinode
;
1228 mlog(0, "Clean up truncate log %llu\n",
1229 (unsigned long long)le64_to_cpu(tl_dinode
->i_blkno
));
1231 ret
= ocfs2_complete_truncate_log_recovery(osb
,
1239 ret
= ocfs2_recover_orphans(osb
, item
->lri_slot
);
1243 qrec
= item
->lri_qrec
;
1245 mlog(0, "Recovering quota files");
1246 ret
= ocfs2_finish_quota_recovery(osb
, qrec
,
1250 /* Recovery info is already freed now */
1256 mlog(0, "Recovery completion\n");
1260 /* NOTE: This function always eats your references to la_dinode and
1261 * tl_dinode, either manually on error, or by passing them to
1262 * ocfs2_complete_recovery */
1263 static void ocfs2_queue_recovery_completion(struct ocfs2_journal
*journal
,
1265 struct ocfs2_dinode
*la_dinode
,
1266 struct ocfs2_dinode
*tl_dinode
,
1267 struct ocfs2_quota_recovery
*qrec
)
1269 struct ocfs2_la_recovery_item
*item
;
1271 item
= kmalloc(sizeof(struct ocfs2_la_recovery_item
), GFP_NOFS
);
1273 /* Though we wish to avoid it, we are in fact safe in
1274 * skipping local alloc cleanup as fsck.ocfs2 is more
1275 * than capable of reclaiming unused space. */
1283 ocfs2_free_quota_recovery(qrec
);
1285 mlog_errno(-ENOMEM
);
1289 INIT_LIST_HEAD(&item
->lri_list
);
1290 item
->lri_la_dinode
= la_dinode
;
1291 item
->lri_slot
= slot_num
;
1292 item
->lri_tl_dinode
= tl_dinode
;
1293 item
->lri_qrec
= qrec
;
1295 spin_lock(&journal
->j_lock
);
1296 list_add_tail(&item
->lri_list
, &journal
->j_la_cleanups
);
1297 queue_work(ocfs2_wq
, &journal
->j_recovery_work
);
1298 spin_unlock(&journal
->j_lock
);
1301 /* Called by the mount code to queue recovery the last part of
1302 * recovery for it's own and offline slot(s). */
1303 void ocfs2_complete_mount_recovery(struct ocfs2_super
*osb
)
1305 struct ocfs2_journal
*journal
= osb
->journal
;
1307 /* No need to queue up our truncate_log as regular cleanup will catch
1309 ocfs2_queue_recovery_completion(journal
, osb
->slot_num
,
1310 osb
->local_alloc_copy
, NULL
, NULL
);
1311 ocfs2_schedule_truncate_log_flush(osb
, 0);
1313 osb
->local_alloc_copy
= NULL
;
1316 /* queue to recover orphan slots for all offline slots */
1317 ocfs2_replay_map_set_state(osb
, REPLAY_NEEDED
);
1318 ocfs2_queue_replay_slots(osb
);
1319 ocfs2_free_replay_slots(osb
);
1322 void ocfs2_complete_quota_recovery(struct ocfs2_super
*osb
)
1324 if (osb
->quota_rec
) {
1325 ocfs2_queue_recovery_completion(osb
->journal
,
1330 osb
->quota_rec
= NULL
;
1334 static int __ocfs2_recovery_thread(void *arg
)
1336 int status
, node_num
, slot_num
;
1337 struct ocfs2_super
*osb
= arg
;
1338 struct ocfs2_recovery_map
*rm
= osb
->recovery_map
;
1339 int *rm_quota
= NULL
;
1340 int rm_quota_used
= 0, i
;
1341 struct ocfs2_quota_recovery
*qrec
;
1345 status
= ocfs2_wait_on_mount(osb
);
1350 rm_quota
= kzalloc(osb
->max_slots
* sizeof(int), GFP_NOFS
);
1356 status
= ocfs2_super_lock(osb
, 1);
1362 status
= ocfs2_compute_replay_slots(osb
);
1366 /* queue recovery for our own slot */
1367 ocfs2_queue_recovery_completion(osb
->journal
, osb
->slot_num
, NULL
,
1370 spin_lock(&osb
->osb_lock
);
1371 while (rm
->rm_used
) {
1372 /* It's always safe to remove entry zero, as we won't
1373 * clear it until ocfs2_recover_node() has succeeded. */
1374 node_num
= rm
->rm_entries
[0];
1375 spin_unlock(&osb
->osb_lock
);
1376 mlog(0, "checking node %d\n", node_num
);
1377 slot_num
= ocfs2_node_num_to_slot(osb
, node_num
);
1378 if (slot_num
== -ENOENT
) {
1380 mlog(0, "no slot for this node, so no recovery"
1384 mlog(0, "node %d was using slot %d\n", node_num
, slot_num
);
1386 /* It is a bit subtle with quota recovery. We cannot do it
1387 * immediately because we have to obtain cluster locks from
1388 * quota files and we also don't want to just skip it because
1389 * then quota usage would be out of sync until some node takes
1390 * the slot. So we remember which nodes need quota recovery
1391 * and when everything else is done, we recover quotas. */
1392 for (i
= 0; i
< rm_quota_used
&& rm_quota
[i
] != slot_num
; i
++);
1393 if (i
== rm_quota_used
)
1394 rm_quota
[rm_quota_used
++] = slot_num
;
1396 status
= ocfs2_recover_node(osb
, node_num
, slot_num
);
1399 ocfs2_recovery_map_clear(osb
, node_num
);
1402 "Error %d recovering node %d on device (%u,%u)!\n",
1404 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1405 mlog(ML_ERROR
, "Volume requires unmount.\n");
1408 spin_lock(&osb
->osb_lock
);
1410 spin_unlock(&osb
->osb_lock
);
1411 mlog(0, "All nodes recovered\n");
1413 /* Refresh all journal recovery generations from disk */
1414 status
= ocfs2_check_journals_nolocks(osb
);
1415 status
= (status
== -EROFS
) ? 0 : status
;
1419 /* Now it is right time to recover quotas... We have to do this under
1420 * superblock lock so that noone can start using the slot (and crash)
1421 * before we recover it */
1422 for (i
= 0; i
< rm_quota_used
; i
++) {
1423 qrec
= ocfs2_begin_quota_recovery(osb
, rm_quota
[i
]);
1425 status
= PTR_ERR(qrec
);
1429 ocfs2_queue_recovery_completion(osb
->journal
, rm_quota
[i
],
1433 ocfs2_super_unlock(osb
, 1);
1435 /* queue recovery for offline slots */
1436 ocfs2_queue_replay_slots(osb
);
1439 mutex_lock(&osb
->recovery_lock
);
1440 if (!status
&& !ocfs2_recovery_completed(osb
)) {
1441 mutex_unlock(&osb
->recovery_lock
);
1445 ocfs2_free_replay_slots(osb
);
1446 osb
->recovery_thread_task
= NULL
;
1447 mb(); /* sync with ocfs2_recovery_thread_running */
1448 wake_up(&osb
->recovery_event
);
1450 mutex_unlock(&osb
->recovery_lock
);
1456 /* no one is callint kthread_stop() for us so the kthread() api
1457 * requires that we call do_exit(). And it isn't exported, but
1458 * complete_and_exit() seems to be a minimal wrapper around it. */
1459 complete_and_exit(NULL
, status
);
1463 void ocfs2_recovery_thread(struct ocfs2_super
*osb
, int node_num
)
1465 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1466 node_num
, osb
->node_num
);
1468 mutex_lock(&osb
->recovery_lock
);
1469 if (osb
->disable_recovery
)
1472 /* People waiting on recovery will wait on
1473 * the recovery map to empty. */
1474 if (ocfs2_recovery_map_set(osb
, node_num
))
1475 mlog(0, "node %d already in recovery map.\n", node_num
);
1477 mlog(0, "starting recovery thread...\n");
1479 if (osb
->recovery_thread_task
)
1482 osb
->recovery_thread_task
= kthread_run(__ocfs2_recovery_thread
, osb
,
1484 if (IS_ERR(osb
->recovery_thread_task
)) {
1485 mlog_errno((int)PTR_ERR(osb
->recovery_thread_task
));
1486 osb
->recovery_thread_task
= NULL
;
1490 mutex_unlock(&osb
->recovery_lock
);
1491 wake_up(&osb
->recovery_event
);
1496 static int ocfs2_read_journal_inode(struct ocfs2_super
*osb
,
1498 struct buffer_head
**bh
,
1499 struct inode
**ret_inode
)
1501 int status
= -EACCES
;
1502 struct inode
*inode
= NULL
;
1504 BUG_ON(slot_num
>= osb
->max_slots
);
1506 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1508 if (!inode
|| is_bad_inode(inode
)) {
1512 SET_INODE_JOURNAL(inode
);
1514 status
= ocfs2_read_inode_block_full(inode
, bh
, OCFS2_BH_IGNORE_CACHE
);
1524 if (status
|| !ret_inode
)
1532 /* Does the actual journal replay and marks the journal inode as
1533 * clean. Will only replay if the journal inode is marked dirty. */
1534 static int ocfs2_replay_journal(struct ocfs2_super
*osb
,
1541 struct inode
*inode
= NULL
;
1542 struct ocfs2_dinode
*fe
;
1543 journal_t
*journal
= NULL
;
1544 struct buffer_head
*bh
= NULL
;
1547 status
= ocfs2_read_journal_inode(osb
, slot_num
, &bh
, &inode
);
1553 fe
= (struct ocfs2_dinode
*)bh
->b_data
;
1554 slot_reco_gen
= ocfs2_get_recovery_generation(fe
);
1559 * As the fs recovery is asynchronous, there is a small chance that
1560 * another node mounted (and recovered) the slot before the recovery
1561 * thread could get the lock. To handle that, we dirty read the journal
1562 * inode for that slot to get the recovery generation. If it is
1563 * different than what we expected, the slot has been recovered.
1564 * If not, it needs recovery.
1566 if (osb
->slot_recovery_generations
[slot_num
] != slot_reco_gen
) {
1567 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num
,
1568 osb
->slot_recovery_generations
[slot_num
], slot_reco_gen
);
1569 osb
->slot_recovery_generations
[slot_num
] = slot_reco_gen
;
1574 /* Continue with recovery as the journal has not yet been recovered */
1576 status
= ocfs2_inode_lock_full(inode
, &bh
, 1, OCFS2_META_LOCK_RECOVERY
);
1578 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status
);
1579 if (status
!= -ERESTARTSYS
)
1580 mlog(ML_ERROR
, "Could not lock journal!\n");
1585 fe
= (struct ocfs2_dinode
*) bh
->b_data
;
1587 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1588 slot_reco_gen
= ocfs2_get_recovery_generation(fe
);
1590 if (!(flags
& OCFS2_JOURNAL_DIRTY_FL
)) {
1591 mlog(0, "No recovery required for node %d\n", node_num
);
1592 /* Refresh recovery generation for the slot */
1593 osb
->slot_recovery_generations
[slot_num
] = slot_reco_gen
;
1597 /* we need to run complete recovery for offline orphan slots */
1598 ocfs2_replay_map_set_state(osb
, REPLAY_NEEDED
);
1600 mlog(ML_NOTICE
, "Recovering node %d from slot %d on device (%u,%u)\n",
1602 MAJOR(osb
->sb
->s_dev
), MINOR(osb
->sb
->s_dev
));
1604 OCFS2_I(inode
)->ip_clusters
= le32_to_cpu(fe
->i_clusters
);
1606 status
= ocfs2_force_read_journal(inode
);
1612 mlog(0, "calling journal_init_inode\n");
1613 journal
= jbd2_journal_init_inode(inode
);
1614 if (journal
== NULL
) {
1615 mlog(ML_ERROR
, "Linux journal layer error\n");
1620 status
= jbd2_journal_load(journal
);
1625 jbd2_journal_destroy(journal
);
1629 ocfs2_clear_journal_error(osb
->sb
, journal
, slot_num
);
1631 /* wipe the journal */
1632 mlog(0, "flushing the journal.\n");
1633 jbd2_journal_lock_updates(journal
);
1634 status
= jbd2_journal_flush(journal
);
1635 jbd2_journal_unlock_updates(journal
);
1639 /* This will mark the node clean */
1640 flags
= le32_to_cpu(fe
->id1
.journal1
.ij_flags
);
1641 flags
&= ~OCFS2_JOURNAL_DIRTY_FL
;
1642 fe
->id1
.journal1
.ij_flags
= cpu_to_le32(flags
);
1644 /* Increment recovery generation to indicate successful recovery */
1645 ocfs2_bump_recovery_generation(fe
);
1646 osb
->slot_recovery_generations
[slot_num
] =
1647 ocfs2_get_recovery_generation(fe
);
1649 ocfs2_compute_meta_ecc(osb
->sb
, bh
->b_data
, &fe
->i_check
);
1650 status
= ocfs2_write_block(osb
, bh
, INODE_CACHE(inode
));
1657 jbd2_journal_destroy(journal
);
1660 /* drop the lock on this nodes journal */
1662 ocfs2_inode_unlock(inode
, 1);
1674 * Do the most important parts of node recovery:
1675 * - Replay it's journal
1676 * - Stamp a clean local allocator file
1677 * - Stamp a clean truncate log
1678 * - Mark the node clean
1680 * If this function completes without error, a node in OCFS2 can be
1681 * said to have been safely recovered. As a result, failure during the
1682 * second part of a nodes recovery process (local alloc recovery) is
1683 * far less concerning.
1685 static int ocfs2_recover_node(struct ocfs2_super
*osb
,
1686 int node_num
, int slot_num
)
1689 struct ocfs2_dinode
*la_copy
= NULL
;
1690 struct ocfs2_dinode
*tl_copy
= NULL
;
1692 mlog_entry("(node_num=%d, slot_num=%d, osb->node_num = %d)\n",
1693 node_num
, slot_num
, osb
->node_num
);
1695 /* Should not ever be called to recover ourselves -- in that
1696 * case we should've called ocfs2_journal_load instead. */
1697 BUG_ON(osb
->node_num
== node_num
);
1699 status
= ocfs2_replay_journal(osb
, node_num
, slot_num
);
1701 if (status
== -EBUSY
) {
1702 mlog(0, "Skipping recovery for slot %u (node %u) "
1703 "as another node has recovered it\n", slot_num
,
1712 /* Stamp a clean local alloc file AFTER recovering the journal... */
1713 status
= ocfs2_begin_local_alloc_recovery(osb
, slot_num
, &la_copy
);
1719 /* An error from begin_truncate_log_recovery is not
1720 * serious enough to warrant halting the rest of
1722 status
= ocfs2_begin_truncate_log_recovery(osb
, slot_num
, &tl_copy
);
1726 /* Likewise, this would be a strange but ultimately not so
1727 * harmful place to get an error... */
1728 status
= ocfs2_clear_slot(osb
, slot_num
);
1732 /* This will kfree the memory pointed to by la_copy and tl_copy */
1733 ocfs2_queue_recovery_completion(osb
->journal
, slot_num
, la_copy
,
1743 /* Test node liveness by trylocking his journal. If we get the lock,
1744 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1745 * still alive (we couldn't get the lock) and < 0 on error. */
1746 static int ocfs2_trylock_journal(struct ocfs2_super
*osb
,
1750 struct inode
*inode
= NULL
;
1752 inode
= ocfs2_get_system_file_inode(osb
, JOURNAL_SYSTEM_INODE
,
1754 if (inode
== NULL
) {
1755 mlog(ML_ERROR
, "access error\n");
1759 if (is_bad_inode(inode
)) {
1760 mlog(ML_ERROR
, "access error (bad inode)\n");
1766 SET_INODE_JOURNAL(inode
);
1768 flags
= OCFS2_META_LOCK_RECOVERY
| OCFS2_META_LOCK_NOQUEUE
;
1769 status
= ocfs2_inode_lock_full(inode
, NULL
, 1, flags
);
1771 if (status
!= -EAGAIN
)
1776 ocfs2_inode_unlock(inode
, 1);
1784 /* Call this underneath ocfs2_super_lock. It also assumes that the
1785 * slot info struct has been updated from disk. */
1786 int ocfs2_mark_dead_nodes(struct ocfs2_super
*osb
)
1788 unsigned int node_num
;
1791 struct buffer_head
*bh
= NULL
;
1792 struct ocfs2_dinode
*di
;
1794 /* This is called with the super block cluster lock, so we
1795 * know that the slot map can't change underneath us. */
1797 for (i
= 0; i
< osb
->max_slots
; i
++) {
1798 /* Read journal inode to get the recovery generation */
1799 status
= ocfs2_read_journal_inode(osb
, i
, &bh
, NULL
);
1804 di
= (struct ocfs2_dinode
*)bh
->b_data
;
1805 gen
= ocfs2_get_recovery_generation(di
);
1809 spin_lock(&osb
->osb_lock
);
1810 osb
->slot_recovery_generations
[i
] = gen
;
1812 mlog(0, "Slot %u recovery generation is %u\n", i
,
1813 osb
->slot_recovery_generations
[i
]);
1815 if (i
== osb
->slot_num
) {
1816 spin_unlock(&osb
->osb_lock
);
1820 status
= ocfs2_slot_to_node_num_locked(osb
, i
, &node_num
);
1821 if (status
== -ENOENT
) {
1822 spin_unlock(&osb
->osb_lock
);
1826 if (__ocfs2_recovery_map_test(osb
, node_num
)) {
1827 spin_unlock(&osb
->osb_lock
);
1830 spin_unlock(&osb
->osb_lock
);
1832 /* Ok, we have a slot occupied by another node which
1833 * is not in the recovery map. We trylock his journal
1834 * file here to test if he's alive. */
1835 status
= ocfs2_trylock_journal(osb
, i
);
1837 /* Since we're called from mount, we know that
1838 * the recovery thread can't race us on
1839 * setting / checking the recovery bits. */
1840 ocfs2_recovery_thread(osb
, node_num
);
1841 } else if ((status
< 0) && (status
!= -EAGAIN
)) {
1854 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1855 * randomness to the timeout to minimize multple nodes firing the timer at the
1858 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1862 get_random_bytes(&time
, sizeof(time
));
1863 time
= ORPHAN_SCAN_SCHEDULE_TIMEOUT
+ (time
% 5000);
1864 return msecs_to_jiffies(time
);
1868 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1869 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1870 * is done to catch any orphans that are left over in orphan directories.
1872 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1873 * seconds. It gets an EX lock on os_lockres and checks sequence number
1874 * stored in LVB. If the sequence number has changed, it means some other
1875 * node has done the scan. This node skips the scan and tracks the
1876 * sequence number. If the sequence number didn't change, it means a scan
1877 * hasn't happened. The node queues a scan and increments the
1878 * sequence number in the LVB.
1880 void ocfs2_queue_orphan_scan(struct ocfs2_super
*osb
)
1882 struct ocfs2_orphan_scan
*os
;
1886 os
= &osb
->osb_orphan_scan
;
1888 if (atomic_read(&os
->os_state
) == ORPHAN_SCAN_INACTIVE
)
1891 status
= ocfs2_orphan_scan_lock(osb
, &seqno
);
1893 if (status
!= -EAGAIN
)
1898 /* Do no queue the tasks if the volume is being umounted */
1899 if (atomic_read(&os
->os_state
) == ORPHAN_SCAN_INACTIVE
)
1902 if (os
->os_seqno
!= seqno
) {
1903 os
->os_seqno
= seqno
;
1907 for (i
= 0; i
< osb
->max_slots
; i
++)
1908 ocfs2_queue_recovery_completion(osb
->journal
, i
, NULL
, NULL
,
1911 * We queued a recovery on orphan slots, increment the sequence
1912 * number and update LVB so other node will skip the scan for a while
1916 os
->os_scantime
= CURRENT_TIME
;
1918 ocfs2_orphan_scan_unlock(osb
, seqno
);
1923 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1924 void ocfs2_orphan_scan_work(struct work_struct
*work
)
1926 struct ocfs2_orphan_scan
*os
;
1927 struct ocfs2_super
*osb
;
1929 os
= container_of(work
, struct ocfs2_orphan_scan
,
1930 os_orphan_scan_work
.work
);
1933 mutex_lock(&os
->os_lock
);
1934 ocfs2_queue_orphan_scan(osb
);
1935 if (atomic_read(&os
->os_state
) == ORPHAN_SCAN_ACTIVE
)
1936 queue_delayed_work(ocfs2_wq
, &os
->os_orphan_scan_work
,
1937 ocfs2_orphan_scan_timeout());
1938 mutex_unlock(&os
->os_lock
);
1941 void ocfs2_orphan_scan_stop(struct ocfs2_super
*osb
)
1943 struct ocfs2_orphan_scan
*os
;
1945 os
= &osb
->osb_orphan_scan
;
1946 if (atomic_read(&os
->os_state
) == ORPHAN_SCAN_ACTIVE
) {
1947 atomic_set(&os
->os_state
, ORPHAN_SCAN_INACTIVE
);
1948 mutex_lock(&os
->os_lock
);
1949 cancel_delayed_work(&os
->os_orphan_scan_work
);
1950 mutex_unlock(&os
->os_lock
);
1954 void ocfs2_orphan_scan_init(struct ocfs2_super
*osb
)
1956 struct ocfs2_orphan_scan
*os
;
1958 os
= &osb
->osb_orphan_scan
;
1962 mutex_init(&os
->os_lock
);
1963 INIT_DELAYED_WORK(&os
->os_orphan_scan_work
, ocfs2_orphan_scan_work
);
1966 void ocfs2_orphan_scan_start(struct ocfs2_super
*osb
)
1968 struct ocfs2_orphan_scan
*os
;
1970 os
= &osb
->osb_orphan_scan
;
1971 os
->os_scantime
= CURRENT_TIME
;
1972 if (ocfs2_is_hard_readonly(osb
) || ocfs2_mount_local(osb
))
1973 atomic_set(&os
->os_state
, ORPHAN_SCAN_INACTIVE
);
1975 atomic_set(&os
->os_state
, ORPHAN_SCAN_ACTIVE
);
1976 queue_delayed_work(ocfs2_wq
, &os
->os_orphan_scan_work
,
1977 ocfs2_orphan_scan_timeout());
1981 struct ocfs2_orphan_filldir_priv
{
1983 struct ocfs2_super
*osb
;
1986 static int ocfs2_orphan_filldir(void *priv
, const char *name
, int name_len
,
1987 loff_t pos
, u64 ino
, unsigned type
)
1989 struct ocfs2_orphan_filldir_priv
*p
= priv
;
1992 if (name_len
== 1 && !strncmp(".", name
, 1))
1994 if (name_len
== 2 && !strncmp("..", name
, 2))
1997 /* Skip bad inodes so that recovery can continue */
1998 iter
= ocfs2_iget(p
->osb
, ino
,
1999 OCFS2_FI_FLAG_ORPHAN_RECOVERY
, 0);
2003 mlog(0, "queue orphan %llu\n",
2004 (unsigned long long)OCFS2_I(iter
)->ip_blkno
);
2005 /* No locking is required for the next_orphan queue as there
2006 * is only ever a single process doing orphan recovery. */
2007 OCFS2_I(iter
)->ip_next_orphan
= p
->head
;
2013 static int ocfs2_queue_orphans(struct ocfs2_super
*osb
,
2015 struct inode
**head
)
2018 struct inode
*orphan_dir_inode
= NULL
;
2019 struct ocfs2_orphan_filldir_priv priv
;
2025 orphan_dir_inode
= ocfs2_get_system_file_inode(osb
,
2026 ORPHAN_DIR_SYSTEM_INODE
,
2028 if (!orphan_dir_inode
) {
2034 mutex_lock(&orphan_dir_inode
->i_mutex
);
2035 status
= ocfs2_inode_lock(orphan_dir_inode
, NULL
, 0);
2041 status
= ocfs2_dir_foreach(orphan_dir_inode
, &pos
, &priv
,
2042 ocfs2_orphan_filldir
);
2051 ocfs2_inode_unlock(orphan_dir_inode
, 0);
2053 mutex_unlock(&orphan_dir_inode
->i_mutex
);
2054 iput(orphan_dir_inode
);
2058 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super
*osb
,
2063 spin_lock(&osb
->osb_lock
);
2064 ret
= !osb
->osb_orphan_wipes
[slot
];
2065 spin_unlock(&osb
->osb_lock
);
2069 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super
*osb
,
2072 spin_lock(&osb
->osb_lock
);
2073 /* Mark ourselves such that new processes in delete_inode()
2074 * know to quit early. */
2075 ocfs2_node_map_set_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
2076 while (osb
->osb_orphan_wipes
[slot
]) {
2077 /* If any processes are already in the middle of an
2078 * orphan wipe on this dir, then we need to wait for
2080 spin_unlock(&osb
->osb_lock
);
2081 wait_event_interruptible(osb
->osb_wipe_event
,
2082 ocfs2_orphan_recovery_can_continue(osb
, slot
));
2083 spin_lock(&osb
->osb_lock
);
2085 spin_unlock(&osb
->osb_lock
);
2088 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super
*osb
,
2091 ocfs2_node_map_clear_bit(osb
, &osb
->osb_recovering_orphan_dirs
, slot
);
2095 * Orphan recovery. Each mounted node has it's own orphan dir which we
2096 * must run during recovery. Our strategy here is to build a list of
2097 * the inodes in the orphan dir and iget/iput them. The VFS does
2098 * (most) of the rest of the work.
2100 * Orphan recovery can happen at any time, not just mount so we have a
2101 * couple of extra considerations.
2103 * - We grab as many inodes as we can under the orphan dir lock -
2104 * doing iget() outside the orphan dir risks getting a reference on
2106 * - We must be sure not to deadlock with other processes on the
2107 * system wanting to run delete_inode(). This can happen when they go
2108 * to lock the orphan dir and the orphan recovery process attempts to
2109 * iget() inside the orphan dir lock. This can be avoided by
2110 * advertising our state to ocfs2_delete_inode().
2112 static int ocfs2_recover_orphans(struct ocfs2_super
*osb
,
2116 struct inode
*inode
= NULL
;
2118 struct ocfs2_inode_info
*oi
;
2120 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot
);
2122 ocfs2_mark_recovering_orphan_dir(osb
, slot
);
2123 ret
= ocfs2_queue_orphans(osb
, slot
, &inode
);
2124 ocfs2_clear_recovering_orphan_dir(osb
, slot
);
2126 /* Error here should be noted, but we want to continue with as
2127 * many queued inodes as we've got. */
2132 oi
= OCFS2_I(inode
);
2133 mlog(0, "iput orphan %llu\n", (unsigned long long)oi
->ip_blkno
);
2135 iter
= oi
->ip_next_orphan
;
2137 spin_lock(&oi
->ip_lock
);
2138 /* The remote delete code may have set these on the
2139 * assumption that the other node would wipe them
2140 * successfully. If they are still in the node's
2141 * orphan dir, we need to reset that state. */
2142 oi
->ip_flags
&= ~(OCFS2_INODE_DELETED
|OCFS2_INODE_SKIP_DELETE
);
2144 /* Set the proper information to get us going into
2145 * ocfs2_delete_inode. */
2146 oi
->ip_flags
|= OCFS2_INODE_MAYBE_ORPHANED
;
2147 spin_unlock(&oi
->ip_lock
);
2157 static int __ocfs2_wait_on_mount(struct ocfs2_super
*osb
, int quota
)
2159 /* This check is good because ocfs2 will wait on our recovery
2160 * thread before changing it to something other than MOUNTED
2162 wait_event(osb
->osb_mount_event
,
2163 (!quota
&& atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED
) ||
2164 atomic_read(&osb
->vol_state
) == VOLUME_MOUNTED_QUOTAS
||
2165 atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
);
2167 /* If there's an error on mount, then we may never get to the
2168 * MOUNTED flag, but this is set right before
2169 * dismount_volume() so we can trust it. */
2170 if (atomic_read(&osb
->vol_state
) == VOLUME_DISABLED
) {
2171 mlog(0, "mount error, exiting!\n");
2178 static int ocfs2_commit_thread(void *arg
)
2181 struct ocfs2_super
*osb
= arg
;
2182 struct ocfs2_journal
*journal
= osb
->journal
;
2184 /* we can trust j_num_trans here because _should_stop() is only set in
2185 * shutdown and nobody other than ourselves should be able to start
2186 * transactions. committing on shutdown might take a few iterations
2187 * as final transactions put deleted inodes on the list */
2188 while (!(kthread_should_stop() &&
2189 atomic_read(&journal
->j_num_trans
) == 0)) {
2191 wait_event_interruptible(osb
->checkpoint_event
,
2192 atomic_read(&journal
->j_num_trans
)
2193 || kthread_should_stop());
2195 status
= ocfs2_commit_cache(osb
);
2199 if (kthread_should_stop() && atomic_read(&journal
->j_num_trans
)){
2201 "commit_thread: %u transactions pending on "
2203 atomic_read(&journal
->j_num_trans
));
2210 /* Reads all the journal inodes without taking any cluster locks. Used
2211 * for hard readonly access to determine whether any journal requires
2212 * recovery. Also used to refresh the recovery generation numbers after
2213 * a journal has been recovered by another node.
2215 int ocfs2_check_journals_nolocks(struct ocfs2_super
*osb
)
2219 struct buffer_head
*di_bh
= NULL
;
2220 struct ocfs2_dinode
*di
;
2221 int journal_dirty
= 0;
2223 for(slot
= 0; slot
< osb
->max_slots
; slot
++) {
2224 ret
= ocfs2_read_journal_inode(osb
, slot
, &di_bh
, NULL
);
2230 di
= (struct ocfs2_dinode
*) di_bh
->b_data
;
2232 osb
->slot_recovery_generations
[slot
] =
2233 ocfs2_get_recovery_generation(di
);
2235 if (le32_to_cpu(di
->id1
.journal1
.ij_flags
) &
2236 OCFS2_JOURNAL_DIRTY_FL
)