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>
33 #include <linux/delay.h>
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"
55 #include "ocfs2_trace.h"
57 DEFINE_SPINLOCK(trans_inc_lock);
59 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
61 static int ocfs2_force_read_journal(struct inode *inode);
62 static int ocfs2_recover_node(struct ocfs2_super *osb,
63 int node_num, int slot_num);
64 static int __ocfs2_recovery_thread(void *arg);
65 static int ocfs2_commit_cache(struct ocfs2_super *osb);
66 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota);
67 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
68 int dirty, int replayed);
69 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
71 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
73 static int ocfs2_commit_thread(void *arg);
74 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
76 struct ocfs2_dinode *la_dinode,
77 struct ocfs2_dinode *tl_dinode,
78 struct ocfs2_quota_recovery *qrec);
80 static inline int ocfs2_wait_on_mount(struct ocfs2_super *osb)
82 return __ocfs2_wait_on_mount(osb, 0);
85 static inline int ocfs2_wait_on_quotas(struct ocfs2_super *osb)
87 return __ocfs2_wait_on_mount(osb, 1);
91 * This replay_map is to track online/offline slots, so we could recover
92 * offline slots during recovery and mount
95 enum ocfs2_replay_state {
96 REPLAY_UNNEEDED = 0, /* Replay is not needed, so ignore this map */
97 REPLAY_NEEDED, /* Replay slots marked in rm_replay_slots */
98 REPLAY_DONE /* Replay was already queued */
101 struct ocfs2_replay_map {
102 unsigned int rm_slots;
103 enum ocfs2_replay_state rm_state;
104 unsigned char rm_replay_slots[0];
107 void ocfs2_replay_map_set_state(struct ocfs2_super *osb, int state)
109 if (!osb->replay_map)
112 /* If we've already queued the replay, we don't have any more to do */
113 if (osb->replay_map->rm_state == REPLAY_DONE)
116 osb->replay_map->rm_state = state;
119 int ocfs2_compute_replay_slots(struct ocfs2_super *osb)
121 struct ocfs2_replay_map *replay_map;
124 /* If replay map is already set, we don't do it again */
128 replay_map = kzalloc(sizeof(struct ocfs2_replay_map) +
129 (osb->max_slots * sizeof(char)), GFP_KERNEL);
136 spin_lock(&osb->osb_lock);
138 replay_map->rm_slots = osb->max_slots;
139 replay_map->rm_state = REPLAY_UNNEEDED;
141 /* set rm_replay_slots for offline slot(s) */
142 for (i = 0; i < replay_map->rm_slots; i++) {
143 if (ocfs2_slot_to_node_num_locked(osb, i, &node_num) == -ENOENT)
144 replay_map->rm_replay_slots[i] = 1;
147 osb->replay_map = replay_map;
148 spin_unlock(&osb->osb_lock);
152 void ocfs2_queue_replay_slots(struct ocfs2_super *osb)
154 struct ocfs2_replay_map *replay_map = osb->replay_map;
160 if (replay_map->rm_state != REPLAY_NEEDED)
163 for (i = 0; i < replay_map->rm_slots; i++)
164 if (replay_map->rm_replay_slots[i])
165 ocfs2_queue_recovery_completion(osb->journal, i, NULL,
167 replay_map->rm_state = REPLAY_DONE;
170 void ocfs2_free_replay_slots(struct ocfs2_super *osb)
172 struct ocfs2_replay_map *replay_map = osb->replay_map;
174 if (!osb->replay_map)
178 osb->replay_map = NULL;
181 int ocfs2_recovery_init(struct ocfs2_super *osb)
183 struct ocfs2_recovery_map *rm;
185 mutex_init(&osb->recovery_lock);
186 osb->disable_recovery = 0;
187 osb->recovery_thread_task = NULL;
188 init_waitqueue_head(&osb->recovery_event);
190 rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
191 osb->max_slots * sizeof(unsigned int),
198 rm->rm_entries = (unsigned int *)((char *)rm +
199 sizeof(struct ocfs2_recovery_map));
200 osb->recovery_map = rm;
205 /* we can't grab the goofy sem lock from inside wait_event, so we use
206 * memory barriers to make sure that we'll see the null task before
208 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
211 return osb->recovery_thread_task != NULL;
214 void ocfs2_recovery_exit(struct ocfs2_super *osb)
216 struct ocfs2_recovery_map *rm;
218 /* disable any new recovery threads and wait for any currently
219 * running ones to exit. Do this before setting the vol_state. */
220 mutex_lock(&osb->recovery_lock);
221 osb->disable_recovery = 1;
222 mutex_unlock(&osb->recovery_lock);
223 wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
225 /* At this point, we know that no more recovery threads can be
226 * launched, so wait for any recovery completion work to
228 flush_workqueue(ocfs2_wq);
231 * Now that recovery is shut down, and the osb is about to be
232 * freed, the osb_lock is not taken here.
234 rm = osb->recovery_map;
235 /* XXX: Should we bug if there are dirty entries? */
240 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
241 unsigned int node_num)
244 struct ocfs2_recovery_map *rm = osb->recovery_map;
246 assert_spin_locked(&osb->osb_lock);
248 for (i = 0; i < rm->rm_used; i++) {
249 if (rm->rm_entries[i] == node_num)
256 /* Behaves like test-and-set. Returns the previous value */
257 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
258 unsigned int node_num)
260 struct ocfs2_recovery_map *rm = osb->recovery_map;
262 spin_lock(&osb->osb_lock);
263 if (__ocfs2_recovery_map_test(osb, node_num)) {
264 spin_unlock(&osb->osb_lock);
268 /* XXX: Can this be exploited? Not from o2dlm... */
269 BUG_ON(rm->rm_used >= osb->max_slots);
271 rm->rm_entries[rm->rm_used] = node_num;
273 spin_unlock(&osb->osb_lock);
278 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
279 unsigned int node_num)
282 struct ocfs2_recovery_map *rm = osb->recovery_map;
284 spin_lock(&osb->osb_lock);
286 for (i = 0; i < rm->rm_used; i++) {
287 if (rm->rm_entries[i] == node_num)
291 if (i < rm->rm_used) {
292 /* XXX: be careful with the pointer math */
293 memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
294 (rm->rm_used - i - 1) * sizeof(unsigned int));
298 spin_unlock(&osb->osb_lock);
301 static int ocfs2_commit_cache(struct ocfs2_super *osb)
304 unsigned int flushed;
305 struct ocfs2_journal *journal = NULL;
307 journal = osb->journal;
309 /* Flush all pending commits and checkpoint the journal. */
310 down_write(&journal->j_trans_barrier);
312 flushed = atomic_read(&journal->j_num_trans);
313 trace_ocfs2_commit_cache_begin(flushed);
315 up_write(&journal->j_trans_barrier);
319 jbd2_journal_lock_updates(journal->j_journal);
320 status = jbd2_journal_flush(journal->j_journal);
321 jbd2_journal_unlock_updates(journal->j_journal);
323 up_write(&journal->j_trans_barrier);
328 ocfs2_inc_trans_id(journal);
330 flushed = atomic_read(&journal->j_num_trans);
331 atomic_set(&journal->j_num_trans, 0);
332 up_write(&journal->j_trans_barrier);
334 trace_ocfs2_commit_cache_end(journal->j_trans_id, flushed);
336 ocfs2_wake_downconvert_thread(osb);
337 wake_up(&journal->j_checkpointed);
342 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
344 journal_t *journal = osb->journal->j_journal;
347 BUG_ON(!osb || !osb->journal->j_journal);
349 if (ocfs2_is_hard_readonly(osb))
350 return ERR_PTR(-EROFS);
352 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
353 BUG_ON(max_buffs <= 0);
355 /* Nested transaction? Just return the handle... */
356 if (journal_current_handle())
357 return jbd2_journal_start(journal, max_buffs);
359 sb_start_intwrite(osb->sb);
361 down_read(&osb->journal->j_trans_barrier);
363 handle = jbd2_journal_start(journal, max_buffs);
364 if (IS_ERR(handle)) {
365 up_read(&osb->journal->j_trans_barrier);
366 sb_end_intwrite(osb->sb);
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);
397 sb_end_intwrite(osb->sb);
404 * 'nblocks' is what you want to add to the current transaction.
406 * This might call jbd2_journal_restart() which will commit dirty buffers
407 * and then restart the transaction. Before calling
408 * ocfs2_extend_trans(), any changed blocks should have been
409 * dirtied. After calling it, all blocks which need to be changed must
410 * go through another set of journal_access/journal_dirty calls.
412 * WARNING: This will not release any semaphores or disk locks taken
413 * during the transaction, so make sure they were taken *before*
414 * start_trans or we'll have ordering deadlocks.
416 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
417 * good because transaction ids haven't yet been recorded on the
418 * cluster locks associated with this handle.
420 int ocfs2_extend_trans(handle_t *handle, int nblocks)
422 int status, old_nblocks;
430 old_nblocks = handle->h_buffer_credits;
432 trace_ocfs2_extend_trans(old_nblocks, nblocks);
434 #ifdef CONFIG_OCFS2_DEBUG_FS
437 status = jbd2_journal_extend(handle, nblocks);
445 trace_ocfs2_extend_trans_restart(old_nblocks + nblocks);
446 status = jbd2_journal_restart(handle,
447 old_nblocks + nblocks);
460 * If we have fewer than thresh credits, extend by OCFS2_MAX_TRANS_DATA.
461 * If that fails, restart the transaction & regain write access for the
462 * buffer head which is used for metadata modifications.
463 * Taken from Ext4: extend_or_restart_transaction()
465 int ocfs2_allocate_extend_trans(handle_t *handle, int thresh)
467 int status, old_nblks;
471 old_nblks = handle->h_buffer_credits;
472 trace_ocfs2_allocate_extend_trans(old_nblks, thresh);
474 if (old_nblks < thresh)
477 status = jbd2_journal_extend(handle, OCFS2_MAX_TRANS_DATA);
484 status = jbd2_journal_restart(handle, OCFS2_MAX_TRANS_DATA);
494 struct ocfs2_triggers {
495 struct jbd2_buffer_trigger_type ot_triggers;
499 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
501 return container_of(triggers, struct ocfs2_triggers, ot_triggers);
504 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
505 struct buffer_head *bh,
506 void *data, size_t size)
508 struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
511 * We aren't guaranteed to have the superblock here, so we
512 * must unconditionally compute the ecc data.
513 * __ocfs2_journal_access() will only set the triggers if
514 * metaecc is enabled.
516 ocfs2_block_check_compute(data, size, data + ot->ot_offset);
520 * Quota blocks have their own trigger because the struct ocfs2_block_check
521 * offset depends on the blocksize.
523 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
524 struct buffer_head *bh,
525 void *data, size_t size)
527 struct ocfs2_disk_dqtrailer *dqt =
528 ocfs2_block_dqtrailer(size, data);
531 * We aren't guaranteed to have the superblock here, so we
532 * must unconditionally compute the ecc data.
533 * __ocfs2_journal_access() will only set the triggers if
534 * metaecc is enabled.
536 ocfs2_block_check_compute(data, size, &dqt->dq_check);
540 * Directory blocks also have their own trigger because the
541 * struct ocfs2_block_check offset depends on the blocksize.
543 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
544 struct buffer_head *bh,
545 void *data, size_t size)
547 struct ocfs2_dir_block_trailer *trailer =
548 ocfs2_dir_trailer_from_size(size, data);
551 * We aren't guaranteed to have the superblock here, so we
552 * must unconditionally compute the ecc data.
553 * __ocfs2_journal_access() will only set the triggers if
554 * metaecc is enabled.
556 ocfs2_block_check_compute(data, size, &trailer->db_check);
559 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
560 struct buffer_head *bh)
563 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
564 "bh->b_blocknr = %llu\n",
566 (unsigned long long)bh->b_blocknr);
568 /* We aren't guaranteed to have the superblock here - but if we
569 * don't, it'll just crash. */
570 ocfs2_error(bh->b_assoc_map->host->i_sb,
571 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
574 static struct ocfs2_triggers di_triggers = {
576 .t_frozen = ocfs2_frozen_trigger,
577 .t_abort = ocfs2_abort_trigger,
579 .ot_offset = offsetof(struct ocfs2_dinode, i_check),
582 static struct ocfs2_triggers eb_triggers = {
584 .t_frozen = ocfs2_frozen_trigger,
585 .t_abort = ocfs2_abort_trigger,
587 .ot_offset = offsetof(struct ocfs2_extent_block, h_check),
590 static struct ocfs2_triggers rb_triggers = {
592 .t_frozen = ocfs2_frozen_trigger,
593 .t_abort = ocfs2_abort_trigger,
595 .ot_offset = offsetof(struct ocfs2_refcount_block, rf_check),
598 static struct ocfs2_triggers gd_triggers = {
600 .t_frozen = ocfs2_frozen_trigger,
601 .t_abort = ocfs2_abort_trigger,
603 .ot_offset = offsetof(struct ocfs2_group_desc, bg_check),
606 static struct ocfs2_triggers db_triggers = {
608 .t_frozen = ocfs2_db_frozen_trigger,
609 .t_abort = ocfs2_abort_trigger,
613 static struct ocfs2_triggers xb_triggers = {
615 .t_frozen = ocfs2_frozen_trigger,
616 .t_abort = ocfs2_abort_trigger,
618 .ot_offset = offsetof(struct ocfs2_xattr_block, xb_check),
621 static struct ocfs2_triggers dq_triggers = {
623 .t_frozen = ocfs2_dq_frozen_trigger,
624 .t_abort = ocfs2_abort_trigger,
628 static struct ocfs2_triggers dr_triggers = {
630 .t_frozen = ocfs2_frozen_trigger,
631 .t_abort = ocfs2_abort_trigger,
633 .ot_offset = offsetof(struct ocfs2_dx_root_block, dr_check),
636 static struct ocfs2_triggers dl_triggers = {
638 .t_frozen = ocfs2_frozen_trigger,
639 .t_abort = ocfs2_abort_trigger,
641 .ot_offset = offsetof(struct ocfs2_dx_leaf, dl_check),
644 static int __ocfs2_journal_access(handle_t *handle,
645 struct ocfs2_caching_info *ci,
646 struct buffer_head *bh,
647 struct ocfs2_triggers *triggers,
651 struct ocfs2_super *osb =
652 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
654 BUG_ON(!ci || !ci->ci_ops);
658 trace_ocfs2_journal_access(
659 (unsigned long long)ocfs2_metadata_cache_owner(ci),
660 (unsigned long long)bh->b_blocknr, type, bh->b_size);
662 /* we can safely remove this assertion after testing. */
663 if (!buffer_uptodate(bh)) {
664 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
665 mlog(ML_ERROR, "b_blocknr=%llu\n",
666 (unsigned long long)bh->b_blocknr);
670 /* Set the current transaction information on the ci so
671 * that the locking code knows whether it can drop it's locks
672 * on this ci or not. We're protected from the commit
673 * thread updating the current transaction id until
674 * ocfs2_commit_trans() because ocfs2_start_trans() took
675 * j_trans_barrier for us. */
676 ocfs2_set_ci_lock_trans(osb->journal, ci);
678 ocfs2_metadata_cache_io_lock(ci);
680 case OCFS2_JOURNAL_ACCESS_CREATE:
681 case OCFS2_JOURNAL_ACCESS_WRITE:
682 status = jbd2_journal_get_write_access(handle, bh);
685 case OCFS2_JOURNAL_ACCESS_UNDO:
686 status = jbd2_journal_get_undo_access(handle, bh);
691 mlog(ML_ERROR, "Unknown access type!\n");
693 if (!status && ocfs2_meta_ecc(osb) && triggers)
694 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
695 ocfs2_metadata_cache_io_unlock(ci);
698 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
704 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
705 struct buffer_head *bh, int type)
707 return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
710 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
711 struct buffer_head *bh, int type)
713 return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
716 int ocfs2_journal_access_rb(handle_t *handle, struct ocfs2_caching_info *ci,
717 struct buffer_head *bh, int type)
719 return __ocfs2_journal_access(handle, ci, bh, &rb_triggers,
723 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
724 struct buffer_head *bh, int type)
726 return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
729 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
730 struct buffer_head *bh, int type)
732 return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
735 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
736 struct buffer_head *bh, int type)
738 return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
741 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
742 struct buffer_head *bh, int type)
744 return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
747 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
748 struct buffer_head *bh, int type)
750 return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
753 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
754 struct buffer_head *bh, int type)
756 return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
759 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
760 struct buffer_head *bh, int type)
762 return __ocfs2_journal_access(handle, ci, bh, NULL, type);
765 void ocfs2_journal_dirty(handle_t *handle, struct buffer_head *bh)
769 trace_ocfs2_journal_dirty((unsigned long long)bh->b_blocknr);
771 status = jbd2_journal_dirty_metadata(handle, bh);
775 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
777 void ocfs2_set_journal_params(struct ocfs2_super *osb)
779 journal_t *journal = osb->journal->j_journal;
780 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
782 if (osb->osb_commit_interval)
783 commit_interval = osb->osb_commit_interval;
785 write_lock(&journal->j_state_lock);
786 journal->j_commit_interval = commit_interval;
787 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
788 journal->j_flags |= JBD2_BARRIER;
790 journal->j_flags &= ~JBD2_BARRIER;
791 write_unlock(&journal->j_state_lock);
794 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
797 struct inode *inode = NULL; /* the journal inode */
798 journal_t *j_journal = NULL;
799 struct ocfs2_dinode *di = NULL;
800 struct buffer_head *bh = NULL;
801 struct ocfs2_super *osb;
806 osb = journal->j_osb;
808 /* already have the inode for our journal */
809 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
816 if (is_bad_inode(inode)) {
817 mlog(ML_ERROR, "access error (bad inode)\n");
824 SET_INODE_JOURNAL(inode);
825 OCFS2_I(inode)->ip_open_count++;
827 /* Skip recovery waits here - journal inode metadata never
828 * changes in a live cluster so it can be considered an
829 * exception to the rule. */
830 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
832 if (status != -ERESTARTSYS)
833 mlog(ML_ERROR, "Could not get lock on journal!\n");
838 di = (struct ocfs2_dinode *)bh->b_data;
840 if (i_size_read(inode) < OCFS2_MIN_JOURNAL_SIZE) {
841 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
847 trace_ocfs2_journal_init(i_size_read(inode),
848 (unsigned long long)inode->i_blocks,
849 OCFS2_I(inode)->ip_clusters);
851 /* call the kernels journal init function now */
852 j_journal = jbd2_journal_init_inode(inode);
853 if (j_journal == NULL) {
854 mlog(ML_ERROR, "Linux journal layer error\n");
859 trace_ocfs2_journal_init_maxlen(j_journal->j_maxlen);
861 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
862 OCFS2_JOURNAL_DIRTY_FL);
864 journal->j_journal = j_journal;
865 journal->j_inode = inode;
868 ocfs2_set_journal_params(osb);
870 journal->j_state = OCFS2_JOURNAL_LOADED;
876 ocfs2_inode_unlock(inode, 1);
879 OCFS2_I(inode)->ip_open_count--;
887 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
889 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
892 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
894 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
897 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
898 int dirty, int replayed)
902 struct ocfs2_journal *journal = osb->journal;
903 struct buffer_head *bh = journal->j_bh;
904 struct ocfs2_dinode *fe;
906 fe = (struct ocfs2_dinode *)bh->b_data;
908 /* The journal bh on the osb always comes from ocfs2_journal_init()
909 * and was validated there inside ocfs2_inode_lock_full(). It's a
910 * code bug if we mess it up. */
911 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
913 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
915 flags |= OCFS2_JOURNAL_DIRTY_FL;
917 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
918 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
921 ocfs2_bump_recovery_generation(fe);
923 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
924 status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
932 * If the journal has been kmalloc'd it needs to be freed after this
935 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
937 struct ocfs2_journal *journal = NULL;
939 struct inode *inode = NULL;
940 int num_running_trans = 0;
944 journal = osb->journal;
948 inode = journal->j_inode;
950 if (journal->j_state != OCFS2_JOURNAL_LOADED)
953 /* need to inc inode use count - jbd2_journal_destroy will iput. */
957 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
958 trace_ocfs2_journal_shutdown(num_running_trans);
960 /* Do a commit_cache here. It will flush our journal, *and*
961 * release any locks that are still held.
962 * set the SHUTDOWN flag and release the trans lock.
963 * the commit thread will take the trans lock for us below. */
964 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
966 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
967 * drop the trans_lock (which we want to hold until we
968 * completely destroy the journal. */
969 if (osb->commit_task) {
970 /* Wait for the commit thread */
971 trace_ocfs2_journal_shutdown_wait(osb->commit_task);
972 kthread_stop(osb->commit_task);
973 osb->commit_task = NULL;
976 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
978 if (ocfs2_mount_local(osb)) {
979 jbd2_journal_lock_updates(journal->j_journal);
980 status = jbd2_journal_flush(journal->j_journal);
981 jbd2_journal_unlock_updates(journal->j_journal);
988 * Do not toggle if flush was unsuccessful otherwise
989 * will leave dirty metadata in a "clean" journal
991 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
996 /* Shutdown the kernel journal system */
997 jbd2_journal_destroy(journal->j_journal);
998 journal->j_journal = NULL;
1000 OCFS2_I(inode)->ip_open_count--;
1002 /* unlock our journal */
1003 ocfs2_inode_unlock(inode, 1);
1005 brelse(journal->j_bh);
1006 journal->j_bh = NULL;
1008 journal->j_state = OCFS2_JOURNAL_FREE;
1010 // up_write(&journal->j_trans_barrier);
1016 static void ocfs2_clear_journal_error(struct super_block *sb,
1022 olderr = jbd2_journal_errno(journal);
1024 mlog(ML_ERROR, "File system error %d recorded in "
1025 "journal %u.\n", olderr, slot);
1026 mlog(ML_ERROR, "File system on device %s needs checking.\n",
1029 jbd2_journal_ack_err(journal);
1030 jbd2_journal_clear_err(journal);
1034 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
1037 struct ocfs2_super *osb;
1041 osb = journal->j_osb;
1043 status = jbd2_journal_load(journal->j_journal);
1045 mlog(ML_ERROR, "Failed to load journal!\n");
1049 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1051 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1057 /* Launch the commit thread */
1059 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1061 if (IS_ERR(osb->commit_task)) {
1062 status = PTR_ERR(osb->commit_task);
1063 osb->commit_task = NULL;
1064 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1065 "error=%d", status);
1069 osb->commit_task = NULL;
1076 /* 'full' flag tells us whether we clear out all blocks or if we just
1077 * mark the journal clean */
1078 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1084 status = jbd2_journal_wipe(journal->j_journal, full);
1090 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1098 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1101 struct ocfs2_recovery_map *rm = osb->recovery_map;
1103 spin_lock(&osb->osb_lock);
1104 empty = (rm->rm_used == 0);
1105 spin_unlock(&osb->osb_lock);
1110 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1112 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1116 * JBD Might read a cached version of another nodes journal file. We
1117 * don't want this as this file changes often and we get no
1118 * notification on those changes. The only way to be sure that we've
1119 * got the most up to date version of those blocks then is to force
1120 * read them off disk. Just searching through the buffer cache won't
1121 * work as there may be pages backing this file which are still marked
1122 * up to date. We know things can't change on this file underneath us
1123 * as we have the lock by now :)
1125 static int ocfs2_force_read_journal(struct inode *inode)
1129 u64 v_blkno, p_blkno, p_blocks, num_blocks;
1130 #define CONCURRENT_JOURNAL_FILL 32ULL
1131 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
1133 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
1135 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
1137 while (v_blkno < num_blocks) {
1138 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1139 &p_blkno, &p_blocks, NULL);
1145 if (p_blocks > CONCURRENT_JOURNAL_FILL)
1146 p_blocks = CONCURRENT_JOURNAL_FILL;
1148 /* We are reading journal data which should not
1149 * be put in the uptodate cache */
1150 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
1151 p_blkno, p_blocks, bhs);
1157 for(i = 0; i < p_blocks; i++) {
1162 v_blkno += p_blocks;
1166 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
1171 struct ocfs2_la_recovery_item {
1172 struct list_head lri_list;
1174 struct ocfs2_dinode *lri_la_dinode;
1175 struct ocfs2_dinode *lri_tl_dinode;
1176 struct ocfs2_quota_recovery *lri_qrec;
1179 /* Does the second half of the recovery process. By this point, the
1180 * node is marked clean and can actually be considered recovered,
1181 * hence it's no longer in the recovery map, but there's still some
1182 * cleanup we can do which shouldn't happen within the recovery thread
1183 * as locking in that context becomes very difficult if we are to take
1184 * recovering nodes into account.
1186 * NOTE: This function can and will sleep on recovery of other nodes
1187 * during cluster locking, just like any other ocfs2 process.
1189 void ocfs2_complete_recovery(struct work_struct *work)
1192 struct ocfs2_journal *journal =
1193 container_of(work, struct ocfs2_journal, j_recovery_work);
1194 struct ocfs2_super *osb = journal->j_osb;
1195 struct ocfs2_dinode *la_dinode, *tl_dinode;
1196 struct ocfs2_la_recovery_item *item, *n;
1197 struct ocfs2_quota_recovery *qrec;
1198 LIST_HEAD(tmp_la_list);
1200 trace_ocfs2_complete_recovery(
1201 (unsigned long long)OCFS2_I(journal->j_inode)->ip_blkno);
1203 spin_lock(&journal->j_lock);
1204 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1205 spin_unlock(&journal->j_lock);
1207 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1208 list_del_init(&item->lri_list);
1210 ocfs2_wait_on_quotas(osb);
1212 la_dinode = item->lri_la_dinode;
1213 tl_dinode = item->lri_tl_dinode;
1214 qrec = item->lri_qrec;
1216 trace_ocfs2_complete_recovery_slot(item->lri_slot,
1217 la_dinode ? le64_to_cpu(la_dinode->i_blkno) : 0,
1218 tl_dinode ? le64_to_cpu(tl_dinode->i_blkno) : 0,
1222 ret = ocfs2_complete_local_alloc_recovery(osb,
1231 ret = ocfs2_complete_truncate_log_recovery(osb,
1239 ret = ocfs2_recover_orphans(osb, item->lri_slot);
1244 ret = ocfs2_finish_quota_recovery(osb, qrec,
1248 /* Recovery info is already freed now */
1254 trace_ocfs2_complete_recovery_end(ret);
1257 /* NOTE: This function always eats your references to la_dinode and
1258 * tl_dinode, either manually on error, or by passing them to
1259 * ocfs2_complete_recovery */
1260 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1262 struct ocfs2_dinode *la_dinode,
1263 struct ocfs2_dinode *tl_dinode,
1264 struct ocfs2_quota_recovery *qrec)
1266 struct ocfs2_la_recovery_item *item;
1268 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1270 /* Though we wish to avoid it, we are in fact safe in
1271 * skipping local alloc cleanup as fsck.ocfs2 is more
1272 * than capable of reclaiming unused space. */
1277 ocfs2_free_quota_recovery(qrec);
1279 mlog_errno(-ENOMEM);
1283 INIT_LIST_HEAD(&item->lri_list);
1284 item->lri_la_dinode = la_dinode;
1285 item->lri_slot = slot_num;
1286 item->lri_tl_dinode = tl_dinode;
1287 item->lri_qrec = qrec;
1289 spin_lock(&journal->j_lock);
1290 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1291 queue_work(ocfs2_wq, &journal->j_recovery_work);
1292 spin_unlock(&journal->j_lock);
1295 /* Called by the mount code to queue recovery the last part of
1296 * recovery for it's own and offline slot(s). */
1297 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1299 struct ocfs2_journal *journal = osb->journal;
1301 if (ocfs2_is_hard_readonly(osb))
1304 /* No need to queue up our truncate_log as regular cleanup will catch
1306 ocfs2_queue_recovery_completion(journal, osb->slot_num,
1307 osb->local_alloc_copy, NULL, NULL);
1308 ocfs2_schedule_truncate_log_flush(osb, 0);
1310 osb->local_alloc_copy = NULL;
1313 /* queue to recover orphan slots for all offline slots */
1314 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1315 ocfs2_queue_replay_slots(osb);
1316 ocfs2_free_replay_slots(osb);
1319 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1321 if (osb->quota_rec) {
1322 ocfs2_queue_recovery_completion(osb->journal,
1327 osb->quota_rec = NULL;
1331 static int __ocfs2_recovery_thread(void *arg)
1333 int status, node_num, slot_num;
1334 struct ocfs2_super *osb = arg;
1335 struct ocfs2_recovery_map *rm = osb->recovery_map;
1336 int *rm_quota = NULL;
1337 int rm_quota_used = 0, i;
1338 struct ocfs2_quota_recovery *qrec;
1340 status = ocfs2_wait_on_mount(osb);
1345 rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1351 status = ocfs2_super_lock(osb, 1);
1357 status = ocfs2_compute_replay_slots(osb);
1361 /* queue recovery for our own slot */
1362 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1365 spin_lock(&osb->osb_lock);
1366 while (rm->rm_used) {
1367 /* It's always safe to remove entry zero, as we won't
1368 * clear it until ocfs2_recover_node() has succeeded. */
1369 node_num = rm->rm_entries[0];
1370 spin_unlock(&osb->osb_lock);
1371 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1372 trace_ocfs2_recovery_thread_node(node_num, slot_num);
1373 if (slot_num == -ENOENT) {
1378 /* It is a bit subtle with quota recovery. We cannot do it
1379 * immediately because we have to obtain cluster locks from
1380 * quota files and we also don't want to just skip it because
1381 * then quota usage would be out of sync until some node takes
1382 * the slot. So we remember which nodes need quota recovery
1383 * and when everything else is done, we recover quotas. */
1384 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1385 if (i == rm_quota_used)
1386 rm_quota[rm_quota_used++] = slot_num;
1388 status = ocfs2_recover_node(osb, node_num, slot_num);
1391 ocfs2_recovery_map_clear(osb, node_num);
1394 "Error %d recovering node %d on device (%u,%u)!\n",
1396 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1397 mlog(ML_ERROR, "Volume requires unmount.\n");
1400 spin_lock(&osb->osb_lock);
1402 spin_unlock(&osb->osb_lock);
1403 trace_ocfs2_recovery_thread_end(status);
1405 /* Refresh all journal recovery generations from disk */
1406 status = ocfs2_check_journals_nolocks(osb);
1407 status = (status == -EROFS) ? 0 : status;
1411 /* Now it is right time to recover quotas... We have to do this under
1412 * superblock lock so that no one can start using the slot (and crash)
1413 * before we recover it */
1414 for (i = 0; i < rm_quota_used; i++) {
1415 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1417 status = PTR_ERR(qrec);
1421 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1425 ocfs2_super_unlock(osb, 1);
1427 /* queue recovery for offline slots */
1428 ocfs2_queue_replay_slots(osb);
1431 mutex_lock(&osb->recovery_lock);
1432 if (!status && !ocfs2_recovery_completed(osb)) {
1433 mutex_unlock(&osb->recovery_lock);
1437 ocfs2_free_replay_slots(osb);
1438 osb->recovery_thread_task = NULL;
1439 mb(); /* sync with ocfs2_recovery_thread_running */
1440 wake_up(&osb->recovery_event);
1442 mutex_unlock(&osb->recovery_lock);
1446 /* no one is callint kthread_stop() for us so the kthread() api
1447 * requires that we call do_exit(). And it isn't exported, but
1448 * complete_and_exit() seems to be a minimal wrapper around it. */
1449 complete_and_exit(NULL, status);
1453 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1455 mutex_lock(&osb->recovery_lock);
1457 trace_ocfs2_recovery_thread(node_num, osb->node_num,
1458 osb->disable_recovery, osb->recovery_thread_task,
1459 osb->disable_recovery ?
1460 -1 : ocfs2_recovery_map_set(osb, node_num));
1462 if (osb->disable_recovery)
1465 if (osb->recovery_thread_task)
1468 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1470 if (IS_ERR(osb->recovery_thread_task)) {
1471 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1472 osb->recovery_thread_task = NULL;
1476 mutex_unlock(&osb->recovery_lock);
1477 wake_up(&osb->recovery_event);
1480 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1482 struct buffer_head **bh,
1483 struct inode **ret_inode)
1485 int status = -EACCES;
1486 struct inode *inode = NULL;
1488 BUG_ON(slot_num >= osb->max_slots);
1490 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1492 if (!inode || is_bad_inode(inode)) {
1496 SET_INODE_JOURNAL(inode);
1498 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1508 if (status || !ret_inode)
1516 /* Does the actual journal replay and marks the journal inode as
1517 * clean. Will only replay if the journal inode is marked dirty. */
1518 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1525 struct inode *inode = NULL;
1526 struct ocfs2_dinode *fe;
1527 journal_t *journal = NULL;
1528 struct buffer_head *bh = NULL;
1531 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1537 fe = (struct ocfs2_dinode *)bh->b_data;
1538 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1543 * As the fs recovery is asynchronous, there is a small chance that
1544 * another node mounted (and recovered) the slot before the recovery
1545 * thread could get the lock. To handle that, we dirty read the journal
1546 * inode for that slot to get the recovery generation. If it is
1547 * different than what we expected, the slot has been recovered.
1548 * If not, it needs recovery.
1550 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1551 trace_ocfs2_replay_journal_recovered(slot_num,
1552 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1553 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1558 /* Continue with recovery as the journal has not yet been recovered */
1560 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1562 trace_ocfs2_replay_journal_lock_err(status);
1563 if (status != -ERESTARTSYS)
1564 mlog(ML_ERROR, "Could not lock journal!\n");
1569 fe = (struct ocfs2_dinode *) bh->b_data;
1571 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1572 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1574 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1575 trace_ocfs2_replay_journal_skip(node_num);
1576 /* Refresh recovery generation for the slot */
1577 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1581 /* we need to run complete recovery for offline orphan slots */
1582 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1584 printk(KERN_NOTICE "ocfs2: Begin replay journal (node %d, slot %d) on "\
1585 "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1586 MINOR(osb->sb->s_dev));
1588 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1590 status = ocfs2_force_read_journal(inode);
1596 journal = jbd2_journal_init_inode(inode);
1597 if (journal == NULL) {
1598 mlog(ML_ERROR, "Linux journal layer error\n");
1603 status = jbd2_journal_load(journal);
1608 jbd2_journal_destroy(journal);
1612 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1614 /* wipe the journal */
1615 jbd2_journal_lock_updates(journal);
1616 status = jbd2_journal_flush(journal);
1617 jbd2_journal_unlock_updates(journal);
1621 /* This will mark the node clean */
1622 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1623 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1624 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1626 /* Increment recovery generation to indicate successful recovery */
1627 ocfs2_bump_recovery_generation(fe);
1628 osb->slot_recovery_generations[slot_num] =
1629 ocfs2_get_recovery_generation(fe);
1631 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1632 status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1639 jbd2_journal_destroy(journal);
1641 printk(KERN_NOTICE "ocfs2: End replay journal (node %d, slot %d) on "\
1642 "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1643 MINOR(osb->sb->s_dev));
1645 /* drop the lock on this nodes journal */
1647 ocfs2_inode_unlock(inode, 1);
1658 * Do the most important parts of node recovery:
1659 * - Replay it's journal
1660 * - Stamp a clean local allocator file
1661 * - Stamp a clean truncate log
1662 * - Mark the node clean
1664 * If this function completes without error, a node in OCFS2 can be
1665 * said to have been safely recovered. As a result, failure during the
1666 * second part of a nodes recovery process (local alloc recovery) is
1667 * far less concerning.
1669 static int ocfs2_recover_node(struct ocfs2_super *osb,
1670 int node_num, int slot_num)
1673 struct ocfs2_dinode *la_copy = NULL;
1674 struct ocfs2_dinode *tl_copy = NULL;
1676 trace_ocfs2_recover_node(node_num, slot_num, osb->node_num);
1678 /* Should not ever be called to recover ourselves -- in that
1679 * case we should've called ocfs2_journal_load instead. */
1680 BUG_ON(osb->node_num == node_num);
1682 status = ocfs2_replay_journal(osb, node_num, slot_num);
1684 if (status == -EBUSY) {
1685 trace_ocfs2_recover_node_skip(slot_num, node_num);
1693 /* Stamp a clean local alloc file AFTER recovering the journal... */
1694 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1700 /* An error from begin_truncate_log_recovery is not
1701 * serious enough to warrant halting the rest of
1703 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1707 /* Likewise, this would be a strange but ultimately not so
1708 * harmful place to get an error... */
1709 status = ocfs2_clear_slot(osb, slot_num);
1713 /* This will kfree the memory pointed to by la_copy and tl_copy */
1714 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1723 /* Test node liveness by trylocking his journal. If we get the lock,
1724 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1725 * still alive (we couldn't get the lock) and < 0 on error. */
1726 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1730 struct inode *inode = NULL;
1732 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1734 if (inode == NULL) {
1735 mlog(ML_ERROR, "access error\n");
1739 if (is_bad_inode(inode)) {
1740 mlog(ML_ERROR, "access error (bad inode)\n");
1746 SET_INODE_JOURNAL(inode);
1748 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1749 status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1751 if (status != -EAGAIN)
1756 ocfs2_inode_unlock(inode, 1);
1764 /* Call this underneath ocfs2_super_lock. It also assumes that the
1765 * slot info struct has been updated from disk. */
1766 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1768 unsigned int node_num;
1771 struct buffer_head *bh = NULL;
1772 struct ocfs2_dinode *di;
1774 /* This is called with the super block cluster lock, so we
1775 * know that the slot map can't change underneath us. */
1777 for (i = 0; i < osb->max_slots; i++) {
1778 /* Read journal inode to get the recovery generation */
1779 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1784 di = (struct ocfs2_dinode *)bh->b_data;
1785 gen = ocfs2_get_recovery_generation(di);
1789 spin_lock(&osb->osb_lock);
1790 osb->slot_recovery_generations[i] = gen;
1792 trace_ocfs2_mark_dead_nodes(i,
1793 osb->slot_recovery_generations[i]);
1795 if (i == osb->slot_num) {
1796 spin_unlock(&osb->osb_lock);
1800 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1801 if (status == -ENOENT) {
1802 spin_unlock(&osb->osb_lock);
1806 if (__ocfs2_recovery_map_test(osb, node_num)) {
1807 spin_unlock(&osb->osb_lock);
1810 spin_unlock(&osb->osb_lock);
1812 /* Ok, we have a slot occupied by another node which
1813 * is not in the recovery map. We trylock his journal
1814 * file here to test if he's alive. */
1815 status = ocfs2_trylock_journal(osb, i);
1817 /* Since we're called from mount, we know that
1818 * the recovery thread can't race us on
1819 * setting / checking the recovery bits. */
1820 ocfs2_recovery_thread(osb, node_num);
1821 } else if ((status < 0) && (status != -EAGAIN)) {
1833 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1834 * randomness to the timeout to minimize multple nodes firing the timer at the
1837 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1841 get_random_bytes(&time, sizeof(time));
1842 time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1843 return msecs_to_jiffies(time);
1847 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1848 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1849 * is done to catch any orphans that are left over in orphan directories.
1851 * It scans all slots, even ones that are in use. It does so to handle the
1852 * case described below:
1854 * Node 1 has an inode it was using. The dentry went away due to memory
1855 * pressure. Node 1 closes the inode, but it's on the free list. The node
1856 * has the open lock.
1857 * Node 2 unlinks the inode. It grabs the dentry lock to notify others,
1858 * but node 1 has no dentry and doesn't get the message. It trylocks the
1859 * open lock, sees that another node has a PR, and does nothing.
1860 * Later node 2 runs its orphan dir. It igets the inode, trylocks the
1861 * open lock, sees the PR still, and does nothing.
1862 * Basically, we have to trigger an orphan iput on node 1. The only way
1863 * for this to happen is if node 1 runs node 2's orphan dir.
1865 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1866 * seconds. It gets an EX lock on os_lockres and checks sequence number
1867 * stored in LVB. If the sequence number has changed, it means some other
1868 * node has done the scan. This node skips the scan and tracks the
1869 * sequence number. If the sequence number didn't change, it means a scan
1870 * hasn't happened. The node queues a scan and increments the
1871 * sequence number in the LVB.
1873 void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1875 struct ocfs2_orphan_scan *os;
1879 os = &osb->osb_orphan_scan;
1881 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1884 trace_ocfs2_queue_orphan_scan_begin(os->os_count, os->os_seqno,
1885 atomic_read(&os->os_state));
1887 status = ocfs2_orphan_scan_lock(osb, &seqno);
1889 if (status != -EAGAIN)
1894 /* Do no queue the tasks if the volume is being umounted */
1895 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1898 if (os->os_seqno != seqno) {
1899 os->os_seqno = seqno;
1903 for (i = 0; i < osb->max_slots; i++)
1904 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1907 * We queued a recovery on orphan slots, increment the sequence
1908 * number and update LVB so other node will skip the scan for a while
1912 os->os_scantime = CURRENT_TIME;
1914 ocfs2_orphan_scan_unlock(osb, seqno);
1916 trace_ocfs2_queue_orphan_scan_end(os->os_count, os->os_seqno,
1917 atomic_read(&os->os_state));
1921 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1922 void ocfs2_orphan_scan_work(struct work_struct *work)
1924 struct ocfs2_orphan_scan *os;
1925 struct ocfs2_super *osb;
1927 os = container_of(work, struct ocfs2_orphan_scan,
1928 os_orphan_scan_work.work);
1931 mutex_lock(&os->os_lock);
1932 ocfs2_queue_orphan_scan(osb);
1933 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1934 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1935 ocfs2_orphan_scan_timeout());
1936 mutex_unlock(&os->os_lock);
1939 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1941 struct ocfs2_orphan_scan *os;
1943 os = &osb->osb_orphan_scan;
1944 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1945 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1946 mutex_lock(&os->os_lock);
1947 cancel_delayed_work(&os->os_orphan_scan_work);
1948 mutex_unlock(&os->os_lock);
1952 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1954 struct ocfs2_orphan_scan *os;
1956 os = &osb->osb_orphan_scan;
1960 mutex_init(&os->os_lock);
1961 INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
1964 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
1966 struct ocfs2_orphan_scan *os;
1968 os = &osb->osb_orphan_scan;
1969 os->os_scantime = CURRENT_TIME;
1970 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1971 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1973 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
1974 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1975 ocfs2_orphan_scan_timeout());
1979 struct ocfs2_orphan_filldir_priv {
1980 struct dir_context ctx;
1982 struct ocfs2_super *osb;
1985 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1986 loff_t pos, u64 ino, unsigned type)
1988 struct ocfs2_orphan_filldir_priv *p = priv;
1991 if (name_len == 1 && !strncmp(".", name, 1))
1993 if (name_len == 2 && !strncmp("..", name, 2))
1996 /* Skip bad inodes so that recovery can continue */
1997 iter = ocfs2_iget(p->osb, ino,
1998 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
2002 trace_ocfs2_orphan_filldir((unsigned long long)OCFS2_I(iter)->ip_blkno);
2003 /* No locking is required for the next_orphan queue as there
2004 * is only ever a single process doing orphan recovery. */
2005 OCFS2_I(iter)->ip_next_orphan = p->head;
2011 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
2013 struct inode **head)
2016 struct inode *orphan_dir_inode = NULL;
2017 struct ocfs2_orphan_filldir_priv priv = {
2018 .ctx.actor = ocfs2_orphan_filldir,
2023 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2024 ORPHAN_DIR_SYSTEM_INODE,
2026 if (!orphan_dir_inode) {
2032 mutex_lock(&orphan_dir_inode->i_mutex);
2033 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2039 status = ocfs2_dir_foreach(orphan_dir_inode, &priv.ctx);
2048 ocfs2_inode_unlock(orphan_dir_inode, 0);
2050 mutex_unlock(&orphan_dir_inode->i_mutex);
2051 iput(orphan_dir_inode);
2055 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2060 spin_lock(&osb->osb_lock);
2061 ret = !osb->osb_orphan_wipes[slot];
2062 spin_unlock(&osb->osb_lock);
2066 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2069 spin_lock(&osb->osb_lock);
2070 /* Mark ourselves such that new processes in delete_inode()
2071 * know to quit early. */
2072 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2073 while (osb->osb_orphan_wipes[slot]) {
2074 /* If any processes are already in the middle of an
2075 * orphan wipe on this dir, then we need to wait for
2077 spin_unlock(&osb->osb_lock);
2078 wait_event_interruptible(osb->osb_wipe_event,
2079 ocfs2_orphan_recovery_can_continue(osb, slot));
2080 spin_lock(&osb->osb_lock);
2082 spin_unlock(&osb->osb_lock);
2085 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2088 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2092 * Orphan recovery. Each mounted node has it's own orphan dir which we
2093 * must run during recovery. Our strategy here is to build a list of
2094 * the inodes in the orphan dir and iget/iput them. The VFS does
2095 * (most) of the rest of the work.
2097 * Orphan recovery can happen at any time, not just mount so we have a
2098 * couple of extra considerations.
2100 * - We grab as many inodes as we can under the orphan dir lock -
2101 * doing iget() outside the orphan dir risks getting a reference on
2103 * - We must be sure not to deadlock with other processes on the
2104 * system wanting to run delete_inode(). This can happen when they go
2105 * to lock the orphan dir and the orphan recovery process attempts to
2106 * iget() inside the orphan dir lock. This can be avoided by
2107 * advertising our state to ocfs2_delete_inode().
2109 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2113 struct inode *inode = NULL;
2115 struct ocfs2_inode_info *oi;
2117 trace_ocfs2_recover_orphans(slot);
2119 ocfs2_mark_recovering_orphan_dir(osb, slot);
2120 ret = ocfs2_queue_orphans(osb, slot, &inode);
2121 ocfs2_clear_recovering_orphan_dir(osb, slot);
2123 /* Error here should be noted, but we want to continue with as
2124 * many queued inodes as we've got. */
2129 oi = OCFS2_I(inode);
2130 trace_ocfs2_recover_orphans_iput(
2131 (unsigned long long)oi->ip_blkno);
2133 iter = oi->ip_next_orphan;
2135 spin_lock(&oi->ip_lock);
2136 /* Set the proper information to get us going into
2137 * ocfs2_delete_inode. */
2138 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2139 spin_unlock(&oi->ip_lock);
2149 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2151 /* This check is good because ocfs2 will wait on our recovery
2152 * thread before changing it to something other than MOUNTED
2154 wait_event(osb->osb_mount_event,
2155 (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2156 atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2157 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2159 /* If there's an error on mount, then we may never get to the
2160 * MOUNTED flag, but this is set right before
2161 * dismount_volume() so we can trust it. */
2162 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2163 trace_ocfs2_wait_on_mount(VOLUME_DISABLED);
2164 mlog(0, "mount error, exiting!\n");
2171 static int ocfs2_commit_thread(void *arg)
2174 struct ocfs2_super *osb = arg;
2175 struct ocfs2_journal *journal = osb->journal;
2177 /* we can trust j_num_trans here because _should_stop() is only set in
2178 * shutdown and nobody other than ourselves should be able to start
2179 * transactions. committing on shutdown might take a few iterations
2180 * as final transactions put deleted inodes on the list */
2181 while (!(kthread_should_stop() &&
2182 atomic_read(&journal->j_num_trans) == 0)) {
2184 wait_event_interruptible(osb->checkpoint_event,
2185 atomic_read(&journal->j_num_trans)
2186 || kthread_should_stop());
2188 status = ocfs2_commit_cache(osb);
2190 static unsigned long abort_warn_time;
2192 /* Warn about this once per minute */
2193 if (printk_timed_ratelimit(&abort_warn_time, 60*HZ))
2194 mlog(ML_ERROR, "status = %d, journal is "
2195 "already aborted.\n", status);
2197 * After ocfs2_commit_cache() fails, j_num_trans has a
2198 * non-zero value. Sleep here to avoid a busy-wait
2201 msleep_interruptible(1000);
2204 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2206 "commit_thread: %u transactions pending on "
2208 atomic_read(&journal->j_num_trans));
2215 /* Reads all the journal inodes without taking any cluster locks. Used
2216 * for hard readonly access to determine whether any journal requires
2217 * recovery. Also used to refresh the recovery generation numbers after
2218 * a journal has been recovered by another node.
2220 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2224 struct buffer_head *di_bh = NULL;
2225 struct ocfs2_dinode *di;
2226 int journal_dirty = 0;
2228 for(slot = 0; slot < osb->max_slots; slot++) {
2229 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2235 di = (struct ocfs2_dinode *) di_bh->b_data;
2237 osb->slot_recovery_generations[slot] =
2238 ocfs2_get_recovery_generation(di);
2240 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2241 OCFS2_JOURNAL_DIRTY_FL)