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"
56 #include "buffer_head_io.h"
57 #include "ocfs2_trace.h"
59 DEFINE_SPINLOCK(trans_inc_lock);
61 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
63 static int ocfs2_force_read_journal(struct inode *inode);
64 static int ocfs2_recover_node(struct ocfs2_super *osb,
65 int node_num, int slot_num);
66 static int __ocfs2_recovery_thread(void *arg);
67 static int ocfs2_commit_cache(struct ocfs2_super *osb);
68 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota);
69 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
70 int dirty, int replayed);
71 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
73 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
75 enum ocfs2_orphan_reco_type orphan_reco_type);
76 static int ocfs2_commit_thread(void *arg);
77 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
79 struct ocfs2_dinode *la_dinode,
80 struct ocfs2_dinode *tl_dinode,
81 struct ocfs2_quota_recovery *qrec,
82 enum ocfs2_orphan_reco_type orphan_reco_type);
84 static inline int ocfs2_wait_on_mount(struct ocfs2_super *osb)
86 return __ocfs2_wait_on_mount(osb, 0);
89 static inline int ocfs2_wait_on_quotas(struct ocfs2_super *osb)
91 return __ocfs2_wait_on_mount(osb, 1);
95 * This replay_map is to track online/offline slots, so we could recover
96 * offline slots during recovery and mount
99 enum ocfs2_replay_state {
100 REPLAY_UNNEEDED = 0, /* Replay is not needed, so ignore this map */
101 REPLAY_NEEDED, /* Replay slots marked in rm_replay_slots */
102 REPLAY_DONE /* Replay was already queued */
105 struct ocfs2_replay_map {
106 unsigned int rm_slots;
107 enum ocfs2_replay_state rm_state;
108 unsigned char rm_replay_slots[0];
111 void ocfs2_replay_map_set_state(struct ocfs2_super *osb, int state)
113 if (!osb->replay_map)
116 /* If we've already queued the replay, we don't have any more to do */
117 if (osb->replay_map->rm_state == REPLAY_DONE)
120 osb->replay_map->rm_state = state;
123 int ocfs2_compute_replay_slots(struct ocfs2_super *osb)
125 struct ocfs2_replay_map *replay_map;
128 /* If replay map is already set, we don't do it again */
132 replay_map = kzalloc(sizeof(struct ocfs2_replay_map) +
133 (osb->max_slots * sizeof(char)), GFP_KERNEL);
140 spin_lock(&osb->osb_lock);
142 replay_map->rm_slots = osb->max_slots;
143 replay_map->rm_state = REPLAY_UNNEEDED;
145 /* set rm_replay_slots for offline slot(s) */
146 for (i = 0; i < replay_map->rm_slots; i++) {
147 if (ocfs2_slot_to_node_num_locked(osb, i, &node_num) == -ENOENT)
148 replay_map->rm_replay_slots[i] = 1;
151 osb->replay_map = replay_map;
152 spin_unlock(&osb->osb_lock);
156 void ocfs2_queue_replay_slots(struct ocfs2_super *osb,
157 enum ocfs2_orphan_reco_type orphan_reco_type)
159 struct ocfs2_replay_map *replay_map = osb->replay_map;
165 if (replay_map->rm_state != REPLAY_NEEDED)
168 for (i = 0; i < replay_map->rm_slots; i++)
169 if (replay_map->rm_replay_slots[i])
170 ocfs2_queue_recovery_completion(osb->journal, i, NULL,
173 replay_map->rm_state = REPLAY_DONE;
176 void ocfs2_free_replay_slots(struct ocfs2_super *osb)
178 struct ocfs2_replay_map *replay_map = osb->replay_map;
180 if (!osb->replay_map)
184 osb->replay_map = NULL;
187 int ocfs2_recovery_init(struct ocfs2_super *osb)
189 struct ocfs2_recovery_map *rm;
191 mutex_init(&osb->recovery_lock);
192 osb->disable_recovery = 0;
193 osb->recovery_thread_task = NULL;
194 init_waitqueue_head(&osb->recovery_event);
196 rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
197 osb->max_slots * sizeof(unsigned int),
204 rm->rm_entries = (unsigned int *)((char *)rm +
205 sizeof(struct ocfs2_recovery_map));
206 osb->recovery_map = rm;
211 /* we can't grab the goofy sem lock from inside wait_event, so we use
212 * memory barriers to make sure that we'll see the null task before
214 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
217 return osb->recovery_thread_task != NULL;
220 void ocfs2_recovery_exit(struct ocfs2_super *osb)
222 struct ocfs2_recovery_map *rm;
224 /* disable any new recovery threads and wait for any currently
225 * running ones to exit. Do this before setting the vol_state. */
226 mutex_lock(&osb->recovery_lock);
227 osb->disable_recovery = 1;
228 mutex_unlock(&osb->recovery_lock);
229 wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
231 /* At this point, we know that no more recovery threads can be
232 * launched, so wait for any recovery completion work to
234 flush_workqueue(ocfs2_wq);
237 * Now that recovery is shut down, and the osb is about to be
238 * freed, the osb_lock is not taken here.
240 rm = osb->recovery_map;
241 /* XXX: Should we bug if there are dirty entries? */
246 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
247 unsigned int node_num)
250 struct ocfs2_recovery_map *rm = osb->recovery_map;
252 assert_spin_locked(&osb->osb_lock);
254 for (i = 0; i < rm->rm_used; i++) {
255 if (rm->rm_entries[i] == node_num)
262 /* Behaves like test-and-set. Returns the previous value */
263 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
264 unsigned int node_num)
266 struct ocfs2_recovery_map *rm = osb->recovery_map;
268 spin_lock(&osb->osb_lock);
269 if (__ocfs2_recovery_map_test(osb, node_num)) {
270 spin_unlock(&osb->osb_lock);
274 /* XXX: Can this be exploited? Not from o2dlm... */
275 BUG_ON(rm->rm_used >= osb->max_slots);
277 rm->rm_entries[rm->rm_used] = node_num;
279 spin_unlock(&osb->osb_lock);
284 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
285 unsigned int node_num)
288 struct ocfs2_recovery_map *rm = osb->recovery_map;
290 spin_lock(&osb->osb_lock);
292 for (i = 0; i < rm->rm_used; i++) {
293 if (rm->rm_entries[i] == node_num)
297 if (i < rm->rm_used) {
298 /* XXX: be careful with the pointer math */
299 memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
300 (rm->rm_used - i - 1) * sizeof(unsigned int));
304 spin_unlock(&osb->osb_lock);
307 static int ocfs2_commit_cache(struct ocfs2_super *osb)
310 unsigned int flushed;
311 struct ocfs2_journal *journal = NULL;
313 journal = osb->journal;
315 /* Flush all pending commits and checkpoint the journal. */
316 down_write(&journal->j_trans_barrier);
318 flushed = atomic_read(&journal->j_num_trans);
319 trace_ocfs2_commit_cache_begin(flushed);
321 up_write(&journal->j_trans_barrier);
325 jbd2_journal_lock_updates(journal->j_journal);
326 status = jbd2_journal_flush(journal->j_journal);
327 jbd2_journal_unlock_updates(journal->j_journal);
329 up_write(&journal->j_trans_barrier);
334 ocfs2_inc_trans_id(journal);
336 flushed = atomic_read(&journal->j_num_trans);
337 atomic_set(&journal->j_num_trans, 0);
338 up_write(&journal->j_trans_barrier);
340 trace_ocfs2_commit_cache_end(journal->j_trans_id, flushed);
342 ocfs2_wake_downconvert_thread(osb);
343 wake_up(&journal->j_checkpointed);
348 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
350 journal_t *journal = osb->journal->j_journal;
353 BUG_ON(!osb || !osb->journal->j_journal);
355 if (ocfs2_is_hard_readonly(osb))
356 return ERR_PTR(-EROFS);
358 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
359 BUG_ON(max_buffs <= 0);
361 /* Nested transaction? Just return the handle... */
362 if (journal_current_handle())
363 return jbd2_journal_start(journal, max_buffs);
365 sb_start_intwrite(osb->sb);
367 down_read(&osb->journal->j_trans_barrier);
369 handle = jbd2_journal_start(journal, max_buffs);
370 if (IS_ERR(handle)) {
371 up_read(&osb->journal->j_trans_barrier);
372 sb_end_intwrite(osb->sb);
374 mlog_errno(PTR_ERR(handle));
376 if (is_journal_aborted(journal)) {
377 ocfs2_abort(osb->sb, "Detected aborted journal");
378 handle = ERR_PTR(-EROFS);
381 if (!ocfs2_mount_local(osb))
382 atomic_inc(&(osb->journal->j_num_trans));
388 int ocfs2_commit_trans(struct ocfs2_super *osb,
392 struct ocfs2_journal *journal = osb->journal;
396 nested = handle->h_ref > 1;
397 ret = jbd2_journal_stop(handle);
402 up_read(&journal->j_trans_barrier);
403 sb_end_intwrite(osb->sb);
410 * 'nblocks' is what you want to add to the current transaction.
412 * This might call jbd2_journal_restart() which will commit dirty buffers
413 * and then restart the transaction. Before calling
414 * ocfs2_extend_trans(), any changed blocks should have been
415 * dirtied. After calling it, all blocks which need to be changed must
416 * go through another set of journal_access/journal_dirty calls.
418 * WARNING: This will not release any semaphores or disk locks taken
419 * during the transaction, so make sure they were taken *before*
420 * start_trans or we'll have ordering deadlocks.
422 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
423 * good because transaction ids haven't yet been recorded on the
424 * cluster locks associated with this handle.
426 int ocfs2_extend_trans(handle_t *handle, int nblocks)
428 int status, old_nblocks;
436 old_nblocks = handle->h_buffer_credits;
438 trace_ocfs2_extend_trans(old_nblocks, nblocks);
440 #ifdef CONFIG_OCFS2_DEBUG_FS
443 status = jbd2_journal_extend(handle, nblocks);
451 trace_ocfs2_extend_trans_restart(old_nblocks + nblocks);
452 status = jbd2_journal_restart(handle,
453 old_nblocks + nblocks);
466 * If we have fewer than thresh credits, extend by OCFS2_MAX_TRANS_DATA.
467 * If that fails, restart the transaction & regain write access for the
468 * buffer head which is used for metadata modifications.
469 * Taken from Ext4: extend_or_restart_transaction()
471 int ocfs2_allocate_extend_trans(handle_t *handle, int thresh)
473 int status, old_nblks;
477 old_nblks = handle->h_buffer_credits;
478 trace_ocfs2_allocate_extend_trans(old_nblks, thresh);
480 if (old_nblks < thresh)
483 status = jbd2_journal_extend(handle, OCFS2_MAX_TRANS_DATA);
490 status = jbd2_journal_restart(handle, OCFS2_MAX_TRANS_DATA);
500 struct ocfs2_triggers {
501 struct jbd2_buffer_trigger_type ot_triggers;
505 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
507 return container_of(triggers, struct ocfs2_triggers, ot_triggers);
510 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
511 struct buffer_head *bh,
512 void *data, size_t size)
514 struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
517 * We aren't guaranteed to have the superblock here, so we
518 * must unconditionally compute the ecc data.
519 * __ocfs2_journal_access() will only set the triggers if
520 * metaecc is enabled.
522 ocfs2_block_check_compute(data, size, data + ot->ot_offset);
526 * Quota blocks have their own trigger because the struct ocfs2_block_check
527 * offset depends on the blocksize.
529 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
530 struct buffer_head *bh,
531 void *data, size_t size)
533 struct ocfs2_disk_dqtrailer *dqt =
534 ocfs2_block_dqtrailer(size, data);
537 * We aren't guaranteed to have the superblock here, so we
538 * must unconditionally compute the ecc data.
539 * __ocfs2_journal_access() will only set the triggers if
540 * metaecc is enabled.
542 ocfs2_block_check_compute(data, size, &dqt->dq_check);
546 * Directory blocks also have their own trigger because the
547 * struct ocfs2_block_check offset depends on the blocksize.
549 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
550 struct buffer_head *bh,
551 void *data, size_t size)
553 struct ocfs2_dir_block_trailer *trailer =
554 ocfs2_dir_trailer_from_size(size, data);
557 * We aren't guaranteed to have the superblock here, so we
558 * must unconditionally compute the ecc data.
559 * __ocfs2_journal_access() will only set the triggers if
560 * metaecc is enabled.
562 ocfs2_block_check_compute(data, size, &trailer->db_check);
565 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
566 struct buffer_head *bh)
569 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
570 "bh->b_blocknr = %llu\n",
572 (unsigned long long)bh->b_blocknr);
574 /* We aren't guaranteed to have the superblock here - but if we
575 * don't, it'll just crash. */
576 ocfs2_error(bh->b_assoc_map->host->i_sb,
577 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
580 static struct ocfs2_triggers di_triggers = {
582 .t_frozen = ocfs2_frozen_trigger,
583 .t_abort = ocfs2_abort_trigger,
585 .ot_offset = offsetof(struct ocfs2_dinode, i_check),
588 static struct ocfs2_triggers eb_triggers = {
590 .t_frozen = ocfs2_frozen_trigger,
591 .t_abort = ocfs2_abort_trigger,
593 .ot_offset = offsetof(struct ocfs2_extent_block, h_check),
596 static struct ocfs2_triggers rb_triggers = {
598 .t_frozen = ocfs2_frozen_trigger,
599 .t_abort = ocfs2_abort_trigger,
601 .ot_offset = offsetof(struct ocfs2_refcount_block, rf_check),
604 static struct ocfs2_triggers gd_triggers = {
606 .t_frozen = ocfs2_frozen_trigger,
607 .t_abort = ocfs2_abort_trigger,
609 .ot_offset = offsetof(struct ocfs2_group_desc, bg_check),
612 static struct ocfs2_triggers db_triggers = {
614 .t_frozen = ocfs2_db_frozen_trigger,
615 .t_abort = ocfs2_abort_trigger,
619 static struct ocfs2_triggers xb_triggers = {
621 .t_frozen = ocfs2_frozen_trigger,
622 .t_abort = ocfs2_abort_trigger,
624 .ot_offset = offsetof(struct ocfs2_xattr_block, xb_check),
627 static struct ocfs2_triggers dq_triggers = {
629 .t_frozen = ocfs2_dq_frozen_trigger,
630 .t_abort = ocfs2_abort_trigger,
634 static struct ocfs2_triggers dr_triggers = {
636 .t_frozen = ocfs2_frozen_trigger,
637 .t_abort = ocfs2_abort_trigger,
639 .ot_offset = offsetof(struct ocfs2_dx_root_block, dr_check),
642 static struct ocfs2_triggers dl_triggers = {
644 .t_frozen = ocfs2_frozen_trigger,
645 .t_abort = ocfs2_abort_trigger,
647 .ot_offset = offsetof(struct ocfs2_dx_leaf, dl_check),
650 static int __ocfs2_journal_access(handle_t *handle,
651 struct ocfs2_caching_info *ci,
652 struct buffer_head *bh,
653 struct ocfs2_triggers *triggers,
657 struct ocfs2_super *osb =
658 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
660 BUG_ON(!ci || !ci->ci_ops);
664 trace_ocfs2_journal_access(
665 (unsigned long long)ocfs2_metadata_cache_owner(ci),
666 (unsigned long long)bh->b_blocknr, type, bh->b_size);
668 /* we can safely remove this assertion after testing. */
669 if (!buffer_uptodate(bh)) {
670 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
671 mlog(ML_ERROR, "b_blocknr=%llu\n",
672 (unsigned long long)bh->b_blocknr);
676 /* Set the current transaction information on the ci so
677 * that the locking code knows whether it can drop it's locks
678 * on this ci or not. We're protected from the commit
679 * thread updating the current transaction id until
680 * ocfs2_commit_trans() because ocfs2_start_trans() took
681 * j_trans_barrier for us. */
682 ocfs2_set_ci_lock_trans(osb->journal, ci);
684 ocfs2_metadata_cache_io_lock(ci);
686 case OCFS2_JOURNAL_ACCESS_CREATE:
687 case OCFS2_JOURNAL_ACCESS_WRITE:
688 status = jbd2_journal_get_write_access(handle, bh);
691 case OCFS2_JOURNAL_ACCESS_UNDO:
692 status = jbd2_journal_get_undo_access(handle, bh);
697 mlog(ML_ERROR, "Unknown access type!\n");
699 if (!status && ocfs2_meta_ecc(osb) && triggers)
700 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
701 ocfs2_metadata_cache_io_unlock(ci);
704 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
710 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
711 struct buffer_head *bh, int type)
713 return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
716 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
717 struct buffer_head *bh, int type)
719 return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
722 int ocfs2_journal_access_rb(handle_t *handle, struct ocfs2_caching_info *ci,
723 struct buffer_head *bh, int type)
725 return __ocfs2_journal_access(handle, ci, bh, &rb_triggers,
729 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
730 struct buffer_head *bh, int type)
732 return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
735 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
736 struct buffer_head *bh, int type)
738 return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
741 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
742 struct buffer_head *bh, int type)
744 return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
747 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
748 struct buffer_head *bh, int type)
750 return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
753 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
754 struct buffer_head *bh, int type)
756 return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
759 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
760 struct buffer_head *bh, int type)
762 return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
765 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
766 struct buffer_head *bh, int type)
768 return __ocfs2_journal_access(handle, ci, bh, NULL, type);
771 void ocfs2_journal_dirty(handle_t *handle, struct buffer_head *bh)
775 trace_ocfs2_journal_dirty((unsigned long long)bh->b_blocknr);
777 status = jbd2_journal_dirty_metadata(handle, bh);
781 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
783 void ocfs2_set_journal_params(struct ocfs2_super *osb)
785 journal_t *journal = osb->journal->j_journal;
786 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
788 if (osb->osb_commit_interval)
789 commit_interval = osb->osb_commit_interval;
791 write_lock(&journal->j_state_lock);
792 journal->j_commit_interval = commit_interval;
793 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
794 journal->j_flags |= JBD2_BARRIER;
796 journal->j_flags &= ~JBD2_BARRIER;
797 write_unlock(&journal->j_state_lock);
800 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
803 struct inode *inode = NULL; /* the journal inode */
804 journal_t *j_journal = NULL;
805 struct ocfs2_dinode *di = NULL;
806 struct buffer_head *bh = NULL;
807 struct ocfs2_super *osb;
812 osb = journal->j_osb;
814 /* already have the inode for our journal */
815 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
822 if (is_bad_inode(inode)) {
823 mlog(ML_ERROR, "access error (bad inode)\n");
830 SET_INODE_JOURNAL(inode);
831 OCFS2_I(inode)->ip_open_count++;
833 /* Skip recovery waits here - journal inode metadata never
834 * changes in a live cluster so it can be considered an
835 * exception to the rule. */
836 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
838 if (status != -ERESTARTSYS)
839 mlog(ML_ERROR, "Could not get lock on journal!\n");
844 di = (struct ocfs2_dinode *)bh->b_data;
846 if (i_size_read(inode) < OCFS2_MIN_JOURNAL_SIZE) {
847 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
853 trace_ocfs2_journal_init(i_size_read(inode),
854 (unsigned long long)inode->i_blocks,
855 OCFS2_I(inode)->ip_clusters);
857 /* call the kernels journal init function now */
858 j_journal = jbd2_journal_init_inode(inode);
859 if (j_journal == NULL) {
860 mlog(ML_ERROR, "Linux journal layer error\n");
865 trace_ocfs2_journal_init_maxlen(j_journal->j_maxlen);
867 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
868 OCFS2_JOURNAL_DIRTY_FL);
870 journal->j_journal = j_journal;
871 journal->j_inode = inode;
874 ocfs2_set_journal_params(osb);
876 journal->j_state = OCFS2_JOURNAL_LOADED;
882 ocfs2_inode_unlock(inode, 1);
885 OCFS2_I(inode)->ip_open_count--;
893 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
895 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
898 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
900 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
903 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
904 int dirty, int replayed)
908 struct ocfs2_journal *journal = osb->journal;
909 struct buffer_head *bh = journal->j_bh;
910 struct ocfs2_dinode *fe;
912 fe = (struct ocfs2_dinode *)bh->b_data;
914 /* The journal bh on the osb always comes from ocfs2_journal_init()
915 * and was validated there inside ocfs2_inode_lock_full(). It's a
916 * code bug if we mess it up. */
917 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
919 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
921 flags |= OCFS2_JOURNAL_DIRTY_FL;
923 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
924 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
927 ocfs2_bump_recovery_generation(fe);
929 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
930 status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
938 * If the journal has been kmalloc'd it needs to be freed after this
941 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
943 struct ocfs2_journal *journal = NULL;
945 struct inode *inode = NULL;
946 int num_running_trans = 0;
950 journal = osb->journal;
954 inode = journal->j_inode;
956 if (journal->j_state != OCFS2_JOURNAL_LOADED)
959 /* need to inc inode use count - jbd2_journal_destroy will iput. */
963 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
964 trace_ocfs2_journal_shutdown(num_running_trans);
966 /* Do a commit_cache here. It will flush our journal, *and*
967 * release any locks that are still held.
968 * set the SHUTDOWN flag and release the trans lock.
969 * the commit thread will take the trans lock for us below. */
970 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
972 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
973 * drop the trans_lock (which we want to hold until we
974 * completely destroy the journal. */
975 if (osb->commit_task) {
976 /* Wait for the commit thread */
977 trace_ocfs2_journal_shutdown_wait(osb->commit_task);
978 kthread_stop(osb->commit_task);
979 osb->commit_task = NULL;
982 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
984 if (ocfs2_mount_local(osb)) {
985 jbd2_journal_lock_updates(journal->j_journal);
986 status = jbd2_journal_flush(journal->j_journal);
987 jbd2_journal_unlock_updates(journal->j_journal);
994 * Do not toggle if flush was unsuccessful otherwise
995 * will leave dirty metadata in a "clean" journal
997 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
1002 /* Shutdown the kernel journal system */
1003 jbd2_journal_destroy(journal->j_journal);
1004 journal->j_journal = NULL;
1006 OCFS2_I(inode)->ip_open_count--;
1008 /* unlock our journal */
1009 ocfs2_inode_unlock(inode, 1);
1011 brelse(journal->j_bh);
1012 journal->j_bh = NULL;
1014 journal->j_state = OCFS2_JOURNAL_FREE;
1016 // up_write(&journal->j_trans_barrier);
1022 static void ocfs2_clear_journal_error(struct super_block *sb,
1028 olderr = jbd2_journal_errno(journal);
1030 mlog(ML_ERROR, "File system error %d recorded in "
1031 "journal %u.\n", olderr, slot);
1032 mlog(ML_ERROR, "File system on device %s needs checking.\n",
1035 jbd2_journal_ack_err(journal);
1036 jbd2_journal_clear_err(journal);
1040 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
1043 struct ocfs2_super *osb;
1047 osb = journal->j_osb;
1049 status = jbd2_journal_load(journal->j_journal);
1051 mlog(ML_ERROR, "Failed to load journal!\n");
1055 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1057 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1063 /* Launch the commit thread */
1065 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1067 if (IS_ERR(osb->commit_task)) {
1068 status = PTR_ERR(osb->commit_task);
1069 osb->commit_task = NULL;
1070 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1071 "error=%d", status);
1075 osb->commit_task = NULL;
1082 /* 'full' flag tells us whether we clear out all blocks or if we just
1083 * mark the journal clean */
1084 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1090 status = jbd2_journal_wipe(journal->j_journal, full);
1096 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1104 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1107 struct ocfs2_recovery_map *rm = osb->recovery_map;
1109 spin_lock(&osb->osb_lock);
1110 empty = (rm->rm_used == 0);
1111 spin_unlock(&osb->osb_lock);
1116 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1118 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1122 * JBD Might read a cached version of another nodes journal file. We
1123 * don't want this as this file changes often and we get no
1124 * notification on those changes. The only way to be sure that we've
1125 * got the most up to date version of those blocks then is to force
1126 * read them off disk. Just searching through the buffer cache won't
1127 * work as there may be pages backing this file which are still marked
1128 * up to date. We know things can't change on this file underneath us
1129 * as we have the lock by now :)
1131 static int ocfs2_force_read_journal(struct inode *inode)
1135 u64 v_blkno, p_blkno, p_blocks, num_blocks;
1136 #define CONCURRENT_JOURNAL_FILL 32ULL
1137 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
1139 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
1141 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
1143 while (v_blkno < num_blocks) {
1144 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1145 &p_blkno, &p_blocks, NULL);
1151 if (p_blocks > CONCURRENT_JOURNAL_FILL)
1152 p_blocks = CONCURRENT_JOURNAL_FILL;
1154 /* We are reading journal data which should not
1155 * be put in the uptodate cache */
1156 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
1157 p_blkno, p_blocks, bhs);
1163 for(i = 0; i < p_blocks; i++) {
1168 v_blkno += p_blocks;
1172 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
1177 struct ocfs2_la_recovery_item {
1178 struct list_head lri_list;
1180 struct ocfs2_dinode *lri_la_dinode;
1181 struct ocfs2_dinode *lri_tl_dinode;
1182 struct ocfs2_quota_recovery *lri_qrec;
1183 enum ocfs2_orphan_reco_type lri_orphan_reco_type;
1186 /* Does the second half of the recovery process. By this point, the
1187 * node is marked clean and can actually be considered recovered,
1188 * hence it's no longer in the recovery map, but there's still some
1189 * cleanup we can do which shouldn't happen within the recovery thread
1190 * as locking in that context becomes very difficult if we are to take
1191 * recovering nodes into account.
1193 * NOTE: This function can and will sleep on recovery of other nodes
1194 * during cluster locking, just like any other ocfs2 process.
1196 void ocfs2_complete_recovery(struct work_struct *work)
1199 struct ocfs2_journal *journal =
1200 container_of(work, struct ocfs2_journal, j_recovery_work);
1201 struct ocfs2_super *osb = journal->j_osb;
1202 struct ocfs2_dinode *la_dinode, *tl_dinode;
1203 struct ocfs2_la_recovery_item *item, *n;
1204 struct ocfs2_quota_recovery *qrec;
1205 enum ocfs2_orphan_reco_type orphan_reco_type;
1206 LIST_HEAD(tmp_la_list);
1208 trace_ocfs2_complete_recovery(
1209 (unsigned long long)OCFS2_I(journal->j_inode)->ip_blkno);
1211 spin_lock(&journal->j_lock);
1212 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1213 spin_unlock(&journal->j_lock);
1215 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1216 list_del_init(&item->lri_list);
1218 ocfs2_wait_on_quotas(osb);
1220 la_dinode = item->lri_la_dinode;
1221 tl_dinode = item->lri_tl_dinode;
1222 qrec = item->lri_qrec;
1223 orphan_reco_type = item->lri_orphan_reco_type;
1225 trace_ocfs2_complete_recovery_slot(item->lri_slot,
1226 la_dinode ? le64_to_cpu(la_dinode->i_blkno) : 0,
1227 tl_dinode ? le64_to_cpu(tl_dinode->i_blkno) : 0,
1231 ret = ocfs2_complete_local_alloc_recovery(osb,
1240 ret = ocfs2_complete_truncate_log_recovery(osb,
1248 ret = ocfs2_recover_orphans(osb, item->lri_slot,
1254 ret = ocfs2_finish_quota_recovery(osb, qrec,
1258 /* Recovery info is already freed now */
1264 trace_ocfs2_complete_recovery_end(ret);
1267 /* NOTE: This function always eats your references to la_dinode and
1268 * tl_dinode, either manually on error, or by passing them to
1269 * ocfs2_complete_recovery */
1270 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1272 struct ocfs2_dinode *la_dinode,
1273 struct ocfs2_dinode *tl_dinode,
1274 struct ocfs2_quota_recovery *qrec,
1275 enum ocfs2_orphan_reco_type orphan_reco_type)
1277 struct ocfs2_la_recovery_item *item;
1279 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1281 /* Though we wish to avoid it, we are in fact safe in
1282 * skipping local alloc cleanup as fsck.ocfs2 is more
1283 * than capable of reclaiming unused space. */
1288 ocfs2_free_quota_recovery(qrec);
1290 mlog_errno(-ENOMEM);
1294 INIT_LIST_HEAD(&item->lri_list);
1295 item->lri_la_dinode = la_dinode;
1296 item->lri_slot = slot_num;
1297 item->lri_tl_dinode = tl_dinode;
1298 item->lri_qrec = qrec;
1299 item->lri_orphan_reco_type = orphan_reco_type;
1301 spin_lock(&journal->j_lock);
1302 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1303 queue_work(ocfs2_wq, &journal->j_recovery_work);
1304 spin_unlock(&journal->j_lock);
1307 /* Called by the mount code to queue recovery the last part of
1308 * recovery for it's own and offline slot(s). */
1309 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1311 struct ocfs2_journal *journal = osb->journal;
1313 if (ocfs2_is_hard_readonly(osb))
1316 /* No need to queue up our truncate_log as regular cleanup will catch
1318 ocfs2_queue_recovery_completion(journal, osb->slot_num,
1319 osb->local_alloc_copy, NULL, NULL,
1320 ORPHAN_NEED_TRUNCATE);
1321 ocfs2_schedule_truncate_log_flush(osb, 0);
1323 osb->local_alloc_copy = NULL;
1326 /* queue to recover orphan slots for all offline slots */
1327 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1328 ocfs2_queue_replay_slots(osb, ORPHAN_NEED_TRUNCATE);
1329 ocfs2_free_replay_slots(osb);
1332 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1334 if (osb->quota_rec) {
1335 ocfs2_queue_recovery_completion(osb->journal,
1340 ORPHAN_NEED_TRUNCATE);
1341 osb->quota_rec = NULL;
1345 static int __ocfs2_recovery_thread(void *arg)
1347 int status, node_num, slot_num;
1348 struct ocfs2_super *osb = arg;
1349 struct ocfs2_recovery_map *rm = osb->recovery_map;
1350 int *rm_quota = NULL;
1351 int rm_quota_used = 0, i;
1352 struct ocfs2_quota_recovery *qrec;
1354 status = ocfs2_wait_on_mount(osb);
1359 rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1365 status = ocfs2_super_lock(osb, 1);
1371 status = ocfs2_compute_replay_slots(osb);
1375 /* queue recovery for our own slot */
1376 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1377 NULL, NULL, ORPHAN_NO_NEED_TRUNCATE);
1379 spin_lock(&osb->osb_lock);
1380 while (rm->rm_used) {
1381 /* It's always safe to remove entry zero, as we won't
1382 * clear it until ocfs2_recover_node() has succeeded. */
1383 node_num = rm->rm_entries[0];
1384 spin_unlock(&osb->osb_lock);
1385 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1386 trace_ocfs2_recovery_thread_node(node_num, slot_num);
1387 if (slot_num == -ENOENT) {
1392 /* It is a bit subtle with quota recovery. We cannot do it
1393 * immediately because we have to obtain cluster locks from
1394 * quota files and we also don't want to just skip it because
1395 * then quota usage would be out of sync until some node takes
1396 * the slot. So we remember which nodes need quota recovery
1397 * and when everything else is done, we recover quotas. */
1398 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1399 if (i == rm_quota_used)
1400 rm_quota[rm_quota_used++] = slot_num;
1402 status = ocfs2_recover_node(osb, node_num, slot_num);
1405 ocfs2_recovery_map_clear(osb, node_num);
1408 "Error %d recovering node %d on device (%u,%u)!\n",
1410 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1411 mlog(ML_ERROR, "Volume requires unmount.\n");
1414 spin_lock(&osb->osb_lock);
1416 spin_unlock(&osb->osb_lock);
1417 trace_ocfs2_recovery_thread_end(status);
1419 /* Refresh all journal recovery generations from disk */
1420 status = ocfs2_check_journals_nolocks(osb);
1421 status = (status == -EROFS) ? 0 : status;
1425 /* Now it is right time to recover quotas... We have to do this under
1426 * superblock lock so that no one can start using the slot (and crash)
1427 * before we recover it */
1428 for (i = 0; i < rm_quota_used; i++) {
1429 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1431 status = PTR_ERR(qrec);
1435 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1437 ORPHAN_NEED_TRUNCATE);
1440 ocfs2_super_unlock(osb, 1);
1442 /* queue recovery for offline slots */
1443 ocfs2_queue_replay_slots(osb, ORPHAN_NEED_TRUNCATE);
1446 mutex_lock(&osb->recovery_lock);
1447 if (!status && !ocfs2_recovery_completed(osb)) {
1448 mutex_unlock(&osb->recovery_lock);
1452 ocfs2_free_replay_slots(osb);
1453 osb->recovery_thread_task = NULL;
1454 mb(); /* sync with ocfs2_recovery_thread_running */
1455 wake_up(&osb->recovery_event);
1457 mutex_unlock(&osb->recovery_lock);
1461 /* no one is callint kthread_stop() for us so the kthread() api
1462 * requires that we call do_exit(). And it isn't exported, but
1463 * complete_and_exit() seems to be a minimal wrapper around it. */
1464 complete_and_exit(NULL, status);
1467 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1469 mutex_lock(&osb->recovery_lock);
1471 trace_ocfs2_recovery_thread(node_num, osb->node_num,
1472 osb->disable_recovery, osb->recovery_thread_task,
1473 osb->disable_recovery ?
1474 -1 : ocfs2_recovery_map_set(osb, node_num));
1476 if (osb->disable_recovery)
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);
1494 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1496 struct buffer_head **bh,
1497 struct inode **ret_inode)
1499 int status = -EACCES;
1500 struct inode *inode = NULL;
1502 BUG_ON(slot_num >= osb->max_slots);
1504 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1506 if (!inode || is_bad_inode(inode)) {
1510 SET_INODE_JOURNAL(inode);
1512 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1522 if (status || !ret_inode)
1530 /* Does the actual journal replay and marks the journal inode as
1531 * clean. Will only replay if the journal inode is marked dirty. */
1532 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1539 struct inode *inode = NULL;
1540 struct ocfs2_dinode *fe;
1541 journal_t *journal = NULL;
1542 struct buffer_head *bh = NULL;
1545 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1551 fe = (struct ocfs2_dinode *)bh->b_data;
1552 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1557 * As the fs recovery is asynchronous, there is a small chance that
1558 * another node mounted (and recovered) the slot before the recovery
1559 * thread could get the lock. To handle that, we dirty read the journal
1560 * inode for that slot to get the recovery generation. If it is
1561 * different than what we expected, the slot has been recovered.
1562 * If not, it needs recovery.
1564 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1565 trace_ocfs2_replay_journal_recovered(slot_num,
1566 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1567 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1572 /* Continue with recovery as the journal has not yet been recovered */
1574 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1576 trace_ocfs2_replay_journal_lock_err(status);
1577 if (status != -ERESTARTSYS)
1578 mlog(ML_ERROR, "Could not lock journal!\n");
1583 fe = (struct ocfs2_dinode *) bh->b_data;
1585 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1586 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1588 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1589 trace_ocfs2_replay_journal_skip(node_num);
1590 /* Refresh recovery generation for the slot */
1591 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1595 /* we need to run complete recovery for offline orphan slots */
1596 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1598 printk(KERN_NOTICE "ocfs2: Begin replay journal (node %d, slot %d) on "\
1599 "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1600 MINOR(osb->sb->s_dev));
1602 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1604 status = ocfs2_force_read_journal(inode);
1610 journal = jbd2_journal_init_inode(inode);
1611 if (journal == NULL) {
1612 mlog(ML_ERROR, "Linux journal layer error\n");
1617 status = jbd2_journal_load(journal);
1622 jbd2_journal_destroy(journal);
1626 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1628 /* wipe the journal */
1629 jbd2_journal_lock_updates(journal);
1630 status = jbd2_journal_flush(journal);
1631 jbd2_journal_unlock_updates(journal);
1635 /* This will mark the node clean */
1636 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1637 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1638 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1640 /* Increment recovery generation to indicate successful recovery */
1641 ocfs2_bump_recovery_generation(fe);
1642 osb->slot_recovery_generations[slot_num] =
1643 ocfs2_get_recovery_generation(fe);
1645 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1646 status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1653 jbd2_journal_destroy(journal);
1655 printk(KERN_NOTICE "ocfs2: End replay journal (node %d, slot %d) on "\
1656 "device (%u,%u)\n", node_num, slot_num, MAJOR(osb->sb->s_dev),
1657 MINOR(osb->sb->s_dev));
1659 /* drop the lock on this nodes journal */
1661 ocfs2_inode_unlock(inode, 1);
1672 * Do the most important parts of node recovery:
1673 * - Replay it's journal
1674 * - Stamp a clean local allocator file
1675 * - Stamp a clean truncate log
1676 * - Mark the node clean
1678 * If this function completes without error, a node in OCFS2 can be
1679 * said to have been safely recovered. As a result, failure during the
1680 * second part of a nodes recovery process (local alloc recovery) is
1681 * far less concerning.
1683 static int ocfs2_recover_node(struct ocfs2_super *osb,
1684 int node_num, int slot_num)
1687 struct ocfs2_dinode *la_copy = NULL;
1688 struct ocfs2_dinode *tl_copy = NULL;
1690 trace_ocfs2_recover_node(node_num, slot_num, osb->node_num);
1692 /* Should not ever be called to recover ourselves -- in that
1693 * case we should've called ocfs2_journal_load instead. */
1694 BUG_ON(osb->node_num == node_num);
1696 status = ocfs2_replay_journal(osb, node_num, slot_num);
1698 if (status == -EBUSY) {
1699 trace_ocfs2_recover_node_skip(slot_num, node_num);
1707 /* Stamp a clean local alloc file AFTER recovering the journal... */
1708 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1714 /* An error from begin_truncate_log_recovery is not
1715 * serious enough to warrant halting the rest of
1717 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1721 /* Likewise, this would be a strange but ultimately not so
1722 * harmful place to get an error... */
1723 status = ocfs2_clear_slot(osb, slot_num);
1727 /* This will kfree the memory pointed to by la_copy and tl_copy */
1728 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1729 tl_copy, NULL, ORPHAN_NEED_TRUNCATE);
1737 /* Test node liveness by trylocking his journal. If we get the lock,
1738 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1739 * still alive (we couldn't get the lock) and < 0 on error. */
1740 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1744 struct inode *inode = NULL;
1746 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1748 if (inode == NULL) {
1749 mlog(ML_ERROR, "access error\n");
1753 if (is_bad_inode(inode)) {
1754 mlog(ML_ERROR, "access error (bad inode)\n");
1760 SET_INODE_JOURNAL(inode);
1762 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1763 status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1765 if (status != -EAGAIN)
1770 ocfs2_inode_unlock(inode, 1);
1778 /* Call this underneath ocfs2_super_lock. It also assumes that the
1779 * slot info struct has been updated from disk. */
1780 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1782 unsigned int node_num;
1785 struct buffer_head *bh = NULL;
1786 struct ocfs2_dinode *di;
1788 /* This is called with the super block cluster lock, so we
1789 * know that the slot map can't change underneath us. */
1791 for (i = 0; i < osb->max_slots; i++) {
1792 /* Read journal inode to get the recovery generation */
1793 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1798 di = (struct ocfs2_dinode *)bh->b_data;
1799 gen = ocfs2_get_recovery_generation(di);
1803 spin_lock(&osb->osb_lock);
1804 osb->slot_recovery_generations[i] = gen;
1806 trace_ocfs2_mark_dead_nodes(i,
1807 osb->slot_recovery_generations[i]);
1809 if (i == osb->slot_num) {
1810 spin_unlock(&osb->osb_lock);
1814 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1815 if (status == -ENOENT) {
1816 spin_unlock(&osb->osb_lock);
1820 if (__ocfs2_recovery_map_test(osb, node_num)) {
1821 spin_unlock(&osb->osb_lock);
1824 spin_unlock(&osb->osb_lock);
1826 /* Ok, we have a slot occupied by another node which
1827 * is not in the recovery map. We trylock his journal
1828 * file here to test if he's alive. */
1829 status = ocfs2_trylock_journal(osb, i);
1831 /* Since we're called from mount, we know that
1832 * the recovery thread can't race us on
1833 * setting / checking the recovery bits. */
1834 ocfs2_recovery_thread(osb, node_num);
1835 } else if ((status < 0) && (status != -EAGAIN)) {
1847 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1848 * randomness to the timeout to minimize multple nodes firing the timer at the
1851 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1855 get_random_bytes(&time, sizeof(time));
1856 time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1857 return msecs_to_jiffies(time);
1861 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1862 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1863 * is done to catch any orphans that are left over in orphan directories.
1865 * It scans all slots, even ones that are in use. It does so to handle the
1866 * case described below:
1868 * Node 1 has an inode it was using. The dentry went away due to memory
1869 * pressure. Node 1 closes the inode, but it's on the free list. The node
1870 * has the open lock.
1871 * Node 2 unlinks the inode. It grabs the dentry lock to notify others,
1872 * but node 1 has no dentry and doesn't get the message. It trylocks the
1873 * open lock, sees that another node has a PR, and does nothing.
1874 * Later node 2 runs its orphan dir. It igets the inode, trylocks the
1875 * open lock, sees the PR still, and does nothing.
1876 * Basically, we have to trigger an orphan iput on node 1. The only way
1877 * for this to happen is if node 1 runs node 2's orphan dir.
1879 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1880 * seconds. It gets an EX lock on os_lockres and checks sequence number
1881 * stored in LVB. If the sequence number has changed, it means some other
1882 * node has done the scan. This node skips the scan and tracks the
1883 * sequence number. If the sequence number didn't change, it means a scan
1884 * hasn't happened. The node queues a scan and increments the
1885 * sequence number in the LVB.
1887 void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1889 struct ocfs2_orphan_scan *os;
1893 os = &osb->osb_orphan_scan;
1895 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1898 trace_ocfs2_queue_orphan_scan_begin(os->os_count, os->os_seqno,
1899 atomic_read(&os->os_state));
1901 status = ocfs2_orphan_scan_lock(osb, &seqno);
1903 if (status != -EAGAIN)
1908 /* Do no queue the tasks if the volume is being umounted */
1909 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1912 if (os->os_seqno != seqno) {
1913 os->os_seqno = seqno;
1917 for (i = 0; i < osb->max_slots; i++)
1918 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1919 NULL, ORPHAN_NO_NEED_TRUNCATE);
1921 * We queued a recovery on orphan slots, increment the sequence
1922 * number and update LVB so other node will skip the scan for a while
1926 os->os_scantime = CURRENT_TIME;
1928 ocfs2_orphan_scan_unlock(osb, seqno);
1930 trace_ocfs2_queue_orphan_scan_end(os->os_count, os->os_seqno,
1931 atomic_read(&os->os_state));
1935 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1936 void ocfs2_orphan_scan_work(struct work_struct *work)
1938 struct ocfs2_orphan_scan *os;
1939 struct ocfs2_super *osb;
1941 os = container_of(work, struct ocfs2_orphan_scan,
1942 os_orphan_scan_work.work);
1945 mutex_lock(&os->os_lock);
1946 ocfs2_queue_orphan_scan(osb);
1947 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1948 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1949 ocfs2_orphan_scan_timeout());
1950 mutex_unlock(&os->os_lock);
1953 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1955 struct ocfs2_orphan_scan *os;
1957 os = &osb->osb_orphan_scan;
1958 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1959 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1960 mutex_lock(&os->os_lock);
1961 cancel_delayed_work(&os->os_orphan_scan_work);
1962 mutex_unlock(&os->os_lock);
1966 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1968 struct ocfs2_orphan_scan *os;
1970 os = &osb->osb_orphan_scan;
1974 mutex_init(&os->os_lock);
1975 INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
1978 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
1980 struct ocfs2_orphan_scan *os;
1982 os = &osb->osb_orphan_scan;
1983 os->os_scantime = CURRENT_TIME;
1984 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1985 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1987 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
1988 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1989 ocfs2_orphan_scan_timeout());
1993 struct ocfs2_orphan_filldir_priv {
1994 struct dir_context ctx;
1996 struct ocfs2_super *osb;
1999 static int ocfs2_orphan_filldir(struct dir_context *ctx, const char *name,
2000 int name_len, loff_t pos, u64 ino,
2003 struct ocfs2_orphan_filldir_priv *p =
2004 container_of(ctx, struct ocfs2_orphan_filldir_priv, ctx);
2007 if (name_len == 1 && !strncmp(".", name, 1))
2009 if (name_len == 2 && !strncmp("..", name, 2))
2012 /* Skip bad inodes so that recovery can continue */
2013 iter = ocfs2_iget(p->osb, ino,
2014 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
2018 /* Skip inodes which are already added to recover list, since dio may
2019 * happen concurrently with unlink/rename */
2020 if (OCFS2_I(iter)->ip_next_orphan) {
2025 trace_ocfs2_orphan_filldir((unsigned long long)OCFS2_I(iter)->ip_blkno);
2026 /* No locking is required for the next_orphan queue as there
2027 * is only ever a single process doing orphan recovery. */
2028 OCFS2_I(iter)->ip_next_orphan = p->head;
2034 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
2036 struct inode **head)
2039 struct inode *orphan_dir_inode = NULL;
2040 struct ocfs2_orphan_filldir_priv priv = {
2041 .ctx.actor = ocfs2_orphan_filldir,
2046 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2047 ORPHAN_DIR_SYSTEM_INODE,
2049 if (!orphan_dir_inode) {
2055 mutex_lock(&orphan_dir_inode->i_mutex);
2056 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2062 status = ocfs2_dir_foreach(orphan_dir_inode, &priv.ctx);
2071 ocfs2_inode_unlock(orphan_dir_inode, 0);
2073 mutex_unlock(&orphan_dir_inode->i_mutex);
2074 iput(orphan_dir_inode);
2078 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2083 spin_lock(&osb->osb_lock);
2084 ret = !osb->osb_orphan_wipes[slot];
2085 spin_unlock(&osb->osb_lock);
2089 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2092 spin_lock(&osb->osb_lock);
2093 /* Mark ourselves such that new processes in delete_inode()
2094 * know to quit early. */
2095 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2096 while (osb->osb_orphan_wipes[slot]) {
2097 /* If any processes are already in the middle of an
2098 * orphan wipe on this dir, then we need to wait for
2100 spin_unlock(&osb->osb_lock);
2101 wait_event_interruptible(osb->osb_wipe_event,
2102 ocfs2_orphan_recovery_can_continue(osb, slot));
2103 spin_lock(&osb->osb_lock);
2105 spin_unlock(&osb->osb_lock);
2108 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2111 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2115 * Orphan recovery. Each mounted node has it's own orphan dir which we
2116 * must run during recovery. Our strategy here is to build a list of
2117 * the inodes in the orphan dir and iget/iput them. The VFS does
2118 * (most) of the rest of the work.
2120 * Orphan recovery can happen at any time, not just mount so we have a
2121 * couple of extra considerations.
2123 * - We grab as many inodes as we can under the orphan dir lock -
2124 * doing iget() outside the orphan dir risks getting a reference on
2126 * - We must be sure not to deadlock with other processes on the
2127 * system wanting to run delete_inode(). This can happen when they go
2128 * to lock the orphan dir and the orphan recovery process attempts to
2129 * iget() inside the orphan dir lock. This can be avoided by
2130 * advertising our state to ocfs2_delete_inode().
2132 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2134 enum ocfs2_orphan_reco_type orphan_reco_type)
2137 struct inode *inode = NULL;
2139 struct ocfs2_inode_info *oi;
2141 trace_ocfs2_recover_orphans(slot);
2143 ocfs2_mark_recovering_orphan_dir(osb, slot);
2144 ret = ocfs2_queue_orphans(osb, slot, &inode);
2145 ocfs2_clear_recovering_orphan_dir(osb, slot);
2147 /* Error here should be noted, but we want to continue with as
2148 * many queued inodes as we've got. */
2153 oi = OCFS2_I(inode);
2154 trace_ocfs2_recover_orphans_iput(
2155 (unsigned long long)oi->ip_blkno);
2157 iter = oi->ip_next_orphan;
2158 oi->ip_next_orphan = NULL;
2161 * We need to take and drop the inode lock to
2162 * force read inode from disk.
2164 ret = ocfs2_inode_lock(inode, NULL, 0);
2169 ocfs2_inode_unlock(inode, 0);
2171 if (inode->i_nlink == 0) {
2172 spin_lock(&oi->ip_lock);
2173 /* Set the proper information to get us going into
2174 * ocfs2_delete_inode. */
2175 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2176 spin_unlock(&oi->ip_lock);
2177 } else if (orphan_reco_type == ORPHAN_NEED_TRUNCATE) {
2178 struct buffer_head *di_bh = NULL;
2180 ret = ocfs2_rw_lock(inode, 1);
2186 ret = ocfs2_inode_lock(inode, &di_bh, 1);
2188 ocfs2_rw_unlock(inode, 1);
2193 ret = ocfs2_truncate_file(inode, di_bh,
2194 i_size_read(inode));
2195 ocfs2_inode_unlock(inode, 1);
2196 ocfs2_rw_unlock(inode, 1);
2204 ret = ocfs2_del_inode_from_orphan(osb, inode, 0, 0);
2208 wake_up(&OCFS2_I(inode)->append_dio_wq);
2209 } /* else if ORPHAN_NO_NEED_TRUNCATE, do nothing */
2220 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2222 /* This check is good because ocfs2 will wait on our recovery
2223 * thread before changing it to something other than MOUNTED
2225 wait_event(osb->osb_mount_event,
2226 (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2227 atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2228 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2230 /* If there's an error on mount, then we may never get to the
2231 * MOUNTED flag, but this is set right before
2232 * dismount_volume() so we can trust it. */
2233 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2234 trace_ocfs2_wait_on_mount(VOLUME_DISABLED);
2235 mlog(0, "mount error, exiting!\n");
2242 static int ocfs2_commit_thread(void *arg)
2245 struct ocfs2_super *osb = arg;
2246 struct ocfs2_journal *journal = osb->journal;
2248 /* we can trust j_num_trans here because _should_stop() is only set in
2249 * shutdown and nobody other than ourselves should be able to start
2250 * transactions. committing on shutdown might take a few iterations
2251 * as final transactions put deleted inodes on the list */
2252 while (!(kthread_should_stop() &&
2253 atomic_read(&journal->j_num_trans) == 0)) {
2255 wait_event_interruptible(osb->checkpoint_event,
2256 atomic_read(&journal->j_num_trans)
2257 || kthread_should_stop());
2259 status = ocfs2_commit_cache(osb);
2261 static unsigned long abort_warn_time;
2263 /* Warn about this once per minute */
2264 if (printk_timed_ratelimit(&abort_warn_time, 60*HZ))
2265 mlog(ML_ERROR, "status = %d, journal is "
2266 "already aborted.\n", status);
2268 * After ocfs2_commit_cache() fails, j_num_trans has a
2269 * non-zero value. Sleep here to avoid a busy-wait
2272 msleep_interruptible(1000);
2275 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2277 "commit_thread: %u transactions pending on "
2279 atomic_read(&journal->j_num_trans));
2286 /* Reads all the journal inodes without taking any cluster locks. Used
2287 * for hard readonly access to determine whether any journal requires
2288 * recovery. Also used to refresh the recovery generation numbers after
2289 * a journal has been recovered by another node.
2291 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2295 struct buffer_head *di_bh = NULL;
2296 struct ocfs2_dinode *di;
2297 int journal_dirty = 0;
2299 for(slot = 0; slot < osb->max_slots; slot++) {
2300 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2306 di = (struct ocfs2_dinode *) di_bh->b_data;
2308 osb->slot_recovery_generations[slot] =
2309 ocfs2_get_recovery_generation(di);
2311 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2312 OCFS2_JOURNAL_DIRTY_FL)