1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Defines functions of journalling api
8 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/kthread.h>
31 #include <linux/time.h>
32 #include <linux/random.h>
34 #define MLOG_MASK_PREFIX ML_JOURNAL
35 #include <cluster/masklog.h>
40 #include "blockcheck.h"
43 #include "extent_map.h"
44 #include "heartbeat.h"
47 #include "localalloc.h"
54 #include "buffer_head_io.h"
56 DEFINE_SPINLOCK(trans_inc_lock);
58 #define ORPHAN_SCAN_SCHEDULE_TIMEOUT 300000
60 static int ocfs2_force_read_journal(struct inode *inode);
61 static int ocfs2_recover_node(struct ocfs2_super *osb,
62 int node_num, int slot_num);
63 static int __ocfs2_recovery_thread(void *arg);
64 static int ocfs2_commit_cache(struct ocfs2_super *osb);
65 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota);
66 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
67 int dirty, int replayed);
68 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
70 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
72 static int ocfs2_commit_thread(void *arg);
73 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
75 struct ocfs2_dinode *la_dinode,
76 struct ocfs2_dinode *tl_dinode,
77 struct ocfs2_quota_recovery *qrec);
79 static inline int ocfs2_wait_on_mount(struct ocfs2_super *osb)
81 return __ocfs2_wait_on_mount(osb, 0);
84 static inline int ocfs2_wait_on_quotas(struct ocfs2_super *osb)
86 return __ocfs2_wait_on_mount(osb, 1);
90 * This replay_map is to track online/offline slots, so we could recover
91 * offline slots during recovery and mount
94 enum ocfs2_replay_state {
95 REPLAY_UNNEEDED = 0, /* Replay is not needed, so ignore this map */
96 REPLAY_NEEDED, /* Replay slots marked in rm_replay_slots */
97 REPLAY_DONE /* Replay was already queued */
100 struct ocfs2_replay_map {
101 unsigned int rm_slots;
102 enum ocfs2_replay_state rm_state;
103 unsigned char rm_replay_slots[0];
106 void ocfs2_replay_map_set_state(struct ocfs2_super *osb, int state)
108 if (!osb->replay_map)
111 /* If we've already queued the replay, we don't have any more to do */
112 if (osb->replay_map->rm_state == REPLAY_DONE)
115 osb->replay_map->rm_state = state;
118 int ocfs2_compute_replay_slots(struct ocfs2_super *osb)
120 struct ocfs2_replay_map *replay_map;
123 /* If replay map is already set, we don't do it again */
127 replay_map = kzalloc(sizeof(struct ocfs2_replay_map) +
128 (osb->max_slots * sizeof(char)), GFP_KERNEL);
135 spin_lock(&osb->osb_lock);
137 replay_map->rm_slots = osb->max_slots;
138 replay_map->rm_state = REPLAY_UNNEEDED;
140 /* set rm_replay_slots for offline slot(s) */
141 for (i = 0; i < replay_map->rm_slots; i++) {
142 if (ocfs2_slot_to_node_num_locked(osb, i, &node_num) == -ENOENT)
143 replay_map->rm_replay_slots[i] = 1;
146 osb->replay_map = replay_map;
147 spin_unlock(&osb->osb_lock);
151 void ocfs2_queue_replay_slots(struct ocfs2_super *osb)
153 struct ocfs2_replay_map *replay_map = osb->replay_map;
159 if (replay_map->rm_state != REPLAY_NEEDED)
162 for (i = 0; i < replay_map->rm_slots; i++)
163 if (replay_map->rm_replay_slots[i])
164 ocfs2_queue_recovery_completion(osb->journal, i, NULL,
166 replay_map->rm_state = REPLAY_DONE;
169 void ocfs2_free_replay_slots(struct ocfs2_super *osb)
171 struct ocfs2_replay_map *replay_map = osb->replay_map;
173 if (!osb->replay_map)
177 osb->replay_map = NULL;
180 int ocfs2_recovery_init(struct ocfs2_super *osb)
182 struct ocfs2_recovery_map *rm;
184 mutex_init(&osb->recovery_lock);
185 osb->disable_recovery = 0;
186 osb->recovery_thread_task = NULL;
187 init_waitqueue_head(&osb->recovery_event);
189 rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
190 osb->max_slots * sizeof(unsigned int),
197 rm->rm_entries = (unsigned int *)((char *)rm +
198 sizeof(struct ocfs2_recovery_map));
199 osb->recovery_map = rm;
204 /* we can't grab the goofy sem lock from inside wait_event, so we use
205 * memory barriers to make sure that we'll see the null task before
207 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
210 return osb->recovery_thread_task != NULL;
213 void ocfs2_recovery_exit(struct ocfs2_super *osb)
215 struct ocfs2_recovery_map *rm;
217 /* disable any new recovery threads and wait for any currently
218 * running ones to exit. Do this before setting the vol_state. */
219 mutex_lock(&osb->recovery_lock);
220 osb->disable_recovery = 1;
221 mutex_unlock(&osb->recovery_lock);
222 wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
224 /* At this point, we know that no more recovery threads can be
225 * launched, so wait for any recovery completion work to
227 flush_workqueue(ocfs2_wq);
230 * Now that recovery is shut down, and the osb is about to be
231 * freed, the osb_lock is not taken here.
233 rm = osb->recovery_map;
234 /* XXX: Should we bug if there are dirty entries? */
239 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
240 unsigned int node_num)
243 struct ocfs2_recovery_map *rm = osb->recovery_map;
245 assert_spin_locked(&osb->osb_lock);
247 for (i = 0; i < rm->rm_used; i++) {
248 if (rm->rm_entries[i] == node_num)
255 /* Behaves like test-and-set. Returns the previous value */
256 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
257 unsigned int node_num)
259 struct ocfs2_recovery_map *rm = osb->recovery_map;
261 spin_lock(&osb->osb_lock);
262 if (__ocfs2_recovery_map_test(osb, node_num)) {
263 spin_unlock(&osb->osb_lock);
267 /* XXX: Can this be exploited? Not from o2dlm... */
268 BUG_ON(rm->rm_used >= osb->max_slots);
270 rm->rm_entries[rm->rm_used] = node_num;
272 spin_unlock(&osb->osb_lock);
277 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
278 unsigned int node_num)
281 struct ocfs2_recovery_map *rm = osb->recovery_map;
283 spin_lock(&osb->osb_lock);
285 for (i = 0; i < rm->rm_used; i++) {
286 if (rm->rm_entries[i] == node_num)
290 if (i < rm->rm_used) {
291 /* XXX: be careful with the pointer math */
292 memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
293 (rm->rm_used - i - 1) * sizeof(unsigned int));
297 spin_unlock(&osb->osb_lock);
300 static int ocfs2_commit_cache(struct ocfs2_super *osb)
303 unsigned int flushed;
304 struct ocfs2_journal *journal = NULL;
306 journal = osb->journal;
308 /* Flush all pending commits and checkpoint the journal. */
309 down_write(&journal->j_trans_barrier);
311 if (atomic_read(&journal->j_num_trans) == 0) {
312 up_write(&journal->j_trans_barrier);
313 mlog(0, "No transactions for me to flush!\n");
317 jbd2_journal_lock_updates(journal->j_journal);
318 status = jbd2_journal_flush(journal->j_journal);
319 jbd2_journal_unlock_updates(journal->j_journal);
321 up_write(&journal->j_trans_barrier);
326 ocfs2_inc_trans_id(journal);
328 flushed = atomic_read(&journal->j_num_trans);
329 atomic_set(&journal->j_num_trans, 0);
330 up_write(&journal->j_trans_barrier);
332 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
333 journal->j_trans_id, flushed);
335 ocfs2_wake_downconvert_thread(osb);
336 wake_up(&journal->j_checkpointed);
342 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 down_read(&osb->journal->j_trans_barrier);
361 handle = jbd2_journal_start(journal, max_buffs);
362 if (IS_ERR(handle)) {
363 up_read(&osb->journal->j_trans_barrier);
365 mlog_errno(PTR_ERR(handle));
367 if (is_journal_aborted(journal)) {
368 ocfs2_abort(osb->sb, "Detected aborted journal");
369 handle = ERR_PTR(-EROFS);
372 if (!ocfs2_mount_local(osb))
373 atomic_inc(&(osb->journal->j_num_trans));
379 int ocfs2_commit_trans(struct ocfs2_super *osb,
383 struct ocfs2_journal *journal = osb->journal;
387 nested = handle->h_ref > 1;
388 ret = jbd2_journal_stop(handle);
393 up_read(&journal->j_trans_barrier);
399 * 'nblocks' is what you want to add to the current transaction.
401 * This might call jbd2_journal_restart() which will commit dirty buffers
402 * and then restart the transaction. Before calling
403 * ocfs2_extend_trans(), any changed blocks should have been
404 * dirtied. After calling it, all blocks which need to be changed must
405 * go through another set of journal_access/journal_dirty calls.
407 * WARNING: This will not release any semaphores or disk locks taken
408 * during the transaction, so make sure they were taken *before*
409 * start_trans or we'll have ordering deadlocks.
411 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
412 * good because transaction ids haven't yet been recorded on the
413 * cluster locks associated with this handle.
415 int ocfs2_extend_trans(handle_t *handle, int nblocks)
417 int status, old_nblocks;
425 old_nblocks = handle->h_buffer_credits;
427 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
429 #ifdef CONFIG_OCFS2_DEBUG_FS
432 status = jbd2_journal_extend(handle, nblocks);
441 "jbd2_journal_extend failed, trying "
442 "jbd2_journal_restart\n");
443 status = jbd2_journal_restart(handle,
444 old_nblocks + nblocks);
458 struct ocfs2_triggers {
459 struct jbd2_buffer_trigger_type ot_triggers;
463 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
465 return container_of(triggers, struct ocfs2_triggers, ot_triggers);
468 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
469 struct buffer_head *bh,
470 void *data, size_t size)
472 struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
475 * We aren't guaranteed to have the superblock here, so we
476 * must unconditionally compute the ecc data.
477 * __ocfs2_journal_access() will only set the triggers if
478 * metaecc is enabled.
480 ocfs2_block_check_compute(data, size, data + ot->ot_offset);
484 * Quota blocks have their own trigger because the struct ocfs2_block_check
485 * offset depends on the blocksize.
487 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
488 struct buffer_head *bh,
489 void *data, size_t size)
491 struct ocfs2_disk_dqtrailer *dqt =
492 ocfs2_block_dqtrailer(size, data);
495 * We aren't guaranteed to have the superblock here, so we
496 * must unconditionally compute the ecc data.
497 * __ocfs2_journal_access() will only set the triggers if
498 * metaecc is enabled.
500 ocfs2_block_check_compute(data, size, &dqt->dq_check);
504 * Directory blocks also have their own trigger because the
505 * struct ocfs2_block_check offset depends on the blocksize.
507 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
508 struct buffer_head *bh,
509 void *data, size_t size)
511 struct ocfs2_dir_block_trailer *trailer =
512 ocfs2_dir_trailer_from_size(size, data);
515 * We aren't guaranteed to have the superblock here, so we
516 * must unconditionally compute the ecc data.
517 * __ocfs2_journal_access() will only set the triggers if
518 * metaecc is enabled.
520 ocfs2_block_check_compute(data, size, &trailer->db_check);
523 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
524 struct buffer_head *bh)
527 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
528 "bh->b_blocknr = %llu\n",
530 (unsigned long long)bh->b_blocknr);
532 /* We aren't guaranteed to have the superblock here - but if we
533 * don't, it'll just crash. */
534 ocfs2_error(bh->b_assoc_map->host->i_sb,
535 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
538 static struct ocfs2_triggers di_triggers = {
540 .t_frozen = ocfs2_frozen_trigger,
541 .t_abort = ocfs2_abort_trigger,
543 .ot_offset = offsetof(struct ocfs2_dinode, i_check),
546 static struct ocfs2_triggers eb_triggers = {
548 .t_frozen = ocfs2_frozen_trigger,
549 .t_abort = ocfs2_abort_trigger,
551 .ot_offset = offsetof(struct ocfs2_extent_block, h_check),
554 static struct ocfs2_triggers rb_triggers = {
556 .t_frozen = ocfs2_frozen_trigger,
557 .t_abort = ocfs2_abort_trigger,
559 .ot_offset = offsetof(struct ocfs2_refcount_block, rf_check),
562 static struct ocfs2_triggers gd_triggers = {
564 .t_frozen = ocfs2_frozen_trigger,
565 .t_abort = ocfs2_abort_trigger,
567 .ot_offset = offsetof(struct ocfs2_group_desc, bg_check),
570 static struct ocfs2_triggers db_triggers = {
572 .t_frozen = ocfs2_db_frozen_trigger,
573 .t_abort = ocfs2_abort_trigger,
577 static struct ocfs2_triggers xb_triggers = {
579 .t_frozen = ocfs2_frozen_trigger,
580 .t_abort = ocfs2_abort_trigger,
582 .ot_offset = offsetof(struct ocfs2_xattr_block, xb_check),
585 static struct ocfs2_triggers dq_triggers = {
587 .t_frozen = ocfs2_dq_frozen_trigger,
588 .t_abort = ocfs2_abort_trigger,
592 static struct ocfs2_triggers dr_triggers = {
594 .t_frozen = ocfs2_frozen_trigger,
595 .t_abort = ocfs2_abort_trigger,
597 .ot_offset = offsetof(struct ocfs2_dx_root_block, dr_check),
600 static struct ocfs2_triggers dl_triggers = {
602 .t_frozen = ocfs2_frozen_trigger,
603 .t_abort = ocfs2_abort_trigger,
605 .ot_offset = offsetof(struct ocfs2_dx_leaf, dl_check),
608 static int __ocfs2_journal_access(handle_t *handle,
609 struct ocfs2_caching_info *ci,
610 struct buffer_head *bh,
611 struct ocfs2_triggers *triggers,
615 struct ocfs2_super *osb =
616 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
618 BUG_ON(!ci || !ci->ci_ops);
622 mlog(0, "bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
623 (unsigned long long)bh->b_blocknr, type,
624 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
625 "OCFS2_JOURNAL_ACCESS_CREATE" :
626 "OCFS2_JOURNAL_ACCESS_WRITE",
629 /* we can safely remove this assertion after testing. */
630 if (!buffer_uptodate(bh)) {
631 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
632 mlog(ML_ERROR, "b_blocknr=%llu\n",
633 (unsigned long long)bh->b_blocknr);
637 /* Set the current transaction information on the ci so
638 * that the locking code knows whether it can drop it's locks
639 * on this ci or not. We're protected from the commit
640 * thread updating the current transaction id until
641 * ocfs2_commit_trans() because ocfs2_start_trans() took
642 * j_trans_barrier for us. */
643 ocfs2_set_ci_lock_trans(osb->journal, ci);
645 ocfs2_metadata_cache_io_lock(ci);
647 case OCFS2_JOURNAL_ACCESS_CREATE:
648 case OCFS2_JOURNAL_ACCESS_WRITE:
649 status = jbd2_journal_get_write_access(handle, bh);
652 case OCFS2_JOURNAL_ACCESS_UNDO:
653 status = jbd2_journal_get_undo_access(handle, bh);
658 mlog(ML_ERROR, "Unknown access type!\n");
660 if (!status && ocfs2_meta_ecc(osb) && triggers)
661 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
662 ocfs2_metadata_cache_io_unlock(ci);
665 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
672 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
673 struct buffer_head *bh, int type)
675 return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
678 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
679 struct buffer_head *bh, int type)
681 return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
684 int ocfs2_journal_access_rb(handle_t *handle, struct ocfs2_caching_info *ci,
685 struct buffer_head *bh, int type)
687 return __ocfs2_journal_access(handle, ci, bh, &rb_triggers,
691 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
692 struct buffer_head *bh, int type)
694 return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
697 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
698 struct buffer_head *bh, int type)
700 return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
703 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
704 struct buffer_head *bh, int type)
706 return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
709 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
710 struct buffer_head *bh, int type)
712 return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
715 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
716 struct buffer_head *bh, int type)
718 return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
721 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
722 struct buffer_head *bh, int type)
724 return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
727 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
728 struct buffer_head *bh, int type)
730 return __ocfs2_journal_access(handle, ci, bh, NULL, type);
733 void ocfs2_journal_dirty(handle_t *handle, struct buffer_head *bh)
737 mlog(0, "(bh->b_blocknr=%llu)\n",
738 (unsigned long long)bh->b_blocknr);
740 status = jbd2_journal_dirty_metadata(handle, bh);
746 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
748 void ocfs2_set_journal_params(struct ocfs2_super *osb)
750 journal_t *journal = osb->journal->j_journal;
751 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
753 if (osb->osb_commit_interval)
754 commit_interval = osb->osb_commit_interval;
756 write_lock(&journal->j_state_lock);
757 journal->j_commit_interval = commit_interval;
758 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
759 journal->j_flags |= JBD2_BARRIER;
761 journal->j_flags &= ~JBD2_BARRIER;
762 write_unlock(&journal->j_state_lock);
765 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
768 struct inode *inode = NULL; /* the journal inode */
769 journal_t *j_journal = NULL;
770 struct ocfs2_dinode *di = NULL;
771 struct buffer_head *bh = NULL;
772 struct ocfs2_super *osb;
777 osb = journal->j_osb;
779 /* already have the inode for our journal */
780 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
787 if (is_bad_inode(inode)) {
788 mlog(ML_ERROR, "access error (bad inode)\n");
795 SET_INODE_JOURNAL(inode);
796 OCFS2_I(inode)->ip_open_count++;
798 /* Skip recovery waits here - journal inode metadata never
799 * changes in a live cluster so it can be considered an
800 * exception to the rule. */
801 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
803 if (status != -ERESTARTSYS)
804 mlog(ML_ERROR, "Could not get lock on journal!\n");
809 di = (struct ocfs2_dinode *)bh->b_data;
811 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
812 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
818 mlog(0, "inode->i_size = %lld\n", inode->i_size);
819 mlog(0, "inode->i_blocks = %llu\n",
820 (unsigned long long)inode->i_blocks);
821 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
823 /* call the kernels journal init function now */
824 j_journal = jbd2_journal_init_inode(inode);
825 if (j_journal == NULL) {
826 mlog(ML_ERROR, "Linux journal layer error\n");
831 mlog(0, "Returned from jbd2_journal_init_inode\n");
832 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
834 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
835 OCFS2_JOURNAL_DIRTY_FL);
837 journal->j_journal = j_journal;
838 journal->j_inode = inode;
841 ocfs2_set_journal_params(osb);
843 journal->j_state = OCFS2_JOURNAL_LOADED;
849 ocfs2_inode_unlock(inode, 1);
852 OCFS2_I(inode)->ip_open_count--;
861 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
863 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
866 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
868 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
871 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
872 int dirty, int replayed)
876 struct ocfs2_journal *journal = osb->journal;
877 struct buffer_head *bh = journal->j_bh;
878 struct ocfs2_dinode *fe;
880 fe = (struct ocfs2_dinode *)bh->b_data;
882 /* The journal bh on the osb always comes from ocfs2_journal_init()
883 * and was validated there inside ocfs2_inode_lock_full(). It's a
884 * code bug if we mess it up. */
885 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
887 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
889 flags |= OCFS2_JOURNAL_DIRTY_FL;
891 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
892 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
895 ocfs2_bump_recovery_generation(fe);
897 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
898 status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
907 * If the journal has been kmalloc'd it needs to be freed after this
910 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
912 struct ocfs2_journal *journal = NULL;
914 struct inode *inode = NULL;
915 int num_running_trans = 0;
919 journal = osb->journal;
923 inode = journal->j_inode;
925 if (journal->j_state != OCFS2_JOURNAL_LOADED)
928 /* need to inc inode use count - jbd2_journal_destroy will iput. */
932 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
933 if (num_running_trans > 0)
934 mlog(0, "Shutting down journal: must wait on %d "
935 "running transactions!\n",
938 /* Do a commit_cache here. It will flush our journal, *and*
939 * release any locks that are still held.
940 * set the SHUTDOWN flag and release the trans lock.
941 * the commit thread will take the trans lock for us below. */
942 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
944 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
945 * drop the trans_lock (which we want to hold until we
946 * completely destroy the journal. */
947 if (osb->commit_task) {
948 /* Wait for the commit thread */
949 mlog(0, "Waiting for ocfs2commit to exit....\n");
950 kthread_stop(osb->commit_task);
951 osb->commit_task = NULL;
954 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
956 if (ocfs2_mount_local(osb)) {
957 jbd2_journal_lock_updates(journal->j_journal);
958 status = jbd2_journal_flush(journal->j_journal);
959 jbd2_journal_unlock_updates(journal->j_journal);
966 * Do not toggle if flush was unsuccessful otherwise
967 * will leave dirty metadata in a "clean" journal
969 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
974 /* Shutdown the kernel journal system */
975 jbd2_journal_destroy(journal->j_journal);
976 journal->j_journal = NULL;
978 OCFS2_I(inode)->ip_open_count--;
980 /* unlock our journal */
981 ocfs2_inode_unlock(inode, 1);
983 brelse(journal->j_bh);
984 journal->j_bh = NULL;
986 journal->j_state = OCFS2_JOURNAL_FREE;
988 // up_write(&journal->j_trans_barrier);
995 static void ocfs2_clear_journal_error(struct super_block *sb,
1001 olderr = jbd2_journal_errno(journal);
1003 mlog(ML_ERROR, "File system error %d recorded in "
1004 "journal %u.\n", olderr, slot);
1005 mlog(ML_ERROR, "File system on device %s needs checking.\n",
1008 jbd2_journal_ack_err(journal);
1009 jbd2_journal_clear_err(journal);
1013 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
1016 struct ocfs2_super *osb;
1020 osb = journal->j_osb;
1022 status = jbd2_journal_load(journal->j_journal);
1024 mlog(ML_ERROR, "Failed to load journal!\n");
1028 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1030 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1036 /* Launch the commit thread */
1038 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1040 if (IS_ERR(osb->commit_task)) {
1041 status = PTR_ERR(osb->commit_task);
1042 osb->commit_task = NULL;
1043 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1044 "error=%d", status);
1048 osb->commit_task = NULL;
1056 /* 'full' flag tells us whether we clear out all blocks or if we just
1057 * mark the journal clean */
1058 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1064 status = jbd2_journal_wipe(journal->j_journal, full);
1070 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1079 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1082 struct ocfs2_recovery_map *rm = osb->recovery_map;
1084 spin_lock(&osb->osb_lock);
1085 empty = (rm->rm_used == 0);
1086 spin_unlock(&osb->osb_lock);
1091 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1093 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1097 * JBD Might read a cached version of another nodes journal file. We
1098 * don't want this as this file changes often and we get no
1099 * notification on those changes. The only way to be sure that we've
1100 * got the most up to date version of those blocks then is to force
1101 * read them off disk. Just searching through the buffer cache won't
1102 * work as there may be pages backing this file which are still marked
1103 * up to date. We know things can't change on this file underneath us
1104 * as we have the lock by now :)
1106 static int ocfs2_force_read_journal(struct inode *inode)
1110 u64 v_blkno, p_blkno, p_blocks, num_blocks;
1111 #define CONCURRENT_JOURNAL_FILL 32ULL
1112 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
1114 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
1116 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
1118 while (v_blkno < num_blocks) {
1119 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1120 &p_blkno, &p_blocks, NULL);
1126 if (p_blocks > CONCURRENT_JOURNAL_FILL)
1127 p_blocks = CONCURRENT_JOURNAL_FILL;
1129 /* We are reading journal data which should not
1130 * be put in the uptodate cache */
1131 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
1132 p_blkno, p_blocks, bhs);
1138 for(i = 0; i < p_blocks; i++) {
1143 v_blkno += p_blocks;
1147 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
1153 struct ocfs2_la_recovery_item {
1154 struct list_head lri_list;
1156 struct ocfs2_dinode *lri_la_dinode;
1157 struct ocfs2_dinode *lri_tl_dinode;
1158 struct ocfs2_quota_recovery *lri_qrec;
1161 /* Does the second half of the recovery process. By this point, the
1162 * node is marked clean and can actually be considered recovered,
1163 * hence it's no longer in the recovery map, but there's still some
1164 * cleanup we can do which shouldn't happen within the recovery thread
1165 * as locking in that context becomes very difficult if we are to take
1166 * recovering nodes into account.
1168 * NOTE: This function can and will sleep on recovery of other nodes
1169 * during cluster locking, just like any other ocfs2 process.
1171 void ocfs2_complete_recovery(struct work_struct *work)
1174 struct ocfs2_journal *journal =
1175 container_of(work, struct ocfs2_journal, j_recovery_work);
1176 struct ocfs2_super *osb = journal->j_osb;
1177 struct ocfs2_dinode *la_dinode, *tl_dinode;
1178 struct ocfs2_la_recovery_item *item, *n;
1179 struct ocfs2_quota_recovery *qrec;
1180 LIST_HEAD(tmp_la_list);
1182 mlog(0, "completing recovery from keventd\n");
1184 spin_lock(&journal->j_lock);
1185 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1186 spin_unlock(&journal->j_lock);
1188 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1189 list_del_init(&item->lri_list);
1191 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
1193 ocfs2_wait_on_quotas(osb);
1195 la_dinode = item->lri_la_dinode;
1197 mlog(0, "Clean up local alloc %llu\n",
1198 (unsigned long long)le64_to_cpu(la_dinode->i_blkno));
1200 ret = ocfs2_complete_local_alloc_recovery(osb,
1208 tl_dinode = item->lri_tl_dinode;
1210 mlog(0, "Clean up truncate log %llu\n",
1211 (unsigned long long)le64_to_cpu(tl_dinode->i_blkno));
1213 ret = ocfs2_complete_truncate_log_recovery(osb,
1221 ret = ocfs2_recover_orphans(osb, item->lri_slot);
1225 qrec = item->lri_qrec;
1227 mlog(0, "Recovering quota files");
1228 ret = ocfs2_finish_quota_recovery(osb, qrec,
1232 /* Recovery info is already freed now */
1238 mlog(0, "Recovery completion\n");
1242 /* NOTE: This function always eats your references to la_dinode and
1243 * tl_dinode, either manually on error, or by passing them to
1244 * ocfs2_complete_recovery */
1245 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1247 struct ocfs2_dinode *la_dinode,
1248 struct ocfs2_dinode *tl_dinode,
1249 struct ocfs2_quota_recovery *qrec)
1251 struct ocfs2_la_recovery_item *item;
1253 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1255 /* Though we wish to avoid it, we are in fact safe in
1256 * skipping local alloc cleanup as fsck.ocfs2 is more
1257 * than capable of reclaiming unused space. */
1265 ocfs2_free_quota_recovery(qrec);
1267 mlog_errno(-ENOMEM);
1271 INIT_LIST_HEAD(&item->lri_list);
1272 item->lri_la_dinode = la_dinode;
1273 item->lri_slot = slot_num;
1274 item->lri_tl_dinode = tl_dinode;
1275 item->lri_qrec = qrec;
1277 spin_lock(&journal->j_lock);
1278 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1279 queue_work(ocfs2_wq, &journal->j_recovery_work);
1280 spin_unlock(&journal->j_lock);
1283 /* Called by the mount code to queue recovery the last part of
1284 * recovery for it's own and offline slot(s). */
1285 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1287 struct ocfs2_journal *journal = osb->journal;
1289 /* No need to queue up our truncate_log as regular cleanup will catch
1291 ocfs2_queue_recovery_completion(journal, osb->slot_num,
1292 osb->local_alloc_copy, NULL, NULL);
1293 ocfs2_schedule_truncate_log_flush(osb, 0);
1295 osb->local_alloc_copy = NULL;
1298 /* queue to recover orphan slots for all offline slots */
1299 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1300 ocfs2_queue_replay_slots(osb);
1301 ocfs2_free_replay_slots(osb);
1304 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1306 if (osb->quota_rec) {
1307 ocfs2_queue_recovery_completion(osb->journal,
1312 osb->quota_rec = NULL;
1316 static int __ocfs2_recovery_thread(void *arg)
1318 int status, node_num, slot_num;
1319 struct ocfs2_super *osb = arg;
1320 struct ocfs2_recovery_map *rm = osb->recovery_map;
1321 int *rm_quota = NULL;
1322 int rm_quota_used = 0, i;
1323 struct ocfs2_quota_recovery *qrec;
1325 status = ocfs2_wait_on_mount(osb);
1330 rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1336 status = ocfs2_super_lock(osb, 1);
1342 status = ocfs2_compute_replay_slots(osb);
1346 /* queue recovery for our own slot */
1347 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1350 spin_lock(&osb->osb_lock);
1351 while (rm->rm_used) {
1352 /* It's always safe to remove entry zero, as we won't
1353 * clear it until ocfs2_recover_node() has succeeded. */
1354 node_num = rm->rm_entries[0];
1355 spin_unlock(&osb->osb_lock);
1356 mlog(0, "checking node %d\n", node_num);
1357 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1358 if (slot_num == -ENOENT) {
1360 mlog(0, "no slot for this node, so no recovery"
1364 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1366 /* It is a bit subtle with quota recovery. We cannot do it
1367 * immediately because we have to obtain cluster locks from
1368 * quota files and we also don't want to just skip it because
1369 * then quota usage would be out of sync until some node takes
1370 * the slot. So we remember which nodes need quota recovery
1371 * and when everything else is done, we recover quotas. */
1372 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1373 if (i == rm_quota_used)
1374 rm_quota[rm_quota_used++] = slot_num;
1376 status = ocfs2_recover_node(osb, node_num, slot_num);
1379 ocfs2_recovery_map_clear(osb, node_num);
1382 "Error %d recovering node %d on device (%u,%u)!\n",
1384 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1385 mlog(ML_ERROR, "Volume requires unmount.\n");
1388 spin_lock(&osb->osb_lock);
1390 spin_unlock(&osb->osb_lock);
1391 mlog(0, "All nodes recovered\n");
1393 /* Refresh all journal recovery generations from disk */
1394 status = ocfs2_check_journals_nolocks(osb);
1395 status = (status == -EROFS) ? 0 : status;
1399 /* Now it is right time to recover quotas... We have to do this under
1400 * superblock lock so that noone can start using the slot (and crash)
1401 * before we recover it */
1402 for (i = 0; i < rm_quota_used; i++) {
1403 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1405 status = PTR_ERR(qrec);
1409 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1413 ocfs2_super_unlock(osb, 1);
1415 /* queue recovery for offline slots */
1416 ocfs2_queue_replay_slots(osb);
1419 mutex_lock(&osb->recovery_lock);
1420 if (!status && !ocfs2_recovery_completed(osb)) {
1421 mutex_unlock(&osb->recovery_lock);
1425 ocfs2_free_replay_slots(osb);
1426 osb->recovery_thread_task = NULL;
1427 mb(); /* sync with ocfs2_recovery_thread_running */
1428 wake_up(&osb->recovery_event);
1430 mutex_unlock(&osb->recovery_lock);
1436 /* no one is callint kthread_stop() for us so the kthread() api
1437 * requires that we call do_exit(). And it isn't exported, but
1438 * complete_and_exit() seems to be a minimal wrapper around it. */
1439 complete_and_exit(NULL, status);
1443 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1445 mlog(0, "(node_num=%d, osb->node_num = %d)\n",
1446 node_num, osb->node_num);
1448 mutex_lock(&osb->recovery_lock);
1449 if (osb->disable_recovery)
1452 /* People waiting on recovery will wait on
1453 * the recovery map to empty. */
1454 if (ocfs2_recovery_map_set(osb, node_num))
1455 mlog(0, "node %d already in recovery map.\n", node_num);
1457 mlog(0, "starting recovery thread...\n");
1459 if (osb->recovery_thread_task)
1462 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1464 if (IS_ERR(osb->recovery_thread_task)) {
1465 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1466 osb->recovery_thread_task = NULL;
1470 mutex_unlock(&osb->recovery_lock);
1471 wake_up(&osb->recovery_event);
1476 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1478 struct buffer_head **bh,
1479 struct inode **ret_inode)
1481 int status = -EACCES;
1482 struct inode *inode = NULL;
1484 BUG_ON(slot_num >= osb->max_slots);
1486 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1488 if (!inode || is_bad_inode(inode)) {
1492 SET_INODE_JOURNAL(inode);
1494 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1504 if (status || !ret_inode)
1512 /* Does the actual journal replay and marks the journal inode as
1513 * clean. Will only replay if the journal inode is marked dirty. */
1514 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1521 struct inode *inode = NULL;
1522 struct ocfs2_dinode *fe;
1523 journal_t *journal = NULL;
1524 struct buffer_head *bh = NULL;
1527 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1533 fe = (struct ocfs2_dinode *)bh->b_data;
1534 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1539 * As the fs recovery is asynchronous, there is a small chance that
1540 * another node mounted (and recovered) the slot before the recovery
1541 * thread could get the lock. To handle that, we dirty read the journal
1542 * inode for that slot to get the recovery generation. If it is
1543 * different than what we expected, the slot has been recovered.
1544 * If not, it needs recovery.
1546 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1547 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num,
1548 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1549 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1554 /* Continue with recovery as the journal has not yet been recovered */
1556 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1558 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status);
1559 if (status != -ERESTARTSYS)
1560 mlog(ML_ERROR, "Could not lock journal!\n");
1565 fe = (struct ocfs2_dinode *) bh->b_data;
1567 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1568 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1570 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1571 mlog(0, "No recovery required for node %d\n", node_num);
1572 /* Refresh recovery generation for the slot */
1573 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1577 /* we need to run complete recovery for offline orphan slots */
1578 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1580 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1582 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1584 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1586 status = ocfs2_force_read_journal(inode);
1592 mlog(0, "calling journal_init_inode\n");
1593 journal = jbd2_journal_init_inode(inode);
1594 if (journal == NULL) {
1595 mlog(ML_ERROR, "Linux journal layer error\n");
1600 status = jbd2_journal_load(journal);
1605 jbd2_journal_destroy(journal);
1609 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1611 /* wipe the journal */
1612 mlog(0, "flushing the journal.\n");
1613 jbd2_journal_lock_updates(journal);
1614 status = jbd2_journal_flush(journal);
1615 jbd2_journal_unlock_updates(journal);
1619 /* This will mark the node clean */
1620 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1621 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1622 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1624 /* Increment recovery generation to indicate successful recovery */
1625 ocfs2_bump_recovery_generation(fe);
1626 osb->slot_recovery_generations[slot_num] =
1627 ocfs2_get_recovery_generation(fe);
1629 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1630 status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1637 jbd2_journal_destroy(journal);
1640 /* drop the lock on this nodes journal */
1642 ocfs2_inode_unlock(inode, 1);
1654 * Do the most important parts of node recovery:
1655 * - Replay it's journal
1656 * - Stamp a clean local allocator file
1657 * - Stamp a clean truncate log
1658 * - Mark the node clean
1660 * If this function completes without error, a node in OCFS2 can be
1661 * said to have been safely recovered. As a result, failure during the
1662 * second part of a nodes recovery process (local alloc recovery) is
1663 * far less concerning.
1665 static int ocfs2_recover_node(struct ocfs2_super *osb,
1666 int node_num, int slot_num)
1669 struct ocfs2_dinode *la_copy = NULL;
1670 struct ocfs2_dinode *tl_copy = NULL;
1672 mlog(0, "(node_num=%d, slot_num=%d, osb->node_num = %d)\n",
1673 node_num, slot_num, osb->node_num);
1675 /* Should not ever be called to recover ourselves -- in that
1676 * case we should've called ocfs2_journal_load instead. */
1677 BUG_ON(osb->node_num == node_num);
1679 status = ocfs2_replay_journal(osb, node_num, slot_num);
1681 if (status == -EBUSY) {
1682 mlog(0, "Skipping recovery for slot %u (node %u) "
1683 "as another node has recovered it\n", slot_num,
1692 /* Stamp a clean local alloc file AFTER recovering the journal... */
1693 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1699 /* An error from begin_truncate_log_recovery is not
1700 * serious enough to warrant halting the rest of
1702 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1706 /* Likewise, this would be a strange but ultimately not so
1707 * harmful place to get an error... */
1708 status = ocfs2_clear_slot(osb, slot_num);
1712 /* This will kfree the memory pointed to by la_copy and tl_copy */
1713 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 mlog(0, "Slot %u recovery generation is %u\n", 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)) {
1834 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1835 * randomness to the timeout to minimize multple nodes firing the timer at the
1838 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1842 get_random_bytes(&time, sizeof(time));
1843 time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1844 return msecs_to_jiffies(time);
1848 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1849 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1850 * is done to catch any orphans that are left over in orphan directories.
1852 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1853 * seconds. It gets an EX lock on os_lockres and checks sequence number
1854 * stored in LVB. If the sequence number has changed, it means some other
1855 * node has done the scan. This node skips the scan and tracks the
1856 * sequence number. If the sequence number didn't change, it means a scan
1857 * hasn't happened. The node queues a scan and increments the
1858 * sequence number in the LVB.
1860 void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1862 struct ocfs2_orphan_scan *os;
1866 os = &osb->osb_orphan_scan;
1868 mlog(0, "Begin orphan scan\n");
1870 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1873 status = ocfs2_orphan_scan_lock(osb, &seqno);
1875 if (status != -EAGAIN)
1880 /* Do no queue the tasks if the volume is being umounted */
1881 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1884 if (os->os_seqno != seqno) {
1885 os->os_seqno = seqno;
1889 for (i = 0; i < osb->max_slots; i++)
1890 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1893 * We queued a recovery on orphan slots, increment the sequence
1894 * number and update LVB so other node will skip the scan for a while
1898 os->os_scantime = CURRENT_TIME;
1900 ocfs2_orphan_scan_unlock(osb, seqno);
1902 mlog(0, "Orphan scan completed\n");
1906 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1907 void ocfs2_orphan_scan_work(struct work_struct *work)
1909 struct ocfs2_orphan_scan *os;
1910 struct ocfs2_super *osb;
1912 os = container_of(work, struct ocfs2_orphan_scan,
1913 os_orphan_scan_work.work);
1916 mutex_lock(&os->os_lock);
1917 ocfs2_queue_orphan_scan(osb);
1918 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1919 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1920 ocfs2_orphan_scan_timeout());
1921 mutex_unlock(&os->os_lock);
1924 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1926 struct ocfs2_orphan_scan *os;
1928 os = &osb->osb_orphan_scan;
1929 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1930 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1931 mutex_lock(&os->os_lock);
1932 cancel_delayed_work(&os->os_orphan_scan_work);
1933 mutex_unlock(&os->os_lock);
1937 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1939 struct ocfs2_orphan_scan *os;
1941 os = &osb->osb_orphan_scan;
1945 mutex_init(&os->os_lock);
1946 INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
1949 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
1951 struct ocfs2_orphan_scan *os;
1953 os = &osb->osb_orphan_scan;
1954 os->os_scantime = CURRENT_TIME;
1955 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1956 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1958 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
1959 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1960 ocfs2_orphan_scan_timeout());
1964 struct ocfs2_orphan_filldir_priv {
1966 struct ocfs2_super *osb;
1969 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1970 loff_t pos, u64 ino, unsigned type)
1972 struct ocfs2_orphan_filldir_priv *p = priv;
1975 if (name_len == 1 && !strncmp(".", name, 1))
1977 if (name_len == 2 && !strncmp("..", name, 2))
1980 /* Skip bad inodes so that recovery can continue */
1981 iter = ocfs2_iget(p->osb, ino,
1982 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
1986 mlog(0, "queue orphan %llu\n",
1987 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1988 /* No locking is required for the next_orphan queue as there
1989 * is only ever a single process doing orphan recovery. */
1990 OCFS2_I(iter)->ip_next_orphan = p->head;
1996 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1998 struct inode **head)
2001 struct inode *orphan_dir_inode = NULL;
2002 struct ocfs2_orphan_filldir_priv priv;
2008 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2009 ORPHAN_DIR_SYSTEM_INODE,
2011 if (!orphan_dir_inode) {
2017 mutex_lock(&orphan_dir_inode->i_mutex);
2018 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2024 status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
2025 ocfs2_orphan_filldir);
2034 ocfs2_inode_unlock(orphan_dir_inode, 0);
2036 mutex_unlock(&orphan_dir_inode->i_mutex);
2037 iput(orphan_dir_inode);
2041 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2046 spin_lock(&osb->osb_lock);
2047 ret = !osb->osb_orphan_wipes[slot];
2048 spin_unlock(&osb->osb_lock);
2052 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2055 spin_lock(&osb->osb_lock);
2056 /* Mark ourselves such that new processes in delete_inode()
2057 * know to quit early. */
2058 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2059 while (osb->osb_orphan_wipes[slot]) {
2060 /* If any processes are already in the middle of an
2061 * orphan wipe on this dir, then we need to wait for
2063 spin_unlock(&osb->osb_lock);
2064 wait_event_interruptible(osb->osb_wipe_event,
2065 ocfs2_orphan_recovery_can_continue(osb, slot));
2066 spin_lock(&osb->osb_lock);
2068 spin_unlock(&osb->osb_lock);
2071 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2074 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2078 * Orphan recovery. Each mounted node has it's own orphan dir which we
2079 * must run during recovery. Our strategy here is to build a list of
2080 * the inodes in the orphan dir and iget/iput them. The VFS does
2081 * (most) of the rest of the work.
2083 * Orphan recovery can happen at any time, not just mount so we have a
2084 * couple of extra considerations.
2086 * - We grab as many inodes as we can under the orphan dir lock -
2087 * doing iget() outside the orphan dir risks getting a reference on
2089 * - We must be sure not to deadlock with other processes on the
2090 * system wanting to run delete_inode(). This can happen when they go
2091 * to lock the orphan dir and the orphan recovery process attempts to
2092 * iget() inside the orphan dir lock. This can be avoided by
2093 * advertising our state to ocfs2_delete_inode().
2095 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2099 struct inode *inode = NULL;
2101 struct ocfs2_inode_info *oi;
2103 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
2105 ocfs2_mark_recovering_orphan_dir(osb, slot);
2106 ret = ocfs2_queue_orphans(osb, slot, &inode);
2107 ocfs2_clear_recovering_orphan_dir(osb, slot);
2109 /* Error here should be noted, but we want to continue with as
2110 * many queued inodes as we've got. */
2115 oi = OCFS2_I(inode);
2116 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
2118 iter = oi->ip_next_orphan;
2120 spin_lock(&oi->ip_lock);
2121 /* The remote delete code may have set these on the
2122 * assumption that the other node would wipe them
2123 * successfully. If they are still in the node's
2124 * orphan dir, we need to reset that state. */
2125 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
2127 /* Set the proper information to get us going into
2128 * ocfs2_delete_inode. */
2129 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2130 spin_unlock(&oi->ip_lock);
2140 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2142 /* This check is good because ocfs2 will wait on our recovery
2143 * thread before changing it to something other than MOUNTED
2145 wait_event(osb->osb_mount_event,
2146 (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2147 atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2148 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2150 /* If there's an error on mount, then we may never get to the
2151 * MOUNTED flag, but this is set right before
2152 * dismount_volume() so we can trust it. */
2153 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2154 mlog(0, "mount error, exiting!\n");
2161 static int ocfs2_commit_thread(void *arg)
2164 struct ocfs2_super *osb = arg;
2165 struct ocfs2_journal *journal = osb->journal;
2167 /* we can trust j_num_trans here because _should_stop() is only set in
2168 * shutdown and nobody other than ourselves should be able to start
2169 * transactions. committing on shutdown might take a few iterations
2170 * as final transactions put deleted inodes on the list */
2171 while (!(kthread_should_stop() &&
2172 atomic_read(&journal->j_num_trans) == 0)) {
2174 wait_event_interruptible(osb->checkpoint_event,
2175 atomic_read(&journal->j_num_trans)
2176 || kthread_should_stop());
2178 status = ocfs2_commit_cache(osb);
2182 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2184 "commit_thread: %u transactions pending on "
2186 atomic_read(&journal->j_num_trans));
2193 /* Reads all the journal inodes without taking any cluster locks. Used
2194 * for hard readonly access to determine whether any journal requires
2195 * recovery. Also used to refresh the recovery generation numbers after
2196 * a journal has been recovered by another node.
2198 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2202 struct buffer_head *di_bh = NULL;
2203 struct ocfs2_dinode *di;
2204 int journal_dirty = 0;
2206 for(slot = 0; slot < osb->max_slots; slot++) {
2207 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2213 di = (struct ocfs2_dinode *) di_bh->b_data;
2215 osb->slot_recovery_generations[slot] =
2216 ocfs2_get_recovery_generation(di);
2218 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2219 OCFS2_JOURNAL_DIRTY_FL)