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 unsigned long old_id;
305 struct ocfs2_journal *journal = NULL;
309 journal = osb->journal;
311 /* Flush all pending commits and checkpoint the journal. */
312 down_write(&journal->j_trans_barrier);
314 if (atomic_read(&journal->j_num_trans) == 0) {
315 up_write(&journal->j_trans_barrier);
316 mlog(0, "No transactions for me to flush!\n");
320 jbd2_journal_lock_updates(journal->j_journal);
321 status = jbd2_journal_flush(journal->j_journal);
322 jbd2_journal_unlock_updates(journal->j_journal);
324 up_write(&journal->j_trans_barrier);
329 old_id = ocfs2_inc_trans_id(journal);
331 flushed = atomic_read(&journal->j_num_trans);
332 atomic_set(&journal->j_num_trans, 0);
333 up_write(&journal->j_trans_barrier);
335 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
336 journal->j_trans_id, flushed);
338 ocfs2_wake_downconvert_thread(osb);
339 wake_up(&journal->j_checkpointed);
345 /* pass it NULL and it will allocate a new handle object for you. If
346 * you pass it a handle however, it may still return error, in which
347 * case it has free'd the passed handle for you. */
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 down_read(&osb->journal->j_trans_barrier);
367 handle = jbd2_journal_start(journal, max_buffs);
368 if (IS_ERR(handle)) {
369 up_read(&osb->journal->j_trans_barrier);
371 mlog_errno(PTR_ERR(handle));
373 if (is_journal_aborted(journal)) {
374 ocfs2_abort(osb->sb, "Detected aborted journal");
375 handle = ERR_PTR(-EROFS);
378 if (!ocfs2_mount_local(osb))
379 atomic_inc(&(osb->journal->j_num_trans));
385 int ocfs2_commit_trans(struct ocfs2_super *osb,
389 struct ocfs2_journal *journal = osb->journal;
393 nested = handle->h_ref > 1;
394 ret = jbd2_journal_stop(handle);
399 up_read(&journal->j_trans_barrier);
405 * 'nblocks' is what you want to add to the current
406 * transaction. extend_trans will either extend the current handle by
407 * nblocks, or commit it and start a new one with nblocks credits.
409 * This might call jbd2_journal_restart() which will commit dirty buffers
410 * and then restart the transaction. Before calling
411 * ocfs2_extend_trans(), any changed blocks should have been
412 * dirtied. After calling it, all blocks which need to be changed must
413 * go through another set of journal_access/journal_dirty calls.
415 * WARNING: This will not release any semaphores or disk locks taken
416 * during the transaction, so make sure they were taken *before*
417 * start_trans or we'll have ordering deadlocks.
419 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
420 * good because transaction ids haven't yet been recorded on the
421 * cluster locks associated with this handle.
423 int ocfs2_extend_trans(handle_t *handle, int nblocks)
432 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
434 #ifdef CONFIG_OCFS2_DEBUG_FS
437 status = jbd2_journal_extend(handle, nblocks);
446 "jbd2_journal_extend failed, trying "
447 "jbd2_journal_restart\n");
448 status = jbd2_journal_restart(handle, nblocks);
462 struct ocfs2_triggers {
463 struct jbd2_buffer_trigger_type ot_triggers;
467 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
469 return container_of(triggers, struct ocfs2_triggers, ot_triggers);
472 static void ocfs2_commit_trigger(struct jbd2_buffer_trigger_type *triggers,
473 struct buffer_head *bh,
474 void *data, size_t size)
476 struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
479 * We aren't guaranteed to have the superblock here, so we
480 * must unconditionally compute the ecc data.
481 * __ocfs2_journal_access() will only set the triggers if
482 * metaecc is enabled.
484 ocfs2_block_check_compute(data, size, data + ot->ot_offset);
488 * Quota blocks have their own trigger because the struct ocfs2_block_check
489 * offset depends on the blocksize.
491 static void ocfs2_dq_commit_trigger(struct jbd2_buffer_trigger_type *triggers,
492 struct buffer_head *bh,
493 void *data, size_t size)
495 struct ocfs2_disk_dqtrailer *dqt =
496 ocfs2_block_dqtrailer(size, data);
499 * We aren't guaranteed to have the superblock here, so we
500 * must unconditionally compute the ecc data.
501 * __ocfs2_journal_access() will only set the triggers if
502 * metaecc is enabled.
504 ocfs2_block_check_compute(data, size, &dqt->dq_check);
508 * Directory blocks also have their own trigger because the
509 * struct ocfs2_block_check offset depends on the blocksize.
511 static void ocfs2_db_commit_trigger(struct jbd2_buffer_trigger_type *triggers,
512 struct buffer_head *bh,
513 void *data, size_t size)
515 struct ocfs2_dir_block_trailer *trailer =
516 ocfs2_dir_trailer_from_size(size, data);
519 * We aren't guaranteed to have the superblock here, so we
520 * must unconditionally compute the ecc data.
521 * __ocfs2_journal_access() will only set the triggers if
522 * metaecc is enabled.
524 ocfs2_block_check_compute(data, size, &trailer->db_check);
527 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
528 struct buffer_head *bh)
531 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
532 "bh->b_blocknr = %llu\n",
534 (unsigned long long)bh->b_blocknr);
536 /* We aren't guaranteed to have the superblock here - but if we
537 * don't, it'll just crash. */
538 ocfs2_error(bh->b_assoc_map->host->i_sb,
539 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
542 static struct ocfs2_triggers di_triggers = {
544 .t_commit = ocfs2_commit_trigger,
545 .t_abort = ocfs2_abort_trigger,
547 .ot_offset = offsetof(struct ocfs2_dinode, i_check),
550 static struct ocfs2_triggers eb_triggers = {
552 .t_commit = ocfs2_commit_trigger,
553 .t_abort = ocfs2_abort_trigger,
555 .ot_offset = offsetof(struct ocfs2_extent_block, h_check),
558 static struct ocfs2_triggers gd_triggers = {
560 .t_commit = ocfs2_commit_trigger,
561 .t_abort = ocfs2_abort_trigger,
563 .ot_offset = offsetof(struct ocfs2_group_desc, bg_check),
566 static struct ocfs2_triggers db_triggers = {
568 .t_commit = ocfs2_db_commit_trigger,
569 .t_abort = ocfs2_abort_trigger,
573 static struct ocfs2_triggers xb_triggers = {
575 .t_commit = ocfs2_commit_trigger,
576 .t_abort = ocfs2_abort_trigger,
578 .ot_offset = offsetof(struct ocfs2_xattr_block, xb_check),
581 static struct ocfs2_triggers dq_triggers = {
583 .t_commit = ocfs2_dq_commit_trigger,
584 .t_abort = ocfs2_abort_trigger,
588 static struct ocfs2_triggers dr_triggers = {
590 .t_commit = ocfs2_commit_trigger,
591 .t_abort = ocfs2_abort_trigger,
593 .ot_offset = offsetof(struct ocfs2_dx_root_block, dr_check),
596 static struct ocfs2_triggers dl_triggers = {
598 .t_commit = ocfs2_commit_trigger,
599 .t_abort = ocfs2_abort_trigger,
601 .ot_offset = offsetof(struct ocfs2_dx_leaf, dl_check),
604 static int __ocfs2_journal_access(handle_t *handle,
605 struct ocfs2_caching_info *ci,
606 struct buffer_head *bh,
607 struct ocfs2_triggers *triggers,
611 struct ocfs2_super *osb =
612 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
614 BUG_ON(!ci || !ci->ci_ops);
618 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
619 (unsigned long long)bh->b_blocknr, type,
620 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
621 "OCFS2_JOURNAL_ACCESS_CREATE" :
622 "OCFS2_JOURNAL_ACCESS_WRITE",
625 /* we can safely remove this assertion after testing. */
626 if (!buffer_uptodate(bh)) {
627 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
628 mlog(ML_ERROR, "b_blocknr=%llu\n",
629 (unsigned long long)bh->b_blocknr);
633 /* Set the current transaction information on the ci so
634 * that the locking code knows whether it can drop it's locks
635 * on this ci or not. We're protected from the commit
636 * thread updating the current transaction id until
637 * ocfs2_commit_trans() because ocfs2_start_trans() took
638 * j_trans_barrier for us. */
639 ocfs2_set_ci_lock_trans(osb->journal, ci);
641 ocfs2_metadata_cache_io_lock(ci);
643 case OCFS2_JOURNAL_ACCESS_CREATE:
644 case OCFS2_JOURNAL_ACCESS_WRITE:
645 status = jbd2_journal_get_write_access(handle, bh);
648 case OCFS2_JOURNAL_ACCESS_UNDO:
649 status = jbd2_journal_get_undo_access(handle, bh);
654 mlog(ML_ERROR, "Uknown access type!\n");
656 if (!status && ocfs2_meta_ecc(osb) && triggers)
657 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
658 ocfs2_metadata_cache_io_unlock(ci);
661 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
668 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
669 struct buffer_head *bh, int type)
671 return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
674 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
675 struct buffer_head *bh, int type)
677 return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
680 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
681 struct buffer_head *bh, int type)
683 return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
686 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
687 struct buffer_head *bh, int type)
689 return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
692 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
693 struct buffer_head *bh, int type)
695 return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
698 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
699 struct buffer_head *bh, int type)
701 return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
704 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
705 struct buffer_head *bh, int type)
707 return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
710 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
711 struct buffer_head *bh, int type)
713 return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
716 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
717 struct buffer_head *bh, int type)
719 return __ocfs2_journal_access(handle, ci, bh, NULL, type);
722 int ocfs2_journal_dirty(handle_t *handle,
723 struct buffer_head *bh)
727 mlog_entry("(bh->b_blocknr=%llu)\n",
728 (unsigned long long)bh->b_blocknr);
730 status = jbd2_journal_dirty_metadata(handle, bh);
732 mlog(ML_ERROR, "Could not dirty metadata buffer. "
733 "(bh->b_blocknr=%llu)\n",
734 (unsigned long long)bh->b_blocknr);
740 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
742 void ocfs2_set_journal_params(struct ocfs2_super *osb)
744 journal_t *journal = osb->journal->j_journal;
745 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
747 if (osb->osb_commit_interval)
748 commit_interval = osb->osb_commit_interval;
750 spin_lock(&journal->j_state_lock);
751 journal->j_commit_interval = commit_interval;
752 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
753 journal->j_flags |= JBD2_BARRIER;
755 journal->j_flags &= ~JBD2_BARRIER;
756 spin_unlock(&journal->j_state_lock);
759 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
762 struct inode *inode = NULL; /* the journal inode */
763 journal_t *j_journal = NULL;
764 struct ocfs2_dinode *di = NULL;
765 struct buffer_head *bh = NULL;
766 struct ocfs2_super *osb;
773 osb = journal->j_osb;
775 /* already have the inode for our journal */
776 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
783 if (is_bad_inode(inode)) {
784 mlog(ML_ERROR, "access error (bad inode)\n");
791 SET_INODE_JOURNAL(inode);
792 OCFS2_I(inode)->ip_open_count++;
794 /* Skip recovery waits here - journal inode metadata never
795 * changes in a live cluster so it can be considered an
796 * exception to the rule. */
797 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
799 if (status != -ERESTARTSYS)
800 mlog(ML_ERROR, "Could not get lock on journal!\n");
805 di = (struct ocfs2_dinode *)bh->b_data;
807 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
808 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
814 mlog(0, "inode->i_size = %lld\n", inode->i_size);
815 mlog(0, "inode->i_blocks = %llu\n",
816 (unsigned long long)inode->i_blocks);
817 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
819 /* call the kernels journal init function now */
820 j_journal = jbd2_journal_init_inode(inode);
821 if (j_journal == NULL) {
822 mlog(ML_ERROR, "Linux journal layer error\n");
827 mlog(0, "Returned from jbd2_journal_init_inode\n");
828 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
830 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
831 OCFS2_JOURNAL_DIRTY_FL);
833 journal->j_journal = j_journal;
834 journal->j_inode = inode;
837 ocfs2_set_journal_params(osb);
839 journal->j_state = OCFS2_JOURNAL_LOADED;
845 ocfs2_inode_unlock(inode, 1);
848 OCFS2_I(inode)->ip_open_count--;
857 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
859 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
862 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
864 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
867 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
868 int dirty, int replayed)
872 struct ocfs2_journal *journal = osb->journal;
873 struct buffer_head *bh = journal->j_bh;
874 struct ocfs2_dinode *fe;
878 fe = (struct ocfs2_dinode *)bh->b_data;
880 /* The journal bh on the osb always comes from ocfs2_journal_init()
881 * and was validated there inside ocfs2_inode_lock_full(). It's a
882 * code bug if we mess it up. */
883 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
885 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
887 flags |= OCFS2_JOURNAL_DIRTY_FL;
889 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
890 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
893 ocfs2_bump_recovery_generation(fe);
895 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
896 status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
905 * If the journal has been kmalloc'd it needs to be freed after this
908 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
910 struct ocfs2_journal *journal = NULL;
912 struct inode *inode = NULL;
913 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;
1022 osb = journal->j_osb;
1024 status = jbd2_journal_load(journal->j_journal);
1026 mlog(ML_ERROR, "Failed to load journal!\n");
1030 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1032 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1038 /* Launch the commit thread */
1040 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1042 if (IS_ERR(osb->commit_task)) {
1043 status = PTR_ERR(osb->commit_task);
1044 osb->commit_task = NULL;
1045 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1046 "error=%d", status);
1050 osb->commit_task = NULL;
1058 /* 'full' flag tells us whether we clear out all blocks or if we just
1059 * mark the journal clean */
1060 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1068 status = jbd2_journal_wipe(journal->j_journal, full);
1074 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1083 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1086 struct ocfs2_recovery_map *rm = osb->recovery_map;
1088 spin_lock(&osb->osb_lock);
1089 empty = (rm->rm_used == 0);
1090 spin_unlock(&osb->osb_lock);
1095 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1097 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1101 * JBD Might read a cached version of another nodes journal file. We
1102 * don't want this as this file changes often and we get no
1103 * notification on those changes. The only way to be sure that we've
1104 * got the most up to date version of those blocks then is to force
1105 * read them off disk. Just searching through the buffer cache won't
1106 * work as there may be pages backing this file which are still marked
1107 * up to date. We know things can't change on this file underneath us
1108 * as we have the lock by now :)
1110 static int ocfs2_force_read_journal(struct inode *inode)
1114 u64 v_blkno, p_blkno, p_blocks, num_blocks;
1115 #define CONCURRENT_JOURNAL_FILL 32ULL
1116 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
1120 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
1122 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
1124 while (v_blkno < num_blocks) {
1125 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1126 &p_blkno, &p_blocks, NULL);
1132 if (p_blocks > CONCURRENT_JOURNAL_FILL)
1133 p_blocks = CONCURRENT_JOURNAL_FILL;
1135 /* We are reading journal data which should not
1136 * be put in the uptodate cache */
1137 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
1138 p_blkno, p_blocks, bhs);
1144 for(i = 0; i < p_blocks; i++) {
1149 v_blkno += p_blocks;
1153 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
1159 struct ocfs2_la_recovery_item {
1160 struct list_head lri_list;
1162 struct ocfs2_dinode *lri_la_dinode;
1163 struct ocfs2_dinode *lri_tl_dinode;
1164 struct ocfs2_quota_recovery *lri_qrec;
1167 /* Does the second half of the recovery process. By this point, the
1168 * node is marked clean and can actually be considered recovered,
1169 * hence it's no longer in the recovery map, but there's still some
1170 * cleanup we can do which shouldn't happen within the recovery thread
1171 * as locking in that context becomes very difficult if we are to take
1172 * recovering nodes into account.
1174 * NOTE: This function can and will sleep on recovery of other nodes
1175 * during cluster locking, just like any other ocfs2 process.
1177 void ocfs2_complete_recovery(struct work_struct *work)
1180 struct ocfs2_journal *journal =
1181 container_of(work, struct ocfs2_journal, j_recovery_work);
1182 struct ocfs2_super *osb = journal->j_osb;
1183 struct ocfs2_dinode *la_dinode, *tl_dinode;
1184 struct ocfs2_la_recovery_item *item, *n;
1185 struct ocfs2_quota_recovery *qrec;
1186 LIST_HEAD(tmp_la_list);
1190 mlog(0, "completing recovery from keventd\n");
1192 spin_lock(&journal->j_lock);
1193 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1194 spin_unlock(&journal->j_lock);
1196 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1197 list_del_init(&item->lri_list);
1199 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
1201 ocfs2_wait_on_quotas(osb);
1203 la_dinode = item->lri_la_dinode;
1205 mlog(0, "Clean up local alloc %llu\n",
1206 (unsigned long long)le64_to_cpu(la_dinode->i_blkno));
1208 ret = ocfs2_complete_local_alloc_recovery(osb,
1216 tl_dinode = item->lri_tl_dinode;
1218 mlog(0, "Clean up truncate log %llu\n",
1219 (unsigned long long)le64_to_cpu(tl_dinode->i_blkno));
1221 ret = ocfs2_complete_truncate_log_recovery(osb,
1229 ret = ocfs2_recover_orphans(osb, item->lri_slot);
1233 qrec = item->lri_qrec;
1235 mlog(0, "Recovering quota files");
1236 ret = ocfs2_finish_quota_recovery(osb, qrec,
1240 /* Recovery info is already freed now */
1246 mlog(0, "Recovery completion\n");
1250 /* NOTE: This function always eats your references to la_dinode and
1251 * tl_dinode, either manually on error, or by passing them to
1252 * ocfs2_complete_recovery */
1253 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1255 struct ocfs2_dinode *la_dinode,
1256 struct ocfs2_dinode *tl_dinode,
1257 struct ocfs2_quota_recovery *qrec)
1259 struct ocfs2_la_recovery_item *item;
1261 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1263 /* Though we wish to avoid it, we are in fact safe in
1264 * skipping local alloc cleanup as fsck.ocfs2 is more
1265 * than capable of reclaiming unused space. */
1273 ocfs2_free_quota_recovery(qrec);
1275 mlog_errno(-ENOMEM);
1279 INIT_LIST_HEAD(&item->lri_list);
1280 item->lri_la_dinode = la_dinode;
1281 item->lri_slot = slot_num;
1282 item->lri_tl_dinode = tl_dinode;
1283 item->lri_qrec = qrec;
1285 spin_lock(&journal->j_lock);
1286 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1287 queue_work(ocfs2_wq, &journal->j_recovery_work);
1288 spin_unlock(&journal->j_lock);
1291 /* Called by the mount code to queue recovery the last part of
1292 * recovery for it's own and offline slot(s). */
1293 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1295 struct ocfs2_journal *journal = osb->journal;
1297 /* No need to queue up our truncate_log as regular cleanup will catch
1299 ocfs2_queue_recovery_completion(journal, osb->slot_num,
1300 osb->local_alloc_copy, NULL, NULL);
1301 ocfs2_schedule_truncate_log_flush(osb, 0);
1303 osb->local_alloc_copy = NULL;
1306 /* queue to recover orphan slots for all offline slots */
1307 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1308 ocfs2_queue_replay_slots(osb);
1309 ocfs2_free_replay_slots(osb);
1312 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1314 if (osb->quota_rec) {
1315 ocfs2_queue_recovery_completion(osb->journal,
1320 osb->quota_rec = NULL;
1324 static int __ocfs2_recovery_thread(void *arg)
1326 int status, node_num, slot_num;
1327 struct ocfs2_super *osb = arg;
1328 struct ocfs2_recovery_map *rm = osb->recovery_map;
1329 int *rm_quota = NULL;
1330 int rm_quota_used = 0, i;
1331 struct ocfs2_quota_recovery *qrec;
1335 status = ocfs2_wait_on_mount(osb);
1340 rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1346 status = ocfs2_super_lock(osb, 1);
1352 status = ocfs2_compute_replay_slots(osb);
1356 /* queue recovery for our own slot */
1357 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1360 spin_lock(&osb->osb_lock);
1361 while (rm->rm_used) {
1362 /* It's always safe to remove entry zero, as we won't
1363 * clear it until ocfs2_recover_node() has succeeded. */
1364 node_num = rm->rm_entries[0];
1365 spin_unlock(&osb->osb_lock);
1366 mlog(0, "checking node %d\n", node_num);
1367 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1368 if (slot_num == -ENOENT) {
1370 mlog(0, "no slot for this node, so no recovery"
1374 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1376 /* It is a bit subtle with quota recovery. We cannot do it
1377 * immediately because we have to obtain cluster locks from
1378 * quota files and we also don't want to just skip it because
1379 * then quota usage would be out of sync until some node takes
1380 * the slot. So we remember which nodes need quota recovery
1381 * and when everything else is done, we recover quotas. */
1382 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1383 if (i == rm_quota_used)
1384 rm_quota[rm_quota_used++] = slot_num;
1386 status = ocfs2_recover_node(osb, node_num, slot_num);
1389 ocfs2_recovery_map_clear(osb, node_num);
1392 "Error %d recovering node %d on device (%u,%u)!\n",
1394 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1395 mlog(ML_ERROR, "Volume requires unmount.\n");
1398 spin_lock(&osb->osb_lock);
1400 spin_unlock(&osb->osb_lock);
1401 mlog(0, "All nodes recovered\n");
1403 /* Refresh all journal recovery generations from disk */
1404 status = ocfs2_check_journals_nolocks(osb);
1405 status = (status == -EROFS) ? 0 : status;
1409 /* Now it is right time to recover quotas... We have to do this under
1410 * superblock lock so that noone can start using the slot (and crash)
1411 * before we recover it */
1412 for (i = 0; i < rm_quota_used; i++) {
1413 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1415 status = PTR_ERR(qrec);
1419 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1423 ocfs2_super_unlock(osb, 1);
1425 /* queue recovery for offline slots */
1426 ocfs2_queue_replay_slots(osb);
1429 mutex_lock(&osb->recovery_lock);
1430 if (!status && !ocfs2_recovery_completed(osb)) {
1431 mutex_unlock(&osb->recovery_lock);
1435 ocfs2_free_replay_slots(osb);
1436 osb->recovery_thread_task = NULL;
1437 mb(); /* sync with ocfs2_recovery_thread_running */
1438 wake_up(&osb->recovery_event);
1440 mutex_unlock(&osb->recovery_lock);
1446 /* no one is callint kthread_stop() for us so the kthread() api
1447 * requires that we call do_exit(). And it isn't exported, but
1448 * complete_and_exit() seems to be a minimal wrapper around it. */
1449 complete_and_exit(NULL, status);
1453 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1455 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1456 node_num, osb->node_num);
1458 mutex_lock(&osb->recovery_lock);
1459 if (osb->disable_recovery)
1462 /* People waiting on recovery will wait on
1463 * the recovery map to empty. */
1464 if (ocfs2_recovery_map_set(osb, node_num))
1465 mlog(0, "node %d already in recovery map.\n", node_num);
1467 mlog(0, "starting recovery thread...\n");
1469 if (osb->recovery_thread_task)
1472 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1474 if (IS_ERR(osb->recovery_thread_task)) {
1475 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1476 osb->recovery_thread_task = NULL;
1480 mutex_unlock(&osb->recovery_lock);
1481 wake_up(&osb->recovery_event);
1486 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1488 struct buffer_head **bh,
1489 struct inode **ret_inode)
1491 int status = -EACCES;
1492 struct inode *inode = NULL;
1494 BUG_ON(slot_num >= osb->max_slots);
1496 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1498 if (!inode || is_bad_inode(inode)) {
1502 SET_INODE_JOURNAL(inode);
1504 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1514 if (status || !ret_inode)
1522 /* Does the actual journal replay and marks the journal inode as
1523 * clean. Will only replay if the journal inode is marked dirty. */
1524 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1531 struct inode *inode = NULL;
1532 struct ocfs2_dinode *fe;
1533 journal_t *journal = NULL;
1534 struct buffer_head *bh = NULL;
1537 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1543 fe = (struct ocfs2_dinode *)bh->b_data;
1544 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1549 * As the fs recovery is asynchronous, there is a small chance that
1550 * another node mounted (and recovered) the slot before the recovery
1551 * thread could get the lock. To handle that, we dirty read the journal
1552 * inode for that slot to get the recovery generation. If it is
1553 * different than what we expected, the slot has been recovered.
1554 * If not, it needs recovery.
1556 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1557 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num,
1558 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1559 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1564 /* Continue with recovery as the journal has not yet been recovered */
1566 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1568 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status);
1569 if (status != -ERESTARTSYS)
1570 mlog(ML_ERROR, "Could not lock journal!\n");
1575 fe = (struct ocfs2_dinode *) bh->b_data;
1577 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1578 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1580 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1581 mlog(0, "No recovery required for node %d\n", node_num);
1582 /* Refresh recovery generation for the slot */
1583 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1587 /* we need to run complete recovery for offline orphan slots */
1588 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1590 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1592 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1594 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1596 status = ocfs2_force_read_journal(inode);
1602 mlog(0, "calling journal_init_inode\n");
1603 journal = jbd2_journal_init_inode(inode);
1604 if (journal == NULL) {
1605 mlog(ML_ERROR, "Linux journal layer error\n");
1610 status = jbd2_journal_load(journal);
1615 jbd2_journal_destroy(journal);
1619 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1621 /* wipe the journal */
1622 mlog(0, "flushing the journal.\n");
1623 jbd2_journal_lock_updates(journal);
1624 status = jbd2_journal_flush(journal);
1625 jbd2_journal_unlock_updates(journal);
1629 /* This will mark the node clean */
1630 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1631 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1632 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1634 /* Increment recovery generation to indicate successful recovery */
1635 ocfs2_bump_recovery_generation(fe);
1636 osb->slot_recovery_generations[slot_num] =
1637 ocfs2_get_recovery_generation(fe);
1639 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1640 status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1647 jbd2_journal_destroy(journal);
1650 /* drop the lock on this nodes journal */
1652 ocfs2_inode_unlock(inode, 1);
1664 * Do the most important parts of node recovery:
1665 * - Replay it's journal
1666 * - Stamp a clean local allocator file
1667 * - Stamp a clean truncate log
1668 * - Mark the node clean
1670 * If this function completes without error, a node in OCFS2 can be
1671 * said to have been safely recovered. As a result, failure during the
1672 * second part of a nodes recovery process (local alloc recovery) is
1673 * far less concerning.
1675 static int ocfs2_recover_node(struct ocfs2_super *osb,
1676 int node_num, int slot_num)
1679 struct ocfs2_dinode *la_copy = NULL;
1680 struct ocfs2_dinode *tl_copy = NULL;
1682 mlog_entry("(node_num=%d, slot_num=%d, osb->node_num = %d)\n",
1683 node_num, slot_num, osb->node_num);
1685 /* Should not ever be called to recover ourselves -- in that
1686 * case we should've called ocfs2_journal_load instead. */
1687 BUG_ON(osb->node_num == node_num);
1689 status = ocfs2_replay_journal(osb, node_num, slot_num);
1691 if (status == -EBUSY) {
1692 mlog(0, "Skipping recovery for slot %u (node %u) "
1693 "as another node has recovered it\n", slot_num,
1702 /* Stamp a clean local alloc file AFTER recovering the journal... */
1703 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1709 /* An error from begin_truncate_log_recovery is not
1710 * serious enough to warrant halting the rest of
1712 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1716 /* Likewise, this would be a strange but ultimately not so
1717 * harmful place to get an error... */
1718 status = ocfs2_clear_slot(osb, slot_num);
1722 /* This will kfree the memory pointed to by la_copy and tl_copy */
1723 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1733 /* Test node liveness by trylocking his journal. If we get the lock,
1734 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1735 * still alive (we couldn't get the lock) and < 0 on error. */
1736 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1740 struct inode *inode = NULL;
1742 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1744 if (inode == NULL) {
1745 mlog(ML_ERROR, "access error\n");
1749 if (is_bad_inode(inode)) {
1750 mlog(ML_ERROR, "access error (bad inode)\n");
1756 SET_INODE_JOURNAL(inode);
1758 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1759 status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1761 if (status != -EAGAIN)
1766 ocfs2_inode_unlock(inode, 1);
1774 /* Call this underneath ocfs2_super_lock. It also assumes that the
1775 * slot info struct has been updated from disk. */
1776 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1778 unsigned int node_num;
1781 struct buffer_head *bh = NULL;
1782 struct ocfs2_dinode *di;
1784 /* This is called with the super block cluster lock, so we
1785 * know that the slot map can't change underneath us. */
1787 for (i = 0; i < osb->max_slots; i++) {
1788 /* Read journal inode to get the recovery generation */
1789 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1794 di = (struct ocfs2_dinode *)bh->b_data;
1795 gen = ocfs2_get_recovery_generation(di);
1799 spin_lock(&osb->osb_lock);
1800 osb->slot_recovery_generations[i] = gen;
1802 mlog(0, "Slot %u recovery generation is %u\n", i,
1803 osb->slot_recovery_generations[i]);
1805 if (i == osb->slot_num) {
1806 spin_unlock(&osb->osb_lock);
1810 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1811 if (status == -ENOENT) {
1812 spin_unlock(&osb->osb_lock);
1816 if (__ocfs2_recovery_map_test(osb, node_num)) {
1817 spin_unlock(&osb->osb_lock);
1820 spin_unlock(&osb->osb_lock);
1822 /* Ok, we have a slot occupied by another node which
1823 * is not in the recovery map. We trylock his journal
1824 * file here to test if he's alive. */
1825 status = ocfs2_trylock_journal(osb, i);
1827 /* Since we're called from mount, we know that
1828 * the recovery thread can't race us on
1829 * setting / checking the recovery bits. */
1830 ocfs2_recovery_thread(osb, node_num);
1831 } else if ((status < 0) && (status != -EAGAIN)) {
1844 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1845 * randomness to the timeout to minimize multple nodes firing the timer at the
1848 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1852 get_random_bytes(&time, sizeof(time));
1853 time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1854 return msecs_to_jiffies(time);
1858 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1859 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1860 * is done to catch any orphans that are left over in orphan directories.
1862 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1863 * seconds. It gets an EX lock on os_lockres and checks sequence number
1864 * stored in LVB. If the sequence number has changed, it means some other
1865 * node has done the scan. This node skips the scan and tracks the
1866 * sequence number. If the sequence number didn't change, it means a scan
1867 * hasn't happened. The node queues a scan and increments the
1868 * sequence number in the LVB.
1870 void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1872 struct ocfs2_orphan_scan *os;
1876 os = &osb->osb_orphan_scan;
1878 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1881 status = ocfs2_orphan_scan_lock(osb, &seqno);
1883 if (status != -EAGAIN)
1888 /* Do no queue the tasks if the volume is being umounted */
1889 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1892 if (os->os_seqno != seqno) {
1893 os->os_seqno = seqno;
1897 for (i = 0; i < osb->max_slots; i++)
1898 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1901 * We queued a recovery on orphan slots, increment the sequence
1902 * number and update LVB so other node will skip the scan for a while
1906 os->os_scantime = CURRENT_TIME;
1908 ocfs2_orphan_scan_unlock(osb, seqno);
1913 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1914 void ocfs2_orphan_scan_work(struct work_struct *work)
1916 struct ocfs2_orphan_scan *os;
1917 struct ocfs2_super *osb;
1919 os = container_of(work, struct ocfs2_orphan_scan,
1920 os_orphan_scan_work.work);
1923 mutex_lock(&os->os_lock);
1924 ocfs2_queue_orphan_scan(osb);
1925 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1926 schedule_delayed_work(&os->os_orphan_scan_work,
1927 ocfs2_orphan_scan_timeout());
1928 mutex_unlock(&os->os_lock);
1931 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1933 struct ocfs2_orphan_scan *os;
1935 os = &osb->osb_orphan_scan;
1936 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1937 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1938 mutex_lock(&os->os_lock);
1939 cancel_delayed_work(&os->os_orphan_scan_work);
1940 mutex_unlock(&os->os_lock);
1944 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1946 struct ocfs2_orphan_scan *os;
1948 os = &osb->osb_orphan_scan;
1952 mutex_init(&os->os_lock);
1953 INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
1956 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
1958 struct ocfs2_orphan_scan *os;
1960 os = &osb->osb_orphan_scan;
1961 os->os_scantime = CURRENT_TIME;
1962 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1963 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1965 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
1966 schedule_delayed_work(&os->os_orphan_scan_work,
1967 ocfs2_orphan_scan_timeout());
1971 struct ocfs2_orphan_filldir_priv {
1973 struct ocfs2_super *osb;
1976 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1977 loff_t pos, u64 ino, unsigned type)
1979 struct ocfs2_orphan_filldir_priv *p = priv;
1982 if (name_len == 1 && !strncmp(".", name, 1))
1984 if (name_len == 2 && !strncmp("..", name, 2))
1987 /* Skip bad inodes so that recovery can continue */
1988 iter = ocfs2_iget(p->osb, ino,
1989 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
1993 mlog(0, "queue orphan %llu\n",
1994 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1995 /* No locking is required for the next_orphan queue as there
1996 * is only ever a single process doing orphan recovery. */
1997 OCFS2_I(iter)->ip_next_orphan = p->head;
2003 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
2005 struct inode **head)
2008 struct inode *orphan_dir_inode = NULL;
2009 struct ocfs2_orphan_filldir_priv priv;
2015 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2016 ORPHAN_DIR_SYSTEM_INODE,
2018 if (!orphan_dir_inode) {
2024 mutex_lock(&orphan_dir_inode->i_mutex);
2025 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2031 status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
2032 ocfs2_orphan_filldir);
2041 ocfs2_inode_unlock(orphan_dir_inode, 0);
2043 mutex_unlock(&orphan_dir_inode->i_mutex);
2044 iput(orphan_dir_inode);
2048 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2053 spin_lock(&osb->osb_lock);
2054 ret = !osb->osb_orphan_wipes[slot];
2055 spin_unlock(&osb->osb_lock);
2059 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2062 spin_lock(&osb->osb_lock);
2063 /* Mark ourselves such that new processes in delete_inode()
2064 * know to quit early. */
2065 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2066 while (osb->osb_orphan_wipes[slot]) {
2067 /* If any processes are already in the middle of an
2068 * orphan wipe on this dir, then we need to wait for
2070 spin_unlock(&osb->osb_lock);
2071 wait_event_interruptible(osb->osb_wipe_event,
2072 ocfs2_orphan_recovery_can_continue(osb, slot));
2073 spin_lock(&osb->osb_lock);
2075 spin_unlock(&osb->osb_lock);
2078 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2081 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2085 * Orphan recovery. Each mounted node has it's own orphan dir which we
2086 * must run during recovery. Our strategy here is to build a list of
2087 * the inodes in the orphan dir and iget/iput them. The VFS does
2088 * (most) of the rest of the work.
2090 * Orphan recovery can happen at any time, not just mount so we have a
2091 * couple of extra considerations.
2093 * - We grab as many inodes as we can under the orphan dir lock -
2094 * doing iget() outside the orphan dir risks getting a reference on
2096 * - We must be sure not to deadlock with other processes on the
2097 * system wanting to run delete_inode(). This can happen when they go
2098 * to lock the orphan dir and the orphan recovery process attempts to
2099 * iget() inside the orphan dir lock. This can be avoided by
2100 * advertising our state to ocfs2_delete_inode().
2102 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2106 struct inode *inode = NULL;
2108 struct ocfs2_inode_info *oi;
2110 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
2112 ocfs2_mark_recovering_orphan_dir(osb, slot);
2113 ret = ocfs2_queue_orphans(osb, slot, &inode);
2114 ocfs2_clear_recovering_orphan_dir(osb, slot);
2116 /* Error here should be noted, but we want to continue with as
2117 * many queued inodes as we've got. */
2122 oi = OCFS2_I(inode);
2123 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
2125 iter = oi->ip_next_orphan;
2127 spin_lock(&oi->ip_lock);
2128 /* The remote delete code may have set these on the
2129 * assumption that the other node would wipe them
2130 * successfully. If they are still in the node's
2131 * orphan dir, we need to reset that state. */
2132 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
2134 /* Set the proper information to get us going into
2135 * ocfs2_delete_inode. */
2136 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2137 spin_unlock(&oi->ip_lock);
2147 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2149 /* This check is good because ocfs2 will wait on our recovery
2150 * thread before changing it to something other than MOUNTED
2152 wait_event(osb->osb_mount_event,
2153 (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2154 atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2155 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2157 /* If there's an error on mount, then we may never get to the
2158 * MOUNTED flag, but this is set right before
2159 * dismount_volume() so we can trust it. */
2160 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2161 mlog(0, "mount error, exiting!\n");
2168 static int ocfs2_commit_thread(void *arg)
2171 struct ocfs2_super *osb = arg;
2172 struct ocfs2_journal *journal = osb->journal;
2174 /* we can trust j_num_trans here because _should_stop() is only set in
2175 * shutdown and nobody other than ourselves should be able to start
2176 * transactions. committing on shutdown might take a few iterations
2177 * as final transactions put deleted inodes on the list */
2178 while (!(kthread_should_stop() &&
2179 atomic_read(&journal->j_num_trans) == 0)) {
2181 wait_event_interruptible(osb->checkpoint_event,
2182 atomic_read(&journal->j_num_trans)
2183 || kthread_should_stop());
2185 status = ocfs2_commit_cache(osb);
2189 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2191 "commit_thread: %u transactions pending on "
2193 atomic_read(&journal->j_num_trans));
2200 /* Reads all the journal inodes without taking any cluster locks. Used
2201 * for hard readonly access to determine whether any journal requires
2202 * recovery. Also used to refresh the recovery generation numbers after
2203 * a journal has been recovered by another node.
2205 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2209 struct buffer_head *di_bh = NULL;
2210 struct ocfs2_dinode *di;
2211 int journal_dirty = 0;
2213 for(slot = 0; slot < osb->max_slots; slot++) {
2214 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2220 di = (struct ocfs2_dinode *) di_bh->b_data;
2222 osb->slot_recovery_generations[slot] =
2223 ocfs2_get_recovery_generation(di);
2225 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2226 OCFS2_JOURNAL_DIRTY_FL)