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;
308 journal = osb->journal;
310 /* Flush all pending commits and checkpoint the journal. */
311 down_write(&journal->j_trans_barrier);
313 if (atomic_read(&journal->j_num_trans) == 0) {
314 up_write(&journal->j_trans_barrier);
315 mlog(0, "No transactions for me to flush!\n");
319 jbd2_journal_lock_updates(journal->j_journal);
320 status = jbd2_journal_flush(journal->j_journal);
321 jbd2_journal_unlock_updates(journal->j_journal);
323 up_write(&journal->j_trans_barrier);
328 ocfs2_inc_trans_id(journal);
330 flushed = atomic_read(&journal->j_num_trans);
331 atomic_set(&journal->j_num_trans, 0);
332 up_write(&journal->j_trans_barrier);
334 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
335 journal->j_trans_id, flushed);
337 ocfs2_wake_downconvert_thread(osb);
338 wake_up(&journal->j_checkpointed);
344 /* pass it NULL and it will allocate a new handle object for you. If
345 * you pass it a handle however, it may still return error, in which
346 * case it has free'd the passed handle for you. */
347 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
349 journal_t *journal = osb->journal->j_journal;
352 BUG_ON(!osb || !osb->journal->j_journal);
354 if (ocfs2_is_hard_readonly(osb))
355 return ERR_PTR(-EROFS);
357 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
358 BUG_ON(max_buffs <= 0);
360 /* Nested transaction? Just return the handle... */
361 if (journal_current_handle())
362 return jbd2_journal_start(journal, max_buffs);
364 down_read(&osb->journal->j_trans_barrier);
366 handle = jbd2_journal_start(journal, max_buffs);
367 if (IS_ERR(handle)) {
368 up_read(&osb->journal->j_trans_barrier);
370 mlog_errno(PTR_ERR(handle));
372 if (is_journal_aborted(journal)) {
373 ocfs2_abort(osb->sb, "Detected aborted journal");
374 handle = ERR_PTR(-EROFS);
377 if (!ocfs2_mount_local(osb))
378 atomic_inc(&(osb->journal->j_num_trans));
384 int ocfs2_commit_trans(struct ocfs2_super *osb,
388 struct ocfs2_journal *journal = osb->journal;
392 nested = handle->h_ref > 1;
393 ret = jbd2_journal_stop(handle);
398 up_read(&journal->j_trans_barrier);
404 * 'nblocks' is what you want to add to the current transaction.
406 * This might call jbd2_journal_restart() which will commit dirty buffers
407 * and then restart the transaction. Before calling
408 * ocfs2_extend_trans(), any changed blocks should have been
409 * dirtied. After calling it, all blocks which need to be changed must
410 * go through another set of journal_access/journal_dirty calls.
412 * WARNING: This will not release any semaphores or disk locks taken
413 * during the transaction, so make sure they were taken *before*
414 * start_trans or we'll have ordering deadlocks.
416 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
417 * good because transaction ids haven't yet been recorded on the
418 * cluster locks associated with this handle.
420 int ocfs2_extend_trans(handle_t *handle, int nblocks)
422 int status, old_nblocks;
430 old_nblocks = handle->h_buffer_credits;
433 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
435 #ifdef CONFIG_OCFS2_DEBUG_FS
438 status = jbd2_journal_extend(handle, nblocks);
447 "jbd2_journal_extend failed, trying "
448 "jbd2_journal_restart\n");
449 status = jbd2_journal_restart(handle,
450 old_nblocks + nblocks);
464 struct ocfs2_triggers {
465 struct jbd2_buffer_trigger_type ot_triggers;
469 static inline struct ocfs2_triggers *to_ocfs2_trigger(struct jbd2_buffer_trigger_type *triggers)
471 return container_of(triggers, struct ocfs2_triggers, ot_triggers);
474 static void ocfs2_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
475 struct buffer_head *bh,
476 void *data, size_t size)
478 struct ocfs2_triggers *ot = to_ocfs2_trigger(triggers);
481 * We aren't guaranteed to have the superblock here, so we
482 * must unconditionally compute the ecc data.
483 * __ocfs2_journal_access() will only set the triggers if
484 * metaecc is enabled.
486 ocfs2_block_check_compute(data, size, data + ot->ot_offset);
490 * Quota blocks have their own trigger because the struct ocfs2_block_check
491 * offset depends on the blocksize.
493 static void ocfs2_dq_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
494 struct buffer_head *bh,
495 void *data, size_t size)
497 struct ocfs2_disk_dqtrailer *dqt =
498 ocfs2_block_dqtrailer(size, data);
501 * We aren't guaranteed to have the superblock here, so we
502 * must unconditionally compute the ecc data.
503 * __ocfs2_journal_access() will only set the triggers if
504 * metaecc is enabled.
506 ocfs2_block_check_compute(data, size, &dqt->dq_check);
510 * Directory blocks also have their own trigger because the
511 * struct ocfs2_block_check offset depends on the blocksize.
513 static void ocfs2_db_frozen_trigger(struct jbd2_buffer_trigger_type *triggers,
514 struct buffer_head *bh,
515 void *data, size_t size)
517 struct ocfs2_dir_block_trailer *trailer =
518 ocfs2_dir_trailer_from_size(size, data);
521 * We aren't guaranteed to have the superblock here, so we
522 * must unconditionally compute the ecc data.
523 * __ocfs2_journal_access() will only set the triggers if
524 * metaecc is enabled.
526 ocfs2_block_check_compute(data, size, &trailer->db_check);
529 static void ocfs2_abort_trigger(struct jbd2_buffer_trigger_type *triggers,
530 struct buffer_head *bh)
533 "ocfs2_abort_trigger called by JBD2. bh = 0x%lx, "
534 "bh->b_blocknr = %llu\n",
536 (unsigned long long)bh->b_blocknr);
538 /* We aren't guaranteed to have the superblock here - but if we
539 * don't, it'll just crash. */
540 ocfs2_error(bh->b_assoc_map->host->i_sb,
541 "JBD2 has aborted our journal, ocfs2 cannot continue\n");
544 static struct ocfs2_triggers di_triggers = {
546 .t_frozen = ocfs2_frozen_trigger,
547 .t_abort = ocfs2_abort_trigger,
549 .ot_offset = offsetof(struct ocfs2_dinode, i_check),
552 static struct ocfs2_triggers eb_triggers = {
554 .t_frozen = ocfs2_frozen_trigger,
555 .t_abort = ocfs2_abort_trigger,
557 .ot_offset = offsetof(struct ocfs2_extent_block, h_check),
560 static struct ocfs2_triggers rb_triggers = {
562 .t_frozen = ocfs2_frozen_trigger,
563 .t_abort = ocfs2_abort_trigger,
565 .ot_offset = offsetof(struct ocfs2_refcount_block, rf_check),
568 static struct ocfs2_triggers gd_triggers = {
570 .t_frozen = ocfs2_frozen_trigger,
571 .t_abort = ocfs2_abort_trigger,
573 .ot_offset = offsetof(struct ocfs2_group_desc, bg_check),
576 static struct ocfs2_triggers db_triggers = {
578 .t_frozen = ocfs2_db_frozen_trigger,
579 .t_abort = ocfs2_abort_trigger,
583 static struct ocfs2_triggers xb_triggers = {
585 .t_frozen = ocfs2_frozen_trigger,
586 .t_abort = ocfs2_abort_trigger,
588 .ot_offset = offsetof(struct ocfs2_xattr_block, xb_check),
591 static struct ocfs2_triggers dq_triggers = {
593 .t_frozen = ocfs2_dq_frozen_trigger,
594 .t_abort = ocfs2_abort_trigger,
598 static struct ocfs2_triggers dr_triggers = {
600 .t_frozen = ocfs2_frozen_trigger,
601 .t_abort = ocfs2_abort_trigger,
603 .ot_offset = offsetof(struct ocfs2_dx_root_block, dr_check),
606 static struct ocfs2_triggers dl_triggers = {
608 .t_frozen = ocfs2_frozen_trigger,
609 .t_abort = ocfs2_abort_trigger,
611 .ot_offset = offsetof(struct ocfs2_dx_leaf, dl_check),
614 static int __ocfs2_journal_access(handle_t *handle,
615 struct ocfs2_caching_info *ci,
616 struct buffer_head *bh,
617 struct ocfs2_triggers *triggers,
621 struct ocfs2_super *osb =
622 OCFS2_SB(ocfs2_metadata_cache_get_super(ci));
624 BUG_ON(!ci || !ci->ci_ops);
628 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
629 (unsigned long long)bh->b_blocknr, type,
630 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
631 "OCFS2_JOURNAL_ACCESS_CREATE" :
632 "OCFS2_JOURNAL_ACCESS_WRITE",
635 /* we can safely remove this assertion after testing. */
636 if (!buffer_uptodate(bh)) {
637 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
638 mlog(ML_ERROR, "b_blocknr=%llu\n",
639 (unsigned long long)bh->b_blocknr);
643 /* Set the current transaction information on the ci so
644 * that the locking code knows whether it can drop it's locks
645 * on this ci or not. We're protected from the commit
646 * thread updating the current transaction id until
647 * ocfs2_commit_trans() because ocfs2_start_trans() took
648 * j_trans_barrier for us. */
649 ocfs2_set_ci_lock_trans(osb->journal, ci);
651 ocfs2_metadata_cache_io_lock(ci);
653 case OCFS2_JOURNAL_ACCESS_CREATE:
654 case OCFS2_JOURNAL_ACCESS_WRITE:
655 status = jbd2_journal_get_write_access(handle, bh);
658 case OCFS2_JOURNAL_ACCESS_UNDO:
659 status = jbd2_journal_get_undo_access(handle, bh);
664 mlog(ML_ERROR, "Unknown access type!\n");
666 if (!status && ocfs2_meta_ecc(osb) && triggers)
667 jbd2_journal_set_triggers(bh, &triggers->ot_triggers);
668 ocfs2_metadata_cache_io_unlock(ci);
671 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
678 int ocfs2_journal_access_di(handle_t *handle, struct ocfs2_caching_info *ci,
679 struct buffer_head *bh, int type)
681 return __ocfs2_journal_access(handle, ci, bh, &di_triggers, type);
684 int ocfs2_journal_access_eb(handle_t *handle, struct ocfs2_caching_info *ci,
685 struct buffer_head *bh, int type)
687 return __ocfs2_journal_access(handle, ci, bh, &eb_triggers, type);
690 int ocfs2_journal_access_rb(handle_t *handle, struct ocfs2_caching_info *ci,
691 struct buffer_head *bh, int type)
693 return __ocfs2_journal_access(handle, ci, bh, &rb_triggers,
697 int ocfs2_journal_access_gd(handle_t *handle, struct ocfs2_caching_info *ci,
698 struct buffer_head *bh, int type)
700 return __ocfs2_journal_access(handle, ci, bh, &gd_triggers, type);
703 int ocfs2_journal_access_db(handle_t *handle, struct ocfs2_caching_info *ci,
704 struct buffer_head *bh, int type)
706 return __ocfs2_journal_access(handle, ci, bh, &db_triggers, type);
709 int ocfs2_journal_access_xb(handle_t *handle, struct ocfs2_caching_info *ci,
710 struct buffer_head *bh, int type)
712 return __ocfs2_journal_access(handle, ci, bh, &xb_triggers, type);
715 int ocfs2_journal_access_dq(handle_t *handle, struct ocfs2_caching_info *ci,
716 struct buffer_head *bh, int type)
718 return __ocfs2_journal_access(handle, ci, bh, &dq_triggers, type);
721 int ocfs2_journal_access_dr(handle_t *handle, struct ocfs2_caching_info *ci,
722 struct buffer_head *bh, int type)
724 return __ocfs2_journal_access(handle, ci, bh, &dr_triggers, type);
727 int ocfs2_journal_access_dl(handle_t *handle, struct ocfs2_caching_info *ci,
728 struct buffer_head *bh, int type)
730 return __ocfs2_journal_access(handle, ci, bh, &dl_triggers, type);
733 int ocfs2_journal_access(handle_t *handle, struct ocfs2_caching_info *ci,
734 struct buffer_head *bh, int type)
736 return __ocfs2_journal_access(handle, ci, bh, NULL, type);
739 void ocfs2_journal_dirty(handle_t *handle, struct buffer_head *bh)
743 mlog_entry("(bh->b_blocknr=%llu)\n",
744 (unsigned long long)bh->b_blocknr);
746 status = jbd2_journal_dirty_metadata(handle, bh);
752 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
754 void ocfs2_set_journal_params(struct ocfs2_super *osb)
756 journal_t *journal = osb->journal->j_journal;
757 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
759 if (osb->osb_commit_interval)
760 commit_interval = osb->osb_commit_interval;
762 write_lock(&journal->j_state_lock);
763 journal->j_commit_interval = commit_interval;
764 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
765 journal->j_flags |= JBD2_BARRIER;
767 journal->j_flags &= ~JBD2_BARRIER;
768 write_unlock(&journal->j_state_lock);
771 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
774 struct inode *inode = NULL; /* the journal inode */
775 journal_t *j_journal = NULL;
776 struct ocfs2_dinode *di = NULL;
777 struct buffer_head *bh = NULL;
778 struct ocfs2_super *osb;
785 osb = journal->j_osb;
787 /* already have the inode for our journal */
788 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
795 if (is_bad_inode(inode)) {
796 mlog(ML_ERROR, "access error (bad inode)\n");
803 SET_INODE_JOURNAL(inode);
804 OCFS2_I(inode)->ip_open_count++;
806 /* Skip recovery waits here - journal inode metadata never
807 * changes in a live cluster so it can be considered an
808 * exception to the rule. */
809 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
811 if (status != -ERESTARTSYS)
812 mlog(ML_ERROR, "Could not get lock on journal!\n");
817 di = (struct ocfs2_dinode *)bh->b_data;
819 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
820 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
826 mlog(0, "inode->i_size = %lld\n", inode->i_size);
827 mlog(0, "inode->i_blocks = %llu\n",
828 (unsigned long long)inode->i_blocks);
829 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
831 /* call the kernels journal init function now */
832 j_journal = jbd2_journal_init_inode(inode);
833 if (j_journal == NULL) {
834 mlog(ML_ERROR, "Linux journal layer error\n");
839 mlog(0, "Returned from jbd2_journal_init_inode\n");
840 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
842 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
843 OCFS2_JOURNAL_DIRTY_FL);
845 journal->j_journal = j_journal;
846 journal->j_inode = inode;
849 ocfs2_set_journal_params(osb);
851 journal->j_state = OCFS2_JOURNAL_LOADED;
857 ocfs2_inode_unlock(inode, 1);
860 OCFS2_I(inode)->ip_open_count--;
869 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
871 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
874 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
876 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
879 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
880 int dirty, int replayed)
884 struct ocfs2_journal *journal = osb->journal;
885 struct buffer_head *bh = journal->j_bh;
886 struct ocfs2_dinode *fe;
890 fe = (struct ocfs2_dinode *)bh->b_data;
892 /* The journal bh on the osb always comes from ocfs2_journal_init()
893 * and was validated there inside ocfs2_inode_lock_full(). It's a
894 * code bug if we mess it up. */
895 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
897 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
899 flags |= OCFS2_JOURNAL_DIRTY_FL;
901 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
902 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
905 ocfs2_bump_recovery_generation(fe);
907 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
908 status = ocfs2_write_block(osb, bh, INODE_CACHE(journal->j_inode));
917 * If the journal has been kmalloc'd it needs to be freed after this
920 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
922 struct ocfs2_journal *journal = NULL;
924 struct inode *inode = NULL;
925 int num_running_trans = 0;
931 journal = osb->journal;
935 inode = journal->j_inode;
937 if (journal->j_state != OCFS2_JOURNAL_LOADED)
940 /* need to inc inode use count - jbd2_journal_destroy will iput. */
944 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
945 if (num_running_trans > 0)
946 mlog(0, "Shutting down journal: must wait on %d "
947 "running transactions!\n",
950 /* Do a commit_cache here. It will flush our journal, *and*
951 * release any locks that are still held.
952 * set the SHUTDOWN flag and release the trans lock.
953 * the commit thread will take the trans lock for us below. */
954 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
956 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
957 * drop the trans_lock (which we want to hold until we
958 * completely destroy the journal. */
959 if (osb->commit_task) {
960 /* Wait for the commit thread */
961 mlog(0, "Waiting for ocfs2commit to exit....\n");
962 kthread_stop(osb->commit_task);
963 osb->commit_task = NULL;
966 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
968 if (ocfs2_mount_local(osb)) {
969 jbd2_journal_lock_updates(journal->j_journal);
970 status = jbd2_journal_flush(journal->j_journal);
971 jbd2_journal_unlock_updates(journal->j_journal);
978 * Do not toggle if flush was unsuccessful otherwise
979 * will leave dirty metadata in a "clean" journal
981 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
986 /* Shutdown the kernel journal system */
987 jbd2_journal_destroy(journal->j_journal);
988 journal->j_journal = NULL;
990 OCFS2_I(inode)->ip_open_count--;
992 /* unlock our journal */
993 ocfs2_inode_unlock(inode, 1);
995 brelse(journal->j_bh);
996 journal->j_bh = NULL;
998 journal->j_state = OCFS2_JOURNAL_FREE;
1000 // up_write(&journal->j_trans_barrier);
1007 static void ocfs2_clear_journal_error(struct super_block *sb,
1013 olderr = jbd2_journal_errno(journal);
1015 mlog(ML_ERROR, "File system error %d recorded in "
1016 "journal %u.\n", olderr, slot);
1017 mlog(ML_ERROR, "File system on device %s needs checking.\n",
1020 jbd2_journal_ack_err(journal);
1021 jbd2_journal_clear_err(journal);
1025 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
1028 struct ocfs2_super *osb;
1034 osb = journal->j_osb;
1036 status = jbd2_journal_load(journal->j_journal);
1038 mlog(ML_ERROR, "Failed to load journal!\n");
1042 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
1044 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
1050 /* Launch the commit thread */
1052 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
1054 if (IS_ERR(osb->commit_task)) {
1055 status = PTR_ERR(osb->commit_task);
1056 osb->commit_task = NULL;
1057 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
1058 "error=%d", status);
1062 osb->commit_task = NULL;
1070 /* 'full' flag tells us whether we clear out all blocks or if we just
1071 * mark the journal clean */
1072 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
1080 status = jbd2_journal_wipe(journal->j_journal, full);
1086 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
1095 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
1098 struct ocfs2_recovery_map *rm = osb->recovery_map;
1100 spin_lock(&osb->osb_lock);
1101 empty = (rm->rm_used == 0);
1102 spin_unlock(&osb->osb_lock);
1107 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
1109 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
1113 * JBD Might read a cached version of another nodes journal file. We
1114 * don't want this as this file changes often and we get no
1115 * notification on those changes. The only way to be sure that we've
1116 * got the most up to date version of those blocks then is to force
1117 * read them off disk. Just searching through the buffer cache won't
1118 * work as there may be pages backing this file which are still marked
1119 * up to date. We know things can't change on this file underneath us
1120 * as we have the lock by now :)
1122 static int ocfs2_force_read_journal(struct inode *inode)
1126 u64 v_blkno, p_blkno, p_blocks, num_blocks;
1127 #define CONCURRENT_JOURNAL_FILL 32ULL
1128 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
1132 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
1134 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
1136 while (v_blkno < num_blocks) {
1137 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
1138 &p_blkno, &p_blocks, NULL);
1144 if (p_blocks > CONCURRENT_JOURNAL_FILL)
1145 p_blocks = CONCURRENT_JOURNAL_FILL;
1147 /* We are reading journal data which should not
1148 * be put in the uptodate cache */
1149 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
1150 p_blkno, p_blocks, bhs);
1156 for(i = 0; i < p_blocks; i++) {
1161 v_blkno += p_blocks;
1165 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
1171 struct ocfs2_la_recovery_item {
1172 struct list_head lri_list;
1174 struct ocfs2_dinode *lri_la_dinode;
1175 struct ocfs2_dinode *lri_tl_dinode;
1176 struct ocfs2_quota_recovery *lri_qrec;
1179 /* Does the second half of the recovery process. By this point, the
1180 * node is marked clean and can actually be considered recovered,
1181 * hence it's no longer in the recovery map, but there's still some
1182 * cleanup we can do which shouldn't happen within the recovery thread
1183 * as locking in that context becomes very difficult if we are to take
1184 * recovering nodes into account.
1186 * NOTE: This function can and will sleep on recovery of other nodes
1187 * during cluster locking, just like any other ocfs2 process.
1189 void ocfs2_complete_recovery(struct work_struct *work)
1192 struct ocfs2_journal *journal =
1193 container_of(work, struct ocfs2_journal, j_recovery_work);
1194 struct ocfs2_super *osb = journal->j_osb;
1195 struct ocfs2_dinode *la_dinode, *tl_dinode;
1196 struct ocfs2_la_recovery_item *item, *n;
1197 struct ocfs2_quota_recovery *qrec;
1198 LIST_HEAD(tmp_la_list);
1202 mlog(0, "completing recovery from keventd\n");
1204 spin_lock(&journal->j_lock);
1205 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
1206 spin_unlock(&journal->j_lock);
1208 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
1209 list_del_init(&item->lri_list);
1211 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
1213 ocfs2_wait_on_quotas(osb);
1215 la_dinode = item->lri_la_dinode;
1217 mlog(0, "Clean up local alloc %llu\n",
1218 (unsigned long long)le64_to_cpu(la_dinode->i_blkno));
1220 ret = ocfs2_complete_local_alloc_recovery(osb,
1228 tl_dinode = item->lri_tl_dinode;
1230 mlog(0, "Clean up truncate log %llu\n",
1231 (unsigned long long)le64_to_cpu(tl_dinode->i_blkno));
1233 ret = ocfs2_complete_truncate_log_recovery(osb,
1241 ret = ocfs2_recover_orphans(osb, item->lri_slot);
1245 qrec = item->lri_qrec;
1247 mlog(0, "Recovering quota files");
1248 ret = ocfs2_finish_quota_recovery(osb, qrec,
1252 /* Recovery info is already freed now */
1258 mlog(0, "Recovery completion\n");
1262 /* NOTE: This function always eats your references to la_dinode and
1263 * tl_dinode, either manually on error, or by passing them to
1264 * ocfs2_complete_recovery */
1265 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
1267 struct ocfs2_dinode *la_dinode,
1268 struct ocfs2_dinode *tl_dinode,
1269 struct ocfs2_quota_recovery *qrec)
1271 struct ocfs2_la_recovery_item *item;
1273 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
1275 /* Though we wish to avoid it, we are in fact safe in
1276 * skipping local alloc cleanup as fsck.ocfs2 is more
1277 * than capable of reclaiming unused space. */
1285 ocfs2_free_quota_recovery(qrec);
1287 mlog_errno(-ENOMEM);
1291 INIT_LIST_HEAD(&item->lri_list);
1292 item->lri_la_dinode = la_dinode;
1293 item->lri_slot = slot_num;
1294 item->lri_tl_dinode = tl_dinode;
1295 item->lri_qrec = qrec;
1297 spin_lock(&journal->j_lock);
1298 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1299 queue_work(ocfs2_wq, &journal->j_recovery_work);
1300 spin_unlock(&journal->j_lock);
1303 /* Called by the mount code to queue recovery the last part of
1304 * recovery for it's own and offline slot(s). */
1305 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1307 struct ocfs2_journal *journal = osb->journal;
1309 /* No need to queue up our truncate_log as regular cleanup will catch
1311 ocfs2_queue_recovery_completion(journal, osb->slot_num,
1312 osb->local_alloc_copy, NULL, NULL);
1313 ocfs2_schedule_truncate_log_flush(osb, 0);
1315 osb->local_alloc_copy = NULL;
1318 /* queue to recover orphan slots for all offline slots */
1319 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1320 ocfs2_queue_replay_slots(osb);
1321 ocfs2_free_replay_slots(osb);
1324 void ocfs2_complete_quota_recovery(struct ocfs2_super *osb)
1326 if (osb->quota_rec) {
1327 ocfs2_queue_recovery_completion(osb->journal,
1332 osb->quota_rec = NULL;
1336 static int __ocfs2_recovery_thread(void *arg)
1338 int status, node_num, slot_num;
1339 struct ocfs2_super *osb = arg;
1340 struct ocfs2_recovery_map *rm = osb->recovery_map;
1341 int *rm_quota = NULL;
1342 int rm_quota_used = 0, i;
1343 struct ocfs2_quota_recovery *qrec;
1347 status = ocfs2_wait_on_mount(osb);
1352 rm_quota = kzalloc(osb->max_slots * sizeof(int), GFP_NOFS);
1358 status = ocfs2_super_lock(osb, 1);
1364 status = ocfs2_compute_replay_slots(osb);
1368 /* queue recovery for our own slot */
1369 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1372 spin_lock(&osb->osb_lock);
1373 while (rm->rm_used) {
1374 /* It's always safe to remove entry zero, as we won't
1375 * clear it until ocfs2_recover_node() has succeeded. */
1376 node_num = rm->rm_entries[0];
1377 spin_unlock(&osb->osb_lock);
1378 mlog(0, "checking node %d\n", node_num);
1379 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1380 if (slot_num == -ENOENT) {
1382 mlog(0, "no slot for this node, so no recovery"
1386 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1388 /* It is a bit subtle with quota recovery. We cannot do it
1389 * immediately because we have to obtain cluster locks from
1390 * quota files and we also don't want to just skip it because
1391 * then quota usage would be out of sync until some node takes
1392 * the slot. So we remember which nodes need quota recovery
1393 * and when everything else is done, we recover quotas. */
1394 for (i = 0; i < rm_quota_used && rm_quota[i] != slot_num; i++);
1395 if (i == rm_quota_used)
1396 rm_quota[rm_quota_used++] = slot_num;
1398 status = ocfs2_recover_node(osb, node_num, slot_num);
1401 ocfs2_recovery_map_clear(osb, node_num);
1404 "Error %d recovering node %d on device (%u,%u)!\n",
1406 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1407 mlog(ML_ERROR, "Volume requires unmount.\n");
1410 spin_lock(&osb->osb_lock);
1412 spin_unlock(&osb->osb_lock);
1413 mlog(0, "All nodes recovered\n");
1415 /* Refresh all journal recovery generations from disk */
1416 status = ocfs2_check_journals_nolocks(osb);
1417 status = (status == -EROFS) ? 0 : status;
1421 /* Now it is right time to recover quotas... We have to do this under
1422 * superblock lock so that noone can start using the slot (and crash)
1423 * before we recover it */
1424 for (i = 0; i < rm_quota_used; i++) {
1425 qrec = ocfs2_begin_quota_recovery(osb, rm_quota[i]);
1427 status = PTR_ERR(qrec);
1431 ocfs2_queue_recovery_completion(osb->journal, rm_quota[i],
1435 ocfs2_super_unlock(osb, 1);
1437 /* queue recovery for offline slots */
1438 ocfs2_queue_replay_slots(osb);
1441 mutex_lock(&osb->recovery_lock);
1442 if (!status && !ocfs2_recovery_completed(osb)) {
1443 mutex_unlock(&osb->recovery_lock);
1447 ocfs2_free_replay_slots(osb);
1448 osb->recovery_thread_task = NULL;
1449 mb(); /* sync with ocfs2_recovery_thread_running */
1450 wake_up(&osb->recovery_event);
1452 mutex_unlock(&osb->recovery_lock);
1458 /* no one is callint kthread_stop() for us so the kthread() api
1459 * requires that we call do_exit(). And it isn't exported, but
1460 * complete_and_exit() seems to be a minimal wrapper around it. */
1461 complete_and_exit(NULL, status);
1465 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1467 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1468 node_num, osb->node_num);
1470 mutex_lock(&osb->recovery_lock);
1471 if (osb->disable_recovery)
1474 /* People waiting on recovery will wait on
1475 * the recovery map to empty. */
1476 if (ocfs2_recovery_map_set(osb, node_num))
1477 mlog(0, "node %d already in recovery map.\n", node_num);
1479 mlog(0, "starting recovery thread...\n");
1481 if (osb->recovery_thread_task)
1484 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1486 if (IS_ERR(osb->recovery_thread_task)) {
1487 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1488 osb->recovery_thread_task = NULL;
1492 mutex_unlock(&osb->recovery_lock);
1493 wake_up(&osb->recovery_event);
1498 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1500 struct buffer_head **bh,
1501 struct inode **ret_inode)
1503 int status = -EACCES;
1504 struct inode *inode = NULL;
1506 BUG_ON(slot_num >= osb->max_slots);
1508 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1510 if (!inode || is_bad_inode(inode)) {
1514 SET_INODE_JOURNAL(inode);
1516 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1526 if (status || !ret_inode)
1534 /* Does the actual journal replay and marks the journal inode as
1535 * clean. Will only replay if the journal inode is marked dirty. */
1536 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1543 struct inode *inode = NULL;
1544 struct ocfs2_dinode *fe;
1545 journal_t *journal = NULL;
1546 struct buffer_head *bh = NULL;
1549 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1555 fe = (struct ocfs2_dinode *)bh->b_data;
1556 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1561 * As the fs recovery is asynchronous, there is a small chance that
1562 * another node mounted (and recovered) the slot before the recovery
1563 * thread could get the lock. To handle that, we dirty read the journal
1564 * inode for that slot to get the recovery generation. If it is
1565 * different than what we expected, the slot has been recovered.
1566 * If not, it needs recovery.
1568 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1569 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num,
1570 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1571 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1576 /* Continue with recovery as the journal has not yet been recovered */
1578 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1580 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status);
1581 if (status != -ERESTARTSYS)
1582 mlog(ML_ERROR, "Could not lock journal!\n");
1587 fe = (struct ocfs2_dinode *) bh->b_data;
1589 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1590 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1592 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1593 mlog(0, "No recovery required for node %d\n", node_num);
1594 /* Refresh recovery generation for the slot */
1595 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1599 /* we need to run complete recovery for offline orphan slots */
1600 ocfs2_replay_map_set_state(osb, REPLAY_NEEDED);
1602 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1604 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1606 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1608 status = ocfs2_force_read_journal(inode);
1614 mlog(0, "calling journal_init_inode\n");
1615 journal = jbd2_journal_init_inode(inode);
1616 if (journal == NULL) {
1617 mlog(ML_ERROR, "Linux journal layer error\n");
1622 status = jbd2_journal_load(journal);
1627 jbd2_journal_destroy(journal);
1631 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1633 /* wipe the journal */
1634 mlog(0, "flushing the journal.\n");
1635 jbd2_journal_lock_updates(journal);
1636 status = jbd2_journal_flush(journal);
1637 jbd2_journal_unlock_updates(journal);
1641 /* This will mark the node clean */
1642 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1643 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1644 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1646 /* Increment recovery generation to indicate successful recovery */
1647 ocfs2_bump_recovery_generation(fe);
1648 osb->slot_recovery_generations[slot_num] =
1649 ocfs2_get_recovery_generation(fe);
1651 ocfs2_compute_meta_ecc(osb->sb, bh->b_data, &fe->i_check);
1652 status = ocfs2_write_block(osb, bh, INODE_CACHE(inode));
1659 jbd2_journal_destroy(journal);
1662 /* drop the lock on this nodes journal */
1664 ocfs2_inode_unlock(inode, 1);
1676 * Do the most important parts of node recovery:
1677 * - Replay it's journal
1678 * - Stamp a clean local allocator file
1679 * - Stamp a clean truncate log
1680 * - Mark the node clean
1682 * If this function completes without error, a node in OCFS2 can be
1683 * said to have been safely recovered. As a result, failure during the
1684 * second part of a nodes recovery process (local alloc recovery) is
1685 * far less concerning.
1687 static int ocfs2_recover_node(struct ocfs2_super *osb,
1688 int node_num, int slot_num)
1691 struct ocfs2_dinode *la_copy = NULL;
1692 struct ocfs2_dinode *tl_copy = NULL;
1694 mlog_entry("(node_num=%d, slot_num=%d, osb->node_num = %d)\n",
1695 node_num, slot_num, osb->node_num);
1697 /* Should not ever be called to recover ourselves -- in that
1698 * case we should've called ocfs2_journal_load instead. */
1699 BUG_ON(osb->node_num == node_num);
1701 status = ocfs2_replay_journal(osb, node_num, slot_num);
1703 if (status == -EBUSY) {
1704 mlog(0, "Skipping recovery for slot %u (node %u) "
1705 "as another node has recovered it\n", slot_num,
1714 /* Stamp a clean local alloc file AFTER recovering the journal... */
1715 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1721 /* An error from begin_truncate_log_recovery is not
1722 * serious enough to warrant halting the rest of
1724 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1728 /* Likewise, this would be a strange but ultimately not so
1729 * harmful place to get an error... */
1730 status = ocfs2_clear_slot(osb, slot_num);
1734 /* This will kfree the memory pointed to by la_copy and tl_copy */
1735 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1745 /* Test node liveness by trylocking his journal. If we get the lock,
1746 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1747 * still alive (we couldn't get the lock) and < 0 on error. */
1748 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1752 struct inode *inode = NULL;
1754 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1756 if (inode == NULL) {
1757 mlog(ML_ERROR, "access error\n");
1761 if (is_bad_inode(inode)) {
1762 mlog(ML_ERROR, "access error (bad inode)\n");
1768 SET_INODE_JOURNAL(inode);
1770 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1771 status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1773 if (status != -EAGAIN)
1778 ocfs2_inode_unlock(inode, 1);
1786 /* Call this underneath ocfs2_super_lock. It also assumes that the
1787 * slot info struct has been updated from disk. */
1788 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1790 unsigned int node_num;
1793 struct buffer_head *bh = NULL;
1794 struct ocfs2_dinode *di;
1796 /* This is called with the super block cluster lock, so we
1797 * know that the slot map can't change underneath us. */
1799 for (i = 0; i < osb->max_slots; i++) {
1800 /* Read journal inode to get the recovery generation */
1801 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1806 di = (struct ocfs2_dinode *)bh->b_data;
1807 gen = ocfs2_get_recovery_generation(di);
1811 spin_lock(&osb->osb_lock);
1812 osb->slot_recovery_generations[i] = gen;
1814 mlog(0, "Slot %u recovery generation is %u\n", i,
1815 osb->slot_recovery_generations[i]);
1817 if (i == osb->slot_num) {
1818 spin_unlock(&osb->osb_lock);
1822 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1823 if (status == -ENOENT) {
1824 spin_unlock(&osb->osb_lock);
1828 if (__ocfs2_recovery_map_test(osb, node_num)) {
1829 spin_unlock(&osb->osb_lock);
1832 spin_unlock(&osb->osb_lock);
1834 /* Ok, we have a slot occupied by another node which
1835 * is not in the recovery map. We trylock his journal
1836 * file here to test if he's alive. */
1837 status = ocfs2_trylock_journal(osb, i);
1839 /* Since we're called from mount, we know that
1840 * the recovery thread can't race us on
1841 * setting / checking the recovery bits. */
1842 ocfs2_recovery_thread(osb, node_num);
1843 } else if ((status < 0) && (status != -EAGAIN)) {
1856 * Scan timer should get fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT. Add some
1857 * randomness to the timeout to minimize multple nodes firing the timer at the
1860 static inline unsigned long ocfs2_orphan_scan_timeout(void)
1864 get_random_bytes(&time, sizeof(time));
1865 time = ORPHAN_SCAN_SCHEDULE_TIMEOUT + (time % 5000);
1866 return msecs_to_jiffies(time);
1870 * ocfs2_queue_orphan_scan calls ocfs2_queue_recovery_completion for
1871 * every slot, queuing a recovery of the slot on the ocfs2_wq thread. This
1872 * is done to catch any orphans that are left over in orphan directories.
1874 * ocfs2_queue_orphan_scan gets called every ORPHAN_SCAN_SCHEDULE_TIMEOUT
1875 * seconds. It gets an EX lock on os_lockres and checks sequence number
1876 * stored in LVB. If the sequence number has changed, it means some other
1877 * node has done the scan. This node skips the scan and tracks the
1878 * sequence number. If the sequence number didn't change, it means a scan
1879 * hasn't happened. The node queues a scan and increments the
1880 * sequence number in the LVB.
1882 void ocfs2_queue_orphan_scan(struct ocfs2_super *osb)
1884 struct ocfs2_orphan_scan *os;
1888 os = &osb->osb_orphan_scan;
1890 mlog(0, "Begin orphan scan\n");
1892 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1895 status = ocfs2_orphan_scan_lock(osb, &seqno);
1897 if (status != -EAGAIN)
1902 /* Do no queue the tasks if the volume is being umounted */
1903 if (atomic_read(&os->os_state) == ORPHAN_SCAN_INACTIVE)
1906 if (os->os_seqno != seqno) {
1907 os->os_seqno = seqno;
1911 for (i = 0; i < osb->max_slots; i++)
1912 ocfs2_queue_recovery_completion(osb->journal, i, NULL, NULL,
1915 * We queued a recovery on orphan slots, increment the sequence
1916 * number and update LVB so other node will skip the scan for a while
1920 os->os_scantime = CURRENT_TIME;
1922 ocfs2_orphan_scan_unlock(osb, seqno);
1924 mlog(0, "Orphan scan completed\n");
1928 /* Worker task that gets fired every ORPHAN_SCAN_SCHEDULE_TIMEOUT millsec */
1929 void ocfs2_orphan_scan_work(struct work_struct *work)
1931 struct ocfs2_orphan_scan *os;
1932 struct ocfs2_super *osb;
1934 os = container_of(work, struct ocfs2_orphan_scan,
1935 os_orphan_scan_work.work);
1938 mutex_lock(&os->os_lock);
1939 ocfs2_queue_orphan_scan(osb);
1940 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE)
1941 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1942 ocfs2_orphan_scan_timeout());
1943 mutex_unlock(&os->os_lock);
1946 void ocfs2_orphan_scan_stop(struct ocfs2_super *osb)
1948 struct ocfs2_orphan_scan *os;
1950 os = &osb->osb_orphan_scan;
1951 if (atomic_read(&os->os_state) == ORPHAN_SCAN_ACTIVE) {
1952 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1953 mutex_lock(&os->os_lock);
1954 cancel_delayed_work(&os->os_orphan_scan_work);
1955 mutex_unlock(&os->os_lock);
1959 void ocfs2_orphan_scan_init(struct ocfs2_super *osb)
1961 struct ocfs2_orphan_scan *os;
1963 os = &osb->osb_orphan_scan;
1967 mutex_init(&os->os_lock);
1968 INIT_DELAYED_WORK(&os->os_orphan_scan_work, ocfs2_orphan_scan_work);
1971 void ocfs2_orphan_scan_start(struct ocfs2_super *osb)
1973 struct ocfs2_orphan_scan *os;
1975 os = &osb->osb_orphan_scan;
1976 os->os_scantime = CURRENT_TIME;
1977 if (ocfs2_is_hard_readonly(osb) || ocfs2_mount_local(osb))
1978 atomic_set(&os->os_state, ORPHAN_SCAN_INACTIVE);
1980 atomic_set(&os->os_state, ORPHAN_SCAN_ACTIVE);
1981 queue_delayed_work(ocfs2_wq, &os->os_orphan_scan_work,
1982 ocfs2_orphan_scan_timeout());
1986 struct ocfs2_orphan_filldir_priv {
1988 struct ocfs2_super *osb;
1991 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1992 loff_t pos, u64 ino, unsigned type)
1994 struct ocfs2_orphan_filldir_priv *p = priv;
1997 if (name_len == 1 && !strncmp(".", name, 1))
1999 if (name_len == 2 && !strncmp("..", name, 2))
2002 /* Skip bad inodes so that recovery can continue */
2003 iter = ocfs2_iget(p->osb, ino,
2004 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
2008 mlog(0, "queue orphan %llu\n",
2009 (unsigned long long)OCFS2_I(iter)->ip_blkno);
2010 /* No locking is required for the next_orphan queue as there
2011 * is only ever a single process doing orphan recovery. */
2012 OCFS2_I(iter)->ip_next_orphan = p->head;
2018 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
2020 struct inode **head)
2023 struct inode *orphan_dir_inode = NULL;
2024 struct ocfs2_orphan_filldir_priv priv;
2030 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
2031 ORPHAN_DIR_SYSTEM_INODE,
2033 if (!orphan_dir_inode) {
2039 mutex_lock(&orphan_dir_inode->i_mutex);
2040 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
2046 status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
2047 ocfs2_orphan_filldir);
2056 ocfs2_inode_unlock(orphan_dir_inode, 0);
2058 mutex_unlock(&orphan_dir_inode->i_mutex);
2059 iput(orphan_dir_inode);
2063 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
2068 spin_lock(&osb->osb_lock);
2069 ret = !osb->osb_orphan_wipes[slot];
2070 spin_unlock(&osb->osb_lock);
2074 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
2077 spin_lock(&osb->osb_lock);
2078 /* Mark ourselves such that new processes in delete_inode()
2079 * know to quit early. */
2080 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2081 while (osb->osb_orphan_wipes[slot]) {
2082 /* If any processes are already in the middle of an
2083 * orphan wipe on this dir, then we need to wait for
2085 spin_unlock(&osb->osb_lock);
2086 wait_event_interruptible(osb->osb_wipe_event,
2087 ocfs2_orphan_recovery_can_continue(osb, slot));
2088 spin_lock(&osb->osb_lock);
2090 spin_unlock(&osb->osb_lock);
2093 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
2096 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
2100 * Orphan recovery. Each mounted node has it's own orphan dir which we
2101 * must run during recovery. Our strategy here is to build a list of
2102 * the inodes in the orphan dir and iget/iput them. The VFS does
2103 * (most) of the rest of the work.
2105 * Orphan recovery can happen at any time, not just mount so we have a
2106 * couple of extra considerations.
2108 * - We grab as many inodes as we can under the orphan dir lock -
2109 * doing iget() outside the orphan dir risks getting a reference on
2111 * - We must be sure not to deadlock with other processes on the
2112 * system wanting to run delete_inode(). This can happen when they go
2113 * to lock the orphan dir and the orphan recovery process attempts to
2114 * iget() inside the orphan dir lock. This can be avoided by
2115 * advertising our state to ocfs2_delete_inode().
2117 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
2121 struct inode *inode = NULL;
2123 struct ocfs2_inode_info *oi;
2125 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
2127 ocfs2_mark_recovering_orphan_dir(osb, slot);
2128 ret = ocfs2_queue_orphans(osb, slot, &inode);
2129 ocfs2_clear_recovering_orphan_dir(osb, slot);
2131 /* Error here should be noted, but we want to continue with as
2132 * many queued inodes as we've got. */
2137 oi = OCFS2_I(inode);
2138 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
2140 iter = oi->ip_next_orphan;
2142 spin_lock(&oi->ip_lock);
2143 /* The remote delete code may have set these on the
2144 * assumption that the other node would wipe them
2145 * successfully. If they are still in the node's
2146 * orphan dir, we need to reset that state. */
2147 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
2149 /* Set the proper information to get us going into
2150 * ocfs2_delete_inode. */
2151 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
2152 spin_unlock(&oi->ip_lock);
2162 static int __ocfs2_wait_on_mount(struct ocfs2_super *osb, int quota)
2164 /* This check is good because ocfs2 will wait on our recovery
2165 * thread before changing it to something other than MOUNTED
2167 wait_event(osb->osb_mount_event,
2168 (!quota && atomic_read(&osb->vol_state) == VOLUME_MOUNTED) ||
2169 atomic_read(&osb->vol_state) == VOLUME_MOUNTED_QUOTAS ||
2170 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
2172 /* If there's an error on mount, then we may never get to the
2173 * MOUNTED flag, but this is set right before
2174 * dismount_volume() so we can trust it. */
2175 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
2176 mlog(0, "mount error, exiting!\n");
2183 static int ocfs2_commit_thread(void *arg)
2186 struct ocfs2_super *osb = arg;
2187 struct ocfs2_journal *journal = osb->journal;
2189 /* we can trust j_num_trans here because _should_stop() is only set in
2190 * shutdown and nobody other than ourselves should be able to start
2191 * transactions. committing on shutdown might take a few iterations
2192 * as final transactions put deleted inodes on the list */
2193 while (!(kthread_should_stop() &&
2194 atomic_read(&journal->j_num_trans) == 0)) {
2196 wait_event_interruptible(osb->checkpoint_event,
2197 atomic_read(&journal->j_num_trans)
2198 || kthread_should_stop());
2200 status = ocfs2_commit_cache(osb);
2204 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
2206 "commit_thread: %u transactions pending on "
2208 atomic_read(&journal->j_num_trans));
2215 /* Reads all the journal inodes without taking any cluster locks. Used
2216 * for hard readonly access to determine whether any journal requires
2217 * recovery. Also used to refresh the recovery generation numbers after
2218 * a journal has been recovered by another node.
2220 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
2224 struct buffer_head *di_bh = NULL;
2225 struct ocfs2_dinode *di;
2226 int journal_dirty = 0;
2228 for(slot = 0; slot < osb->max_slots; slot++) {
2229 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
2235 di = (struct ocfs2_dinode *) di_bh->b_data;
2237 osb->slot_recovery_generations[slot] =
2238 ocfs2_get_recovery_generation(di);
2240 if (le32_to_cpu(di->id1.journal1.ij_flags) &
2241 OCFS2_JOURNAL_DIRTY_FL)