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>
32 #define MLOG_MASK_PREFIX ML_JOURNAL
33 #include <cluster/masklog.h>
39 #include "extent_map.h"
40 #include "heartbeat.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
52 DEFINE_SPINLOCK(trans_inc_lock);
54 static int ocfs2_force_read_journal(struct inode *inode);
55 static int ocfs2_recover_node(struct ocfs2_super *osb,
57 static int __ocfs2_recovery_thread(void *arg);
58 static int ocfs2_commit_cache(struct ocfs2_super *osb);
59 static int ocfs2_wait_on_mount(struct ocfs2_super *osb);
60 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
62 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
64 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
66 static int ocfs2_commit_thread(void *arg);
68 static int ocfs2_commit_cache(struct ocfs2_super *osb)
73 struct ocfs2_journal *journal = NULL;
77 journal = osb->journal;
79 /* Flush all pending commits and checkpoint the journal. */
80 down_write(&journal->j_trans_barrier);
82 if (atomic_read(&journal->j_num_trans) == 0) {
83 up_write(&journal->j_trans_barrier);
84 mlog(0, "No transactions for me to flush!\n");
88 journal_lock_updates(journal->j_journal);
89 status = journal_flush(journal->j_journal);
90 journal_unlock_updates(journal->j_journal);
92 up_write(&journal->j_trans_barrier);
97 old_id = ocfs2_inc_trans_id(journal);
99 flushed = atomic_read(&journal->j_num_trans);
100 atomic_set(&journal->j_num_trans, 0);
101 up_write(&journal->j_trans_barrier);
103 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
104 journal->j_trans_id, flushed);
106 ocfs2_kick_vote_thread(osb);
107 wake_up(&journal->j_checkpointed);
113 /* pass it NULL and it will allocate a new handle object for you. If
114 * you pass it a handle however, it may still return error, in which
115 * case it has free'd the passed handle for you. */
116 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
118 journal_t *journal = osb->journal->j_journal;
121 BUG_ON(!osb || !osb->journal->j_journal);
123 if (ocfs2_is_hard_readonly(osb))
124 return ERR_PTR(-EROFS);
126 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
127 BUG_ON(max_buffs <= 0);
129 /* JBD might support this, but our journalling code doesn't yet. */
130 if (journal_current_handle()) {
131 mlog(ML_ERROR, "Recursive transaction attempted!\n");
135 down_read(&osb->journal->j_trans_barrier);
137 handle = journal_start(journal, max_buffs);
138 if (IS_ERR(handle)) {
139 up_read(&osb->journal->j_trans_barrier);
141 mlog_errno(PTR_ERR(handle));
143 if (is_journal_aborted(journal)) {
144 ocfs2_abort(osb->sb, "Detected aborted journal");
145 handle = ERR_PTR(-EROFS);
148 if (!ocfs2_mount_local(osb))
149 atomic_inc(&(osb->journal->j_num_trans));
155 int ocfs2_commit_trans(struct ocfs2_super *osb,
159 struct ocfs2_journal *journal = osb->journal;
163 ret = journal_stop(handle);
167 up_read(&journal->j_trans_barrier);
173 * 'nblocks' is what you want to add to the current
174 * transaction. extend_trans will either extend the current handle by
175 * nblocks, or commit it and start a new one with nblocks credits.
177 * WARNING: This will not release any semaphores or disk locks taken
178 * during the transaction, so make sure they were taken *before*
179 * start_trans or we'll have ordering deadlocks.
181 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
182 * good because transaction ids haven't yet been recorded on the
183 * cluster locks associated with this handle.
185 int ocfs2_extend_trans(handle_t *handle, int nblocks)
194 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
196 status = journal_extend(handle, nblocks);
203 mlog(0, "journal_extend failed, trying journal_restart\n");
204 status = journal_restart(handle, nblocks);
218 int ocfs2_journal_access(handle_t *handle,
220 struct buffer_head *bh,
229 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
230 (unsigned long long)bh->b_blocknr, type,
231 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
232 "OCFS2_JOURNAL_ACCESS_CREATE" :
233 "OCFS2_JOURNAL_ACCESS_WRITE",
236 /* we can safely remove this assertion after testing. */
237 if (!buffer_uptodate(bh)) {
238 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
239 mlog(ML_ERROR, "b_blocknr=%llu\n",
240 (unsigned long long)bh->b_blocknr);
244 /* Set the current transaction information on the inode so
245 * that the locking code knows whether it can drop it's locks
246 * on this inode or not. We're protected from the commit
247 * thread updating the current transaction id until
248 * ocfs2_commit_trans() because ocfs2_start_trans() took
249 * j_trans_barrier for us. */
250 ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
252 mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
254 case OCFS2_JOURNAL_ACCESS_CREATE:
255 case OCFS2_JOURNAL_ACCESS_WRITE:
256 status = journal_get_write_access(handle, bh);
259 case OCFS2_JOURNAL_ACCESS_UNDO:
260 status = journal_get_undo_access(handle, bh);
265 mlog(ML_ERROR, "Uknown access type!\n");
267 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
270 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
277 int ocfs2_journal_dirty(handle_t *handle,
278 struct buffer_head *bh)
282 mlog_entry("(bh->b_blocknr=%llu)\n",
283 (unsigned long long)bh->b_blocknr);
285 status = journal_dirty_metadata(handle, bh);
287 mlog(ML_ERROR, "Could not dirty metadata buffer. "
288 "(bh->b_blocknr=%llu)\n",
289 (unsigned long long)bh->b_blocknr);
295 int ocfs2_journal_dirty_data(handle_t *handle,
296 struct buffer_head *bh)
298 int err = journal_dirty_data(handle, bh);
301 /* TODO: When we can handle it, abort the handle and go RO on
307 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
309 void ocfs2_set_journal_params(struct ocfs2_super *osb)
311 journal_t *journal = osb->journal->j_journal;
313 spin_lock(&journal->j_state_lock);
314 journal->j_commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
315 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
316 journal->j_flags |= JFS_BARRIER;
318 journal->j_flags &= ~JFS_BARRIER;
319 spin_unlock(&journal->j_state_lock);
322 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
325 struct inode *inode = NULL; /* the journal inode */
326 journal_t *j_journal = NULL;
327 struct ocfs2_dinode *di = NULL;
328 struct buffer_head *bh = NULL;
329 struct ocfs2_super *osb;
336 osb = journal->j_osb;
338 /* already have the inode for our journal */
339 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
346 if (is_bad_inode(inode)) {
347 mlog(ML_ERROR, "access error (bad inode)\n");
354 SET_INODE_JOURNAL(inode);
355 OCFS2_I(inode)->ip_open_count++;
357 /* Skip recovery waits here - journal inode metadata never
358 * changes in a live cluster so it can be considered an
359 * exception to the rule. */
360 status = ocfs2_meta_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
362 if (status != -ERESTARTSYS)
363 mlog(ML_ERROR, "Could not get lock on journal!\n");
368 di = (struct ocfs2_dinode *)bh->b_data;
370 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
371 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
377 mlog(0, "inode->i_size = %lld\n", inode->i_size);
378 mlog(0, "inode->i_blocks = %llu\n",
379 (unsigned long long)inode->i_blocks);
380 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
382 /* call the kernels journal init function now */
383 j_journal = journal_init_inode(inode);
384 if (j_journal == NULL) {
385 mlog(ML_ERROR, "Linux journal layer error\n");
390 mlog(0, "Returned from journal_init_inode\n");
391 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
393 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
394 OCFS2_JOURNAL_DIRTY_FL);
396 journal->j_journal = j_journal;
397 journal->j_inode = inode;
400 ocfs2_set_journal_params(osb);
402 journal->j_state = OCFS2_JOURNAL_LOADED;
408 ocfs2_meta_unlock(inode, 1);
412 OCFS2_I(inode)->ip_open_count--;
421 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
426 struct ocfs2_journal *journal = osb->journal;
427 struct buffer_head *bh = journal->j_bh;
428 struct ocfs2_dinode *fe;
432 fe = (struct ocfs2_dinode *)bh->b_data;
433 if (!OCFS2_IS_VALID_DINODE(fe)) {
434 /* This is called from startup/shutdown which will
435 * handle the errors in a specific manner, so no need
436 * to call ocfs2_error() here. */
437 mlog(ML_ERROR, "Journal dinode %llu has invalid "
438 "signature: %.*s", (unsigned long long)fe->i_blkno, 7,
444 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
446 flags |= OCFS2_JOURNAL_DIRTY_FL;
448 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
449 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
451 status = ocfs2_write_block(osb, bh, journal->j_inode);
461 * If the journal has been kmalloc'd it needs to be freed after this
464 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
466 struct ocfs2_journal *journal = NULL;
468 struct inode *inode = NULL;
469 int num_running_trans = 0;
475 journal = osb->journal;
479 inode = journal->j_inode;
481 if (journal->j_state != OCFS2_JOURNAL_LOADED)
484 /* need to inc inode use count as journal_destroy will iput. */
488 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
489 if (num_running_trans > 0)
490 mlog(0, "Shutting down journal: must wait on %d "
491 "running transactions!\n",
494 /* Do a commit_cache here. It will flush our journal, *and*
495 * release any locks that are still held.
496 * set the SHUTDOWN flag and release the trans lock.
497 * the commit thread will take the trans lock for us below. */
498 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
500 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
501 * drop the trans_lock (which we want to hold until we
502 * completely destroy the journal. */
503 if (osb->commit_task) {
504 /* Wait for the commit thread */
505 mlog(0, "Waiting for ocfs2commit to exit....\n");
506 kthread_stop(osb->commit_task);
507 osb->commit_task = NULL;
510 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
512 if (ocfs2_mount_local(osb)) {
513 journal_lock_updates(journal->j_journal);
514 status = journal_flush(journal->j_journal);
515 journal_unlock_updates(journal->j_journal);
522 * Do not toggle if flush was unsuccessful otherwise
523 * will leave dirty metadata in a "clean" journal
525 status = ocfs2_journal_toggle_dirty(osb, 0);
530 /* Shutdown the kernel journal system */
531 journal_destroy(journal->j_journal);
533 OCFS2_I(inode)->ip_open_count--;
535 /* unlock our journal */
536 ocfs2_meta_unlock(inode, 1);
538 brelse(journal->j_bh);
539 journal->j_bh = NULL;
541 journal->j_state = OCFS2_JOURNAL_FREE;
543 // up_write(&journal->j_trans_barrier);
550 static void ocfs2_clear_journal_error(struct super_block *sb,
556 olderr = journal_errno(journal);
558 mlog(ML_ERROR, "File system error %d recorded in "
559 "journal %u.\n", olderr, slot);
560 mlog(ML_ERROR, "File system on device %s needs checking.\n",
563 journal_ack_err(journal);
564 journal_clear_err(journal);
568 int ocfs2_journal_load(struct ocfs2_journal *journal, int local)
571 struct ocfs2_super *osb;
578 osb = journal->j_osb;
580 status = journal_load(journal->j_journal);
582 mlog(ML_ERROR, "Failed to load journal!\n");
586 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
588 status = ocfs2_journal_toggle_dirty(osb, 1);
594 /* Launch the commit thread */
596 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
598 if (IS_ERR(osb->commit_task)) {
599 status = PTR_ERR(osb->commit_task);
600 osb->commit_task = NULL;
601 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
606 osb->commit_task = NULL;
614 /* 'full' flag tells us whether we clear out all blocks or if we just
615 * mark the journal clean */
616 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
624 status = journal_wipe(journal->j_journal, full);
630 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
640 * JBD Might read a cached version of another nodes journal file. We
641 * don't want this as this file changes often and we get no
642 * notification on those changes. The only way to be sure that we've
643 * got the most up to date version of those blocks then is to force
644 * read them off disk. Just searching through the buffer cache won't
645 * work as there may be pages backing this file which are still marked
646 * up to date. We know things can't change on this file underneath us
647 * as we have the lock by now :)
649 static int ocfs2_force_read_journal(struct inode *inode)
653 u64 v_blkno, p_blkno;
654 #define CONCURRENT_JOURNAL_FILL 32
655 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
659 BUG_ON(inode->i_blocks !=
660 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
662 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
664 mlog(0, "Force reading %llu blocks\n",
665 (unsigned long long)(inode->i_blocks >>
666 (inode->i_sb->s_blocksize_bits - 9)));
670 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
672 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
673 &p_blkno, &p_blocks);
679 if (p_blocks > CONCURRENT_JOURNAL_FILL)
680 p_blocks = CONCURRENT_JOURNAL_FILL;
682 /* We are reading journal data which should not
683 * be put in the uptodate cache */
684 status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
685 p_blkno, p_blocks, bhs, 0,
692 for(i = 0; i < p_blocks; i++) {
701 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
708 struct ocfs2_la_recovery_item {
709 struct list_head lri_list;
711 struct ocfs2_dinode *lri_la_dinode;
712 struct ocfs2_dinode *lri_tl_dinode;
715 /* Does the second half of the recovery process. By this point, the
716 * node is marked clean and can actually be considered recovered,
717 * hence it's no longer in the recovery map, but there's still some
718 * cleanup we can do which shouldn't happen within the recovery thread
719 * as locking in that context becomes very difficult if we are to take
720 * recovering nodes into account.
722 * NOTE: This function can and will sleep on recovery of other nodes
723 * during cluster locking, just like any other ocfs2 process.
725 void ocfs2_complete_recovery(struct work_struct *work)
728 struct ocfs2_journal *journal =
729 container_of(work, struct ocfs2_journal, j_recovery_work);
730 struct ocfs2_super *osb = journal->j_osb;
731 struct ocfs2_dinode *la_dinode, *tl_dinode;
732 struct ocfs2_la_recovery_item *item;
733 struct list_head *p, *n;
734 LIST_HEAD(tmp_la_list);
738 mlog(0, "completing recovery from keventd\n");
740 spin_lock(&journal->j_lock);
741 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
742 spin_unlock(&journal->j_lock);
744 list_for_each_safe(p, n, &tmp_la_list) {
745 item = list_entry(p, struct ocfs2_la_recovery_item, lri_list);
746 list_del_init(&item->lri_list);
748 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
750 la_dinode = item->lri_la_dinode;
752 mlog(0, "Clean up local alloc %llu\n",
753 (unsigned long long)la_dinode->i_blkno);
755 ret = ocfs2_complete_local_alloc_recovery(osb,
763 tl_dinode = item->lri_tl_dinode;
765 mlog(0, "Clean up truncate log %llu\n",
766 (unsigned long long)tl_dinode->i_blkno);
768 ret = ocfs2_complete_truncate_log_recovery(osb,
776 ret = ocfs2_recover_orphans(osb, item->lri_slot);
783 mlog(0, "Recovery completion\n");
787 /* NOTE: This function always eats your references to la_dinode and
788 * tl_dinode, either manually on error, or by passing them to
789 * ocfs2_complete_recovery */
790 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
792 struct ocfs2_dinode *la_dinode,
793 struct ocfs2_dinode *tl_dinode)
795 struct ocfs2_la_recovery_item *item;
797 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
799 /* Though we wish to avoid it, we are in fact safe in
800 * skipping local alloc cleanup as fsck.ocfs2 is more
801 * than capable of reclaiming unused space. */
812 INIT_LIST_HEAD(&item->lri_list);
813 item->lri_la_dinode = la_dinode;
814 item->lri_slot = slot_num;
815 item->lri_tl_dinode = tl_dinode;
817 spin_lock(&journal->j_lock);
818 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
819 queue_work(ocfs2_wq, &journal->j_recovery_work);
820 spin_unlock(&journal->j_lock);
823 /* Called by the mount code to queue recovery the last part of
824 * recovery for it's own slot. */
825 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
827 struct ocfs2_journal *journal = osb->journal;
830 /* No need to queue up our truncate_log as regular
831 * cleanup will catch that. */
832 ocfs2_queue_recovery_completion(journal,
834 osb->local_alloc_copy,
836 ocfs2_schedule_truncate_log_flush(osb, 0);
838 osb->local_alloc_copy = NULL;
843 static int __ocfs2_recovery_thread(void *arg)
845 int status, node_num;
846 struct ocfs2_super *osb = arg;
850 status = ocfs2_wait_on_mount(osb);
856 status = ocfs2_super_lock(osb, 1);
862 while(!ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
863 node_num = ocfs2_node_map_first_set_bit(osb,
865 if (node_num == O2NM_INVALID_NODE_NUM) {
866 mlog(0, "Out of nodes to recover.\n");
870 status = ocfs2_recover_node(osb, node_num);
873 "Error %d recovering node %d on device (%u,%u)!\n",
875 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
876 mlog(ML_ERROR, "Volume requires unmount.\n");
880 ocfs2_recovery_map_clear(osb, node_num);
882 ocfs2_super_unlock(osb, 1);
884 /* We always run recovery on our own orphan dir - the dead
885 * node(s) may have voted "no" on an inode delete earlier. A
886 * revote is therefore required. */
887 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
891 mutex_lock(&osb->recovery_lock);
893 !ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
894 mutex_unlock(&osb->recovery_lock);
898 osb->recovery_thread_task = NULL;
899 mb(); /* sync with ocfs2_recovery_thread_running */
900 wake_up(&osb->recovery_event);
902 mutex_unlock(&osb->recovery_lock);
905 /* no one is callint kthread_stop() for us so the kthread() api
906 * requires that we call do_exit(). And it isn't exported, but
907 * complete_and_exit() seems to be a minimal wrapper around it. */
908 complete_and_exit(NULL, status);
912 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
914 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
915 node_num, osb->node_num);
917 mutex_lock(&osb->recovery_lock);
918 if (osb->disable_recovery)
921 /* People waiting on recovery will wait on
922 * the recovery map to empty. */
923 if (!ocfs2_recovery_map_set(osb, node_num))
924 mlog(0, "node %d already be in recovery.\n", node_num);
926 mlog(0, "starting recovery thread...\n");
928 if (osb->recovery_thread_task)
931 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
933 if (IS_ERR(osb->recovery_thread_task)) {
934 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
935 osb->recovery_thread_task = NULL;
939 mutex_unlock(&osb->recovery_lock);
940 wake_up(&osb->recovery_event);
945 /* Does the actual journal replay and marks the journal inode as
946 * clean. Will only replay if the journal inode is marked dirty. */
947 static int ocfs2_replay_journal(struct ocfs2_super *osb,
954 struct inode *inode = NULL;
955 struct ocfs2_dinode *fe;
956 journal_t *journal = NULL;
957 struct buffer_head *bh = NULL;
959 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
966 if (is_bad_inode(inode)) {
973 SET_INODE_JOURNAL(inode);
975 status = ocfs2_meta_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
977 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status);
978 if (status != -ERESTARTSYS)
979 mlog(ML_ERROR, "Could not lock journal!\n");
984 fe = (struct ocfs2_dinode *) bh->b_data;
986 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
988 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
989 mlog(0, "No recovery required for node %d\n", node_num);
993 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
995 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
997 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
999 status = ocfs2_force_read_journal(inode);
1005 mlog(0, "calling journal_init_inode\n");
1006 journal = journal_init_inode(inode);
1007 if (journal == NULL) {
1008 mlog(ML_ERROR, "Linux journal layer error\n");
1013 status = journal_load(journal);
1018 journal_destroy(journal);
1022 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1024 /* wipe the journal */
1025 mlog(0, "flushing the journal.\n");
1026 journal_lock_updates(journal);
1027 status = journal_flush(journal);
1028 journal_unlock_updates(journal);
1032 /* This will mark the node clean */
1033 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1034 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1035 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1037 status = ocfs2_write_block(osb, bh, inode);
1044 journal_destroy(journal);
1047 /* drop the lock on this nodes journal */
1049 ocfs2_meta_unlock(inode, 1);
1062 * Do the most important parts of node recovery:
1063 * - Replay it's journal
1064 * - Stamp a clean local allocator file
1065 * - Stamp a clean truncate log
1066 * - Mark the node clean
1068 * If this function completes without error, a node in OCFS2 can be
1069 * said to have been safely recovered. As a result, failure during the
1070 * second part of a nodes recovery process (local alloc recovery) is
1071 * far less concerning.
1073 static int ocfs2_recover_node(struct ocfs2_super *osb,
1078 struct ocfs2_slot_info *si = osb->slot_info;
1079 struct ocfs2_dinode *la_copy = NULL;
1080 struct ocfs2_dinode *tl_copy = NULL;
1082 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1083 node_num, osb->node_num);
1085 mlog(0, "checking node %d\n", node_num);
1087 /* Should not ever be called to recover ourselves -- in that
1088 * case we should've called ocfs2_journal_load instead. */
1089 BUG_ON(osb->node_num == node_num);
1091 slot_num = ocfs2_node_num_to_slot(si, node_num);
1092 if (slot_num == OCFS2_INVALID_SLOT) {
1094 mlog(0, "no slot for this node, so no recovery required.\n");
1098 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1100 status = ocfs2_replay_journal(osb, node_num, slot_num);
1106 /* Stamp a clean local alloc file AFTER recovering the journal... */
1107 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1113 /* An error from begin_truncate_log_recovery is not
1114 * serious enough to warrant halting the rest of
1116 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1120 /* Likewise, this would be a strange but ultimately not so
1121 * harmful place to get an error... */
1122 ocfs2_clear_slot(si, slot_num);
1123 status = ocfs2_update_disk_slots(osb, si);
1127 /* This will kfree the memory pointed to by la_copy and tl_copy */
1128 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1138 /* Test node liveness by trylocking his journal. If we get the lock,
1139 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1140 * still alive (we couldn't get the lock) and < 0 on error. */
1141 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1145 struct inode *inode = NULL;
1147 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1149 if (inode == NULL) {
1150 mlog(ML_ERROR, "access error\n");
1154 if (is_bad_inode(inode)) {
1155 mlog(ML_ERROR, "access error (bad inode)\n");
1161 SET_INODE_JOURNAL(inode);
1163 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1164 status = ocfs2_meta_lock_full(inode, NULL, 1, flags);
1166 if (status != -EAGAIN)
1171 ocfs2_meta_unlock(inode, 1);
1179 /* Call this underneath ocfs2_super_lock. It also assumes that the
1180 * slot info struct has been updated from disk. */
1181 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1183 int status, i, node_num;
1184 struct ocfs2_slot_info *si = osb->slot_info;
1186 /* This is called with the super block cluster lock, so we
1187 * know that the slot map can't change underneath us. */
1189 spin_lock(&si->si_lock);
1190 for(i = 0; i < si->si_num_slots; i++) {
1191 if (i == osb->slot_num)
1193 if (ocfs2_is_empty_slot(si, i))
1196 node_num = si->si_global_node_nums[i];
1197 if (ocfs2_node_map_test_bit(osb, &osb->recovery_map, node_num))
1199 spin_unlock(&si->si_lock);
1201 /* Ok, we have a slot occupied by another node which
1202 * is not in the recovery map. We trylock his journal
1203 * file here to test if he's alive. */
1204 status = ocfs2_trylock_journal(osb, i);
1206 /* Since we're called from mount, we know that
1207 * the recovery thread can't race us on
1208 * setting / checking the recovery bits. */
1209 ocfs2_recovery_thread(osb, node_num);
1210 } else if ((status < 0) && (status != -EAGAIN)) {
1215 spin_lock(&si->si_lock);
1217 spin_unlock(&si->si_lock);
1225 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1227 struct inode **head)
1230 struct inode *orphan_dir_inode = NULL;
1232 unsigned long offset, blk, local;
1233 struct buffer_head *bh = NULL;
1234 struct ocfs2_dir_entry *de;
1235 struct super_block *sb = osb->sb;
1237 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1238 ORPHAN_DIR_SYSTEM_INODE,
1240 if (!orphan_dir_inode) {
1246 mutex_lock(&orphan_dir_inode->i_mutex);
1247 status = ocfs2_meta_lock(orphan_dir_inode, NULL, 0);
1255 while(offset < i_size_read(orphan_dir_inode)) {
1256 blk = offset >> sb->s_blocksize_bits;
1258 bh = ocfs2_bread(orphan_dir_inode, blk, &status, 0);
1269 while(offset < i_size_read(orphan_dir_inode)
1270 && local < sb->s_blocksize) {
1271 de = (struct ocfs2_dir_entry *) (bh->b_data + local);
1273 if (!ocfs2_check_dir_entry(orphan_dir_inode,
1281 local += le16_to_cpu(de->rec_len);
1282 offset += le16_to_cpu(de->rec_len);
1284 /* I guess we silently fail on no inode? */
1285 if (!le64_to_cpu(de->inode))
1287 if (de->file_type > OCFS2_FT_MAX) {
1289 "block %llu contains invalid de: "
1290 "inode = %llu, rec_len = %u, "
1291 "name_len = %u, file_type = %u, "
1293 (unsigned long long)bh->b_blocknr,
1294 (unsigned long long)le64_to_cpu(de->inode),
1295 le16_to_cpu(de->rec_len),
1302 if (de->name_len == 1 && !strncmp(".", de->name, 1))
1304 if (de->name_len == 2 && !strncmp("..", de->name, 2))
1307 iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
1308 OCFS2_FI_FLAG_ORPHAN_RECOVERY);
1312 mlog(0, "queue orphan %llu\n",
1313 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1314 /* No locking is required for the next_orphan
1315 * queue as there is only ever a single
1316 * process doing orphan recovery. */
1317 OCFS2_I(iter)->ip_next_orphan = *head;
1324 ocfs2_meta_unlock(orphan_dir_inode, 0);
1326 mutex_unlock(&orphan_dir_inode->i_mutex);
1327 iput(orphan_dir_inode);
1331 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1336 spin_lock(&osb->osb_lock);
1337 ret = !osb->osb_orphan_wipes[slot];
1338 spin_unlock(&osb->osb_lock);
1342 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1345 spin_lock(&osb->osb_lock);
1346 /* Mark ourselves such that new processes in delete_inode()
1347 * know to quit early. */
1348 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1349 while (osb->osb_orphan_wipes[slot]) {
1350 /* If any processes are already in the middle of an
1351 * orphan wipe on this dir, then we need to wait for
1353 spin_unlock(&osb->osb_lock);
1354 wait_event_interruptible(osb->osb_wipe_event,
1355 ocfs2_orphan_recovery_can_continue(osb, slot));
1356 spin_lock(&osb->osb_lock);
1358 spin_unlock(&osb->osb_lock);
1361 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1364 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1368 * Orphan recovery. Each mounted node has it's own orphan dir which we
1369 * must run during recovery. Our strategy here is to build a list of
1370 * the inodes in the orphan dir and iget/iput them. The VFS does
1371 * (most) of the rest of the work.
1373 * Orphan recovery can happen at any time, not just mount so we have a
1374 * couple of extra considerations.
1376 * - We grab as many inodes as we can under the orphan dir lock -
1377 * doing iget() outside the orphan dir risks getting a reference on
1379 * - We must be sure not to deadlock with other processes on the
1380 * system wanting to run delete_inode(). This can happen when they go
1381 * to lock the orphan dir and the orphan recovery process attempts to
1382 * iget() inside the orphan dir lock. This can be avoided by
1383 * advertising our state to ocfs2_delete_inode().
1385 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1389 struct inode *inode = NULL;
1391 struct ocfs2_inode_info *oi;
1393 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1395 ocfs2_mark_recovering_orphan_dir(osb, slot);
1396 ret = ocfs2_queue_orphans(osb, slot, &inode);
1397 ocfs2_clear_recovering_orphan_dir(osb, slot);
1399 /* Error here should be noted, but we want to continue with as
1400 * many queued inodes as we've got. */
1405 oi = OCFS2_I(inode);
1406 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1408 iter = oi->ip_next_orphan;
1410 spin_lock(&oi->ip_lock);
1411 /* Delete voting may have set these on the assumption
1412 * that the other node would wipe them successfully.
1413 * If they are still in the node's orphan dir, we need
1414 * to reset that state. */
1415 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1417 /* Set the proper information to get us going into
1418 * ocfs2_delete_inode. */
1419 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1420 spin_unlock(&oi->ip_lock);
1430 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1432 /* This check is good because ocfs2 will wait on our recovery
1433 * thread before changing it to something other than MOUNTED
1435 wait_event(osb->osb_mount_event,
1436 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1437 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1439 /* If there's an error on mount, then we may never get to the
1440 * MOUNTED flag, but this is set right before
1441 * dismount_volume() so we can trust it. */
1442 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1443 mlog(0, "mount error, exiting!\n");
1450 static int ocfs2_commit_thread(void *arg)
1453 struct ocfs2_super *osb = arg;
1454 struct ocfs2_journal *journal = osb->journal;
1456 /* we can trust j_num_trans here because _should_stop() is only set in
1457 * shutdown and nobody other than ourselves should be able to start
1458 * transactions. committing on shutdown might take a few iterations
1459 * as final transactions put deleted inodes on the list */
1460 while (!(kthread_should_stop() &&
1461 atomic_read(&journal->j_num_trans) == 0)) {
1463 wait_event_interruptible(osb->checkpoint_event,
1464 atomic_read(&journal->j_num_trans)
1465 || kthread_should_stop());
1467 status = ocfs2_commit_cache(osb);
1471 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1473 "commit_thread: %u transactions pending on "
1475 atomic_read(&journal->j_num_trans));
1482 /* Look for a dirty journal without taking any cluster locks. Used for
1483 * hard readonly access to determine whether the file system journals
1484 * require recovery. */
1485 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1489 struct buffer_head *di_bh;
1490 struct ocfs2_dinode *di;
1491 struct inode *journal = NULL;
1493 for(slot = 0; slot < osb->max_slots; slot++) {
1494 journal = ocfs2_get_system_file_inode(osb,
1495 JOURNAL_SYSTEM_INODE,
1497 if (!journal || is_bad_inode(journal)) {
1504 ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1511 di = (struct ocfs2_dinode *) di_bh->b_data;
1513 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1514 OCFS2_JOURNAL_DIRTY_FL)