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>
40 #include "extent_map.h"
41 #include "heartbeat.h"
44 #include "localalloc.h"
49 #include "buffer_head_io.h"
51 DEFINE_SPINLOCK(trans_inc_lock);
53 static int ocfs2_force_read_journal(struct inode *inode);
54 static int ocfs2_recover_node(struct ocfs2_super *osb,
56 static int __ocfs2_recovery_thread(void *arg);
57 static int ocfs2_commit_cache(struct ocfs2_super *osb);
58 static int ocfs2_wait_on_mount(struct ocfs2_super *osb);
59 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
60 int dirty, int replayed);
61 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
63 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
65 static int ocfs2_commit_thread(void *arg);
69 * The recovery_list is a simple linked list of node numbers to recover.
70 * It is protected by the recovery_lock.
73 struct ocfs2_recovery_map {
75 unsigned int *rm_entries;
78 int ocfs2_recovery_init(struct ocfs2_super *osb)
80 struct ocfs2_recovery_map *rm;
82 mutex_init(&osb->recovery_lock);
83 osb->disable_recovery = 0;
84 osb->recovery_thread_task = NULL;
85 init_waitqueue_head(&osb->recovery_event);
87 rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
88 osb->max_slots * sizeof(unsigned int),
95 rm->rm_entries = (unsigned int *)((char *)rm +
96 sizeof(struct ocfs2_recovery_map));
97 osb->recovery_map = rm;
102 /* we can't grab the goofy sem lock from inside wait_event, so we use
103 * memory barriers to make sure that we'll see the null task before
105 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
108 return osb->recovery_thread_task != NULL;
111 void ocfs2_recovery_exit(struct ocfs2_super *osb)
113 struct ocfs2_recovery_map *rm;
115 /* disable any new recovery threads and wait for any currently
116 * running ones to exit. Do this before setting the vol_state. */
117 mutex_lock(&osb->recovery_lock);
118 osb->disable_recovery = 1;
119 mutex_unlock(&osb->recovery_lock);
120 wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
122 /* At this point, we know that no more recovery threads can be
123 * launched, so wait for any recovery completion work to
125 flush_workqueue(ocfs2_wq);
128 * Now that recovery is shut down, and the osb is about to be
129 * freed, the osb_lock is not taken here.
131 rm = osb->recovery_map;
132 /* XXX: Should we bug if there are dirty entries? */
137 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
138 unsigned int node_num)
141 struct ocfs2_recovery_map *rm = osb->recovery_map;
143 assert_spin_locked(&osb->osb_lock);
145 for (i = 0; i < rm->rm_used; i++) {
146 if (rm->rm_entries[i] == node_num)
153 /* Behaves like test-and-set. Returns the previous value */
154 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
155 unsigned int node_num)
157 struct ocfs2_recovery_map *rm = osb->recovery_map;
159 spin_lock(&osb->osb_lock);
160 if (__ocfs2_recovery_map_test(osb, node_num)) {
161 spin_unlock(&osb->osb_lock);
165 /* XXX: Can this be exploited? Not from o2dlm... */
166 BUG_ON(rm->rm_used >= osb->max_slots);
168 rm->rm_entries[rm->rm_used] = node_num;
170 spin_unlock(&osb->osb_lock);
175 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
176 unsigned int node_num)
179 struct ocfs2_recovery_map *rm = osb->recovery_map;
181 spin_lock(&osb->osb_lock);
183 for (i = 0; i < rm->rm_used; i++) {
184 if (rm->rm_entries[i] == node_num)
188 if (i < rm->rm_used) {
189 /* XXX: be careful with the pointer math */
190 memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
191 (rm->rm_used - i - 1) * sizeof(unsigned int));
195 spin_unlock(&osb->osb_lock);
198 static int ocfs2_commit_cache(struct ocfs2_super *osb)
201 unsigned int flushed;
202 unsigned long old_id;
203 struct ocfs2_journal *journal = NULL;
207 journal = osb->journal;
209 /* Flush all pending commits and checkpoint the journal. */
210 down_write(&journal->j_trans_barrier);
212 if (atomic_read(&journal->j_num_trans) == 0) {
213 up_write(&journal->j_trans_barrier);
214 mlog(0, "No transactions for me to flush!\n");
218 jbd2_journal_lock_updates(journal->j_journal);
219 status = jbd2_journal_flush(journal->j_journal);
220 jbd2_journal_unlock_updates(journal->j_journal);
222 up_write(&journal->j_trans_barrier);
227 old_id = ocfs2_inc_trans_id(journal);
229 flushed = atomic_read(&journal->j_num_trans);
230 atomic_set(&journal->j_num_trans, 0);
231 up_write(&journal->j_trans_barrier);
233 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
234 journal->j_trans_id, flushed);
236 ocfs2_wake_downconvert_thread(osb);
237 wake_up(&journal->j_checkpointed);
243 /* pass it NULL and it will allocate a new handle object for you. If
244 * you pass it a handle however, it may still return error, in which
245 * case it has free'd the passed handle for you. */
246 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
248 journal_t *journal = osb->journal->j_journal;
251 BUG_ON(!osb || !osb->journal->j_journal);
253 if (ocfs2_is_hard_readonly(osb))
254 return ERR_PTR(-EROFS);
256 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
257 BUG_ON(max_buffs <= 0);
259 /* Nested transaction? Just return the handle... */
260 if (journal_current_handle())
261 return jbd2_journal_start(journal, max_buffs);
263 down_read(&osb->journal->j_trans_barrier);
265 handle = jbd2_journal_start(journal, max_buffs);
266 if (IS_ERR(handle)) {
267 up_read(&osb->journal->j_trans_barrier);
269 mlog_errno(PTR_ERR(handle));
271 if (is_journal_aborted(journal)) {
272 ocfs2_abort(osb->sb, "Detected aborted journal");
273 handle = ERR_PTR(-EROFS);
276 if (!ocfs2_mount_local(osb))
277 atomic_inc(&(osb->journal->j_num_trans));
283 int ocfs2_commit_trans(struct ocfs2_super *osb,
287 struct ocfs2_journal *journal = osb->journal;
291 nested = handle->h_ref > 1;
292 ret = jbd2_journal_stop(handle);
297 up_read(&journal->j_trans_barrier);
303 * 'nblocks' is what you want to add to the current
304 * transaction. extend_trans will either extend the current handle by
305 * nblocks, or commit it and start a new one with nblocks credits.
307 * This might call jbd2_journal_restart() which will commit dirty buffers
308 * and then restart the transaction. Before calling
309 * ocfs2_extend_trans(), any changed blocks should have been
310 * dirtied. After calling it, all blocks which need to be changed must
311 * go through another set of journal_access/journal_dirty calls.
313 * WARNING: This will not release any semaphores or disk locks taken
314 * during the transaction, so make sure they were taken *before*
315 * start_trans or we'll have ordering deadlocks.
317 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
318 * good because transaction ids haven't yet been recorded on the
319 * cluster locks associated with this handle.
321 int ocfs2_extend_trans(handle_t *handle, int nblocks)
330 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
332 #ifdef CONFIG_OCFS2_DEBUG_FS
335 status = jbd2_journal_extend(handle, nblocks);
344 "jbd2_journal_extend failed, trying "
345 "jbd2_journal_restart\n");
346 status = jbd2_journal_restart(handle, nblocks);
360 int ocfs2_journal_access(handle_t *handle,
362 struct buffer_head *bh,
371 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
372 (unsigned long long)bh->b_blocknr, type,
373 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
374 "OCFS2_JOURNAL_ACCESS_CREATE" :
375 "OCFS2_JOURNAL_ACCESS_WRITE",
378 /* we can safely remove this assertion after testing. */
379 if (!buffer_uptodate(bh)) {
380 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
381 mlog(ML_ERROR, "b_blocknr=%llu\n",
382 (unsigned long long)bh->b_blocknr);
386 /* Set the current transaction information on the inode so
387 * that the locking code knows whether it can drop it's locks
388 * on this inode or not. We're protected from the commit
389 * thread updating the current transaction id until
390 * ocfs2_commit_trans() because ocfs2_start_trans() took
391 * j_trans_barrier for us. */
392 ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
394 mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
396 case OCFS2_JOURNAL_ACCESS_CREATE:
397 case OCFS2_JOURNAL_ACCESS_WRITE:
398 status = jbd2_journal_get_write_access(handle, bh);
401 case OCFS2_JOURNAL_ACCESS_UNDO:
402 status = jbd2_journal_get_undo_access(handle, bh);
407 mlog(ML_ERROR, "Uknown access type!\n");
409 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
412 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
419 int ocfs2_journal_dirty(handle_t *handle,
420 struct buffer_head *bh)
424 mlog_entry("(bh->b_blocknr=%llu)\n",
425 (unsigned long long)bh->b_blocknr);
427 status = jbd2_journal_dirty_metadata(handle, bh);
429 mlog(ML_ERROR, "Could not dirty metadata buffer. "
430 "(bh->b_blocknr=%llu)\n",
431 (unsigned long long)bh->b_blocknr);
437 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * JBD2_DEFAULT_MAX_COMMIT_AGE)
439 void ocfs2_set_journal_params(struct ocfs2_super *osb)
441 journal_t *journal = osb->journal->j_journal;
442 unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
444 if (osb->osb_commit_interval)
445 commit_interval = osb->osb_commit_interval;
447 spin_lock(&journal->j_state_lock);
448 journal->j_commit_interval = commit_interval;
449 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
450 journal->j_flags |= JBD2_BARRIER;
452 journal->j_flags &= ~JBD2_BARRIER;
453 spin_unlock(&journal->j_state_lock);
456 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
459 struct inode *inode = NULL; /* the journal inode */
460 journal_t *j_journal = NULL;
461 struct ocfs2_dinode *di = NULL;
462 struct buffer_head *bh = NULL;
463 struct ocfs2_super *osb;
470 osb = journal->j_osb;
472 /* already have the inode for our journal */
473 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
480 if (is_bad_inode(inode)) {
481 mlog(ML_ERROR, "access error (bad inode)\n");
488 SET_INODE_JOURNAL(inode);
489 OCFS2_I(inode)->ip_open_count++;
491 /* Skip recovery waits here - journal inode metadata never
492 * changes in a live cluster so it can be considered an
493 * exception to the rule. */
494 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
496 if (status != -ERESTARTSYS)
497 mlog(ML_ERROR, "Could not get lock on journal!\n");
502 di = (struct ocfs2_dinode *)bh->b_data;
504 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
505 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
511 mlog(0, "inode->i_size = %lld\n", inode->i_size);
512 mlog(0, "inode->i_blocks = %llu\n",
513 (unsigned long long)inode->i_blocks);
514 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
516 /* call the kernels journal init function now */
517 j_journal = jbd2_journal_init_inode(inode);
518 if (j_journal == NULL) {
519 mlog(ML_ERROR, "Linux journal layer error\n");
524 mlog(0, "Returned from jbd2_journal_init_inode\n");
525 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
527 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
528 OCFS2_JOURNAL_DIRTY_FL);
530 journal->j_journal = j_journal;
531 journal->j_inode = inode;
534 ocfs2_set_journal_params(osb);
536 journal->j_state = OCFS2_JOURNAL_LOADED;
542 ocfs2_inode_unlock(inode, 1);
545 OCFS2_I(inode)->ip_open_count--;
554 static void ocfs2_bump_recovery_generation(struct ocfs2_dinode *di)
556 le32_add_cpu(&(di->id1.journal1.ij_recovery_generation), 1);
559 static u32 ocfs2_get_recovery_generation(struct ocfs2_dinode *di)
561 return le32_to_cpu(di->id1.journal1.ij_recovery_generation);
564 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
565 int dirty, int replayed)
569 struct ocfs2_journal *journal = osb->journal;
570 struct buffer_head *bh = journal->j_bh;
571 struct ocfs2_dinode *fe;
575 fe = (struct ocfs2_dinode *)bh->b_data;
577 /* The journal bh on the osb always comes from ocfs2_journal_init()
578 * and was validated there inside ocfs2_inode_lock_full(). It's a
579 * code bug if we mess it up. */
580 BUG_ON(!OCFS2_IS_VALID_DINODE(fe));
582 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
584 flags |= OCFS2_JOURNAL_DIRTY_FL;
586 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
587 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
590 ocfs2_bump_recovery_generation(fe);
592 status = ocfs2_write_block(osb, bh, journal->j_inode);
601 * If the journal has been kmalloc'd it needs to be freed after this
604 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
606 struct ocfs2_journal *journal = NULL;
608 struct inode *inode = NULL;
609 int num_running_trans = 0;
615 journal = osb->journal;
619 inode = journal->j_inode;
621 if (journal->j_state != OCFS2_JOURNAL_LOADED)
624 /* need to inc inode use count - jbd2_journal_destroy will iput. */
628 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
629 if (num_running_trans > 0)
630 mlog(0, "Shutting down journal: must wait on %d "
631 "running transactions!\n",
634 /* Do a commit_cache here. It will flush our journal, *and*
635 * release any locks that are still held.
636 * set the SHUTDOWN flag and release the trans lock.
637 * the commit thread will take the trans lock for us below. */
638 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
640 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
641 * drop the trans_lock (which we want to hold until we
642 * completely destroy the journal. */
643 if (osb->commit_task) {
644 /* Wait for the commit thread */
645 mlog(0, "Waiting for ocfs2commit to exit....\n");
646 kthread_stop(osb->commit_task);
647 osb->commit_task = NULL;
650 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
652 if (ocfs2_mount_local(osb)) {
653 jbd2_journal_lock_updates(journal->j_journal);
654 status = jbd2_journal_flush(journal->j_journal);
655 jbd2_journal_unlock_updates(journal->j_journal);
662 * Do not toggle if flush was unsuccessful otherwise
663 * will leave dirty metadata in a "clean" journal
665 status = ocfs2_journal_toggle_dirty(osb, 0, 0);
670 /* Shutdown the kernel journal system */
671 jbd2_journal_destroy(journal->j_journal);
672 journal->j_journal = NULL;
674 OCFS2_I(inode)->ip_open_count--;
676 /* unlock our journal */
677 ocfs2_inode_unlock(inode, 1);
679 brelse(journal->j_bh);
680 journal->j_bh = NULL;
682 journal->j_state = OCFS2_JOURNAL_FREE;
684 // up_write(&journal->j_trans_barrier);
691 static void ocfs2_clear_journal_error(struct super_block *sb,
697 olderr = jbd2_journal_errno(journal);
699 mlog(ML_ERROR, "File system error %d recorded in "
700 "journal %u.\n", olderr, slot);
701 mlog(ML_ERROR, "File system on device %s needs checking.\n",
704 jbd2_journal_ack_err(journal);
705 jbd2_journal_clear_err(journal);
709 int ocfs2_journal_load(struct ocfs2_journal *journal, int local, int replayed)
712 struct ocfs2_super *osb;
718 osb = journal->j_osb;
720 status = jbd2_journal_load(journal->j_journal);
722 mlog(ML_ERROR, "Failed to load journal!\n");
726 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
728 status = ocfs2_journal_toggle_dirty(osb, 1, replayed);
734 /* Launch the commit thread */
736 osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
738 if (IS_ERR(osb->commit_task)) {
739 status = PTR_ERR(osb->commit_task);
740 osb->commit_task = NULL;
741 mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
746 osb->commit_task = NULL;
754 /* 'full' flag tells us whether we clear out all blocks or if we just
755 * mark the journal clean */
756 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
764 status = jbd2_journal_wipe(journal->j_journal, full);
770 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0, 0);
779 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
782 struct ocfs2_recovery_map *rm = osb->recovery_map;
784 spin_lock(&osb->osb_lock);
785 empty = (rm->rm_used == 0);
786 spin_unlock(&osb->osb_lock);
791 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
793 wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
797 * JBD Might read a cached version of another nodes journal file. We
798 * don't want this as this file changes often and we get no
799 * notification on those changes. The only way to be sure that we've
800 * got the most up to date version of those blocks then is to force
801 * read them off disk. Just searching through the buffer cache won't
802 * work as there may be pages backing this file which are still marked
803 * up to date. We know things can't change on this file underneath us
804 * as we have the lock by now :)
806 static int ocfs2_force_read_journal(struct inode *inode)
810 u64 v_blkno, p_blkno, p_blocks, num_blocks;
811 #define CONCURRENT_JOURNAL_FILL 32ULL
812 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
816 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
818 num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
820 while (v_blkno < num_blocks) {
821 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
822 &p_blkno, &p_blocks, NULL);
828 if (p_blocks > CONCURRENT_JOURNAL_FILL)
829 p_blocks = CONCURRENT_JOURNAL_FILL;
831 /* We are reading journal data which should not
832 * be put in the uptodate cache */
833 status = ocfs2_read_blocks_sync(OCFS2_SB(inode->i_sb),
834 p_blkno, p_blocks, bhs);
840 for(i = 0; i < p_blocks; i++) {
849 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
855 struct ocfs2_la_recovery_item {
856 struct list_head lri_list;
858 struct ocfs2_dinode *lri_la_dinode;
859 struct ocfs2_dinode *lri_tl_dinode;
862 /* Does the second half of the recovery process. By this point, the
863 * node is marked clean and can actually be considered recovered,
864 * hence it's no longer in the recovery map, but there's still some
865 * cleanup we can do which shouldn't happen within the recovery thread
866 * as locking in that context becomes very difficult if we are to take
867 * recovering nodes into account.
869 * NOTE: This function can and will sleep on recovery of other nodes
870 * during cluster locking, just like any other ocfs2 process.
872 void ocfs2_complete_recovery(struct work_struct *work)
875 struct ocfs2_journal *journal =
876 container_of(work, struct ocfs2_journal, j_recovery_work);
877 struct ocfs2_super *osb = journal->j_osb;
878 struct ocfs2_dinode *la_dinode, *tl_dinode;
879 struct ocfs2_la_recovery_item *item, *n;
880 LIST_HEAD(tmp_la_list);
884 mlog(0, "completing recovery from keventd\n");
886 spin_lock(&journal->j_lock);
887 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
888 spin_unlock(&journal->j_lock);
890 list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
891 list_del_init(&item->lri_list);
893 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
895 la_dinode = item->lri_la_dinode;
897 mlog(0, "Clean up local alloc %llu\n",
898 (unsigned long long)le64_to_cpu(la_dinode->i_blkno));
900 ret = ocfs2_complete_local_alloc_recovery(osb,
908 tl_dinode = item->lri_tl_dinode;
910 mlog(0, "Clean up truncate log %llu\n",
911 (unsigned long long)le64_to_cpu(tl_dinode->i_blkno));
913 ret = ocfs2_complete_truncate_log_recovery(osb,
921 ret = ocfs2_recover_orphans(osb, item->lri_slot);
928 mlog(0, "Recovery completion\n");
932 /* NOTE: This function always eats your references to la_dinode and
933 * tl_dinode, either manually on error, or by passing them to
934 * ocfs2_complete_recovery */
935 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
937 struct ocfs2_dinode *la_dinode,
938 struct ocfs2_dinode *tl_dinode)
940 struct ocfs2_la_recovery_item *item;
942 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
944 /* Though we wish to avoid it, we are in fact safe in
945 * skipping local alloc cleanup as fsck.ocfs2 is more
946 * than capable of reclaiming unused space. */
957 INIT_LIST_HEAD(&item->lri_list);
958 item->lri_la_dinode = la_dinode;
959 item->lri_slot = slot_num;
960 item->lri_tl_dinode = tl_dinode;
962 spin_lock(&journal->j_lock);
963 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
964 queue_work(ocfs2_wq, &journal->j_recovery_work);
965 spin_unlock(&journal->j_lock);
968 /* Called by the mount code to queue recovery the last part of
969 * recovery for it's own slot. */
970 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
972 struct ocfs2_journal *journal = osb->journal;
975 /* No need to queue up our truncate_log as regular
976 * cleanup will catch that. */
977 ocfs2_queue_recovery_completion(journal,
979 osb->local_alloc_copy,
981 ocfs2_schedule_truncate_log_flush(osb, 0);
983 osb->local_alloc_copy = NULL;
988 static int __ocfs2_recovery_thread(void *arg)
990 int status, node_num;
991 struct ocfs2_super *osb = arg;
992 struct ocfs2_recovery_map *rm = osb->recovery_map;
996 status = ocfs2_wait_on_mount(osb);
1002 status = ocfs2_super_lock(osb, 1);
1008 spin_lock(&osb->osb_lock);
1009 while (rm->rm_used) {
1010 /* It's always safe to remove entry zero, as we won't
1011 * clear it until ocfs2_recover_node() has succeeded. */
1012 node_num = rm->rm_entries[0];
1013 spin_unlock(&osb->osb_lock);
1015 status = ocfs2_recover_node(osb, node_num);
1017 ocfs2_recovery_map_clear(osb, node_num);
1020 "Error %d recovering node %d on device (%u,%u)!\n",
1022 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1023 mlog(ML_ERROR, "Volume requires unmount.\n");
1026 spin_lock(&osb->osb_lock);
1028 spin_unlock(&osb->osb_lock);
1029 mlog(0, "All nodes recovered\n");
1031 /* Refresh all journal recovery generations from disk */
1032 status = ocfs2_check_journals_nolocks(osb);
1033 status = (status == -EROFS) ? 0 : status;
1037 ocfs2_super_unlock(osb, 1);
1039 /* We always run recovery on our own orphan dir - the dead
1040 * node(s) may have disallowd a previos inode delete. Re-processing
1041 * is therefore required. */
1042 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1046 mutex_lock(&osb->recovery_lock);
1047 if (!status && !ocfs2_recovery_completed(osb)) {
1048 mutex_unlock(&osb->recovery_lock);
1052 osb->recovery_thread_task = NULL;
1053 mb(); /* sync with ocfs2_recovery_thread_running */
1054 wake_up(&osb->recovery_event);
1056 mutex_unlock(&osb->recovery_lock);
1059 /* no one is callint kthread_stop() for us so the kthread() api
1060 * requires that we call do_exit(). And it isn't exported, but
1061 * complete_and_exit() seems to be a minimal wrapper around it. */
1062 complete_and_exit(NULL, status);
1066 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1068 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1069 node_num, osb->node_num);
1071 mutex_lock(&osb->recovery_lock);
1072 if (osb->disable_recovery)
1075 /* People waiting on recovery will wait on
1076 * the recovery map to empty. */
1077 if (ocfs2_recovery_map_set(osb, node_num))
1078 mlog(0, "node %d already in recovery map.\n", node_num);
1080 mlog(0, "starting recovery thread...\n");
1082 if (osb->recovery_thread_task)
1085 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1087 if (IS_ERR(osb->recovery_thread_task)) {
1088 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1089 osb->recovery_thread_task = NULL;
1093 mutex_unlock(&osb->recovery_lock);
1094 wake_up(&osb->recovery_event);
1099 static int ocfs2_read_journal_inode(struct ocfs2_super *osb,
1101 struct buffer_head **bh,
1102 struct inode **ret_inode)
1104 int status = -EACCES;
1105 struct inode *inode = NULL;
1107 BUG_ON(slot_num >= osb->max_slots);
1109 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1111 if (!inode || is_bad_inode(inode)) {
1115 SET_INODE_JOURNAL(inode);
1117 status = ocfs2_read_inode_block_full(inode, bh, OCFS2_BH_IGNORE_CACHE);
1127 if (status || !ret_inode)
1135 /* Does the actual journal replay and marks the journal inode as
1136 * clean. Will only replay if the journal inode is marked dirty. */
1137 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1144 struct inode *inode = NULL;
1145 struct ocfs2_dinode *fe;
1146 journal_t *journal = NULL;
1147 struct buffer_head *bh = NULL;
1150 status = ocfs2_read_journal_inode(osb, slot_num, &bh, &inode);
1156 fe = (struct ocfs2_dinode *)bh->b_data;
1157 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1162 * As the fs recovery is asynchronous, there is a small chance that
1163 * another node mounted (and recovered) the slot before the recovery
1164 * thread could get the lock. To handle that, we dirty read the journal
1165 * inode for that slot to get the recovery generation. If it is
1166 * different than what we expected, the slot has been recovered.
1167 * If not, it needs recovery.
1169 if (osb->slot_recovery_generations[slot_num] != slot_reco_gen) {
1170 mlog(0, "Slot %u already recovered (old/new=%u/%u)\n", slot_num,
1171 osb->slot_recovery_generations[slot_num], slot_reco_gen);
1172 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1177 /* Continue with recovery as the journal has not yet been recovered */
1179 status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1181 mlog(0, "status returned from ocfs2_inode_lock=%d\n", status);
1182 if (status != -ERESTARTSYS)
1183 mlog(ML_ERROR, "Could not lock journal!\n");
1188 fe = (struct ocfs2_dinode *) bh->b_data;
1190 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1191 slot_reco_gen = ocfs2_get_recovery_generation(fe);
1193 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1194 mlog(0, "No recovery required for node %d\n", node_num);
1195 /* Refresh recovery generation for the slot */
1196 osb->slot_recovery_generations[slot_num] = slot_reco_gen;
1200 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1202 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1204 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1206 status = ocfs2_force_read_journal(inode);
1212 mlog(0, "calling journal_init_inode\n");
1213 journal = jbd2_journal_init_inode(inode);
1214 if (journal == NULL) {
1215 mlog(ML_ERROR, "Linux journal layer error\n");
1220 status = jbd2_journal_load(journal);
1225 jbd2_journal_destroy(journal);
1229 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1231 /* wipe the journal */
1232 mlog(0, "flushing the journal.\n");
1233 jbd2_journal_lock_updates(journal);
1234 status = jbd2_journal_flush(journal);
1235 jbd2_journal_unlock_updates(journal);
1239 /* This will mark the node clean */
1240 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1241 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1242 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1244 /* Increment recovery generation to indicate successful recovery */
1245 ocfs2_bump_recovery_generation(fe);
1246 osb->slot_recovery_generations[slot_num] =
1247 ocfs2_get_recovery_generation(fe);
1249 status = ocfs2_write_block(osb, bh, inode);
1256 jbd2_journal_destroy(journal);
1259 /* drop the lock on this nodes journal */
1261 ocfs2_inode_unlock(inode, 1);
1273 * Do the most important parts of node recovery:
1274 * - Replay it's journal
1275 * - Stamp a clean local allocator file
1276 * - Stamp a clean truncate log
1277 * - Mark the node clean
1279 * If this function completes without error, a node in OCFS2 can be
1280 * said to have been safely recovered. As a result, failure during the
1281 * second part of a nodes recovery process (local alloc recovery) is
1282 * far less concerning.
1284 static int ocfs2_recover_node(struct ocfs2_super *osb,
1289 struct ocfs2_dinode *la_copy = NULL;
1290 struct ocfs2_dinode *tl_copy = NULL;
1292 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1293 node_num, osb->node_num);
1295 mlog(0, "checking node %d\n", node_num);
1297 /* Should not ever be called to recover ourselves -- in that
1298 * case we should've called ocfs2_journal_load instead. */
1299 BUG_ON(osb->node_num == node_num);
1301 slot_num = ocfs2_node_num_to_slot(osb, node_num);
1302 if (slot_num == -ENOENT) {
1304 mlog(0, "no slot for this node, so no recovery required.\n");
1308 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1310 status = ocfs2_replay_journal(osb, node_num, slot_num);
1312 if (status == -EBUSY) {
1313 mlog(0, "Skipping recovery for slot %u (node %u) "
1314 "as another node has recovered it\n", slot_num,
1323 /* Stamp a clean local alloc file AFTER recovering the journal... */
1324 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1330 /* An error from begin_truncate_log_recovery is not
1331 * serious enough to warrant halting the rest of
1333 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1337 /* Likewise, this would be a strange but ultimately not so
1338 * harmful place to get an error... */
1339 status = ocfs2_clear_slot(osb, slot_num);
1343 /* This will kfree the memory pointed to by la_copy and tl_copy */
1344 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1354 /* Test node liveness by trylocking his journal. If we get the lock,
1355 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1356 * still alive (we couldn't get the lock) and < 0 on error. */
1357 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1361 struct inode *inode = NULL;
1363 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1365 if (inode == NULL) {
1366 mlog(ML_ERROR, "access error\n");
1370 if (is_bad_inode(inode)) {
1371 mlog(ML_ERROR, "access error (bad inode)\n");
1377 SET_INODE_JOURNAL(inode);
1379 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1380 status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1382 if (status != -EAGAIN)
1387 ocfs2_inode_unlock(inode, 1);
1395 /* Call this underneath ocfs2_super_lock. It also assumes that the
1396 * slot info struct has been updated from disk. */
1397 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1399 unsigned int node_num;
1402 struct buffer_head *bh = NULL;
1403 struct ocfs2_dinode *di;
1405 /* This is called with the super block cluster lock, so we
1406 * know that the slot map can't change underneath us. */
1408 for (i = 0; i < osb->max_slots; i++) {
1409 /* Read journal inode to get the recovery generation */
1410 status = ocfs2_read_journal_inode(osb, i, &bh, NULL);
1415 di = (struct ocfs2_dinode *)bh->b_data;
1416 gen = ocfs2_get_recovery_generation(di);
1420 spin_lock(&osb->osb_lock);
1421 osb->slot_recovery_generations[i] = gen;
1423 mlog(0, "Slot %u recovery generation is %u\n", i,
1424 osb->slot_recovery_generations[i]);
1426 if (i == osb->slot_num) {
1427 spin_unlock(&osb->osb_lock);
1431 status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1432 if (status == -ENOENT) {
1433 spin_unlock(&osb->osb_lock);
1437 if (__ocfs2_recovery_map_test(osb, node_num)) {
1438 spin_unlock(&osb->osb_lock);
1441 spin_unlock(&osb->osb_lock);
1443 /* Ok, we have a slot occupied by another node which
1444 * is not in the recovery map. We trylock his journal
1445 * file here to test if he's alive. */
1446 status = ocfs2_trylock_journal(osb, i);
1448 /* Since we're called from mount, we know that
1449 * the recovery thread can't race us on
1450 * setting / checking the recovery bits. */
1451 ocfs2_recovery_thread(osb, node_num);
1452 } else if ((status < 0) && (status != -EAGAIN)) {
1464 struct ocfs2_orphan_filldir_priv {
1466 struct ocfs2_super *osb;
1469 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1470 loff_t pos, u64 ino, unsigned type)
1472 struct ocfs2_orphan_filldir_priv *p = priv;
1475 if (name_len == 1 && !strncmp(".", name, 1))
1477 if (name_len == 2 && !strncmp("..", name, 2))
1480 /* Skip bad inodes so that recovery can continue */
1481 iter = ocfs2_iget(p->osb, ino,
1482 OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
1486 mlog(0, "queue orphan %llu\n",
1487 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1488 /* No locking is required for the next_orphan queue as there
1489 * is only ever a single process doing orphan recovery. */
1490 OCFS2_I(iter)->ip_next_orphan = p->head;
1496 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1498 struct inode **head)
1501 struct inode *orphan_dir_inode = NULL;
1502 struct ocfs2_orphan_filldir_priv priv;
1508 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1509 ORPHAN_DIR_SYSTEM_INODE,
1511 if (!orphan_dir_inode) {
1517 mutex_lock(&orphan_dir_inode->i_mutex);
1518 status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
1524 status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
1525 ocfs2_orphan_filldir);
1534 ocfs2_inode_unlock(orphan_dir_inode, 0);
1536 mutex_unlock(&orphan_dir_inode->i_mutex);
1537 iput(orphan_dir_inode);
1541 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1546 spin_lock(&osb->osb_lock);
1547 ret = !osb->osb_orphan_wipes[slot];
1548 spin_unlock(&osb->osb_lock);
1552 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1555 spin_lock(&osb->osb_lock);
1556 /* Mark ourselves such that new processes in delete_inode()
1557 * know to quit early. */
1558 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1559 while (osb->osb_orphan_wipes[slot]) {
1560 /* If any processes are already in the middle of an
1561 * orphan wipe on this dir, then we need to wait for
1563 spin_unlock(&osb->osb_lock);
1564 wait_event_interruptible(osb->osb_wipe_event,
1565 ocfs2_orphan_recovery_can_continue(osb, slot));
1566 spin_lock(&osb->osb_lock);
1568 spin_unlock(&osb->osb_lock);
1571 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1574 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1578 * Orphan recovery. Each mounted node has it's own orphan dir which we
1579 * must run during recovery. Our strategy here is to build a list of
1580 * the inodes in the orphan dir and iget/iput them. The VFS does
1581 * (most) of the rest of the work.
1583 * Orphan recovery can happen at any time, not just mount so we have a
1584 * couple of extra considerations.
1586 * - We grab as many inodes as we can under the orphan dir lock -
1587 * doing iget() outside the orphan dir risks getting a reference on
1589 * - We must be sure not to deadlock with other processes on the
1590 * system wanting to run delete_inode(). This can happen when they go
1591 * to lock the orphan dir and the orphan recovery process attempts to
1592 * iget() inside the orphan dir lock. This can be avoided by
1593 * advertising our state to ocfs2_delete_inode().
1595 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1599 struct inode *inode = NULL;
1601 struct ocfs2_inode_info *oi;
1603 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1605 ocfs2_mark_recovering_orphan_dir(osb, slot);
1606 ret = ocfs2_queue_orphans(osb, slot, &inode);
1607 ocfs2_clear_recovering_orphan_dir(osb, slot);
1609 /* Error here should be noted, but we want to continue with as
1610 * many queued inodes as we've got. */
1615 oi = OCFS2_I(inode);
1616 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1618 iter = oi->ip_next_orphan;
1620 spin_lock(&oi->ip_lock);
1621 /* The remote delete code may have set these on the
1622 * assumption that the other node would wipe them
1623 * successfully. If they are still in the node's
1624 * orphan dir, we need to reset that state. */
1625 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1627 /* Set the proper information to get us going into
1628 * ocfs2_delete_inode. */
1629 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1630 spin_unlock(&oi->ip_lock);
1640 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1642 /* This check is good because ocfs2 will wait on our recovery
1643 * thread before changing it to something other than MOUNTED
1645 wait_event(osb->osb_mount_event,
1646 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1647 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1649 /* If there's an error on mount, then we may never get to the
1650 * MOUNTED flag, but this is set right before
1651 * dismount_volume() so we can trust it. */
1652 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1653 mlog(0, "mount error, exiting!\n");
1660 static int ocfs2_commit_thread(void *arg)
1663 struct ocfs2_super *osb = arg;
1664 struct ocfs2_journal *journal = osb->journal;
1666 /* we can trust j_num_trans here because _should_stop() is only set in
1667 * shutdown and nobody other than ourselves should be able to start
1668 * transactions. committing on shutdown might take a few iterations
1669 * as final transactions put deleted inodes on the list */
1670 while (!(kthread_should_stop() &&
1671 atomic_read(&journal->j_num_trans) == 0)) {
1673 wait_event_interruptible(osb->checkpoint_event,
1674 atomic_read(&journal->j_num_trans)
1675 || kthread_should_stop());
1677 status = ocfs2_commit_cache(osb);
1681 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1683 "commit_thread: %u transactions pending on "
1685 atomic_read(&journal->j_num_trans));
1692 /* Reads all the journal inodes without taking any cluster locks. Used
1693 * for hard readonly access to determine whether any journal requires
1694 * recovery. Also used to refresh the recovery generation numbers after
1695 * a journal has been recovered by another node.
1697 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1701 struct buffer_head *di_bh = NULL;
1702 struct ocfs2_dinode *di;
1703 int journal_dirty = 0;
1705 for(slot = 0; slot < osb->max_slots; slot++) {
1706 ret = ocfs2_read_journal_inode(osb, slot, &di_bh, NULL);
1712 di = (struct ocfs2_dinode *) di_bh->b_data;
1714 osb->slot_recovery_generations[slot] =
1715 ocfs2_get_recovery_generation(di);
1717 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1718 OCFS2_JOURNAL_DIRTY_FL)