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 void ocfs2_handle_cleanup_locks(struct ocfs2_journal *journal,
61 struct ocfs2_journal_handle *handle);
62 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle *handle);
63 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
65 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
67 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
69 static int ocfs2_commit_thread(void *arg);
71 static int ocfs2_commit_cache(struct ocfs2_super *osb)
76 struct ocfs2_journal *journal = NULL;
80 journal = osb->journal;
82 /* Flush all pending commits and checkpoint the journal. */
83 down_write(&journal->j_trans_barrier);
85 if (atomic_read(&journal->j_num_trans) == 0) {
86 up_write(&journal->j_trans_barrier);
87 mlog(0, "No transactions for me to flush!\n");
91 journal_lock_updates(journal->j_journal);
92 status = journal_flush(journal->j_journal);
93 journal_unlock_updates(journal->j_journal);
95 up_write(&journal->j_trans_barrier);
100 old_id = ocfs2_inc_trans_id(journal);
102 flushed = atomic_read(&journal->j_num_trans);
103 atomic_set(&journal->j_num_trans, 0);
104 up_write(&journal->j_trans_barrier);
106 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
107 journal->j_trans_id, flushed);
109 ocfs2_kick_vote_thread(osb);
110 wake_up(&journal->j_checkpointed);
116 struct ocfs2_journal_handle *ocfs2_alloc_handle(struct ocfs2_super *osb)
118 struct ocfs2_journal_handle *retval = NULL;
120 retval = kcalloc(1, sizeof(*retval), GFP_NOFS);
122 mlog(ML_ERROR, "Failed to allocate memory for journal "
127 retval->num_locks = 0;
128 retval->k_handle = NULL;
130 INIT_LIST_HEAD(&retval->locks);
131 INIT_LIST_HEAD(&retval->inode_list);
132 retval->journal = osb->journal;
137 /* pass it NULL and it will allocate a new handle object for you. If
138 * you pass it a handle however, it may still return error, in which
139 * case it has free'd the passed handle for you. */
140 struct ocfs2_journal_handle *ocfs2_start_trans(struct ocfs2_super *osb,
141 struct ocfs2_journal_handle *handle,
145 journal_t *journal = osb->journal->j_journal;
147 mlog_entry("(max_buffs = %d)\n", max_buffs);
149 BUG_ON(!osb || !osb->journal->j_journal);
151 if (ocfs2_is_hard_readonly(osb)) {
156 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
157 BUG_ON(max_buffs <= 0);
159 /* JBD might support this, but our journalling code doesn't yet. */
160 if (journal_current_handle()) {
161 mlog(ML_ERROR, "Recursive transaction attempted!\n");
166 handle = ocfs2_alloc_handle(osb);
169 mlog(ML_ERROR, "Failed to allocate memory for journal "
174 down_read(&osb->journal->j_trans_barrier);
176 /* actually start the transaction now */
177 handle->k_handle = journal_start(journal, max_buffs);
178 if (IS_ERR(handle->k_handle)) {
179 up_read(&osb->journal->j_trans_barrier);
181 ret = PTR_ERR(handle->k_handle);
182 handle->k_handle = NULL;
185 if (is_journal_aborted(journal)) {
186 ocfs2_abort(osb->sb, "Detected aborted journal");
192 atomic_inc(&(osb->journal->j_num_trans));
193 handle->flags |= OCFS2_HANDLE_STARTED;
195 mlog_exit_ptr(handle);
200 ocfs2_commit_unstarted_handle(handle); /* will kfree handle */
206 void ocfs2_handle_add_inode(struct ocfs2_journal_handle *handle,
212 atomic_inc(&inode->i_count);
214 /* we're obviously changing it... */
215 mutex_lock(&inode->i_mutex);
218 BUG_ON(OCFS2_I(inode)->ip_handle);
219 BUG_ON(!list_empty(&OCFS2_I(inode)->ip_handle_list));
221 OCFS2_I(inode)->ip_handle = handle;
222 list_move_tail(&(OCFS2_I(inode)->ip_handle_list), &(handle->inode_list));
225 static void ocfs2_handle_unlock_inodes(struct ocfs2_journal_handle *handle)
227 struct list_head *p, *n;
229 struct ocfs2_inode_info *oi;
231 list_for_each_safe(p, n, &handle->inode_list) {
232 oi = list_entry(p, struct ocfs2_inode_info,
234 inode = &oi->vfs_inode;
236 OCFS2_I(inode)->ip_handle = NULL;
237 list_del_init(&OCFS2_I(inode)->ip_handle_list);
239 mutex_unlock(&inode->i_mutex);
244 /* This is trivial so we do it out of the main commit
245 * paths. Beware, it can be called from start_trans too! */
246 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle *handle)
250 BUG_ON(handle->flags & OCFS2_HANDLE_STARTED);
252 ocfs2_handle_unlock_inodes(handle);
253 /* You are allowed to add journal locks before the transaction
255 ocfs2_handle_cleanup_locks(handle->journal, handle);
262 void ocfs2_commit_trans(struct ocfs2_journal_handle *handle)
264 handle_t *jbd_handle;
266 struct ocfs2_journal *journal = handle->journal;
272 if (!(handle->flags & OCFS2_HANDLE_STARTED)) {
273 ocfs2_commit_unstarted_handle(handle);
278 /* release inode semaphores we took during this transaction */
279 ocfs2_handle_unlock_inodes(handle);
281 /* ocfs2_extend_trans may have had to call journal_restart
282 * which will always commit the transaction, but may return
283 * error for any number of reasons. If this is the case, we
284 * clear k_handle as it's not valid any more. */
285 if (handle->k_handle) {
286 jbd_handle = handle->k_handle;
288 if (handle->flags & OCFS2_HANDLE_SYNC)
289 jbd_handle->h_sync = 1;
291 jbd_handle->h_sync = 0;
293 /* actually stop the transaction. if we've set h_sync,
294 * it'll have been committed when we return */
295 retval = journal_stop(jbd_handle);
298 mlog(ML_ERROR, "Could not commit transaction\n");
302 handle->k_handle = NULL; /* it's been free'd in journal_stop */
305 ocfs2_handle_cleanup_locks(journal, handle);
307 up_read(&journal->j_trans_barrier);
314 * 'nblocks' is what you want to add to the current
315 * transaction. extend_trans will either extend the current handle by
316 * nblocks, or commit it and start a new one with nblocks credits.
318 * WARNING: This will not release any semaphores or disk locks taken
319 * during the transaction, so make sure they were taken *before*
320 * start_trans or we'll have ordering deadlocks.
322 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
323 * good because transaction ids haven't yet been recorded on the
324 * cluster locks associated with this handle.
326 int ocfs2_extend_trans(struct ocfs2_journal_handle *handle,
332 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
337 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
339 status = journal_extend(handle->k_handle, nblocks);
346 mlog(0, "journal_extend failed, trying journal_restart\n");
347 status = journal_restart(handle->k_handle, nblocks);
349 handle->k_handle = NULL;
362 int ocfs2_journal_access(struct ocfs2_journal_handle *handle,
364 struct buffer_head *bh,
372 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
374 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
375 (unsigned long long)bh->b_blocknr, type,
376 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
377 "OCFS2_JOURNAL_ACCESS_CREATE" :
378 "OCFS2_JOURNAL_ACCESS_WRITE",
381 /* we can safely remove this assertion after testing. */
382 if (!buffer_uptodate(bh)) {
383 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
384 mlog(ML_ERROR, "b_blocknr=%llu\n",
385 (unsigned long long)bh->b_blocknr);
389 /* Set the current transaction information on the inode so
390 * that the locking code knows whether it can drop it's locks
391 * on this inode or not. We're protected from the commit
392 * thread updating the current transaction id until
393 * ocfs2_commit_trans() because ocfs2_start_trans() took
394 * j_trans_barrier for us. */
395 ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
397 mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
399 case OCFS2_JOURNAL_ACCESS_CREATE:
400 case OCFS2_JOURNAL_ACCESS_WRITE:
401 status = journal_get_write_access(handle->k_handle, bh);
404 case OCFS2_JOURNAL_ACCESS_UNDO:
405 status = journal_get_undo_access(handle->k_handle, bh);
410 mlog(ML_ERROR, "Uknown access type!\n");
412 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
415 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
422 int ocfs2_journal_dirty(struct ocfs2_journal_handle *handle,
423 struct buffer_head *bh)
427 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
429 mlog_entry("(bh->b_blocknr=%llu)\n",
430 (unsigned long long)bh->b_blocknr);
432 status = journal_dirty_metadata(handle->k_handle, bh);
434 mlog(ML_ERROR, "Could not dirty metadata buffer. "
435 "(bh->b_blocknr=%llu)\n",
436 (unsigned long long)bh->b_blocknr);
442 int ocfs2_journal_dirty_data(handle_t *handle,
443 struct buffer_head *bh)
445 int err = journal_dirty_data(handle, bh);
448 /* TODO: When we can handle it, abort the handle and go RO on
454 /* We always assume you're adding a metadata lock at level 'ex' */
455 int ocfs2_handle_add_lock(struct ocfs2_journal_handle *handle,
459 struct ocfs2_journal_lock *lock;
463 lock = kmem_cache_alloc(ocfs2_lock_cache, GFP_NOFS);
472 lock->jl_inode = inode;
474 list_add_tail(&(lock->jl_lock_list), &(handle->locks));
483 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal *journal,
484 struct ocfs2_journal_handle *handle)
486 struct list_head *p, *n;
487 struct ocfs2_journal_lock *lock;
490 list_for_each_safe(p, n, &(handle->locks)) {
491 lock = list_entry(p, struct ocfs2_journal_lock,
493 list_del(&lock->jl_lock_list);
496 inode = lock->jl_inode;
497 ocfs2_meta_unlock(inode, 1);
498 if (atomic_read(&inode->i_count) == 1)
500 "Inode %llu, I'm doing a last iput for!",
501 (unsigned long long)OCFS2_I(inode)->ip_blkno);
503 kmem_cache_free(ocfs2_lock_cache, lock);
507 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
509 void ocfs2_set_journal_params(struct ocfs2_super *osb)
511 journal_t *journal = osb->journal->j_journal;
513 spin_lock(&journal->j_state_lock);
514 journal->j_commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
515 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
516 journal->j_flags |= JFS_BARRIER;
518 journal->j_flags &= ~JFS_BARRIER;
519 spin_unlock(&journal->j_state_lock);
522 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
525 struct inode *inode = NULL; /* the journal inode */
526 journal_t *j_journal = NULL;
527 struct ocfs2_dinode *di = NULL;
528 struct buffer_head *bh = NULL;
529 struct ocfs2_super *osb;
536 osb = journal->j_osb;
538 /* already have the inode for our journal */
539 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
546 if (is_bad_inode(inode)) {
547 mlog(ML_ERROR, "access error (bad inode)\n");
554 SET_INODE_JOURNAL(inode);
555 OCFS2_I(inode)->ip_open_count++;
557 /* Skip recovery waits here - journal inode metadata never
558 * changes in a live cluster so it can be considered an
559 * exception to the rule. */
560 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
561 OCFS2_META_LOCK_RECOVERY);
563 if (status != -ERESTARTSYS)
564 mlog(ML_ERROR, "Could not get lock on journal!\n");
569 di = (struct ocfs2_dinode *)bh->b_data;
571 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
572 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
578 mlog(0, "inode->i_size = %lld\n", inode->i_size);
579 mlog(0, "inode->i_blocks = %llu\n",
580 (unsigned long long)inode->i_blocks);
581 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
583 /* call the kernels journal init function now */
584 j_journal = journal_init_inode(inode);
585 if (j_journal == NULL) {
586 mlog(ML_ERROR, "Linux journal layer error\n");
591 mlog(0, "Returned from journal_init_inode\n");
592 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
594 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
595 OCFS2_JOURNAL_DIRTY_FL);
597 journal->j_journal = j_journal;
598 journal->j_inode = inode;
601 ocfs2_set_journal_params(osb);
603 journal->j_state = OCFS2_JOURNAL_LOADED;
609 ocfs2_meta_unlock(inode, 1);
613 OCFS2_I(inode)->ip_open_count--;
622 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
627 struct ocfs2_journal *journal = osb->journal;
628 struct buffer_head *bh = journal->j_bh;
629 struct ocfs2_dinode *fe;
633 fe = (struct ocfs2_dinode *)bh->b_data;
634 if (!OCFS2_IS_VALID_DINODE(fe)) {
635 /* This is called from startup/shutdown which will
636 * handle the errors in a specific manner, so no need
637 * to call ocfs2_error() here. */
638 mlog(ML_ERROR, "Journal dinode %llu has invalid "
639 "signature: %.*s", (unsigned long long)fe->i_blkno, 7,
645 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
647 flags |= OCFS2_JOURNAL_DIRTY_FL;
649 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
650 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
652 status = ocfs2_write_block(osb, bh, journal->j_inode);
662 * If the journal has been kmalloc'd it needs to be freed after this
665 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
667 struct ocfs2_journal *journal = NULL;
669 struct inode *inode = NULL;
670 int num_running_trans = 0;
676 journal = osb->journal;
680 inode = journal->j_inode;
682 if (journal->j_state != OCFS2_JOURNAL_LOADED)
685 /* need to inc inode use count as journal_destroy will iput. */
689 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
690 if (num_running_trans > 0)
691 mlog(0, "Shutting down journal: must wait on %d "
692 "running transactions!\n",
695 /* Do a commit_cache here. It will flush our journal, *and*
696 * release any locks that are still held.
697 * set the SHUTDOWN flag and release the trans lock.
698 * the commit thread will take the trans lock for us below. */
699 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
701 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
702 * drop the trans_lock (which we want to hold until we
703 * completely destroy the journal. */
704 if (osb->commit_task) {
705 /* Wait for the commit thread */
706 mlog(0, "Waiting for ocfs2commit to exit....\n");
707 kthread_stop(osb->commit_task);
708 osb->commit_task = NULL;
711 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
713 status = ocfs2_journal_toggle_dirty(osb, 0);
717 /* Shutdown the kernel journal system */
718 journal_destroy(journal->j_journal);
720 OCFS2_I(inode)->ip_open_count--;
722 /* unlock our journal */
723 ocfs2_meta_unlock(inode, 1);
725 brelse(journal->j_bh);
726 journal->j_bh = NULL;
728 journal->j_state = OCFS2_JOURNAL_FREE;
730 // up_write(&journal->j_trans_barrier);
737 static void ocfs2_clear_journal_error(struct super_block *sb,
743 olderr = journal_errno(journal);
745 mlog(ML_ERROR, "File system error %d recorded in "
746 "journal %u.\n", olderr, slot);
747 mlog(ML_ERROR, "File system on device %s needs checking.\n",
750 journal_ack_err(journal);
751 journal_clear_err(journal);
755 int ocfs2_journal_load(struct ocfs2_journal *journal)
758 struct ocfs2_super *osb;
765 osb = journal->j_osb;
767 status = journal_load(journal->j_journal);
769 mlog(ML_ERROR, "Failed to load journal!\n");
773 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
775 status = ocfs2_journal_toggle_dirty(osb, 1);
781 /* Launch the commit thread */
782 osb->commit_task = kthread_run(ocfs2_commit_thread, osb, "ocfs2cmt");
783 if (IS_ERR(osb->commit_task)) {
784 status = PTR_ERR(osb->commit_task);
785 osb->commit_task = NULL;
786 mlog(ML_ERROR, "unable to launch ocfs2commit thread, error=%d",
797 /* 'full' flag tells us whether we clear out all blocks or if we just
798 * mark the journal clean */
799 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
807 status = journal_wipe(journal->j_journal, full);
813 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
823 * JBD Might read a cached version of another nodes journal file. We
824 * don't want this as this file changes often and we get no
825 * notification on those changes. The only way to be sure that we've
826 * got the most up to date version of those blocks then is to force
827 * read them off disk. Just searching through the buffer cache won't
828 * work as there may be pages backing this file which are still marked
829 * up to date. We know things can't change on this file underneath us
830 * as we have the lock by now :)
832 static int ocfs2_force_read_journal(struct inode *inode)
836 u64 v_blkno, p_blkno;
837 #define CONCURRENT_JOURNAL_FILL 32
838 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
842 BUG_ON(inode->i_blocks !=
843 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
845 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
847 mlog(0, "Force reading %llu blocks\n",
848 (unsigned long long)(inode->i_blocks >>
849 (inode->i_sb->s_blocksize_bits - 9)));
853 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
855 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
863 if (p_blocks > CONCURRENT_JOURNAL_FILL)
864 p_blocks = CONCURRENT_JOURNAL_FILL;
866 /* We are reading journal data which should not
867 * be put in the uptodate cache */
868 status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
869 p_blkno, p_blocks, bhs, 0,
876 for(i = 0; i < p_blocks; i++) {
885 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
892 struct ocfs2_la_recovery_item {
893 struct list_head lri_list;
895 struct ocfs2_dinode *lri_la_dinode;
896 struct ocfs2_dinode *lri_tl_dinode;
899 /* Does the second half of the recovery process. By this point, the
900 * node is marked clean and can actually be considered recovered,
901 * hence it's no longer in the recovery map, but there's still some
902 * cleanup we can do which shouldn't happen within the recovery thread
903 * as locking in that context becomes very difficult if we are to take
904 * recovering nodes into account.
906 * NOTE: This function can and will sleep on recovery of other nodes
907 * during cluster locking, just like any other ocfs2 process.
909 void ocfs2_complete_recovery(void *data)
912 struct ocfs2_super *osb = data;
913 struct ocfs2_journal *journal = osb->journal;
914 struct ocfs2_dinode *la_dinode, *tl_dinode;
915 struct ocfs2_la_recovery_item *item;
916 struct list_head *p, *n;
917 LIST_HEAD(tmp_la_list);
921 mlog(0, "completing recovery from keventd\n");
923 spin_lock(&journal->j_lock);
924 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
925 spin_unlock(&journal->j_lock);
927 list_for_each_safe(p, n, &tmp_la_list) {
928 item = list_entry(p, struct ocfs2_la_recovery_item, lri_list);
929 list_del_init(&item->lri_list);
931 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
933 la_dinode = item->lri_la_dinode;
935 mlog(0, "Clean up local alloc %llu\n",
936 (unsigned long long)la_dinode->i_blkno);
938 ret = ocfs2_complete_local_alloc_recovery(osb,
946 tl_dinode = item->lri_tl_dinode;
948 mlog(0, "Clean up truncate log %llu\n",
949 (unsigned long long)tl_dinode->i_blkno);
951 ret = ocfs2_complete_truncate_log_recovery(osb,
959 ret = ocfs2_recover_orphans(osb, item->lri_slot);
966 mlog(0, "Recovery completion\n");
970 /* NOTE: This function always eats your references to la_dinode and
971 * tl_dinode, either manually on error, or by passing them to
972 * ocfs2_complete_recovery */
973 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
975 struct ocfs2_dinode *la_dinode,
976 struct ocfs2_dinode *tl_dinode)
978 struct ocfs2_la_recovery_item *item;
980 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
982 /* Though we wish to avoid it, we are in fact safe in
983 * skipping local alloc cleanup as fsck.ocfs2 is more
984 * than capable of reclaiming unused space. */
995 INIT_LIST_HEAD(&item->lri_list);
996 item->lri_la_dinode = la_dinode;
997 item->lri_slot = slot_num;
998 item->lri_tl_dinode = tl_dinode;
1000 spin_lock(&journal->j_lock);
1001 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1002 queue_work(ocfs2_wq, &journal->j_recovery_work);
1003 spin_unlock(&journal->j_lock);
1006 /* Called by the mount code to queue recovery the last part of
1007 * recovery for it's own slot. */
1008 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1010 struct ocfs2_journal *journal = osb->journal;
1013 /* No need to queue up our truncate_log as regular
1014 * cleanup will catch that. */
1015 ocfs2_queue_recovery_completion(journal,
1017 osb->local_alloc_copy,
1019 ocfs2_schedule_truncate_log_flush(osb, 0);
1021 osb->local_alloc_copy = NULL;
1026 static int __ocfs2_recovery_thread(void *arg)
1028 int status, node_num;
1029 struct ocfs2_super *osb = arg;
1033 status = ocfs2_wait_on_mount(osb);
1039 status = ocfs2_super_lock(osb, 1);
1045 while(!ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1046 node_num = ocfs2_node_map_first_set_bit(osb,
1047 &osb->recovery_map);
1048 if (node_num == O2NM_INVALID_NODE_NUM) {
1049 mlog(0, "Out of nodes to recover.\n");
1053 status = ocfs2_recover_node(osb, node_num);
1056 "Error %d recovering node %d on device (%u,%u)!\n",
1058 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1059 mlog(ML_ERROR, "Volume requires unmount.\n");
1063 ocfs2_recovery_map_clear(osb, node_num);
1065 ocfs2_super_unlock(osb, 1);
1067 /* We always run recovery on our own orphan dir - the dead
1068 * node(s) may have voted "no" on an inode delete earlier. A
1069 * revote is therefore required. */
1070 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1074 mutex_lock(&osb->recovery_lock);
1076 !ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1077 mutex_unlock(&osb->recovery_lock);
1081 osb->recovery_thread_task = NULL;
1082 mb(); /* sync with ocfs2_recovery_thread_running */
1083 wake_up(&osb->recovery_event);
1085 mutex_unlock(&osb->recovery_lock);
1088 /* no one is callint kthread_stop() for us so the kthread() api
1089 * requires that we call do_exit(). And it isn't exported, but
1090 * complete_and_exit() seems to be a minimal wrapper around it. */
1091 complete_and_exit(NULL, status);
1095 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1097 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1098 node_num, osb->node_num);
1100 mutex_lock(&osb->recovery_lock);
1101 if (osb->disable_recovery)
1104 /* People waiting on recovery will wait on
1105 * the recovery map to empty. */
1106 if (!ocfs2_recovery_map_set(osb, node_num))
1107 mlog(0, "node %d already be in recovery.\n", node_num);
1109 mlog(0, "starting recovery thread...\n");
1111 if (osb->recovery_thread_task)
1114 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1116 if (IS_ERR(osb->recovery_thread_task)) {
1117 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1118 osb->recovery_thread_task = NULL;
1122 mutex_unlock(&osb->recovery_lock);
1123 wake_up(&osb->recovery_event);
1128 /* Does the actual journal replay and marks the journal inode as
1129 * clean. Will only replay if the journal inode is marked dirty. */
1130 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1137 struct inode *inode = NULL;
1138 struct ocfs2_dinode *fe;
1139 journal_t *journal = NULL;
1140 struct buffer_head *bh = NULL;
1142 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1144 if (inode == NULL) {
1149 if (is_bad_inode(inode)) {
1156 SET_INODE_JOURNAL(inode);
1158 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
1159 OCFS2_META_LOCK_RECOVERY);
1161 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status);
1162 if (status != -ERESTARTSYS)
1163 mlog(ML_ERROR, "Could not lock journal!\n");
1168 fe = (struct ocfs2_dinode *) bh->b_data;
1170 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1172 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1173 mlog(0, "No recovery required for node %d\n", node_num);
1177 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1179 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1181 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1183 status = ocfs2_force_read_journal(inode);
1189 mlog(0, "calling journal_init_inode\n");
1190 journal = journal_init_inode(inode);
1191 if (journal == NULL) {
1192 mlog(ML_ERROR, "Linux journal layer error\n");
1197 status = journal_load(journal);
1202 journal_destroy(journal);
1206 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1208 /* wipe the journal */
1209 mlog(0, "flushing the journal.\n");
1210 journal_lock_updates(journal);
1211 status = journal_flush(journal);
1212 journal_unlock_updates(journal);
1216 /* This will mark the node clean */
1217 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1218 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1219 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1221 status = ocfs2_write_block(osb, bh, inode);
1228 journal_destroy(journal);
1231 /* drop the lock on this nodes journal */
1233 ocfs2_meta_unlock(inode, 1);
1246 * Do the most important parts of node recovery:
1247 * - Replay it's journal
1248 * - Stamp a clean local allocator file
1249 * - Stamp a clean truncate log
1250 * - Mark the node clean
1252 * If this function completes without error, a node in OCFS2 can be
1253 * said to have been safely recovered. As a result, failure during the
1254 * second part of a nodes recovery process (local alloc recovery) is
1255 * far less concerning.
1257 static int ocfs2_recover_node(struct ocfs2_super *osb,
1262 struct ocfs2_slot_info *si = osb->slot_info;
1263 struct ocfs2_dinode *la_copy = NULL;
1264 struct ocfs2_dinode *tl_copy = NULL;
1266 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1267 node_num, osb->node_num);
1269 mlog(0, "checking node %d\n", node_num);
1271 /* Should not ever be called to recover ourselves -- in that
1272 * case we should've called ocfs2_journal_load instead. */
1273 BUG_ON(osb->node_num == node_num);
1275 slot_num = ocfs2_node_num_to_slot(si, node_num);
1276 if (slot_num == OCFS2_INVALID_SLOT) {
1278 mlog(0, "no slot for this node, so no recovery required.\n");
1282 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1284 status = ocfs2_replay_journal(osb, node_num, slot_num);
1290 /* Stamp a clean local alloc file AFTER recovering the journal... */
1291 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1297 /* An error from begin_truncate_log_recovery is not
1298 * serious enough to warrant halting the rest of
1300 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1304 /* Likewise, this would be a strange but ultimately not so
1305 * harmful place to get an error... */
1306 ocfs2_clear_slot(si, slot_num);
1307 status = ocfs2_update_disk_slots(osb, si);
1311 /* This will kfree the memory pointed to by la_copy and tl_copy */
1312 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1322 /* Test node liveness by trylocking his journal. If we get the lock,
1323 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1324 * still alive (we couldn't get the lock) and < 0 on error. */
1325 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1329 struct inode *inode = NULL;
1331 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1333 if (inode == NULL) {
1334 mlog(ML_ERROR, "access error\n");
1338 if (is_bad_inode(inode)) {
1339 mlog(ML_ERROR, "access error (bad inode)\n");
1345 SET_INODE_JOURNAL(inode);
1347 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1348 status = ocfs2_meta_lock_full(inode, NULL, NULL, 1, flags);
1350 if (status != -EAGAIN)
1355 ocfs2_meta_unlock(inode, 1);
1363 /* Call this underneath ocfs2_super_lock. It also assumes that the
1364 * slot info struct has been updated from disk. */
1365 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1367 int status, i, node_num;
1368 struct ocfs2_slot_info *si = osb->slot_info;
1370 /* This is called with the super block cluster lock, so we
1371 * know that the slot map can't change underneath us. */
1373 spin_lock(&si->si_lock);
1374 for(i = 0; i < si->si_num_slots; i++) {
1375 if (i == osb->slot_num)
1377 if (ocfs2_is_empty_slot(si, i))
1380 node_num = si->si_global_node_nums[i];
1381 if (ocfs2_node_map_test_bit(osb, &osb->recovery_map, node_num))
1383 spin_unlock(&si->si_lock);
1385 /* Ok, we have a slot occupied by another node which
1386 * is not in the recovery map. We trylock his journal
1387 * file here to test if he's alive. */
1388 status = ocfs2_trylock_journal(osb, i);
1390 /* Since we're called from mount, we know that
1391 * the recovery thread can't race us on
1392 * setting / checking the recovery bits. */
1393 ocfs2_recovery_thread(osb, node_num);
1394 } else if ((status < 0) && (status != -EAGAIN)) {
1399 spin_lock(&si->si_lock);
1401 spin_unlock(&si->si_lock);
1409 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1411 struct inode **head)
1414 struct inode *orphan_dir_inode = NULL;
1416 unsigned long offset, blk, local;
1417 struct buffer_head *bh = NULL;
1418 struct ocfs2_dir_entry *de;
1419 struct super_block *sb = osb->sb;
1421 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1422 ORPHAN_DIR_SYSTEM_INODE,
1424 if (!orphan_dir_inode) {
1430 mutex_lock(&orphan_dir_inode->i_mutex);
1431 status = ocfs2_meta_lock(orphan_dir_inode, NULL, NULL, 0);
1439 while(offset < i_size_read(orphan_dir_inode)) {
1440 blk = offset >> sb->s_blocksize_bits;
1442 bh = ocfs2_bread(orphan_dir_inode, blk, &status, 0);
1453 while(offset < i_size_read(orphan_dir_inode)
1454 && local < sb->s_blocksize) {
1455 de = (struct ocfs2_dir_entry *) (bh->b_data + local);
1457 if (!ocfs2_check_dir_entry(orphan_dir_inode,
1465 local += le16_to_cpu(de->rec_len);
1466 offset += le16_to_cpu(de->rec_len);
1468 /* I guess we silently fail on no inode? */
1469 if (!le64_to_cpu(de->inode))
1471 if (de->file_type > OCFS2_FT_MAX) {
1473 "block %llu contains invalid de: "
1474 "inode = %llu, rec_len = %u, "
1475 "name_len = %u, file_type = %u, "
1477 (unsigned long long)bh->b_blocknr,
1478 (unsigned long long)le64_to_cpu(de->inode),
1479 le16_to_cpu(de->rec_len),
1486 if (de->name_len == 1 && !strncmp(".", de->name, 1))
1488 if (de->name_len == 2 && !strncmp("..", de->name, 2))
1491 iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
1492 OCFS2_FI_FLAG_NOLOCK);
1496 mlog(0, "queue orphan %llu\n",
1497 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1498 /* No locking is required for the next_orphan
1499 * queue as there is only ever a single
1500 * process doing orphan recovery. */
1501 OCFS2_I(iter)->ip_next_orphan = *head;
1508 ocfs2_meta_unlock(orphan_dir_inode, 0);
1510 mutex_unlock(&orphan_dir_inode->i_mutex);
1511 iput(orphan_dir_inode);
1515 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1520 spin_lock(&osb->osb_lock);
1521 ret = !osb->osb_orphan_wipes[slot];
1522 spin_unlock(&osb->osb_lock);
1526 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1529 spin_lock(&osb->osb_lock);
1530 /* Mark ourselves such that new processes in delete_inode()
1531 * know to quit early. */
1532 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1533 while (osb->osb_orphan_wipes[slot]) {
1534 /* If any processes are already in the middle of an
1535 * orphan wipe on this dir, then we need to wait for
1537 spin_unlock(&osb->osb_lock);
1538 wait_event_interruptible(osb->osb_wipe_event,
1539 ocfs2_orphan_recovery_can_continue(osb, slot));
1540 spin_lock(&osb->osb_lock);
1542 spin_unlock(&osb->osb_lock);
1545 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1548 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1552 * Orphan recovery. Each mounted node has it's own orphan dir which we
1553 * must run during recovery. Our strategy here is to build a list of
1554 * the inodes in the orphan dir and iget/iput them. The VFS does
1555 * (most) of the rest of the work.
1557 * Orphan recovery can happen at any time, not just mount so we have a
1558 * couple of extra considerations.
1560 * - We grab as many inodes as we can under the orphan dir lock -
1561 * doing iget() outside the orphan dir risks getting a reference on
1563 * - We must be sure not to deadlock with other processes on the
1564 * system wanting to run delete_inode(). This can happen when they go
1565 * to lock the orphan dir and the orphan recovery process attempts to
1566 * iget() inside the orphan dir lock. This can be avoided by
1567 * advertising our state to ocfs2_delete_inode().
1569 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1573 struct inode *inode = NULL;
1575 struct ocfs2_inode_info *oi;
1577 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1579 ocfs2_mark_recovering_orphan_dir(osb, slot);
1580 ret = ocfs2_queue_orphans(osb, slot, &inode);
1581 ocfs2_clear_recovering_orphan_dir(osb, slot);
1583 /* Error here should be noted, but we want to continue with as
1584 * many queued inodes as we've got. */
1589 oi = OCFS2_I(inode);
1590 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1592 iter = oi->ip_next_orphan;
1594 spin_lock(&oi->ip_lock);
1595 /* Delete voting may have set these on the assumption
1596 * that the other node would wipe them successfully.
1597 * If they are still in the node's orphan dir, we need
1598 * to reset that state. */
1599 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1601 /* Set the proper information to get us going into
1602 * ocfs2_delete_inode. */
1603 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1604 oi->ip_orphaned_slot = slot;
1605 spin_unlock(&oi->ip_lock);
1615 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1617 /* This check is good because ocfs2 will wait on our recovery
1618 * thread before changing it to something other than MOUNTED
1620 wait_event(osb->osb_mount_event,
1621 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1622 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1624 /* If there's an error on mount, then we may never get to the
1625 * MOUNTED flag, but this is set right before
1626 * dismount_volume() so we can trust it. */
1627 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1628 mlog(0, "mount error, exiting!\n");
1635 static int ocfs2_commit_thread(void *arg)
1638 struct ocfs2_super *osb = arg;
1639 struct ocfs2_journal *journal = osb->journal;
1641 /* we can trust j_num_trans here because _should_stop() is only set in
1642 * shutdown and nobody other than ourselves should be able to start
1643 * transactions. committing on shutdown might take a few iterations
1644 * as final transactions put deleted inodes on the list */
1645 while (!(kthread_should_stop() &&
1646 atomic_read(&journal->j_num_trans) == 0)) {
1648 wait_event_interruptible(osb->checkpoint_event,
1649 atomic_read(&journal->j_num_trans)
1650 || kthread_should_stop());
1652 status = ocfs2_commit_cache(osb);
1656 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1658 "commit_thread: %u transactions pending on "
1660 atomic_read(&journal->j_num_trans));
1667 /* Look for a dirty journal without taking any cluster locks. Used for
1668 * hard readonly access to determine whether the file system journals
1669 * require recovery. */
1670 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1674 struct buffer_head *di_bh;
1675 struct ocfs2_dinode *di;
1676 struct inode *journal = NULL;
1678 for(slot = 0; slot < osb->max_slots; slot++) {
1679 journal = ocfs2_get_system_file_inode(osb,
1680 JOURNAL_SYSTEM_INODE,
1682 if (!journal || is_bad_inode(journal)) {
1689 ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1696 di = (struct ocfs2_dinode *) di_bh->b_data;
1698 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1699 OCFS2_JOURNAL_DIRTY_FL)