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(handle_t *handle, int nblocks)
335 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
337 status = journal_extend(handle, nblocks);
344 mlog(0, "journal_extend failed, trying journal_restart\n");
345 status = journal_restart(handle, nblocks);
359 int ocfs2_journal_access(struct ocfs2_journal_handle *handle,
361 struct buffer_head *bh,
369 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
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 = journal_get_write_access(handle->k_handle, bh);
401 case OCFS2_JOURNAL_ACCESS_UNDO:
402 status = journal_get_undo_access(handle->k_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(struct ocfs2_journal_handle *handle,
420 struct buffer_head *bh)
424 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
426 mlog_entry("(bh->b_blocknr=%llu)\n",
427 (unsigned long long)bh->b_blocknr);
429 status = journal_dirty_metadata(handle->k_handle, bh);
431 mlog(ML_ERROR, "Could not dirty metadata buffer. "
432 "(bh->b_blocknr=%llu)\n",
433 (unsigned long long)bh->b_blocknr);
439 int ocfs2_journal_dirty_data(handle_t *handle,
440 struct buffer_head *bh)
442 int err = journal_dirty_data(handle, bh);
445 /* TODO: When we can handle it, abort the handle and go RO on
451 /* We always assume you're adding a metadata lock at level 'ex' */
452 int ocfs2_handle_add_lock(struct ocfs2_journal_handle *handle,
456 struct ocfs2_journal_lock *lock;
460 lock = kmem_cache_alloc(ocfs2_lock_cache, GFP_NOFS);
469 lock->jl_inode = inode;
471 list_add_tail(&(lock->jl_lock_list), &(handle->locks));
480 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal *journal,
481 struct ocfs2_journal_handle *handle)
483 struct list_head *p, *n;
484 struct ocfs2_journal_lock *lock;
487 list_for_each_safe(p, n, &(handle->locks)) {
488 lock = list_entry(p, struct ocfs2_journal_lock,
490 list_del(&lock->jl_lock_list);
493 inode = lock->jl_inode;
494 ocfs2_meta_unlock(inode, 1);
495 if (atomic_read(&inode->i_count) == 1)
497 "Inode %llu, I'm doing a last iput for!",
498 (unsigned long long)OCFS2_I(inode)->ip_blkno);
500 kmem_cache_free(ocfs2_lock_cache, lock);
504 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
506 void ocfs2_set_journal_params(struct ocfs2_super *osb)
508 journal_t *journal = osb->journal->j_journal;
510 spin_lock(&journal->j_state_lock);
511 journal->j_commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
512 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
513 journal->j_flags |= JFS_BARRIER;
515 journal->j_flags &= ~JFS_BARRIER;
516 spin_unlock(&journal->j_state_lock);
519 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
522 struct inode *inode = NULL; /* the journal inode */
523 journal_t *j_journal = NULL;
524 struct ocfs2_dinode *di = NULL;
525 struct buffer_head *bh = NULL;
526 struct ocfs2_super *osb;
533 osb = journal->j_osb;
535 /* already have the inode for our journal */
536 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
543 if (is_bad_inode(inode)) {
544 mlog(ML_ERROR, "access error (bad inode)\n");
551 SET_INODE_JOURNAL(inode);
552 OCFS2_I(inode)->ip_open_count++;
554 /* Skip recovery waits here - journal inode metadata never
555 * changes in a live cluster so it can be considered an
556 * exception to the rule. */
557 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
558 OCFS2_META_LOCK_RECOVERY);
560 if (status != -ERESTARTSYS)
561 mlog(ML_ERROR, "Could not get lock on journal!\n");
566 di = (struct ocfs2_dinode *)bh->b_data;
568 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
569 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
575 mlog(0, "inode->i_size = %lld\n", inode->i_size);
576 mlog(0, "inode->i_blocks = %llu\n",
577 (unsigned long long)inode->i_blocks);
578 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
580 /* call the kernels journal init function now */
581 j_journal = journal_init_inode(inode);
582 if (j_journal == NULL) {
583 mlog(ML_ERROR, "Linux journal layer error\n");
588 mlog(0, "Returned from journal_init_inode\n");
589 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
591 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
592 OCFS2_JOURNAL_DIRTY_FL);
594 journal->j_journal = j_journal;
595 journal->j_inode = inode;
598 ocfs2_set_journal_params(osb);
600 journal->j_state = OCFS2_JOURNAL_LOADED;
606 ocfs2_meta_unlock(inode, 1);
610 OCFS2_I(inode)->ip_open_count--;
619 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
624 struct ocfs2_journal *journal = osb->journal;
625 struct buffer_head *bh = journal->j_bh;
626 struct ocfs2_dinode *fe;
630 fe = (struct ocfs2_dinode *)bh->b_data;
631 if (!OCFS2_IS_VALID_DINODE(fe)) {
632 /* This is called from startup/shutdown which will
633 * handle the errors in a specific manner, so no need
634 * to call ocfs2_error() here. */
635 mlog(ML_ERROR, "Journal dinode %llu has invalid "
636 "signature: %.*s", (unsigned long long)fe->i_blkno, 7,
642 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
644 flags |= OCFS2_JOURNAL_DIRTY_FL;
646 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
647 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
649 status = ocfs2_write_block(osb, bh, journal->j_inode);
659 * If the journal has been kmalloc'd it needs to be freed after this
662 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
664 struct ocfs2_journal *journal = NULL;
666 struct inode *inode = NULL;
667 int num_running_trans = 0;
673 journal = osb->journal;
677 inode = journal->j_inode;
679 if (journal->j_state != OCFS2_JOURNAL_LOADED)
682 /* need to inc inode use count as journal_destroy will iput. */
686 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
687 if (num_running_trans > 0)
688 mlog(0, "Shutting down journal: must wait on %d "
689 "running transactions!\n",
692 /* Do a commit_cache here. It will flush our journal, *and*
693 * release any locks that are still held.
694 * set the SHUTDOWN flag and release the trans lock.
695 * the commit thread will take the trans lock for us below. */
696 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
698 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
699 * drop the trans_lock (which we want to hold until we
700 * completely destroy the journal. */
701 if (osb->commit_task) {
702 /* Wait for the commit thread */
703 mlog(0, "Waiting for ocfs2commit to exit....\n");
704 kthread_stop(osb->commit_task);
705 osb->commit_task = NULL;
708 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
710 status = ocfs2_journal_toggle_dirty(osb, 0);
714 /* Shutdown the kernel journal system */
715 journal_destroy(journal->j_journal);
717 OCFS2_I(inode)->ip_open_count--;
719 /* unlock our journal */
720 ocfs2_meta_unlock(inode, 1);
722 brelse(journal->j_bh);
723 journal->j_bh = NULL;
725 journal->j_state = OCFS2_JOURNAL_FREE;
727 // up_write(&journal->j_trans_barrier);
734 static void ocfs2_clear_journal_error(struct super_block *sb,
740 olderr = journal_errno(journal);
742 mlog(ML_ERROR, "File system error %d recorded in "
743 "journal %u.\n", olderr, slot);
744 mlog(ML_ERROR, "File system on device %s needs checking.\n",
747 journal_ack_err(journal);
748 journal_clear_err(journal);
752 int ocfs2_journal_load(struct ocfs2_journal *journal)
755 struct ocfs2_super *osb;
762 osb = journal->j_osb;
764 status = journal_load(journal->j_journal);
766 mlog(ML_ERROR, "Failed to load journal!\n");
770 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
772 status = ocfs2_journal_toggle_dirty(osb, 1);
778 /* Launch the commit thread */
779 osb->commit_task = kthread_run(ocfs2_commit_thread, osb, "ocfs2cmt");
780 if (IS_ERR(osb->commit_task)) {
781 status = PTR_ERR(osb->commit_task);
782 osb->commit_task = NULL;
783 mlog(ML_ERROR, "unable to launch ocfs2commit thread, error=%d",
794 /* 'full' flag tells us whether we clear out all blocks or if we just
795 * mark the journal clean */
796 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
804 status = journal_wipe(journal->j_journal, full);
810 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
820 * JBD Might read a cached version of another nodes journal file. We
821 * don't want this as this file changes often and we get no
822 * notification on those changes. The only way to be sure that we've
823 * got the most up to date version of those blocks then is to force
824 * read them off disk. Just searching through the buffer cache won't
825 * work as there may be pages backing this file which are still marked
826 * up to date. We know things can't change on this file underneath us
827 * as we have the lock by now :)
829 static int ocfs2_force_read_journal(struct inode *inode)
833 u64 v_blkno, p_blkno;
834 #define CONCURRENT_JOURNAL_FILL 32
835 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
839 BUG_ON(inode->i_blocks !=
840 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
842 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
844 mlog(0, "Force reading %llu blocks\n",
845 (unsigned long long)(inode->i_blocks >>
846 (inode->i_sb->s_blocksize_bits - 9)));
850 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
852 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
860 if (p_blocks > CONCURRENT_JOURNAL_FILL)
861 p_blocks = CONCURRENT_JOURNAL_FILL;
863 /* We are reading journal data which should not
864 * be put in the uptodate cache */
865 status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
866 p_blkno, p_blocks, bhs, 0,
873 for(i = 0; i < p_blocks; i++) {
882 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
889 struct ocfs2_la_recovery_item {
890 struct list_head lri_list;
892 struct ocfs2_dinode *lri_la_dinode;
893 struct ocfs2_dinode *lri_tl_dinode;
896 /* Does the second half of the recovery process. By this point, the
897 * node is marked clean and can actually be considered recovered,
898 * hence it's no longer in the recovery map, but there's still some
899 * cleanup we can do which shouldn't happen within the recovery thread
900 * as locking in that context becomes very difficult if we are to take
901 * recovering nodes into account.
903 * NOTE: This function can and will sleep on recovery of other nodes
904 * during cluster locking, just like any other ocfs2 process.
906 void ocfs2_complete_recovery(void *data)
909 struct ocfs2_super *osb = data;
910 struct ocfs2_journal *journal = osb->journal;
911 struct ocfs2_dinode *la_dinode, *tl_dinode;
912 struct ocfs2_la_recovery_item *item;
913 struct list_head *p, *n;
914 LIST_HEAD(tmp_la_list);
918 mlog(0, "completing recovery from keventd\n");
920 spin_lock(&journal->j_lock);
921 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
922 spin_unlock(&journal->j_lock);
924 list_for_each_safe(p, n, &tmp_la_list) {
925 item = list_entry(p, struct ocfs2_la_recovery_item, lri_list);
926 list_del_init(&item->lri_list);
928 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
930 la_dinode = item->lri_la_dinode;
932 mlog(0, "Clean up local alloc %llu\n",
933 (unsigned long long)la_dinode->i_blkno);
935 ret = ocfs2_complete_local_alloc_recovery(osb,
943 tl_dinode = item->lri_tl_dinode;
945 mlog(0, "Clean up truncate log %llu\n",
946 (unsigned long long)tl_dinode->i_blkno);
948 ret = ocfs2_complete_truncate_log_recovery(osb,
956 ret = ocfs2_recover_orphans(osb, item->lri_slot);
963 mlog(0, "Recovery completion\n");
967 /* NOTE: This function always eats your references to la_dinode and
968 * tl_dinode, either manually on error, or by passing them to
969 * ocfs2_complete_recovery */
970 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
972 struct ocfs2_dinode *la_dinode,
973 struct ocfs2_dinode *tl_dinode)
975 struct ocfs2_la_recovery_item *item;
977 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
979 /* Though we wish to avoid it, we are in fact safe in
980 * skipping local alloc cleanup as fsck.ocfs2 is more
981 * than capable of reclaiming unused space. */
992 INIT_LIST_HEAD(&item->lri_list);
993 item->lri_la_dinode = la_dinode;
994 item->lri_slot = slot_num;
995 item->lri_tl_dinode = tl_dinode;
997 spin_lock(&journal->j_lock);
998 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
999 queue_work(ocfs2_wq, &journal->j_recovery_work);
1000 spin_unlock(&journal->j_lock);
1003 /* Called by the mount code to queue recovery the last part of
1004 * recovery for it's own slot. */
1005 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1007 struct ocfs2_journal *journal = osb->journal;
1010 /* No need to queue up our truncate_log as regular
1011 * cleanup will catch that. */
1012 ocfs2_queue_recovery_completion(journal,
1014 osb->local_alloc_copy,
1016 ocfs2_schedule_truncate_log_flush(osb, 0);
1018 osb->local_alloc_copy = NULL;
1023 static int __ocfs2_recovery_thread(void *arg)
1025 int status, node_num;
1026 struct ocfs2_super *osb = arg;
1030 status = ocfs2_wait_on_mount(osb);
1036 status = ocfs2_super_lock(osb, 1);
1042 while(!ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1043 node_num = ocfs2_node_map_first_set_bit(osb,
1044 &osb->recovery_map);
1045 if (node_num == O2NM_INVALID_NODE_NUM) {
1046 mlog(0, "Out of nodes to recover.\n");
1050 status = ocfs2_recover_node(osb, node_num);
1053 "Error %d recovering node %d on device (%u,%u)!\n",
1055 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1056 mlog(ML_ERROR, "Volume requires unmount.\n");
1060 ocfs2_recovery_map_clear(osb, node_num);
1062 ocfs2_super_unlock(osb, 1);
1064 /* We always run recovery on our own orphan dir - the dead
1065 * node(s) may have voted "no" on an inode delete earlier. A
1066 * revote is therefore required. */
1067 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1071 mutex_lock(&osb->recovery_lock);
1073 !ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1074 mutex_unlock(&osb->recovery_lock);
1078 osb->recovery_thread_task = NULL;
1079 mb(); /* sync with ocfs2_recovery_thread_running */
1080 wake_up(&osb->recovery_event);
1082 mutex_unlock(&osb->recovery_lock);
1085 /* no one is callint kthread_stop() for us so the kthread() api
1086 * requires that we call do_exit(). And it isn't exported, but
1087 * complete_and_exit() seems to be a minimal wrapper around it. */
1088 complete_and_exit(NULL, status);
1092 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1094 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1095 node_num, osb->node_num);
1097 mutex_lock(&osb->recovery_lock);
1098 if (osb->disable_recovery)
1101 /* People waiting on recovery will wait on
1102 * the recovery map to empty. */
1103 if (!ocfs2_recovery_map_set(osb, node_num))
1104 mlog(0, "node %d already be in recovery.\n", node_num);
1106 mlog(0, "starting recovery thread...\n");
1108 if (osb->recovery_thread_task)
1111 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1113 if (IS_ERR(osb->recovery_thread_task)) {
1114 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1115 osb->recovery_thread_task = NULL;
1119 mutex_unlock(&osb->recovery_lock);
1120 wake_up(&osb->recovery_event);
1125 /* Does the actual journal replay and marks the journal inode as
1126 * clean. Will only replay if the journal inode is marked dirty. */
1127 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1134 struct inode *inode = NULL;
1135 struct ocfs2_dinode *fe;
1136 journal_t *journal = NULL;
1137 struct buffer_head *bh = NULL;
1139 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1141 if (inode == NULL) {
1146 if (is_bad_inode(inode)) {
1153 SET_INODE_JOURNAL(inode);
1155 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
1156 OCFS2_META_LOCK_RECOVERY);
1158 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status);
1159 if (status != -ERESTARTSYS)
1160 mlog(ML_ERROR, "Could not lock journal!\n");
1165 fe = (struct ocfs2_dinode *) bh->b_data;
1167 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1169 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1170 mlog(0, "No recovery required for node %d\n", node_num);
1174 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1176 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1178 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1180 status = ocfs2_force_read_journal(inode);
1186 mlog(0, "calling journal_init_inode\n");
1187 journal = journal_init_inode(inode);
1188 if (journal == NULL) {
1189 mlog(ML_ERROR, "Linux journal layer error\n");
1194 status = journal_load(journal);
1199 journal_destroy(journal);
1203 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1205 /* wipe the journal */
1206 mlog(0, "flushing the journal.\n");
1207 journal_lock_updates(journal);
1208 status = journal_flush(journal);
1209 journal_unlock_updates(journal);
1213 /* This will mark the node clean */
1214 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1215 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1216 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1218 status = ocfs2_write_block(osb, bh, inode);
1225 journal_destroy(journal);
1228 /* drop the lock on this nodes journal */
1230 ocfs2_meta_unlock(inode, 1);
1243 * Do the most important parts of node recovery:
1244 * - Replay it's journal
1245 * - Stamp a clean local allocator file
1246 * - Stamp a clean truncate log
1247 * - Mark the node clean
1249 * If this function completes without error, a node in OCFS2 can be
1250 * said to have been safely recovered. As a result, failure during the
1251 * second part of a nodes recovery process (local alloc recovery) is
1252 * far less concerning.
1254 static int ocfs2_recover_node(struct ocfs2_super *osb,
1259 struct ocfs2_slot_info *si = osb->slot_info;
1260 struct ocfs2_dinode *la_copy = NULL;
1261 struct ocfs2_dinode *tl_copy = NULL;
1263 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1264 node_num, osb->node_num);
1266 mlog(0, "checking node %d\n", node_num);
1268 /* Should not ever be called to recover ourselves -- in that
1269 * case we should've called ocfs2_journal_load instead. */
1270 BUG_ON(osb->node_num == node_num);
1272 slot_num = ocfs2_node_num_to_slot(si, node_num);
1273 if (slot_num == OCFS2_INVALID_SLOT) {
1275 mlog(0, "no slot for this node, so no recovery required.\n");
1279 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1281 status = ocfs2_replay_journal(osb, node_num, slot_num);
1287 /* Stamp a clean local alloc file AFTER recovering the journal... */
1288 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1294 /* An error from begin_truncate_log_recovery is not
1295 * serious enough to warrant halting the rest of
1297 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1301 /* Likewise, this would be a strange but ultimately not so
1302 * harmful place to get an error... */
1303 ocfs2_clear_slot(si, slot_num);
1304 status = ocfs2_update_disk_slots(osb, si);
1308 /* This will kfree the memory pointed to by la_copy and tl_copy */
1309 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1319 /* Test node liveness by trylocking his journal. If we get the lock,
1320 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1321 * still alive (we couldn't get the lock) and < 0 on error. */
1322 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1326 struct inode *inode = NULL;
1328 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1330 if (inode == NULL) {
1331 mlog(ML_ERROR, "access error\n");
1335 if (is_bad_inode(inode)) {
1336 mlog(ML_ERROR, "access error (bad inode)\n");
1342 SET_INODE_JOURNAL(inode);
1344 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1345 status = ocfs2_meta_lock_full(inode, NULL, NULL, 1, flags);
1347 if (status != -EAGAIN)
1352 ocfs2_meta_unlock(inode, 1);
1360 /* Call this underneath ocfs2_super_lock. It also assumes that the
1361 * slot info struct has been updated from disk. */
1362 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1364 int status, i, node_num;
1365 struct ocfs2_slot_info *si = osb->slot_info;
1367 /* This is called with the super block cluster lock, so we
1368 * know that the slot map can't change underneath us. */
1370 spin_lock(&si->si_lock);
1371 for(i = 0; i < si->si_num_slots; i++) {
1372 if (i == osb->slot_num)
1374 if (ocfs2_is_empty_slot(si, i))
1377 node_num = si->si_global_node_nums[i];
1378 if (ocfs2_node_map_test_bit(osb, &osb->recovery_map, node_num))
1380 spin_unlock(&si->si_lock);
1382 /* Ok, we have a slot occupied by another node which
1383 * is not in the recovery map. We trylock his journal
1384 * file here to test if he's alive. */
1385 status = ocfs2_trylock_journal(osb, i);
1387 /* Since we're called from mount, we know that
1388 * the recovery thread can't race us on
1389 * setting / checking the recovery bits. */
1390 ocfs2_recovery_thread(osb, node_num);
1391 } else if ((status < 0) && (status != -EAGAIN)) {
1396 spin_lock(&si->si_lock);
1398 spin_unlock(&si->si_lock);
1406 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1408 struct inode **head)
1411 struct inode *orphan_dir_inode = NULL;
1413 unsigned long offset, blk, local;
1414 struct buffer_head *bh = NULL;
1415 struct ocfs2_dir_entry *de;
1416 struct super_block *sb = osb->sb;
1418 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1419 ORPHAN_DIR_SYSTEM_INODE,
1421 if (!orphan_dir_inode) {
1427 mutex_lock(&orphan_dir_inode->i_mutex);
1428 status = ocfs2_meta_lock(orphan_dir_inode, NULL, NULL, 0);
1436 while(offset < i_size_read(orphan_dir_inode)) {
1437 blk = offset >> sb->s_blocksize_bits;
1439 bh = ocfs2_bread(orphan_dir_inode, blk, &status, 0);
1450 while(offset < i_size_read(orphan_dir_inode)
1451 && local < sb->s_blocksize) {
1452 de = (struct ocfs2_dir_entry *) (bh->b_data + local);
1454 if (!ocfs2_check_dir_entry(orphan_dir_inode,
1462 local += le16_to_cpu(de->rec_len);
1463 offset += le16_to_cpu(de->rec_len);
1465 /* I guess we silently fail on no inode? */
1466 if (!le64_to_cpu(de->inode))
1468 if (de->file_type > OCFS2_FT_MAX) {
1470 "block %llu contains invalid de: "
1471 "inode = %llu, rec_len = %u, "
1472 "name_len = %u, file_type = %u, "
1474 (unsigned long long)bh->b_blocknr,
1475 (unsigned long long)le64_to_cpu(de->inode),
1476 le16_to_cpu(de->rec_len),
1483 if (de->name_len == 1 && !strncmp(".", de->name, 1))
1485 if (de->name_len == 2 && !strncmp("..", de->name, 2))
1488 iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
1489 OCFS2_FI_FLAG_NOLOCK);
1493 mlog(0, "queue orphan %llu\n",
1494 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1495 /* No locking is required for the next_orphan
1496 * queue as there is only ever a single
1497 * process doing orphan recovery. */
1498 OCFS2_I(iter)->ip_next_orphan = *head;
1505 ocfs2_meta_unlock(orphan_dir_inode, 0);
1507 mutex_unlock(&orphan_dir_inode->i_mutex);
1508 iput(orphan_dir_inode);
1512 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1517 spin_lock(&osb->osb_lock);
1518 ret = !osb->osb_orphan_wipes[slot];
1519 spin_unlock(&osb->osb_lock);
1523 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1526 spin_lock(&osb->osb_lock);
1527 /* Mark ourselves such that new processes in delete_inode()
1528 * know to quit early. */
1529 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1530 while (osb->osb_orphan_wipes[slot]) {
1531 /* If any processes are already in the middle of an
1532 * orphan wipe on this dir, then we need to wait for
1534 spin_unlock(&osb->osb_lock);
1535 wait_event_interruptible(osb->osb_wipe_event,
1536 ocfs2_orphan_recovery_can_continue(osb, slot));
1537 spin_lock(&osb->osb_lock);
1539 spin_unlock(&osb->osb_lock);
1542 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1545 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1549 * Orphan recovery. Each mounted node has it's own orphan dir which we
1550 * must run during recovery. Our strategy here is to build a list of
1551 * the inodes in the orphan dir and iget/iput them. The VFS does
1552 * (most) of the rest of the work.
1554 * Orphan recovery can happen at any time, not just mount so we have a
1555 * couple of extra considerations.
1557 * - We grab as many inodes as we can under the orphan dir lock -
1558 * doing iget() outside the orphan dir risks getting a reference on
1560 * - We must be sure not to deadlock with other processes on the
1561 * system wanting to run delete_inode(). This can happen when they go
1562 * to lock the orphan dir and the orphan recovery process attempts to
1563 * iget() inside the orphan dir lock. This can be avoided by
1564 * advertising our state to ocfs2_delete_inode().
1566 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1570 struct inode *inode = NULL;
1572 struct ocfs2_inode_info *oi;
1574 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1576 ocfs2_mark_recovering_orphan_dir(osb, slot);
1577 ret = ocfs2_queue_orphans(osb, slot, &inode);
1578 ocfs2_clear_recovering_orphan_dir(osb, slot);
1580 /* Error here should be noted, but we want to continue with as
1581 * many queued inodes as we've got. */
1586 oi = OCFS2_I(inode);
1587 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1589 iter = oi->ip_next_orphan;
1591 spin_lock(&oi->ip_lock);
1592 /* Delete voting may have set these on the assumption
1593 * that the other node would wipe them successfully.
1594 * If they are still in the node's orphan dir, we need
1595 * to reset that state. */
1596 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1598 /* Set the proper information to get us going into
1599 * ocfs2_delete_inode. */
1600 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1601 oi->ip_orphaned_slot = slot;
1602 spin_unlock(&oi->ip_lock);
1612 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1614 /* This check is good because ocfs2 will wait on our recovery
1615 * thread before changing it to something other than MOUNTED
1617 wait_event(osb->osb_mount_event,
1618 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1619 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1621 /* If there's an error on mount, then we may never get to the
1622 * MOUNTED flag, but this is set right before
1623 * dismount_volume() so we can trust it. */
1624 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1625 mlog(0, "mount error, exiting!\n");
1632 static int ocfs2_commit_thread(void *arg)
1635 struct ocfs2_super *osb = arg;
1636 struct ocfs2_journal *journal = osb->journal;
1638 /* we can trust j_num_trans here because _should_stop() is only set in
1639 * shutdown and nobody other than ourselves should be able to start
1640 * transactions. committing on shutdown might take a few iterations
1641 * as final transactions put deleted inodes on the list */
1642 while (!(kthread_should_stop() &&
1643 atomic_read(&journal->j_num_trans) == 0)) {
1645 wait_event_interruptible(osb->checkpoint_event,
1646 atomic_read(&journal->j_num_trans)
1647 || kthread_should_stop());
1649 status = ocfs2_commit_cache(osb);
1653 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1655 "commit_thread: %u transactions pending on "
1657 atomic_read(&journal->j_num_trans));
1664 /* Look for a dirty journal without taking any cluster locks. Used for
1665 * hard readonly access to determine whether the file system journals
1666 * require recovery. */
1667 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1671 struct buffer_head *di_bh;
1672 struct ocfs2_dinode *di;
1673 struct inode *journal = NULL;
1675 for(slot = 0; slot < osb->max_slots; slot++) {
1676 journal = ocfs2_get_system_file_inode(osb,
1677 JOURNAL_SYSTEM_INODE,
1679 if (!journal || is_bad_inode(journal)) {
1686 ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1693 di = (struct ocfs2_dinode *) di_bh->b_data;
1695 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1696 OCFS2_JOURNAL_DIRTY_FL)