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 spinlock_t trans_inc_lock = SPIN_LOCK_UNLOCKED;
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->max_buffs = 0;
128 retval->num_locks = 0;
129 retval->k_handle = NULL;
131 INIT_LIST_HEAD(&retval->locks);
132 INIT_LIST_HEAD(&retval->inode_list);
133 retval->journal = osb->journal;
138 /* pass it NULL and it will allocate a new handle object for you. If
139 * you pass it a handle however, it may still return error, in which
140 * case it has free'd the passed handle for you. */
141 struct ocfs2_journal_handle *ocfs2_start_trans(struct ocfs2_super *osb,
142 struct ocfs2_journal_handle *handle,
146 journal_t *journal = osb->journal->j_journal;
148 mlog_entry("(max_buffs = %d)\n", max_buffs);
150 BUG_ON(!osb || !osb->journal->j_journal);
152 if (ocfs2_is_hard_readonly(osb)) {
157 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
158 BUG_ON(max_buffs <= 0);
160 /* JBD might support this, but our journalling code doesn't yet. */
161 if (journal_current_handle()) {
162 mlog(ML_ERROR, "Recursive transaction attempted!\n");
167 handle = ocfs2_alloc_handle(osb);
170 mlog(ML_ERROR, "Failed to allocate memory for journal "
175 handle->max_buffs = max_buffs;
177 down_read(&osb->journal->j_trans_barrier);
179 /* actually start the transaction now */
180 handle->k_handle = journal_start(journal, max_buffs);
181 if (IS_ERR(handle->k_handle)) {
182 up_read(&osb->journal->j_trans_barrier);
184 ret = PTR_ERR(handle->k_handle);
185 handle->k_handle = NULL;
188 if (is_journal_aborted(journal)) {
189 ocfs2_abort(osb->sb, "Detected aborted journal");
195 atomic_inc(&(osb->journal->j_num_trans));
196 handle->flags |= OCFS2_HANDLE_STARTED;
198 mlog_exit_ptr(handle);
203 ocfs2_commit_unstarted_handle(handle); /* will kfree handle */
209 void ocfs2_handle_add_inode(struct ocfs2_journal_handle *handle,
215 atomic_inc(&inode->i_count);
217 /* we're obviously changing it... */
218 mutex_lock(&inode->i_mutex);
221 BUG_ON(OCFS2_I(inode)->ip_handle);
222 BUG_ON(!list_empty(&OCFS2_I(inode)->ip_handle_list));
224 OCFS2_I(inode)->ip_handle = handle;
225 list_del(&(OCFS2_I(inode)->ip_handle_list));
226 list_add_tail(&(OCFS2_I(inode)->ip_handle_list), &(handle->inode_list));
229 static void ocfs2_handle_unlock_inodes(struct ocfs2_journal_handle *handle)
231 struct list_head *p, *n;
233 struct ocfs2_inode_info *oi;
235 list_for_each_safe(p, n, &handle->inode_list) {
236 oi = list_entry(p, struct ocfs2_inode_info,
238 inode = &oi->vfs_inode;
240 OCFS2_I(inode)->ip_handle = NULL;
241 list_del_init(&OCFS2_I(inode)->ip_handle_list);
243 mutex_unlock(&inode->i_mutex);
248 /* This is trivial so we do it out of the main commit
249 * paths. Beware, it can be called from start_trans too! */
250 static void ocfs2_commit_unstarted_handle(struct ocfs2_journal_handle *handle)
254 BUG_ON(handle->flags & OCFS2_HANDLE_STARTED);
256 ocfs2_handle_unlock_inodes(handle);
257 /* You are allowed to add journal locks before the transaction
259 ocfs2_handle_cleanup_locks(handle->journal, handle);
266 void ocfs2_commit_trans(struct ocfs2_journal_handle *handle)
268 handle_t *jbd_handle;
270 struct ocfs2_journal *journal = handle->journal;
276 if (!(handle->flags & OCFS2_HANDLE_STARTED)) {
277 ocfs2_commit_unstarted_handle(handle);
282 /* release inode semaphores we took during this transaction */
283 ocfs2_handle_unlock_inodes(handle);
285 /* ocfs2_extend_trans may have had to call journal_restart
286 * which will always commit the transaction, but may return
287 * error for any number of reasons. If this is the case, we
288 * clear k_handle as it's not valid any more. */
289 if (handle->k_handle) {
290 jbd_handle = handle->k_handle;
292 if (handle->flags & OCFS2_HANDLE_SYNC)
293 jbd_handle->h_sync = 1;
295 jbd_handle->h_sync = 0;
297 /* actually stop the transaction. if we've set h_sync,
298 * it'll have been committed when we return */
299 retval = journal_stop(jbd_handle);
302 mlog(ML_ERROR, "Could not commit transaction\n");
306 handle->k_handle = NULL; /* it's been free'd in journal_stop */
309 ocfs2_handle_cleanup_locks(journal, handle);
311 up_read(&journal->j_trans_barrier);
318 * 'nblocks' is what you want to add to the current
319 * transaction. extend_trans will either extend the current handle by
320 * nblocks, or commit it and start a new one with nblocks credits.
322 * WARNING: This will not release any semaphores or disk locks taken
323 * during the transaction, so make sure they were taken *before*
324 * start_trans or we'll have ordering deadlocks.
326 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
327 * good because transaction ids haven't yet been recorded on the
328 * cluster locks associated with this handle.
330 int ocfs2_extend_trans(struct ocfs2_journal_handle *handle,
336 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
341 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
343 status = journal_extend(handle->k_handle, nblocks);
350 mlog(0, "journal_extend failed, trying journal_restart\n");
351 status = journal_restart(handle->k_handle, nblocks);
353 handle->k_handle = NULL;
357 handle->max_buffs = nblocks;
359 handle->max_buffs += nblocks;
368 int ocfs2_journal_access(struct ocfs2_journal_handle *handle,
370 struct buffer_head *bh,
378 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
380 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
381 (unsigned long long)bh->b_blocknr, type,
382 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
383 "OCFS2_JOURNAL_ACCESS_CREATE" :
384 "OCFS2_JOURNAL_ACCESS_WRITE",
387 /* we can safely remove this assertion after testing. */
388 if (!buffer_uptodate(bh)) {
389 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
390 mlog(ML_ERROR, "b_blocknr=%llu\n",
391 (unsigned long long)bh->b_blocknr);
395 /* Set the current transaction information on the inode so
396 * that the locking code knows whether it can drop it's locks
397 * on this inode or not. We're protected from the commit
398 * thread updating the current transaction id until
399 * ocfs2_commit_trans() because ocfs2_start_trans() took
400 * j_trans_barrier for us. */
401 ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
403 mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
405 case OCFS2_JOURNAL_ACCESS_CREATE:
406 case OCFS2_JOURNAL_ACCESS_WRITE:
407 status = journal_get_write_access(handle->k_handle, bh);
410 case OCFS2_JOURNAL_ACCESS_UNDO:
411 status = journal_get_undo_access(handle->k_handle, bh);
416 mlog(ML_ERROR, "Uknown access type!\n");
418 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
421 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
428 int ocfs2_journal_dirty(struct ocfs2_journal_handle *handle,
429 struct buffer_head *bh)
433 BUG_ON(!(handle->flags & OCFS2_HANDLE_STARTED));
435 mlog_entry("(bh->b_blocknr=%llu)\n",
436 (unsigned long long)bh->b_blocknr);
438 status = journal_dirty_metadata(handle->k_handle, bh);
440 mlog(ML_ERROR, "Could not dirty metadata buffer. "
441 "(bh->b_blocknr=%llu)\n",
442 (unsigned long long)bh->b_blocknr);
448 int ocfs2_journal_dirty_data(handle_t *handle,
449 struct buffer_head *bh)
451 int err = journal_dirty_data(handle, bh);
454 /* TODO: When we can handle it, abort the handle and go RO on
460 /* We always assume you're adding a metadata lock at level 'ex' */
461 int ocfs2_handle_add_lock(struct ocfs2_journal_handle *handle,
465 struct ocfs2_journal_lock *lock;
469 lock = kmem_cache_alloc(ocfs2_lock_cache, GFP_NOFS);
478 lock->jl_inode = inode;
480 list_add_tail(&(lock->jl_lock_list), &(handle->locks));
489 static void ocfs2_handle_cleanup_locks(struct ocfs2_journal *journal,
490 struct ocfs2_journal_handle *handle)
492 struct list_head *p, *n;
493 struct ocfs2_journal_lock *lock;
496 list_for_each_safe(p, n, &(handle->locks)) {
497 lock = list_entry(p, struct ocfs2_journal_lock,
499 list_del(&lock->jl_lock_list);
502 inode = lock->jl_inode;
503 ocfs2_meta_unlock(inode, 1);
504 if (atomic_read(&inode->i_count) == 1)
506 "Inode %llu, I'm doing a last iput for!",
507 (unsigned long long)OCFS2_I(inode)->ip_blkno);
509 kmem_cache_free(ocfs2_lock_cache, lock);
513 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
515 void ocfs2_set_journal_params(struct ocfs2_super *osb)
517 journal_t *journal = osb->journal->j_journal;
519 spin_lock(&journal->j_state_lock);
520 journal->j_commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
521 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
522 journal->j_flags |= JFS_BARRIER;
524 journal->j_flags &= ~JFS_BARRIER;
525 spin_unlock(&journal->j_state_lock);
528 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
531 struct inode *inode = NULL; /* the journal inode */
532 journal_t *j_journal = NULL;
533 struct ocfs2_dinode *di = NULL;
534 struct buffer_head *bh = NULL;
535 struct ocfs2_super *osb;
542 osb = journal->j_osb;
544 /* already have the inode for our journal */
545 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
552 if (is_bad_inode(inode)) {
553 mlog(ML_ERROR, "access error (bad inode)\n");
560 SET_INODE_JOURNAL(inode);
561 OCFS2_I(inode)->ip_open_count++;
563 /* Skip recovery waits here - journal inode metadata never
564 * changes in a live cluster so it can be considered an
565 * exception to the rule. */
566 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
567 OCFS2_META_LOCK_RECOVERY);
569 if (status != -ERESTARTSYS)
570 mlog(ML_ERROR, "Could not get lock on journal!\n");
575 di = (struct ocfs2_dinode *)bh->b_data;
577 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
578 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
584 mlog(0, "inode->i_size = %lld\n", inode->i_size);
585 mlog(0, "inode->i_blocks = %llu\n",
586 (unsigned long long)inode->i_blocks);
587 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
589 /* call the kernels journal init function now */
590 j_journal = journal_init_inode(inode);
591 if (j_journal == NULL) {
592 mlog(ML_ERROR, "Linux journal layer error\n");
597 mlog(0, "Returned from journal_init_inode\n");
598 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
600 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
601 OCFS2_JOURNAL_DIRTY_FL);
603 journal->j_journal = j_journal;
604 journal->j_inode = inode;
607 ocfs2_set_journal_params(osb);
609 journal->j_state = OCFS2_JOURNAL_LOADED;
615 ocfs2_meta_unlock(inode, 1);
619 OCFS2_I(inode)->ip_open_count--;
628 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
633 struct ocfs2_journal *journal = osb->journal;
634 struct buffer_head *bh = journal->j_bh;
635 struct ocfs2_dinode *fe;
639 fe = (struct ocfs2_dinode *)bh->b_data;
640 if (!OCFS2_IS_VALID_DINODE(fe)) {
641 /* This is called from startup/shutdown which will
642 * handle the errors in a specific manner, so no need
643 * to call ocfs2_error() here. */
644 mlog(ML_ERROR, "Journal dinode %llu has invalid "
645 "signature: %.*s", (unsigned long long)fe->i_blkno, 7,
651 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
653 flags |= OCFS2_JOURNAL_DIRTY_FL;
655 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
656 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
658 status = ocfs2_write_block(osb, bh, journal->j_inode);
668 * If the journal has been kmalloc'd it needs to be freed after this
671 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
673 struct ocfs2_journal *journal = NULL;
675 struct inode *inode = NULL;
676 int num_running_trans = 0;
682 journal = osb->journal;
686 inode = journal->j_inode;
688 if (journal->j_state != OCFS2_JOURNAL_LOADED)
691 /* need to inc inode use count as journal_destroy will iput. */
695 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
696 if (num_running_trans > 0)
697 mlog(0, "Shutting down journal: must wait on %d "
698 "running transactions!\n",
701 /* Do a commit_cache here. It will flush our journal, *and*
702 * release any locks that are still held.
703 * set the SHUTDOWN flag and release the trans lock.
704 * the commit thread will take the trans lock for us below. */
705 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
707 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
708 * drop the trans_lock (which we want to hold until we
709 * completely destroy the journal. */
710 if (osb->commit_task) {
711 /* Wait for the commit thread */
712 mlog(0, "Waiting for ocfs2commit to exit....\n");
713 kthread_stop(osb->commit_task);
714 osb->commit_task = NULL;
717 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
719 status = ocfs2_journal_toggle_dirty(osb, 0);
723 /* Shutdown the kernel journal system */
724 journal_destroy(journal->j_journal);
726 OCFS2_I(inode)->ip_open_count--;
728 /* unlock our journal */
729 ocfs2_meta_unlock(inode, 1);
731 brelse(journal->j_bh);
732 journal->j_bh = NULL;
734 journal->j_state = OCFS2_JOURNAL_FREE;
736 // up_write(&journal->j_trans_barrier);
743 static void ocfs2_clear_journal_error(struct super_block *sb,
749 olderr = journal_errno(journal);
751 mlog(ML_ERROR, "File system error %d recorded in "
752 "journal %u.\n", olderr, slot);
753 mlog(ML_ERROR, "File system on device %s needs checking.\n",
756 journal_ack_err(journal);
757 journal_clear_err(journal);
761 int ocfs2_journal_load(struct ocfs2_journal *journal)
764 struct ocfs2_super *osb;
771 osb = journal->j_osb;
773 status = journal_load(journal->j_journal);
775 mlog(ML_ERROR, "Failed to load journal!\n");
779 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
781 status = ocfs2_journal_toggle_dirty(osb, 1);
787 /* Launch the commit thread */
788 osb->commit_task = kthread_run(ocfs2_commit_thread, osb, "ocfs2cmt-%d",
790 if (IS_ERR(osb->commit_task)) {
791 status = PTR_ERR(osb->commit_task);
792 osb->commit_task = NULL;
793 mlog(ML_ERROR, "unable to launch ocfs2commit thread, error=%d",
804 /* 'full' flag tells us whether we clear out all blocks or if we just
805 * mark the journal clean */
806 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
814 status = journal_wipe(journal->j_journal, full);
820 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
830 * JBD Might read a cached version of another nodes journal file. We
831 * don't want this as this file changes often and we get no
832 * notification on those changes. The only way to be sure that we've
833 * got the most up to date version of those blocks then is to force
834 * read them off disk. Just searching through the buffer cache won't
835 * work as there may be pages backing this file which are still marked
836 * up to date. We know things can't change on this file underneath us
837 * as we have the lock by now :)
839 static int ocfs2_force_read_journal(struct inode *inode)
843 u64 v_blkno, p_blkno;
844 #define CONCURRENT_JOURNAL_FILL 32
845 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
849 BUG_ON(inode->i_blocks !=
850 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
852 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
854 mlog(0, "Force reading %llu blocks\n",
855 (unsigned long long)(inode->i_blocks >>
856 (inode->i_sb->s_blocksize_bits - 9)));
860 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
862 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
870 if (p_blocks > CONCURRENT_JOURNAL_FILL)
871 p_blocks = CONCURRENT_JOURNAL_FILL;
873 /* We are reading journal data which should not
874 * be put in the uptodate cache */
875 status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
876 p_blkno, p_blocks, bhs, 0,
883 for(i = 0; i < p_blocks; i++) {
892 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
899 struct ocfs2_la_recovery_item {
900 struct list_head lri_list;
902 struct ocfs2_dinode *lri_la_dinode;
903 struct ocfs2_dinode *lri_tl_dinode;
906 /* Does the second half of the recovery process. By this point, the
907 * node is marked clean and can actually be considered recovered,
908 * hence it's no longer in the recovery map, but there's still some
909 * cleanup we can do which shouldn't happen within the recovery thread
910 * as locking in that context becomes very difficult if we are to take
911 * recovering nodes into account.
913 * NOTE: This function can and will sleep on recovery of other nodes
914 * during cluster locking, just like any other ocfs2 process.
916 void ocfs2_complete_recovery(void *data)
919 struct ocfs2_super *osb = data;
920 struct ocfs2_journal *journal = osb->journal;
921 struct ocfs2_dinode *la_dinode, *tl_dinode;
922 struct ocfs2_la_recovery_item *item;
923 struct list_head *p, *n;
924 LIST_HEAD(tmp_la_list);
928 mlog(0, "completing recovery from keventd\n");
930 spin_lock(&journal->j_lock);
931 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
932 spin_unlock(&journal->j_lock);
934 list_for_each_safe(p, n, &tmp_la_list) {
935 item = list_entry(p, struct ocfs2_la_recovery_item, lri_list);
936 list_del_init(&item->lri_list);
938 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
940 la_dinode = item->lri_la_dinode;
942 mlog(0, "Clean up local alloc %llu\n",
943 (unsigned long long)la_dinode->i_blkno);
945 ret = ocfs2_complete_local_alloc_recovery(osb,
953 tl_dinode = item->lri_tl_dinode;
955 mlog(0, "Clean up truncate log %llu\n",
956 (unsigned long long)tl_dinode->i_blkno);
958 ret = ocfs2_complete_truncate_log_recovery(osb,
966 ret = ocfs2_recover_orphans(osb, item->lri_slot);
973 mlog(0, "Recovery completion\n");
977 /* NOTE: This function always eats your references to la_dinode and
978 * tl_dinode, either manually on error, or by passing them to
979 * ocfs2_complete_recovery */
980 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
982 struct ocfs2_dinode *la_dinode,
983 struct ocfs2_dinode *tl_dinode)
985 struct ocfs2_la_recovery_item *item;
987 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
989 /* Though we wish to avoid it, we are in fact safe in
990 * skipping local alloc cleanup as fsck.ocfs2 is more
991 * than capable of reclaiming unused space. */
1002 INIT_LIST_HEAD(&item->lri_list);
1003 item->lri_la_dinode = la_dinode;
1004 item->lri_slot = slot_num;
1005 item->lri_tl_dinode = tl_dinode;
1007 spin_lock(&journal->j_lock);
1008 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
1009 queue_work(ocfs2_wq, &journal->j_recovery_work);
1010 spin_unlock(&journal->j_lock);
1013 /* Called by the mount code to queue recovery the last part of
1014 * recovery for it's own slot. */
1015 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
1017 struct ocfs2_journal *journal = osb->journal;
1020 /* No need to queue up our truncate_log as regular
1021 * cleanup will catch that. */
1022 ocfs2_queue_recovery_completion(journal,
1024 osb->local_alloc_copy,
1026 ocfs2_schedule_truncate_log_flush(osb, 0);
1028 osb->local_alloc_copy = NULL;
1033 static int __ocfs2_recovery_thread(void *arg)
1035 int status, node_num;
1036 struct ocfs2_super *osb = arg;
1040 status = ocfs2_wait_on_mount(osb);
1046 status = ocfs2_super_lock(osb, 1);
1052 while(!ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1053 node_num = ocfs2_node_map_first_set_bit(osb,
1054 &osb->recovery_map);
1055 if (node_num == O2NM_INVALID_NODE_NUM) {
1056 mlog(0, "Out of nodes to recover.\n");
1060 status = ocfs2_recover_node(osb, node_num);
1063 "Error %d recovering node %d on device (%u,%u)!\n",
1065 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1066 mlog(ML_ERROR, "Volume requires unmount.\n");
1070 ocfs2_recovery_map_clear(osb, node_num);
1072 ocfs2_super_unlock(osb, 1);
1074 /* We always run recovery on our own orphan dir - the dead
1075 * node(s) may have voted "no" on an inode delete earlier. A
1076 * revote is therefore required. */
1077 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1081 mutex_lock(&osb->recovery_lock);
1083 !ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
1084 mutex_unlock(&osb->recovery_lock);
1088 osb->recovery_thread_task = NULL;
1089 mb(); /* sync with ocfs2_recovery_thread_running */
1090 wake_up(&osb->recovery_event);
1092 mutex_unlock(&osb->recovery_lock);
1095 /* no one is callint kthread_stop() for us so the kthread() api
1096 * requires that we call do_exit(). And it isn't exported, but
1097 * complete_and_exit() seems to be a minimal wrapper around it. */
1098 complete_and_exit(NULL, status);
1102 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1104 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1105 node_num, osb->node_num);
1107 mutex_lock(&osb->recovery_lock);
1108 if (osb->disable_recovery)
1111 /* People waiting on recovery will wait on
1112 * the recovery map to empty. */
1113 if (!ocfs2_recovery_map_set(osb, node_num))
1114 mlog(0, "node %d already be in recovery.\n", node_num);
1116 mlog(0, "starting recovery thread...\n");
1118 if (osb->recovery_thread_task)
1121 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
1122 "ocfs2rec-%d", osb->osb_id);
1123 if (IS_ERR(osb->recovery_thread_task)) {
1124 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1125 osb->recovery_thread_task = NULL;
1129 mutex_unlock(&osb->recovery_lock);
1130 wake_up(&osb->recovery_event);
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;
1149 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1151 if (inode == NULL) {
1156 if (is_bad_inode(inode)) {
1163 SET_INODE_JOURNAL(inode);
1165 status = ocfs2_meta_lock_full(inode, NULL, &bh, 1,
1166 OCFS2_META_LOCK_RECOVERY);
1168 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status);
1169 if (status != -ERESTARTSYS)
1170 mlog(ML_ERROR, "Could not lock journal!\n");
1175 fe = (struct ocfs2_dinode *) bh->b_data;
1177 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1179 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1180 mlog(0, "No recovery required for node %d\n", node_num);
1184 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1186 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1188 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1190 status = ocfs2_force_read_journal(inode);
1196 mlog(0, "calling journal_init_inode\n");
1197 journal = journal_init_inode(inode);
1198 if (journal == NULL) {
1199 mlog(ML_ERROR, "Linux journal layer error\n");
1204 status = journal_load(journal);
1209 journal_destroy(journal);
1213 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1215 /* wipe the journal */
1216 mlog(0, "flushing the journal.\n");
1217 journal_lock_updates(journal);
1218 status = journal_flush(journal);
1219 journal_unlock_updates(journal);
1223 /* This will mark the node clean */
1224 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1225 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1226 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1228 status = ocfs2_write_block(osb, bh, inode);
1235 journal_destroy(journal);
1238 /* drop the lock on this nodes journal */
1240 ocfs2_meta_unlock(inode, 1);
1253 * Do the most important parts of node recovery:
1254 * - Replay it's journal
1255 * - Stamp a clean local allocator file
1256 * - Stamp a clean truncate log
1257 * - Mark the node clean
1259 * If this function completes without error, a node in OCFS2 can be
1260 * said to have been safely recovered. As a result, failure during the
1261 * second part of a nodes recovery process (local alloc recovery) is
1262 * far less concerning.
1264 static int ocfs2_recover_node(struct ocfs2_super *osb,
1269 struct ocfs2_slot_info *si = osb->slot_info;
1270 struct ocfs2_dinode *la_copy = NULL;
1271 struct ocfs2_dinode *tl_copy = NULL;
1273 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1274 node_num, osb->node_num);
1276 mlog(0, "checking node %d\n", node_num);
1278 /* Should not ever be called to recover ourselves -- in that
1279 * case we should've called ocfs2_journal_load instead. */
1280 BUG_ON(osb->node_num == node_num);
1282 slot_num = ocfs2_node_num_to_slot(si, node_num);
1283 if (slot_num == OCFS2_INVALID_SLOT) {
1285 mlog(0, "no slot for this node, so no recovery required.\n");
1289 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1291 status = ocfs2_replay_journal(osb, node_num, slot_num);
1297 /* Stamp a clean local alloc file AFTER recovering the journal... */
1298 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1304 /* An error from begin_truncate_log_recovery is not
1305 * serious enough to warrant halting the rest of
1307 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1311 /* Likewise, this would be a strange but ultimately not so
1312 * harmful place to get an error... */
1313 ocfs2_clear_slot(si, slot_num);
1314 status = ocfs2_update_disk_slots(osb, si);
1318 /* This will kfree the memory pointed to by la_copy and tl_copy */
1319 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1329 /* Test node liveness by trylocking his journal. If we get the lock,
1330 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1331 * still alive (we couldn't get the lock) and < 0 on error. */
1332 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1336 struct inode *inode = NULL;
1338 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1340 if (inode == NULL) {
1341 mlog(ML_ERROR, "access error\n");
1345 if (is_bad_inode(inode)) {
1346 mlog(ML_ERROR, "access error (bad inode)\n");
1352 SET_INODE_JOURNAL(inode);
1354 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1355 status = ocfs2_meta_lock_full(inode, NULL, NULL, 1, flags);
1357 if (status != -EAGAIN)
1362 ocfs2_meta_unlock(inode, 1);
1370 /* Call this underneath ocfs2_super_lock. It also assumes that the
1371 * slot info struct has been updated from disk. */
1372 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1374 int status, i, node_num;
1375 struct ocfs2_slot_info *si = osb->slot_info;
1377 /* This is called with the super block cluster lock, so we
1378 * know that the slot map can't change underneath us. */
1380 spin_lock(&si->si_lock);
1381 for(i = 0; i < si->si_num_slots; i++) {
1382 if (i == osb->slot_num)
1384 if (ocfs2_is_empty_slot(si, i))
1387 node_num = si->si_global_node_nums[i];
1388 if (ocfs2_node_map_test_bit(osb, &osb->recovery_map, node_num))
1390 spin_unlock(&si->si_lock);
1392 /* Ok, we have a slot occupied by another node which
1393 * is not in the recovery map. We trylock his journal
1394 * file here to test if he's alive. */
1395 status = ocfs2_trylock_journal(osb, i);
1397 /* Since we're called from mount, we know that
1398 * the recovery thread can't race us on
1399 * setting / checking the recovery bits. */
1400 ocfs2_recovery_thread(osb, node_num);
1401 } else if ((status < 0) && (status != -EAGAIN)) {
1406 spin_lock(&si->si_lock);
1408 spin_unlock(&si->si_lock);
1416 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1418 struct inode **head)
1421 struct inode *orphan_dir_inode = NULL;
1423 unsigned long offset, blk, local;
1424 struct buffer_head *bh = NULL;
1425 struct ocfs2_dir_entry *de;
1426 struct super_block *sb = osb->sb;
1428 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1429 ORPHAN_DIR_SYSTEM_INODE,
1431 if (!orphan_dir_inode) {
1437 mutex_lock(&orphan_dir_inode->i_mutex);
1438 status = ocfs2_meta_lock(orphan_dir_inode, NULL, NULL, 0);
1446 while(offset < i_size_read(orphan_dir_inode)) {
1447 blk = offset >> sb->s_blocksize_bits;
1449 bh = ocfs2_bread(orphan_dir_inode, blk, &status, 0);
1460 while(offset < i_size_read(orphan_dir_inode)
1461 && local < sb->s_blocksize) {
1462 de = (struct ocfs2_dir_entry *) (bh->b_data + local);
1464 if (!ocfs2_check_dir_entry(orphan_dir_inode,
1472 local += le16_to_cpu(de->rec_len);
1473 offset += le16_to_cpu(de->rec_len);
1475 /* I guess we silently fail on no inode? */
1476 if (!le64_to_cpu(de->inode))
1478 if (de->file_type > OCFS2_FT_MAX) {
1480 "block %llu contains invalid de: "
1481 "inode = %llu, rec_len = %u, "
1482 "name_len = %u, file_type = %u, "
1484 (unsigned long long)bh->b_blocknr,
1485 (unsigned long long)le64_to_cpu(de->inode),
1486 le16_to_cpu(de->rec_len),
1493 if (de->name_len == 1 && !strncmp(".", de->name, 1))
1495 if (de->name_len == 2 && !strncmp("..", de->name, 2))
1498 iter = ocfs2_iget(osb, le64_to_cpu(de->inode));
1502 mlog(0, "queue orphan %llu\n",
1503 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1504 /* No locking is required for the next_orphan
1505 * queue as there is only ever a single
1506 * process doing orphan recovery. */
1507 OCFS2_I(iter)->ip_next_orphan = *head;
1514 ocfs2_meta_unlock(orphan_dir_inode, 0);
1516 mutex_unlock(&orphan_dir_inode->i_mutex);
1517 iput(orphan_dir_inode);
1521 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1526 spin_lock(&osb->osb_lock);
1527 ret = !osb->osb_orphan_wipes[slot];
1528 spin_unlock(&osb->osb_lock);
1532 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1535 spin_lock(&osb->osb_lock);
1536 /* Mark ourselves such that new processes in delete_inode()
1537 * know to quit early. */
1538 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1539 while (osb->osb_orphan_wipes[slot]) {
1540 /* If any processes are already in the middle of an
1541 * orphan wipe on this dir, then we need to wait for
1543 spin_unlock(&osb->osb_lock);
1544 wait_event_interruptible(osb->osb_wipe_event,
1545 ocfs2_orphan_recovery_can_continue(osb, slot));
1546 spin_lock(&osb->osb_lock);
1548 spin_unlock(&osb->osb_lock);
1551 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1554 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1558 * Orphan recovery. Each mounted node has it's own orphan dir which we
1559 * must run during recovery. Our strategy here is to build a list of
1560 * the inodes in the orphan dir and iget/iput them. The VFS does
1561 * (most) of the rest of the work.
1563 * Orphan recovery can happen at any time, not just mount so we have a
1564 * couple of extra considerations.
1566 * - We grab as many inodes as we can under the orphan dir lock -
1567 * doing iget() outside the orphan dir risks getting a reference on
1569 * - We must be sure not to deadlock with other processes on the
1570 * system wanting to run delete_inode(). This can happen when they go
1571 * to lock the orphan dir and the orphan recovery process attempts to
1572 * iget() inside the orphan dir lock. This can be avoided by
1573 * advertising our state to ocfs2_delete_inode().
1575 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1579 struct inode *inode = NULL;
1581 struct ocfs2_inode_info *oi;
1583 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1585 ocfs2_mark_recovering_orphan_dir(osb, slot);
1586 ret = ocfs2_queue_orphans(osb, slot, &inode);
1587 ocfs2_clear_recovering_orphan_dir(osb, slot);
1589 /* Error here should be noted, but we want to continue with as
1590 * many queued inodes as we've got. */
1595 oi = OCFS2_I(inode);
1596 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1598 iter = oi->ip_next_orphan;
1600 spin_lock(&oi->ip_lock);
1601 /* Delete voting may have set these on the assumption
1602 * that the other node would wipe them successfully.
1603 * If they are still in the node's orphan dir, we need
1604 * to reset that state. */
1605 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1607 /* Set the proper information to get us going into
1608 * ocfs2_delete_inode. */
1609 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1610 oi->ip_orphaned_slot = slot;
1611 spin_unlock(&oi->ip_lock);
1621 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1623 /* This check is good because ocfs2 will wait on our recovery
1624 * thread before changing it to something other than MOUNTED
1626 wait_event(osb->osb_mount_event,
1627 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1628 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1630 /* If there's an error on mount, then we may never get to the
1631 * MOUNTED flag, but this is set right before
1632 * dismount_volume() so we can trust it. */
1633 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1634 mlog(0, "mount error, exiting!\n");
1641 static int ocfs2_commit_thread(void *arg)
1644 struct ocfs2_super *osb = arg;
1645 struct ocfs2_journal *journal = osb->journal;
1647 /* we can trust j_num_trans here because _should_stop() is only set in
1648 * shutdown and nobody other than ourselves should be able to start
1649 * transactions. committing on shutdown might take a few iterations
1650 * as final transactions put deleted inodes on the list */
1651 while (!(kthread_should_stop() &&
1652 atomic_read(&journal->j_num_trans) == 0)) {
1654 wait_event_interruptible(osb->checkpoint_event,
1655 atomic_read(&journal->j_num_trans)
1656 || kthread_should_stop());
1658 status = ocfs2_commit_cache(osb);
1662 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1664 "commit_thread: %u transactions pending on "
1666 atomic_read(&journal->j_num_trans));
1673 /* Look for a dirty journal without taking any cluster locks. Used for
1674 * hard readonly access to determine whether the file system journals
1675 * require recovery. */
1676 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1680 struct buffer_head *di_bh;
1681 struct ocfs2_dinode *di;
1682 struct inode *journal = NULL;
1684 for(slot = 0; slot < osb->max_slots; slot++) {
1685 journal = ocfs2_get_system_file_inode(osb,
1686 JOURNAL_SYSTEM_INODE,
1688 if (!journal || is_bad_inode(journal)) {
1695 ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1702 di = (struct ocfs2_dinode *) di_bh->b_data;
1704 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1705 OCFS2_JOURNAL_DIRTY_FL)