1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Defines functions of journalling api
8 * Copyright (C) 2003, 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/kthread.h>
32 #define MLOG_MASK_PREFIX ML_JOURNAL
33 #include <cluster/masklog.h>
39 #include "extent_map.h"
40 #include "heartbeat.h"
43 #include "localalloc.h"
50 #include "buffer_head_io.h"
52 DEFINE_SPINLOCK(trans_inc_lock);
54 static int ocfs2_force_read_journal(struct inode *inode);
55 static int ocfs2_recover_node(struct ocfs2_super *osb,
57 static int __ocfs2_recovery_thread(void *arg);
58 static int ocfs2_commit_cache(struct ocfs2_super *osb);
59 static int ocfs2_wait_on_mount(struct ocfs2_super *osb);
60 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
62 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
64 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
66 static int ocfs2_commit_thread(void *arg);
68 static int ocfs2_commit_cache(struct ocfs2_super *osb)
73 struct ocfs2_journal *journal = NULL;
77 journal = osb->journal;
79 /* Flush all pending commits and checkpoint the journal. */
80 down_write(&journal->j_trans_barrier);
82 if (atomic_read(&journal->j_num_trans) == 0) {
83 up_write(&journal->j_trans_barrier);
84 mlog(0, "No transactions for me to flush!\n");
88 journal_lock_updates(journal->j_journal);
89 status = journal_flush(journal->j_journal);
90 journal_unlock_updates(journal->j_journal);
92 up_write(&journal->j_trans_barrier);
97 old_id = ocfs2_inc_trans_id(journal);
99 flushed = atomic_read(&journal->j_num_trans);
100 atomic_set(&journal->j_num_trans, 0);
101 up_write(&journal->j_trans_barrier);
103 mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
104 journal->j_trans_id, flushed);
106 ocfs2_kick_vote_thread(osb);
107 wake_up(&journal->j_checkpointed);
113 static struct ocfs2_journal_handle *ocfs2_alloc_handle(struct ocfs2_super *osb)
115 struct ocfs2_journal_handle *retval = NULL;
117 retval = kcalloc(1, sizeof(*retval), GFP_NOFS);
119 mlog(ML_ERROR, "Failed to allocate memory for journal "
123 retval->k_handle = NULL;
128 /* pass it NULL and it will allocate a new handle object for you. If
129 * you pass it a handle however, it may still return error, in which
130 * case it has free'd the passed handle for you. */
131 struct ocfs2_journal_handle *ocfs2_start_trans(struct ocfs2_super *osb,
132 struct ocfs2_journal_handle *handle,
136 journal_t *journal = osb->journal->j_journal;
138 mlog_entry("(max_buffs = %d)\n", max_buffs);
140 BUG_ON(!osb || !osb->journal->j_journal);
142 if (ocfs2_is_hard_readonly(osb)) {
147 BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
148 BUG_ON(max_buffs <= 0);
150 /* JBD might support this, but our journalling code doesn't yet. */
151 if (journal_current_handle()) {
152 mlog(ML_ERROR, "Recursive transaction attempted!\n");
157 handle = ocfs2_alloc_handle(osb);
160 mlog(ML_ERROR, "Failed to allocate memory for journal "
165 down_read(&osb->journal->j_trans_barrier);
167 /* actually start the transaction now */
168 handle->k_handle = journal_start(journal, max_buffs);
169 if (IS_ERR(handle->k_handle)) {
170 up_read(&osb->journal->j_trans_barrier);
172 ret = PTR_ERR(handle->k_handle);
173 handle->k_handle = NULL;
176 if (is_journal_aborted(journal)) {
177 ocfs2_abort(osb->sb, "Detected aborted journal");
183 atomic_inc(&(osb->journal->j_num_trans));
185 mlog_exit_ptr(handle);
196 void ocfs2_commit_trans(struct ocfs2_super *osb,
197 struct ocfs2_journal_handle *handle)
199 handle_t *jbd_handle;
201 struct ocfs2_journal *journal = osb->journal;
207 if (!handle->k_handle) {
213 /* ocfs2_extend_trans may have had to call journal_restart
214 * which will always commit the transaction, but may return
215 * error for any number of reasons. If this is the case, we
216 * clear k_handle as it's not valid any more. */
217 if (handle->k_handle) {
218 jbd_handle = handle->k_handle;
220 /* actually stop the transaction. if we've set h_sync,
221 * it'll have been committed when we return */
222 retval = journal_stop(jbd_handle);
225 mlog(ML_ERROR, "Could not commit transaction\n");
229 handle->k_handle = NULL; /* it's been free'd in journal_stop */
232 up_read(&journal->j_trans_barrier);
239 * 'nblocks' is what you want to add to the current
240 * transaction. extend_trans will either extend the current handle by
241 * nblocks, or commit it and start a new one with nblocks credits.
243 * WARNING: This will not release any semaphores or disk locks taken
244 * during the transaction, so make sure they were taken *before*
245 * start_trans or we'll have ordering deadlocks.
247 * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
248 * good because transaction ids haven't yet been recorded on the
249 * cluster locks associated with this handle.
251 int ocfs2_extend_trans(handle_t *handle, int nblocks)
260 mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
262 status = journal_extend(handle, nblocks);
269 mlog(0, "journal_extend failed, trying journal_restart\n");
270 status = journal_restart(handle, nblocks);
284 int ocfs2_journal_access(struct ocfs2_journal_handle *handle,
286 struct buffer_head *bh,
295 mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
296 (unsigned long long)bh->b_blocknr, type,
297 (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
298 "OCFS2_JOURNAL_ACCESS_CREATE" :
299 "OCFS2_JOURNAL_ACCESS_WRITE",
302 /* we can safely remove this assertion after testing. */
303 if (!buffer_uptodate(bh)) {
304 mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
305 mlog(ML_ERROR, "b_blocknr=%llu\n",
306 (unsigned long long)bh->b_blocknr);
310 /* Set the current transaction information on the inode so
311 * that the locking code knows whether it can drop it's locks
312 * on this inode or not. We're protected from the commit
313 * thread updating the current transaction id until
314 * ocfs2_commit_trans() because ocfs2_start_trans() took
315 * j_trans_barrier for us. */
316 ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
318 mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
320 case OCFS2_JOURNAL_ACCESS_CREATE:
321 case OCFS2_JOURNAL_ACCESS_WRITE:
322 status = journal_get_write_access(handle->k_handle, bh);
325 case OCFS2_JOURNAL_ACCESS_UNDO:
326 status = journal_get_undo_access(handle->k_handle, bh);
331 mlog(ML_ERROR, "Uknown access type!\n");
333 mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
336 mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
343 int ocfs2_journal_dirty(struct ocfs2_journal_handle *handle,
344 struct buffer_head *bh)
348 mlog_entry("(bh->b_blocknr=%llu)\n",
349 (unsigned long long)bh->b_blocknr);
351 status = journal_dirty_metadata(handle->k_handle, bh);
353 mlog(ML_ERROR, "Could not dirty metadata buffer. "
354 "(bh->b_blocknr=%llu)\n",
355 (unsigned long long)bh->b_blocknr);
361 int ocfs2_journal_dirty_data(handle_t *handle,
362 struct buffer_head *bh)
364 int err = journal_dirty_data(handle, bh);
367 /* TODO: When we can handle it, abort the handle and go RO on
373 #define OCFS2_DEFAULT_COMMIT_INTERVAL (HZ * 5)
375 void ocfs2_set_journal_params(struct ocfs2_super *osb)
377 journal_t *journal = osb->journal->j_journal;
379 spin_lock(&journal->j_state_lock);
380 journal->j_commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
381 if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
382 journal->j_flags |= JFS_BARRIER;
384 journal->j_flags &= ~JFS_BARRIER;
385 spin_unlock(&journal->j_state_lock);
388 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
391 struct inode *inode = NULL; /* the journal inode */
392 journal_t *j_journal = NULL;
393 struct ocfs2_dinode *di = NULL;
394 struct buffer_head *bh = NULL;
395 struct ocfs2_super *osb;
402 osb = journal->j_osb;
404 /* already have the inode for our journal */
405 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
412 if (is_bad_inode(inode)) {
413 mlog(ML_ERROR, "access error (bad inode)\n");
420 SET_INODE_JOURNAL(inode);
421 OCFS2_I(inode)->ip_open_count++;
423 /* Skip recovery waits here - journal inode metadata never
424 * changes in a live cluster so it can be considered an
425 * exception to the rule. */
426 status = ocfs2_meta_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
428 if (status != -ERESTARTSYS)
429 mlog(ML_ERROR, "Could not get lock on journal!\n");
434 di = (struct ocfs2_dinode *)bh->b_data;
436 if (inode->i_size < OCFS2_MIN_JOURNAL_SIZE) {
437 mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
443 mlog(0, "inode->i_size = %lld\n", inode->i_size);
444 mlog(0, "inode->i_blocks = %llu\n",
445 (unsigned long long)inode->i_blocks);
446 mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
448 /* call the kernels journal init function now */
449 j_journal = journal_init_inode(inode);
450 if (j_journal == NULL) {
451 mlog(ML_ERROR, "Linux journal layer error\n");
456 mlog(0, "Returned from journal_init_inode\n");
457 mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
459 *dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
460 OCFS2_JOURNAL_DIRTY_FL);
462 journal->j_journal = j_journal;
463 journal->j_inode = inode;
466 ocfs2_set_journal_params(osb);
468 journal->j_state = OCFS2_JOURNAL_LOADED;
474 ocfs2_meta_unlock(inode, 1);
478 OCFS2_I(inode)->ip_open_count--;
487 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
492 struct ocfs2_journal *journal = osb->journal;
493 struct buffer_head *bh = journal->j_bh;
494 struct ocfs2_dinode *fe;
498 fe = (struct ocfs2_dinode *)bh->b_data;
499 if (!OCFS2_IS_VALID_DINODE(fe)) {
500 /* This is called from startup/shutdown which will
501 * handle the errors in a specific manner, so no need
502 * to call ocfs2_error() here. */
503 mlog(ML_ERROR, "Journal dinode %llu has invalid "
504 "signature: %.*s", (unsigned long long)fe->i_blkno, 7,
510 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
512 flags |= OCFS2_JOURNAL_DIRTY_FL;
514 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
515 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
517 status = ocfs2_write_block(osb, bh, journal->j_inode);
527 * If the journal has been kmalloc'd it needs to be freed after this
530 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
532 struct ocfs2_journal *journal = NULL;
534 struct inode *inode = NULL;
535 int num_running_trans = 0;
541 journal = osb->journal;
545 inode = journal->j_inode;
547 if (journal->j_state != OCFS2_JOURNAL_LOADED)
550 /* need to inc inode use count as journal_destroy will iput. */
554 num_running_trans = atomic_read(&(osb->journal->j_num_trans));
555 if (num_running_trans > 0)
556 mlog(0, "Shutting down journal: must wait on %d "
557 "running transactions!\n",
560 /* Do a commit_cache here. It will flush our journal, *and*
561 * release any locks that are still held.
562 * set the SHUTDOWN flag and release the trans lock.
563 * the commit thread will take the trans lock for us below. */
564 journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
566 /* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
567 * drop the trans_lock (which we want to hold until we
568 * completely destroy the journal. */
569 if (osb->commit_task) {
570 /* Wait for the commit thread */
571 mlog(0, "Waiting for ocfs2commit to exit....\n");
572 kthread_stop(osb->commit_task);
573 osb->commit_task = NULL;
576 BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
578 status = ocfs2_journal_toggle_dirty(osb, 0);
582 /* Shutdown the kernel journal system */
583 journal_destroy(journal->j_journal);
585 OCFS2_I(inode)->ip_open_count--;
587 /* unlock our journal */
588 ocfs2_meta_unlock(inode, 1);
590 brelse(journal->j_bh);
591 journal->j_bh = NULL;
593 journal->j_state = OCFS2_JOURNAL_FREE;
595 // up_write(&journal->j_trans_barrier);
602 static void ocfs2_clear_journal_error(struct super_block *sb,
608 olderr = journal_errno(journal);
610 mlog(ML_ERROR, "File system error %d recorded in "
611 "journal %u.\n", olderr, slot);
612 mlog(ML_ERROR, "File system on device %s needs checking.\n",
615 journal_ack_err(journal);
616 journal_clear_err(journal);
620 int ocfs2_journal_load(struct ocfs2_journal *journal)
623 struct ocfs2_super *osb;
630 osb = journal->j_osb;
632 status = journal_load(journal->j_journal);
634 mlog(ML_ERROR, "Failed to load journal!\n");
638 ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
640 status = ocfs2_journal_toggle_dirty(osb, 1);
646 /* Launch the commit thread */
647 osb->commit_task = kthread_run(ocfs2_commit_thread, osb, "ocfs2cmt");
648 if (IS_ERR(osb->commit_task)) {
649 status = PTR_ERR(osb->commit_task);
650 osb->commit_task = NULL;
651 mlog(ML_ERROR, "unable to launch ocfs2commit thread, error=%d",
662 /* 'full' flag tells us whether we clear out all blocks or if we just
663 * mark the journal clean */
664 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
672 status = journal_wipe(journal->j_journal, full);
678 status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
688 * JBD Might read a cached version of another nodes journal file. We
689 * don't want this as this file changes often and we get no
690 * notification on those changes. The only way to be sure that we've
691 * got the most up to date version of those blocks then is to force
692 * read them off disk. Just searching through the buffer cache won't
693 * work as there may be pages backing this file which are still marked
694 * up to date. We know things can't change on this file underneath us
695 * as we have the lock by now :)
697 static int ocfs2_force_read_journal(struct inode *inode)
701 u64 v_blkno, p_blkno;
702 #define CONCURRENT_JOURNAL_FILL 32
703 struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
707 BUG_ON(inode->i_blocks !=
708 ocfs2_align_bytes_to_sectors(i_size_read(inode)));
710 memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
712 mlog(0, "Force reading %llu blocks\n",
713 (unsigned long long)(inode->i_blocks >>
714 (inode->i_sb->s_blocksize_bits - 9)));
718 (inode->i_blocks >> (inode->i_sb->s_blocksize_bits - 9))) {
720 status = ocfs2_extent_map_get_blocks(inode, v_blkno,
728 if (p_blocks > CONCURRENT_JOURNAL_FILL)
729 p_blocks = CONCURRENT_JOURNAL_FILL;
731 /* We are reading journal data which should not
732 * be put in the uptodate cache */
733 status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
734 p_blkno, p_blocks, bhs, 0,
741 for(i = 0; i < p_blocks; i++) {
750 for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
757 struct ocfs2_la_recovery_item {
758 struct list_head lri_list;
760 struct ocfs2_dinode *lri_la_dinode;
761 struct ocfs2_dinode *lri_tl_dinode;
764 /* Does the second half of the recovery process. By this point, the
765 * node is marked clean and can actually be considered recovered,
766 * hence it's no longer in the recovery map, but there's still some
767 * cleanup we can do which shouldn't happen within the recovery thread
768 * as locking in that context becomes very difficult if we are to take
769 * recovering nodes into account.
771 * NOTE: This function can and will sleep on recovery of other nodes
772 * during cluster locking, just like any other ocfs2 process.
774 void ocfs2_complete_recovery(void *data)
777 struct ocfs2_super *osb = data;
778 struct ocfs2_journal *journal = osb->journal;
779 struct ocfs2_dinode *la_dinode, *tl_dinode;
780 struct ocfs2_la_recovery_item *item;
781 struct list_head *p, *n;
782 LIST_HEAD(tmp_la_list);
786 mlog(0, "completing recovery from keventd\n");
788 spin_lock(&journal->j_lock);
789 list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
790 spin_unlock(&journal->j_lock);
792 list_for_each_safe(p, n, &tmp_la_list) {
793 item = list_entry(p, struct ocfs2_la_recovery_item, lri_list);
794 list_del_init(&item->lri_list);
796 mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
798 la_dinode = item->lri_la_dinode;
800 mlog(0, "Clean up local alloc %llu\n",
801 (unsigned long long)la_dinode->i_blkno);
803 ret = ocfs2_complete_local_alloc_recovery(osb,
811 tl_dinode = item->lri_tl_dinode;
813 mlog(0, "Clean up truncate log %llu\n",
814 (unsigned long long)tl_dinode->i_blkno);
816 ret = ocfs2_complete_truncate_log_recovery(osb,
824 ret = ocfs2_recover_orphans(osb, item->lri_slot);
831 mlog(0, "Recovery completion\n");
835 /* NOTE: This function always eats your references to la_dinode and
836 * tl_dinode, either manually on error, or by passing them to
837 * ocfs2_complete_recovery */
838 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
840 struct ocfs2_dinode *la_dinode,
841 struct ocfs2_dinode *tl_dinode)
843 struct ocfs2_la_recovery_item *item;
845 item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
847 /* Though we wish to avoid it, we are in fact safe in
848 * skipping local alloc cleanup as fsck.ocfs2 is more
849 * than capable of reclaiming unused space. */
860 INIT_LIST_HEAD(&item->lri_list);
861 item->lri_la_dinode = la_dinode;
862 item->lri_slot = slot_num;
863 item->lri_tl_dinode = tl_dinode;
865 spin_lock(&journal->j_lock);
866 list_add_tail(&item->lri_list, &journal->j_la_cleanups);
867 queue_work(ocfs2_wq, &journal->j_recovery_work);
868 spin_unlock(&journal->j_lock);
871 /* Called by the mount code to queue recovery the last part of
872 * recovery for it's own slot. */
873 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
875 struct ocfs2_journal *journal = osb->journal;
878 /* No need to queue up our truncate_log as regular
879 * cleanup will catch that. */
880 ocfs2_queue_recovery_completion(journal,
882 osb->local_alloc_copy,
884 ocfs2_schedule_truncate_log_flush(osb, 0);
886 osb->local_alloc_copy = NULL;
891 static int __ocfs2_recovery_thread(void *arg)
893 int status, node_num;
894 struct ocfs2_super *osb = arg;
898 status = ocfs2_wait_on_mount(osb);
904 status = ocfs2_super_lock(osb, 1);
910 while(!ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
911 node_num = ocfs2_node_map_first_set_bit(osb,
913 if (node_num == O2NM_INVALID_NODE_NUM) {
914 mlog(0, "Out of nodes to recover.\n");
918 status = ocfs2_recover_node(osb, node_num);
921 "Error %d recovering node %d on device (%u,%u)!\n",
923 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
924 mlog(ML_ERROR, "Volume requires unmount.\n");
928 ocfs2_recovery_map_clear(osb, node_num);
930 ocfs2_super_unlock(osb, 1);
932 /* We always run recovery on our own orphan dir - the dead
933 * node(s) may have voted "no" on an inode delete earlier. A
934 * revote is therefore required. */
935 ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
939 mutex_lock(&osb->recovery_lock);
941 !ocfs2_node_map_is_empty(osb, &osb->recovery_map)) {
942 mutex_unlock(&osb->recovery_lock);
946 osb->recovery_thread_task = NULL;
947 mb(); /* sync with ocfs2_recovery_thread_running */
948 wake_up(&osb->recovery_event);
950 mutex_unlock(&osb->recovery_lock);
953 /* no one is callint kthread_stop() for us so the kthread() api
954 * requires that we call do_exit(). And it isn't exported, but
955 * complete_and_exit() seems to be a minimal wrapper around it. */
956 complete_and_exit(NULL, status);
960 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
962 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
963 node_num, osb->node_num);
965 mutex_lock(&osb->recovery_lock);
966 if (osb->disable_recovery)
969 /* People waiting on recovery will wait on
970 * the recovery map to empty. */
971 if (!ocfs2_recovery_map_set(osb, node_num))
972 mlog(0, "node %d already be in recovery.\n", node_num);
974 mlog(0, "starting recovery thread...\n");
976 if (osb->recovery_thread_task)
979 osb->recovery_thread_task = kthread_run(__ocfs2_recovery_thread, osb,
981 if (IS_ERR(osb->recovery_thread_task)) {
982 mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
983 osb->recovery_thread_task = NULL;
987 mutex_unlock(&osb->recovery_lock);
988 wake_up(&osb->recovery_event);
993 /* Does the actual journal replay and marks the journal inode as
994 * clean. Will only replay if the journal inode is marked dirty. */
995 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1002 struct inode *inode = NULL;
1003 struct ocfs2_dinode *fe;
1004 journal_t *journal = NULL;
1005 struct buffer_head *bh = NULL;
1007 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1009 if (inode == NULL) {
1014 if (is_bad_inode(inode)) {
1021 SET_INODE_JOURNAL(inode);
1023 status = ocfs2_meta_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1025 mlog(0, "status returned from ocfs2_meta_lock=%d\n", status);
1026 if (status != -ERESTARTSYS)
1027 mlog(ML_ERROR, "Could not lock journal!\n");
1032 fe = (struct ocfs2_dinode *) bh->b_data;
1034 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1036 if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1037 mlog(0, "No recovery required for node %d\n", node_num);
1041 mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1043 MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1045 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1047 status = ocfs2_force_read_journal(inode);
1053 mlog(0, "calling journal_init_inode\n");
1054 journal = journal_init_inode(inode);
1055 if (journal == NULL) {
1056 mlog(ML_ERROR, "Linux journal layer error\n");
1061 status = journal_load(journal);
1066 journal_destroy(journal);
1070 ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1072 /* wipe the journal */
1073 mlog(0, "flushing the journal.\n");
1074 journal_lock_updates(journal);
1075 status = journal_flush(journal);
1076 journal_unlock_updates(journal);
1080 /* This will mark the node clean */
1081 flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1082 flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1083 fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1085 status = ocfs2_write_block(osb, bh, inode);
1092 journal_destroy(journal);
1095 /* drop the lock on this nodes journal */
1097 ocfs2_meta_unlock(inode, 1);
1110 * Do the most important parts of node recovery:
1111 * - Replay it's journal
1112 * - Stamp a clean local allocator file
1113 * - Stamp a clean truncate log
1114 * - Mark the node clean
1116 * If this function completes without error, a node in OCFS2 can be
1117 * said to have been safely recovered. As a result, failure during the
1118 * second part of a nodes recovery process (local alloc recovery) is
1119 * far less concerning.
1121 static int ocfs2_recover_node(struct ocfs2_super *osb,
1126 struct ocfs2_slot_info *si = osb->slot_info;
1127 struct ocfs2_dinode *la_copy = NULL;
1128 struct ocfs2_dinode *tl_copy = NULL;
1130 mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1131 node_num, osb->node_num);
1133 mlog(0, "checking node %d\n", node_num);
1135 /* Should not ever be called to recover ourselves -- in that
1136 * case we should've called ocfs2_journal_load instead. */
1137 BUG_ON(osb->node_num == node_num);
1139 slot_num = ocfs2_node_num_to_slot(si, node_num);
1140 if (slot_num == OCFS2_INVALID_SLOT) {
1142 mlog(0, "no slot for this node, so no recovery required.\n");
1146 mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1148 status = ocfs2_replay_journal(osb, node_num, slot_num);
1154 /* Stamp a clean local alloc file AFTER recovering the journal... */
1155 status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1161 /* An error from begin_truncate_log_recovery is not
1162 * serious enough to warrant halting the rest of
1164 status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1168 /* Likewise, this would be a strange but ultimately not so
1169 * harmful place to get an error... */
1170 ocfs2_clear_slot(si, slot_num);
1171 status = ocfs2_update_disk_slots(osb, si);
1175 /* This will kfree the memory pointed to by la_copy and tl_copy */
1176 ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1186 /* Test node liveness by trylocking his journal. If we get the lock,
1187 * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1188 * still alive (we couldn't get the lock) and < 0 on error. */
1189 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1193 struct inode *inode = NULL;
1195 inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1197 if (inode == NULL) {
1198 mlog(ML_ERROR, "access error\n");
1202 if (is_bad_inode(inode)) {
1203 mlog(ML_ERROR, "access error (bad inode)\n");
1209 SET_INODE_JOURNAL(inode);
1211 flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1212 status = ocfs2_meta_lock_full(inode, NULL, 1, flags);
1214 if (status != -EAGAIN)
1219 ocfs2_meta_unlock(inode, 1);
1227 /* Call this underneath ocfs2_super_lock. It also assumes that the
1228 * slot info struct has been updated from disk. */
1229 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1231 int status, i, node_num;
1232 struct ocfs2_slot_info *si = osb->slot_info;
1234 /* This is called with the super block cluster lock, so we
1235 * know that the slot map can't change underneath us. */
1237 spin_lock(&si->si_lock);
1238 for(i = 0; i < si->si_num_slots; i++) {
1239 if (i == osb->slot_num)
1241 if (ocfs2_is_empty_slot(si, i))
1244 node_num = si->si_global_node_nums[i];
1245 if (ocfs2_node_map_test_bit(osb, &osb->recovery_map, node_num))
1247 spin_unlock(&si->si_lock);
1249 /* Ok, we have a slot occupied by another node which
1250 * is not in the recovery map. We trylock his journal
1251 * file here to test if he's alive. */
1252 status = ocfs2_trylock_journal(osb, i);
1254 /* Since we're called from mount, we know that
1255 * the recovery thread can't race us on
1256 * setting / checking the recovery bits. */
1257 ocfs2_recovery_thread(osb, node_num);
1258 } else if ((status < 0) && (status != -EAGAIN)) {
1263 spin_lock(&si->si_lock);
1265 spin_unlock(&si->si_lock);
1273 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1275 struct inode **head)
1278 struct inode *orphan_dir_inode = NULL;
1280 unsigned long offset, blk, local;
1281 struct buffer_head *bh = NULL;
1282 struct ocfs2_dir_entry *de;
1283 struct super_block *sb = osb->sb;
1285 orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1286 ORPHAN_DIR_SYSTEM_INODE,
1288 if (!orphan_dir_inode) {
1294 mutex_lock(&orphan_dir_inode->i_mutex);
1295 status = ocfs2_meta_lock(orphan_dir_inode, NULL, 0);
1303 while(offset < i_size_read(orphan_dir_inode)) {
1304 blk = offset >> sb->s_blocksize_bits;
1306 bh = ocfs2_bread(orphan_dir_inode, blk, &status, 0);
1317 while(offset < i_size_read(orphan_dir_inode)
1318 && local < sb->s_blocksize) {
1319 de = (struct ocfs2_dir_entry *) (bh->b_data + local);
1321 if (!ocfs2_check_dir_entry(orphan_dir_inode,
1329 local += le16_to_cpu(de->rec_len);
1330 offset += le16_to_cpu(de->rec_len);
1332 /* I guess we silently fail on no inode? */
1333 if (!le64_to_cpu(de->inode))
1335 if (de->file_type > OCFS2_FT_MAX) {
1337 "block %llu contains invalid de: "
1338 "inode = %llu, rec_len = %u, "
1339 "name_len = %u, file_type = %u, "
1341 (unsigned long long)bh->b_blocknr,
1342 (unsigned long long)le64_to_cpu(de->inode),
1343 le16_to_cpu(de->rec_len),
1350 if (de->name_len == 1 && !strncmp(".", de->name, 1))
1352 if (de->name_len == 2 && !strncmp("..", de->name, 2))
1355 iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
1356 OCFS2_FI_FLAG_NOLOCK);
1360 mlog(0, "queue orphan %llu\n",
1361 (unsigned long long)OCFS2_I(iter)->ip_blkno);
1362 /* No locking is required for the next_orphan
1363 * queue as there is only ever a single
1364 * process doing orphan recovery. */
1365 OCFS2_I(iter)->ip_next_orphan = *head;
1372 ocfs2_meta_unlock(orphan_dir_inode, 0);
1374 mutex_unlock(&orphan_dir_inode->i_mutex);
1375 iput(orphan_dir_inode);
1379 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1384 spin_lock(&osb->osb_lock);
1385 ret = !osb->osb_orphan_wipes[slot];
1386 spin_unlock(&osb->osb_lock);
1390 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1393 spin_lock(&osb->osb_lock);
1394 /* Mark ourselves such that new processes in delete_inode()
1395 * know to quit early. */
1396 ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1397 while (osb->osb_orphan_wipes[slot]) {
1398 /* If any processes are already in the middle of an
1399 * orphan wipe on this dir, then we need to wait for
1401 spin_unlock(&osb->osb_lock);
1402 wait_event_interruptible(osb->osb_wipe_event,
1403 ocfs2_orphan_recovery_can_continue(osb, slot));
1404 spin_lock(&osb->osb_lock);
1406 spin_unlock(&osb->osb_lock);
1409 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1412 ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1416 * Orphan recovery. Each mounted node has it's own orphan dir which we
1417 * must run during recovery. Our strategy here is to build a list of
1418 * the inodes in the orphan dir and iget/iput them. The VFS does
1419 * (most) of the rest of the work.
1421 * Orphan recovery can happen at any time, not just mount so we have a
1422 * couple of extra considerations.
1424 * - We grab as many inodes as we can under the orphan dir lock -
1425 * doing iget() outside the orphan dir risks getting a reference on
1427 * - We must be sure not to deadlock with other processes on the
1428 * system wanting to run delete_inode(). This can happen when they go
1429 * to lock the orphan dir and the orphan recovery process attempts to
1430 * iget() inside the orphan dir lock. This can be avoided by
1431 * advertising our state to ocfs2_delete_inode().
1433 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1437 struct inode *inode = NULL;
1439 struct ocfs2_inode_info *oi;
1441 mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1443 ocfs2_mark_recovering_orphan_dir(osb, slot);
1444 ret = ocfs2_queue_orphans(osb, slot, &inode);
1445 ocfs2_clear_recovering_orphan_dir(osb, slot);
1447 /* Error here should be noted, but we want to continue with as
1448 * many queued inodes as we've got. */
1453 oi = OCFS2_I(inode);
1454 mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1456 iter = oi->ip_next_orphan;
1458 spin_lock(&oi->ip_lock);
1459 /* Delete voting may have set these on the assumption
1460 * that the other node would wipe them successfully.
1461 * If they are still in the node's orphan dir, we need
1462 * to reset that state. */
1463 oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1465 /* Set the proper information to get us going into
1466 * ocfs2_delete_inode. */
1467 oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1468 oi->ip_orphaned_slot = slot;
1469 spin_unlock(&oi->ip_lock);
1479 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1481 /* This check is good because ocfs2 will wait on our recovery
1482 * thread before changing it to something other than MOUNTED
1484 wait_event(osb->osb_mount_event,
1485 atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1486 atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1488 /* If there's an error on mount, then we may never get to the
1489 * MOUNTED flag, but this is set right before
1490 * dismount_volume() so we can trust it. */
1491 if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1492 mlog(0, "mount error, exiting!\n");
1499 static int ocfs2_commit_thread(void *arg)
1502 struct ocfs2_super *osb = arg;
1503 struct ocfs2_journal *journal = osb->journal;
1505 /* we can trust j_num_trans here because _should_stop() is only set in
1506 * shutdown and nobody other than ourselves should be able to start
1507 * transactions. committing on shutdown might take a few iterations
1508 * as final transactions put deleted inodes on the list */
1509 while (!(kthread_should_stop() &&
1510 atomic_read(&journal->j_num_trans) == 0)) {
1512 wait_event_interruptible(osb->checkpoint_event,
1513 atomic_read(&journal->j_num_trans)
1514 || kthread_should_stop());
1516 status = ocfs2_commit_cache(osb);
1520 if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1522 "commit_thread: %u transactions pending on "
1524 atomic_read(&journal->j_num_trans));
1531 /* Look for a dirty journal without taking any cluster locks. Used for
1532 * hard readonly access to determine whether the file system journals
1533 * require recovery. */
1534 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1538 struct buffer_head *di_bh;
1539 struct ocfs2_dinode *di;
1540 struct inode *journal = NULL;
1542 for(slot = 0; slot < osb->max_slots; slot++) {
1543 journal = ocfs2_get_system_file_inode(osb,
1544 JOURNAL_SYSTEM_INODE,
1546 if (!journal || is_bad_inode(journal)) {
1553 ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1560 di = (struct ocfs2_dinode *) di_bh->b_data;
1562 if (le32_to_cpu(di->id1.journal1.ij_flags) &
1563 OCFS2_JOURNAL_DIRTY_FL)