xref: /openbmc/linux/fs/ocfs2/journal.c (revision 22246614)
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * journal.c
5  *
6  * Defines functions of journalling api
7  *
8  * Copyright (C) 2003, 2004 Oracle.  All rights reserved.
9  *
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.
14  *
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.
19  *
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.
24  */
25 
26 #include <linux/fs.h>
27 #include <linux/types.h>
28 #include <linux/slab.h>
29 #include <linux/highmem.h>
30 #include <linux/kthread.h>
31 
32 #define MLOG_MASK_PREFIX ML_JOURNAL
33 #include <cluster/masklog.h>
34 
35 #include "ocfs2.h"
36 
37 #include "alloc.h"
38 #include "dir.h"
39 #include "dlmglue.h"
40 #include "extent_map.h"
41 #include "heartbeat.h"
42 #include "inode.h"
43 #include "journal.h"
44 #include "localalloc.h"
45 #include "slot_map.h"
46 #include "super.h"
47 #include "sysfile.h"
48 
49 #include "buffer_head_io.h"
50 
51 DEFINE_SPINLOCK(trans_inc_lock);
52 
53 static int ocfs2_force_read_journal(struct inode *inode);
54 static int ocfs2_recover_node(struct ocfs2_super *osb,
55 			      int node_num);
56 static int __ocfs2_recovery_thread(void *arg);
57 static int ocfs2_commit_cache(struct ocfs2_super *osb);
58 static int ocfs2_wait_on_mount(struct ocfs2_super *osb);
59 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
60 				      int dirty);
61 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
62 				 int slot_num);
63 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
64 				 int slot);
65 static int ocfs2_commit_thread(void *arg);
66 
67 
68 /*
69  * The recovery_list is a simple linked list of node numbers to recover.
70  * It is protected by the recovery_lock.
71  */
72 
73 struct ocfs2_recovery_map {
74 	unsigned int rm_used;
75 	unsigned int *rm_entries;
76 };
77 
78 int ocfs2_recovery_init(struct ocfs2_super *osb)
79 {
80 	struct ocfs2_recovery_map *rm;
81 
82 	mutex_init(&osb->recovery_lock);
83 	osb->disable_recovery = 0;
84 	osb->recovery_thread_task = NULL;
85 	init_waitqueue_head(&osb->recovery_event);
86 
87 	rm = kzalloc(sizeof(struct ocfs2_recovery_map) +
88 		     osb->max_slots * sizeof(unsigned int),
89 		     GFP_KERNEL);
90 	if (!rm) {
91 		mlog_errno(-ENOMEM);
92 		return -ENOMEM;
93 	}
94 
95 	rm->rm_entries = (unsigned int *)((char *)rm +
96 					  sizeof(struct ocfs2_recovery_map));
97 	osb->recovery_map = rm;
98 
99 	return 0;
100 }
101 
102 /* we can't grab the goofy sem lock from inside wait_event, so we use
103  * memory barriers to make sure that we'll see the null task before
104  * being woken up */
105 static int ocfs2_recovery_thread_running(struct ocfs2_super *osb)
106 {
107 	mb();
108 	return osb->recovery_thread_task != NULL;
109 }
110 
111 void ocfs2_recovery_exit(struct ocfs2_super *osb)
112 {
113 	struct ocfs2_recovery_map *rm;
114 
115 	/* disable any new recovery threads and wait for any currently
116 	 * running ones to exit. Do this before setting the vol_state. */
117 	mutex_lock(&osb->recovery_lock);
118 	osb->disable_recovery = 1;
119 	mutex_unlock(&osb->recovery_lock);
120 	wait_event(osb->recovery_event, !ocfs2_recovery_thread_running(osb));
121 
122 	/* At this point, we know that no more recovery threads can be
123 	 * launched, so wait for any recovery completion work to
124 	 * complete. */
125 	flush_workqueue(ocfs2_wq);
126 
127 	/*
128 	 * Now that recovery is shut down, and the osb is about to be
129 	 * freed,  the osb_lock is not taken here.
130 	 */
131 	rm = osb->recovery_map;
132 	/* XXX: Should we bug if there are dirty entries? */
133 
134 	kfree(rm);
135 }
136 
137 static int __ocfs2_recovery_map_test(struct ocfs2_super *osb,
138 				     unsigned int node_num)
139 {
140 	int i;
141 	struct ocfs2_recovery_map *rm = osb->recovery_map;
142 
143 	assert_spin_locked(&osb->osb_lock);
144 
145 	for (i = 0; i < rm->rm_used; i++) {
146 		if (rm->rm_entries[i] == node_num)
147 			return 1;
148 	}
149 
150 	return 0;
151 }
152 
153 /* Behaves like test-and-set.  Returns the previous value */
154 static int ocfs2_recovery_map_set(struct ocfs2_super *osb,
155 				  unsigned int node_num)
156 {
157 	struct ocfs2_recovery_map *rm = osb->recovery_map;
158 
159 	spin_lock(&osb->osb_lock);
160 	if (__ocfs2_recovery_map_test(osb, node_num)) {
161 		spin_unlock(&osb->osb_lock);
162 		return 1;
163 	}
164 
165 	/* XXX: Can this be exploited? Not from o2dlm... */
166 	BUG_ON(rm->rm_used >= osb->max_slots);
167 
168 	rm->rm_entries[rm->rm_used] = node_num;
169 	rm->rm_used++;
170 	spin_unlock(&osb->osb_lock);
171 
172 	return 0;
173 }
174 
175 static void ocfs2_recovery_map_clear(struct ocfs2_super *osb,
176 				     unsigned int node_num)
177 {
178 	int i;
179 	struct ocfs2_recovery_map *rm = osb->recovery_map;
180 
181 	spin_lock(&osb->osb_lock);
182 
183 	for (i = 0; i < rm->rm_used; i++) {
184 		if (rm->rm_entries[i] == node_num)
185 			break;
186 	}
187 
188 	if (i < rm->rm_used) {
189 		/* XXX: be careful with the pointer math */
190 		memmove(&(rm->rm_entries[i]), &(rm->rm_entries[i + 1]),
191 			(rm->rm_used - i - 1) * sizeof(unsigned int));
192 		rm->rm_used--;
193 	}
194 
195 	spin_unlock(&osb->osb_lock);
196 }
197 
198 static int ocfs2_commit_cache(struct ocfs2_super *osb)
199 {
200 	int status = 0;
201 	unsigned int flushed;
202 	unsigned long old_id;
203 	struct ocfs2_journal *journal = NULL;
204 
205 	mlog_entry_void();
206 
207 	journal = osb->journal;
208 
209 	/* Flush all pending commits and checkpoint the journal. */
210 	down_write(&journal->j_trans_barrier);
211 
212 	if (atomic_read(&journal->j_num_trans) == 0) {
213 		up_write(&journal->j_trans_barrier);
214 		mlog(0, "No transactions for me to flush!\n");
215 		goto finally;
216 	}
217 
218 	journal_lock_updates(journal->j_journal);
219 	status = journal_flush(journal->j_journal);
220 	journal_unlock_updates(journal->j_journal);
221 	if (status < 0) {
222 		up_write(&journal->j_trans_barrier);
223 		mlog_errno(status);
224 		goto finally;
225 	}
226 
227 	old_id = ocfs2_inc_trans_id(journal);
228 
229 	flushed = atomic_read(&journal->j_num_trans);
230 	atomic_set(&journal->j_num_trans, 0);
231 	up_write(&journal->j_trans_barrier);
232 
233 	mlog(0, "commit_thread: flushed transaction %lu (%u handles)\n",
234 	     journal->j_trans_id, flushed);
235 
236 	ocfs2_wake_downconvert_thread(osb);
237 	wake_up(&journal->j_checkpointed);
238 finally:
239 	mlog_exit(status);
240 	return status;
241 }
242 
243 /* pass it NULL and it will allocate a new handle object for you.  If
244  * you pass it a handle however, it may still return error, in which
245  * case it has free'd the passed handle for you. */
246 handle_t *ocfs2_start_trans(struct ocfs2_super *osb, int max_buffs)
247 {
248 	journal_t *journal = osb->journal->j_journal;
249 	handle_t *handle;
250 
251 	BUG_ON(!osb || !osb->journal->j_journal);
252 
253 	if (ocfs2_is_hard_readonly(osb))
254 		return ERR_PTR(-EROFS);
255 
256 	BUG_ON(osb->journal->j_state == OCFS2_JOURNAL_FREE);
257 	BUG_ON(max_buffs <= 0);
258 
259 	/* JBD might support this, but our journalling code doesn't yet. */
260 	if (journal_current_handle()) {
261 		mlog(ML_ERROR, "Recursive transaction attempted!\n");
262 		BUG();
263 	}
264 
265 	down_read(&osb->journal->j_trans_barrier);
266 
267 	handle = journal_start(journal, max_buffs);
268 	if (IS_ERR(handle)) {
269 		up_read(&osb->journal->j_trans_barrier);
270 
271 		mlog_errno(PTR_ERR(handle));
272 
273 		if (is_journal_aborted(journal)) {
274 			ocfs2_abort(osb->sb, "Detected aborted journal");
275 			handle = ERR_PTR(-EROFS);
276 		}
277 	} else {
278 		if (!ocfs2_mount_local(osb))
279 			atomic_inc(&(osb->journal->j_num_trans));
280 	}
281 
282 	return handle;
283 }
284 
285 int ocfs2_commit_trans(struct ocfs2_super *osb,
286 		       handle_t *handle)
287 {
288 	int ret;
289 	struct ocfs2_journal *journal = osb->journal;
290 
291 	BUG_ON(!handle);
292 
293 	ret = journal_stop(handle);
294 	if (ret < 0)
295 		mlog_errno(ret);
296 
297 	up_read(&journal->j_trans_barrier);
298 
299 	return ret;
300 }
301 
302 /*
303  * 'nblocks' is what you want to add to the current
304  * transaction. extend_trans will either extend the current handle by
305  * nblocks, or commit it and start a new one with nblocks credits.
306  *
307  * This might call journal_restart() which will commit dirty buffers
308  * and then restart the transaction. Before calling
309  * ocfs2_extend_trans(), any changed blocks should have been
310  * dirtied. After calling it, all blocks which need to be changed must
311  * go through another set of journal_access/journal_dirty calls.
312  *
313  * WARNING: This will not release any semaphores or disk locks taken
314  * during the transaction, so make sure they were taken *before*
315  * start_trans or we'll have ordering deadlocks.
316  *
317  * WARNING2: Note that we do *not* drop j_trans_barrier here. This is
318  * good because transaction ids haven't yet been recorded on the
319  * cluster locks associated with this handle.
320  */
321 int ocfs2_extend_trans(handle_t *handle, int nblocks)
322 {
323 	int status;
324 
325 	BUG_ON(!handle);
326 	BUG_ON(!nblocks);
327 
328 	mlog_entry_void();
329 
330 	mlog(0, "Trying to extend transaction by %d blocks\n", nblocks);
331 
332 #ifdef OCFS2_DEBUG_FS
333 	status = 1;
334 #else
335 	status = journal_extend(handle, nblocks);
336 	if (status < 0) {
337 		mlog_errno(status);
338 		goto bail;
339 	}
340 #endif
341 
342 	if (status > 0) {
343 		mlog(0, "journal_extend failed, trying journal_restart\n");
344 		status = journal_restart(handle, nblocks);
345 		if (status < 0) {
346 			mlog_errno(status);
347 			goto bail;
348 		}
349 	}
350 
351 	status = 0;
352 bail:
353 
354 	mlog_exit(status);
355 	return status;
356 }
357 
358 int ocfs2_journal_access(handle_t *handle,
359 			 struct inode *inode,
360 			 struct buffer_head *bh,
361 			 int type)
362 {
363 	int status;
364 
365 	BUG_ON(!inode);
366 	BUG_ON(!handle);
367 	BUG_ON(!bh);
368 
369 	mlog_entry("bh->b_blocknr=%llu, type=%d (\"%s\"), bh->b_size = %zu\n",
370 		   (unsigned long long)bh->b_blocknr, type,
371 		   (type == OCFS2_JOURNAL_ACCESS_CREATE) ?
372 		   "OCFS2_JOURNAL_ACCESS_CREATE" :
373 		   "OCFS2_JOURNAL_ACCESS_WRITE",
374 		   bh->b_size);
375 
376 	/* we can safely remove this assertion after testing. */
377 	if (!buffer_uptodate(bh)) {
378 		mlog(ML_ERROR, "giving me a buffer that's not uptodate!\n");
379 		mlog(ML_ERROR, "b_blocknr=%llu\n",
380 		     (unsigned long long)bh->b_blocknr);
381 		BUG();
382 	}
383 
384 	/* Set the current transaction information on the inode so
385 	 * that the locking code knows whether it can drop it's locks
386 	 * on this inode or not. We're protected from the commit
387 	 * thread updating the current transaction id until
388 	 * ocfs2_commit_trans() because ocfs2_start_trans() took
389 	 * j_trans_barrier for us. */
390 	ocfs2_set_inode_lock_trans(OCFS2_SB(inode->i_sb)->journal, inode);
391 
392 	mutex_lock(&OCFS2_I(inode)->ip_io_mutex);
393 	switch (type) {
394 	case OCFS2_JOURNAL_ACCESS_CREATE:
395 	case OCFS2_JOURNAL_ACCESS_WRITE:
396 		status = journal_get_write_access(handle, bh);
397 		break;
398 
399 	case OCFS2_JOURNAL_ACCESS_UNDO:
400 		status = journal_get_undo_access(handle, bh);
401 		break;
402 
403 	default:
404 		status = -EINVAL;
405 		mlog(ML_ERROR, "Uknown access type!\n");
406 	}
407 	mutex_unlock(&OCFS2_I(inode)->ip_io_mutex);
408 
409 	if (status < 0)
410 		mlog(ML_ERROR, "Error %d getting %d access to buffer!\n",
411 		     status, type);
412 
413 	mlog_exit(status);
414 	return status;
415 }
416 
417 int ocfs2_journal_dirty(handle_t *handle,
418 			struct buffer_head *bh)
419 {
420 	int status;
421 
422 	mlog_entry("(bh->b_blocknr=%llu)\n",
423 		   (unsigned long long)bh->b_blocknr);
424 
425 	status = journal_dirty_metadata(handle, bh);
426 	if (status < 0)
427 		mlog(ML_ERROR, "Could not dirty metadata buffer. "
428 		     "(bh->b_blocknr=%llu)\n",
429 		     (unsigned long long)bh->b_blocknr);
430 
431 	mlog_exit(status);
432 	return status;
433 }
434 
435 int ocfs2_journal_dirty_data(handle_t *handle,
436 			     struct buffer_head *bh)
437 {
438 	int err = journal_dirty_data(handle, bh);
439 	if (err)
440 		mlog_errno(err);
441 	/* TODO: When we can handle it, abort the handle and go RO on
442 	 * error here. */
443 
444 	return err;
445 }
446 
447 #define OCFS2_DEFAULT_COMMIT_INTERVAL 	(HZ * JBD_DEFAULT_MAX_COMMIT_AGE)
448 
449 void ocfs2_set_journal_params(struct ocfs2_super *osb)
450 {
451 	journal_t *journal = osb->journal->j_journal;
452 	unsigned long commit_interval = OCFS2_DEFAULT_COMMIT_INTERVAL;
453 
454 	if (osb->osb_commit_interval)
455 		commit_interval = osb->osb_commit_interval;
456 
457 	spin_lock(&journal->j_state_lock);
458 	journal->j_commit_interval = commit_interval;
459 	if (osb->s_mount_opt & OCFS2_MOUNT_BARRIER)
460 		journal->j_flags |= JFS_BARRIER;
461 	else
462 		journal->j_flags &= ~JFS_BARRIER;
463 	spin_unlock(&journal->j_state_lock);
464 }
465 
466 int ocfs2_journal_init(struct ocfs2_journal *journal, int *dirty)
467 {
468 	int status = -1;
469 	struct inode *inode = NULL; /* the journal inode */
470 	journal_t *j_journal = NULL;
471 	struct ocfs2_dinode *di = NULL;
472 	struct buffer_head *bh = NULL;
473 	struct ocfs2_super *osb;
474 	int inode_lock = 0;
475 
476 	mlog_entry_void();
477 
478 	BUG_ON(!journal);
479 
480 	osb = journal->j_osb;
481 
482 	/* already have the inode for our journal */
483 	inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
484 					    osb->slot_num);
485 	if (inode == NULL) {
486 		status = -EACCES;
487 		mlog_errno(status);
488 		goto done;
489 	}
490 	if (is_bad_inode(inode)) {
491 		mlog(ML_ERROR, "access error (bad inode)\n");
492 		iput(inode);
493 		inode = NULL;
494 		status = -EACCES;
495 		goto done;
496 	}
497 
498 	SET_INODE_JOURNAL(inode);
499 	OCFS2_I(inode)->ip_open_count++;
500 
501 	/* Skip recovery waits here - journal inode metadata never
502 	 * changes in a live cluster so it can be considered an
503 	 * exception to the rule. */
504 	status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
505 	if (status < 0) {
506 		if (status != -ERESTARTSYS)
507 			mlog(ML_ERROR, "Could not get lock on journal!\n");
508 		goto done;
509 	}
510 
511 	inode_lock = 1;
512 	di = (struct ocfs2_dinode *)bh->b_data;
513 
514 	if (inode->i_size <  OCFS2_MIN_JOURNAL_SIZE) {
515 		mlog(ML_ERROR, "Journal file size (%lld) is too small!\n",
516 		     inode->i_size);
517 		status = -EINVAL;
518 		goto done;
519 	}
520 
521 	mlog(0, "inode->i_size = %lld\n", inode->i_size);
522 	mlog(0, "inode->i_blocks = %llu\n",
523 			(unsigned long long)inode->i_blocks);
524 	mlog(0, "inode->ip_clusters = %u\n", OCFS2_I(inode)->ip_clusters);
525 
526 	/* call the kernels journal init function now */
527 	j_journal = journal_init_inode(inode);
528 	if (j_journal == NULL) {
529 		mlog(ML_ERROR, "Linux journal layer error\n");
530 		status = -EINVAL;
531 		goto done;
532 	}
533 
534 	mlog(0, "Returned from journal_init_inode\n");
535 	mlog(0, "j_journal->j_maxlen = %u\n", j_journal->j_maxlen);
536 
537 	*dirty = (le32_to_cpu(di->id1.journal1.ij_flags) &
538 		  OCFS2_JOURNAL_DIRTY_FL);
539 
540 	journal->j_journal = j_journal;
541 	journal->j_inode = inode;
542 	journal->j_bh = bh;
543 
544 	ocfs2_set_journal_params(osb);
545 
546 	journal->j_state = OCFS2_JOURNAL_LOADED;
547 
548 	status = 0;
549 done:
550 	if (status < 0) {
551 		if (inode_lock)
552 			ocfs2_inode_unlock(inode, 1);
553 		if (bh != NULL)
554 			brelse(bh);
555 		if (inode) {
556 			OCFS2_I(inode)->ip_open_count--;
557 			iput(inode);
558 		}
559 	}
560 
561 	mlog_exit(status);
562 	return status;
563 }
564 
565 static int ocfs2_journal_toggle_dirty(struct ocfs2_super *osb,
566 				      int dirty)
567 {
568 	int status;
569 	unsigned int flags;
570 	struct ocfs2_journal *journal = osb->journal;
571 	struct buffer_head *bh = journal->j_bh;
572 	struct ocfs2_dinode *fe;
573 
574 	mlog_entry_void();
575 
576 	fe = (struct ocfs2_dinode *)bh->b_data;
577 	if (!OCFS2_IS_VALID_DINODE(fe)) {
578 		/* This is called from startup/shutdown which will
579 		 * handle the errors in a specific manner, so no need
580 		 * to call ocfs2_error() here. */
581 		mlog(ML_ERROR, "Journal dinode %llu  has invalid "
582 		     "signature: %.*s",
583 		     (unsigned long long)le64_to_cpu(fe->i_blkno), 7,
584 		     fe->i_signature);
585 		status = -EIO;
586 		goto out;
587 	}
588 
589 	flags = le32_to_cpu(fe->id1.journal1.ij_flags);
590 	if (dirty)
591 		flags |= OCFS2_JOURNAL_DIRTY_FL;
592 	else
593 		flags &= ~OCFS2_JOURNAL_DIRTY_FL;
594 	fe->id1.journal1.ij_flags = cpu_to_le32(flags);
595 
596 	status = ocfs2_write_block(osb, bh, journal->j_inode);
597 	if (status < 0)
598 		mlog_errno(status);
599 
600 out:
601 	mlog_exit(status);
602 	return status;
603 }
604 
605 /*
606  * If the journal has been kmalloc'd it needs to be freed after this
607  * call.
608  */
609 void ocfs2_journal_shutdown(struct ocfs2_super *osb)
610 {
611 	struct ocfs2_journal *journal = NULL;
612 	int status = 0;
613 	struct inode *inode = NULL;
614 	int num_running_trans = 0;
615 
616 	mlog_entry_void();
617 
618 	BUG_ON(!osb);
619 
620 	journal = osb->journal;
621 	if (!journal)
622 		goto done;
623 
624 	inode = journal->j_inode;
625 
626 	if (journal->j_state != OCFS2_JOURNAL_LOADED)
627 		goto done;
628 
629 	/* need to inc inode use count as journal_destroy will iput. */
630 	if (!igrab(inode))
631 		BUG();
632 
633 	num_running_trans = atomic_read(&(osb->journal->j_num_trans));
634 	if (num_running_trans > 0)
635 		mlog(0, "Shutting down journal: must wait on %d "
636 		     "running transactions!\n",
637 		     num_running_trans);
638 
639 	/* Do a commit_cache here. It will flush our journal, *and*
640 	 * release any locks that are still held.
641 	 * set the SHUTDOWN flag and release the trans lock.
642 	 * the commit thread will take the trans lock for us below. */
643 	journal->j_state = OCFS2_JOURNAL_IN_SHUTDOWN;
644 
645 	/* The OCFS2_JOURNAL_IN_SHUTDOWN will signal to commit_cache to not
646 	 * drop the trans_lock (which we want to hold until we
647 	 * completely destroy the journal. */
648 	if (osb->commit_task) {
649 		/* Wait for the commit thread */
650 		mlog(0, "Waiting for ocfs2commit to exit....\n");
651 		kthread_stop(osb->commit_task);
652 		osb->commit_task = NULL;
653 	}
654 
655 	BUG_ON(atomic_read(&(osb->journal->j_num_trans)) != 0);
656 
657 	if (ocfs2_mount_local(osb)) {
658 		journal_lock_updates(journal->j_journal);
659 		status = journal_flush(journal->j_journal);
660 		journal_unlock_updates(journal->j_journal);
661 		if (status < 0)
662 			mlog_errno(status);
663 	}
664 
665 	if (status == 0) {
666 		/*
667 		 * Do not toggle if flush was unsuccessful otherwise
668 		 * will leave dirty metadata in a "clean" journal
669 		 */
670 		status = ocfs2_journal_toggle_dirty(osb, 0);
671 		if (status < 0)
672 			mlog_errno(status);
673 	}
674 
675 	/* Shutdown the kernel journal system */
676 	journal_destroy(journal->j_journal);
677 
678 	OCFS2_I(inode)->ip_open_count--;
679 
680 	/* unlock our journal */
681 	ocfs2_inode_unlock(inode, 1);
682 
683 	brelse(journal->j_bh);
684 	journal->j_bh = NULL;
685 
686 	journal->j_state = OCFS2_JOURNAL_FREE;
687 
688 //	up_write(&journal->j_trans_barrier);
689 done:
690 	if (inode)
691 		iput(inode);
692 	mlog_exit_void();
693 }
694 
695 static void ocfs2_clear_journal_error(struct super_block *sb,
696 				      journal_t *journal,
697 				      int slot)
698 {
699 	int olderr;
700 
701 	olderr = journal_errno(journal);
702 	if (olderr) {
703 		mlog(ML_ERROR, "File system error %d recorded in "
704 		     "journal %u.\n", olderr, slot);
705 		mlog(ML_ERROR, "File system on device %s needs checking.\n",
706 		     sb->s_id);
707 
708 		journal_ack_err(journal);
709 		journal_clear_err(journal);
710 	}
711 }
712 
713 int ocfs2_journal_load(struct ocfs2_journal *journal, int local)
714 {
715 	int status = 0;
716 	struct ocfs2_super *osb;
717 
718 	mlog_entry_void();
719 
720 	BUG_ON(!journal);
721 
722 	osb = journal->j_osb;
723 
724 	status = journal_load(journal->j_journal);
725 	if (status < 0) {
726 		mlog(ML_ERROR, "Failed to load journal!\n");
727 		goto done;
728 	}
729 
730 	ocfs2_clear_journal_error(osb->sb, journal->j_journal, osb->slot_num);
731 
732 	status = ocfs2_journal_toggle_dirty(osb, 1);
733 	if (status < 0) {
734 		mlog_errno(status);
735 		goto done;
736 	}
737 
738 	/* Launch the commit thread */
739 	if (!local) {
740 		osb->commit_task = kthread_run(ocfs2_commit_thread, osb,
741 					       "ocfs2cmt");
742 		if (IS_ERR(osb->commit_task)) {
743 			status = PTR_ERR(osb->commit_task);
744 			osb->commit_task = NULL;
745 			mlog(ML_ERROR, "unable to launch ocfs2commit thread, "
746 			     "error=%d", status);
747 			goto done;
748 		}
749 	} else
750 		osb->commit_task = NULL;
751 
752 done:
753 	mlog_exit(status);
754 	return status;
755 }
756 
757 
758 /* 'full' flag tells us whether we clear out all blocks or if we just
759  * mark the journal clean */
760 int ocfs2_journal_wipe(struct ocfs2_journal *journal, int full)
761 {
762 	int status;
763 
764 	mlog_entry_void();
765 
766 	BUG_ON(!journal);
767 
768 	status = journal_wipe(journal->j_journal, full);
769 	if (status < 0) {
770 		mlog_errno(status);
771 		goto bail;
772 	}
773 
774 	status = ocfs2_journal_toggle_dirty(journal->j_osb, 0);
775 	if (status < 0)
776 		mlog_errno(status);
777 
778 bail:
779 	mlog_exit(status);
780 	return status;
781 }
782 
783 static int ocfs2_recovery_completed(struct ocfs2_super *osb)
784 {
785 	int empty;
786 	struct ocfs2_recovery_map *rm = osb->recovery_map;
787 
788 	spin_lock(&osb->osb_lock);
789 	empty = (rm->rm_used == 0);
790 	spin_unlock(&osb->osb_lock);
791 
792 	return empty;
793 }
794 
795 void ocfs2_wait_for_recovery(struct ocfs2_super *osb)
796 {
797 	wait_event(osb->recovery_event, ocfs2_recovery_completed(osb));
798 }
799 
800 /*
801  * JBD Might read a cached version of another nodes journal file. We
802  * don't want this as this file changes often and we get no
803  * notification on those changes. The only way to be sure that we've
804  * got the most up to date version of those blocks then is to force
805  * read them off disk. Just searching through the buffer cache won't
806  * work as there may be pages backing this file which are still marked
807  * up to date. We know things can't change on this file underneath us
808  * as we have the lock by now :)
809  */
810 static int ocfs2_force_read_journal(struct inode *inode)
811 {
812 	int status = 0;
813 	int i;
814 	u64 v_blkno, p_blkno, p_blocks, num_blocks;
815 #define CONCURRENT_JOURNAL_FILL 32ULL
816 	struct buffer_head *bhs[CONCURRENT_JOURNAL_FILL];
817 
818 	mlog_entry_void();
819 
820 	memset(bhs, 0, sizeof(struct buffer_head *) * CONCURRENT_JOURNAL_FILL);
821 
822 	num_blocks = ocfs2_blocks_for_bytes(inode->i_sb, inode->i_size);
823 	v_blkno = 0;
824 	while (v_blkno < num_blocks) {
825 		status = ocfs2_extent_map_get_blocks(inode, v_blkno,
826 						     &p_blkno, &p_blocks, NULL);
827 		if (status < 0) {
828 			mlog_errno(status);
829 			goto bail;
830 		}
831 
832 		if (p_blocks > CONCURRENT_JOURNAL_FILL)
833 			p_blocks = CONCURRENT_JOURNAL_FILL;
834 
835 		/* We are reading journal data which should not
836 		 * be put in the uptodate cache */
837 		status = ocfs2_read_blocks(OCFS2_SB(inode->i_sb),
838 					   p_blkno, p_blocks, bhs, 0,
839 					   NULL);
840 		if (status < 0) {
841 			mlog_errno(status);
842 			goto bail;
843 		}
844 
845 		for(i = 0; i < p_blocks; i++) {
846 			brelse(bhs[i]);
847 			bhs[i] = NULL;
848 		}
849 
850 		v_blkno += p_blocks;
851 	}
852 
853 bail:
854 	for(i = 0; i < CONCURRENT_JOURNAL_FILL; i++)
855 		if (bhs[i])
856 			brelse(bhs[i]);
857 	mlog_exit(status);
858 	return status;
859 }
860 
861 struct ocfs2_la_recovery_item {
862 	struct list_head	lri_list;
863 	int			lri_slot;
864 	struct ocfs2_dinode	*lri_la_dinode;
865 	struct ocfs2_dinode	*lri_tl_dinode;
866 };
867 
868 /* Does the second half of the recovery process. By this point, the
869  * node is marked clean and can actually be considered recovered,
870  * hence it's no longer in the recovery map, but there's still some
871  * cleanup we can do which shouldn't happen within the recovery thread
872  * as locking in that context becomes very difficult if we are to take
873  * recovering nodes into account.
874  *
875  * NOTE: This function can and will sleep on recovery of other nodes
876  * during cluster locking, just like any other ocfs2 process.
877  */
878 void ocfs2_complete_recovery(struct work_struct *work)
879 {
880 	int ret;
881 	struct ocfs2_journal *journal =
882 		container_of(work, struct ocfs2_journal, j_recovery_work);
883 	struct ocfs2_super *osb = journal->j_osb;
884 	struct ocfs2_dinode *la_dinode, *tl_dinode;
885 	struct ocfs2_la_recovery_item *item, *n;
886 	LIST_HEAD(tmp_la_list);
887 
888 	mlog_entry_void();
889 
890 	mlog(0, "completing recovery from keventd\n");
891 
892 	spin_lock(&journal->j_lock);
893 	list_splice_init(&journal->j_la_cleanups, &tmp_la_list);
894 	spin_unlock(&journal->j_lock);
895 
896 	list_for_each_entry_safe(item, n, &tmp_la_list, lri_list) {
897 		list_del_init(&item->lri_list);
898 
899 		mlog(0, "Complete recovery for slot %d\n", item->lri_slot);
900 
901 		la_dinode = item->lri_la_dinode;
902 		if (la_dinode) {
903 			mlog(0, "Clean up local alloc %llu\n",
904 			     (unsigned long long)le64_to_cpu(la_dinode->i_blkno));
905 
906 			ret = ocfs2_complete_local_alloc_recovery(osb,
907 								  la_dinode);
908 			if (ret < 0)
909 				mlog_errno(ret);
910 
911 			kfree(la_dinode);
912 		}
913 
914 		tl_dinode = item->lri_tl_dinode;
915 		if (tl_dinode) {
916 			mlog(0, "Clean up truncate log %llu\n",
917 			     (unsigned long long)le64_to_cpu(tl_dinode->i_blkno));
918 
919 			ret = ocfs2_complete_truncate_log_recovery(osb,
920 								   tl_dinode);
921 			if (ret < 0)
922 				mlog_errno(ret);
923 
924 			kfree(tl_dinode);
925 		}
926 
927 		ret = ocfs2_recover_orphans(osb, item->lri_slot);
928 		if (ret < 0)
929 			mlog_errno(ret);
930 
931 		kfree(item);
932 	}
933 
934 	mlog(0, "Recovery completion\n");
935 	mlog_exit_void();
936 }
937 
938 /* NOTE: This function always eats your references to la_dinode and
939  * tl_dinode, either manually on error, or by passing them to
940  * ocfs2_complete_recovery */
941 static void ocfs2_queue_recovery_completion(struct ocfs2_journal *journal,
942 					    int slot_num,
943 					    struct ocfs2_dinode *la_dinode,
944 					    struct ocfs2_dinode *tl_dinode)
945 {
946 	struct ocfs2_la_recovery_item *item;
947 
948 	item = kmalloc(sizeof(struct ocfs2_la_recovery_item), GFP_NOFS);
949 	if (!item) {
950 		/* Though we wish to avoid it, we are in fact safe in
951 		 * skipping local alloc cleanup as fsck.ocfs2 is more
952 		 * than capable of reclaiming unused space. */
953 		if (la_dinode)
954 			kfree(la_dinode);
955 
956 		if (tl_dinode)
957 			kfree(tl_dinode);
958 
959 		mlog_errno(-ENOMEM);
960 		return;
961 	}
962 
963 	INIT_LIST_HEAD(&item->lri_list);
964 	item->lri_la_dinode = la_dinode;
965 	item->lri_slot = slot_num;
966 	item->lri_tl_dinode = tl_dinode;
967 
968 	spin_lock(&journal->j_lock);
969 	list_add_tail(&item->lri_list, &journal->j_la_cleanups);
970 	queue_work(ocfs2_wq, &journal->j_recovery_work);
971 	spin_unlock(&journal->j_lock);
972 }
973 
974 /* Called by the mount code to queue recovery the last part of
975  * recovery for it's own slot. */
976 void ocfs2_complete_mount_recovery(struct ocfs2_super *osb)
977 {
978 	struct ocfs2_journal *journal = osb->journal;
979 
980 	if (osb->dirty) {
981 		/* No need to queue up our truncate_log as regular
982 		 * cleanup will catch that. */
983 		ocfs2_queue_recovery_completion(journal,
984 						osb->slot_num,
985 						osb->local_alloc_copy,
986 						NULL);
987 		ocfs2_schedule_truncate_log_flush(osb, 0);
988 
989 		osb->local_alloc_copy = NULL;
990 		osb->dirty = 0;
991 	}
992 }
993 
994 static int __ocfs2_recovery_thread(void *arg)
995 {
996 	int status, node_num;
997 	struct ocfs2_super *osb = arg;
998 	struct ocfs2_recovery_map *rm = osb->recovery_map;
999 
1000 	mlog_entry_void();
1001 
1002 	status = ocfs2_wait_on_mount(osb);
1003 	if (status < 0) {
1004 		goto bail;
1005 	}
1006 
1007 restart:
1008 	status = ocfs2_super_lock(osb, 1);
1009 	if (status < 0) {
1010 		mlog_errno(status);
1011 		goto bail;
1012 	}
1013 
1014 	spin_lock(&osb->osb_lock);
1015 	while (rm->rm_used) {
1016 		/* It's always safe to remove entry zero, as we won't
1017 		 * clear it until ocfs2_recover_node() has succeeded. */
1018 		node_num = rm->rm_entries[0];
1019 		spin_unlock(&osb->osb_lock);
1020 
1021 		status = ocfs2_recover_node(osb, node_num);
1022 		if (!status) {
1023 			ocfs2_recovery_map_clear(osb, node_num);
1024 		} else {
1025 			mlog(ML_ERROR,
1026 			     "Error %d recovering node %d on device (%u,%u)!\n",
1027 			     status, node_num,
1028 			     MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1029 			mlog(ML_ERROR, "Volume requires unmount.\n");
1030 		}
1031 
1032 		spin_lock(&osb->osb_lock);
1033 	}
1034 	spin_unlock(&osb->osb_lock);
1035 	mlog(0, "All nodes recovered\n");
1036 
1037 	ocfs2_super_unlock(osb, 1);
1038 
1039 	/* We always run recovery on our own orphan dir - the dead
1040 	 * node(s) may have disallowd a previos inode delete. Re-processing
1041 	 * is therefore required. */
1042 	ocfs2_queue_recovery_completion(osb->journal, osb->slot_num, NULL,
1043 					NULL);
1044 
1045 bail:
1046 	mutex_lock(&osb->recovery_lock);
1047 	if (!status && !ocfs2_recovery_completed(osb)) {
1048 		mutex_unlock(&osb->recovery_lock);
1049 		goto restart;
1050 	}
1051 
1052 	osb->recovery_thread_task = NULL;
1053 	mb(); /* sync with ocfs2_recovery_thread_running */
1054 	wake_up(&osb->recovery_event);
1055 
1056 	mutex_unlock(&osb->recovery_lock);
1057 
1058 	mlog_exit(status);
1059 	/* no one is callint kthread_stop() for us so the kthread() api
1060 	 * requires that we call do_exit().  And it isn't exported, but
1061 	 * complete_and_exit() seems to be a minimal wrapper around it. */
1062 	complete_and_exit(NULL, status);
1063 	return status;
1064 }
1065 
1066 void ocfs2_recovery_thread(struct ocfs2_super *osb, int node_num)
1067 {
1068 	mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1069 		   node_num, osb->node_num);
1070 
1071 	mutex_lock(&osb->recovery_lock);
1072 	if (osb->disable_recovery)
1073 		goto out;
1074 
1075 	/* People waiting on recovery will wait on
1076 	 * the recovery map to empty. */
1077 	if (ocfs2_recovery_map_set(osb, node_num))
1078 		mlog(0, "node %d already in recovery map.\n", node_num);
1079 
1080 	mlog(0, "starting recovery thread...\n");
1081 
1082 	if (osb->recovery_thread_task)
1083 		goto out;
1084 
1085 	osb->recovery_thread_task =  kthread_run(__ocfs2_recovery_thread, osb,
1086 						 "ocfs2rec");
1087 	if (IS_ERR(osb->recovery_thread_task)) {
1088 		mlog_errno((int)PTR_ERR(osb->recovery_thread_task));
1089 		osb->recovery_thread_task = NULL;
1090 	}
1091 
1092 out:
1093 	mutex_unlock(&osb->recovery_lock);
1094 	wake_up(&osb->recovery_event);
1095 
1096 	mlog_exit_void();
1097 }
1098 
1099 /* Does the actual journal replay and marks the journal inode as
1100  * clean. Will only replay if the journal inode is marked dirty. */
1101 static int ocfs2_replay_journal(struct ocfs2_super *osb,
1102 				int node_num,
1103 				int slot_num)
1104 {
1105 	int status;
1106 	int got_lock = 0;
1107 	unsigned int flags;
1108 	struct inode *inode = NULL;
1109 	struct ocfs2_dinode *fe;
1110 	journal_t *journal = NULL;
1111 	struct buffer_head *bh = NULL;
1112 
1113 	inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1114 					    slot_num);
1115 	if (inode == NULL) {
1116 		status = -EACCES;
1117 		mlog_errno(status);
1118 		goto done;
1119 	}
1120 	if (is_bad_inode(inode)) {
1121 		status = -EACCES;
1122 		iput(inode);
1123 		inode = NULL;
1124 		mlog_errno(status);
1125 		goto done;
1126 	}
1127 	SET_INODE_JOURNAL(inode);
1128 
1129 	status = ocfs2_inode_lock_full(inode, &bh, 1, OCFS2_META_LOCK_RECOVERY);
1130 	if (status < 0) {
1131 		mlog(0, "status returned from ocfs2_inode_lock=%d\n", status);
1132 		if (status != -ERESTARTSYS)
1133 			mlog(ML_ERROR, "Could not lock journal!\n");
1134 		goto done;
1135 	}
1136 	got_lock = 1;
1137 
1138 	fe = (struct ocfs2_dinode *) bh->b_data;
1139 
1140 	flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1141 
1142 	if (!(flags & OCFS2_JOURNAL_DIRTY_FL)) {
1143 		mlog(0, "No recovery required for node %d\n", node_num);
1144 		goto done;
1145 	}
1146 
1147 	mlog(ML_NOTICE, "Recovering node %d from slot %d on device (%u,%u)\n",
1148 	     node_num, slot_num,
1149 	     MAJOR(osb->sb->s_dev), MINOR(osb->sb->s_dev));
1150 
1151 	OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
1152 
1153 	status = ocfs2_force_read_journal(inode);
1154 	if (status < 0) {
1155 		mlog_errno(status);
1156 		goto done;
1157 	}
1158 
1159 	mlog(0, "calling journal_init_inode\n");
1160 	journal = journal_init_inode(inode);
1161 	if (journal == NULL) {
1162 		mlog(ML_ERROR, "Linux journal layer error\n");
1163 		status = -EIO;
1164 		goto done;
1165 	}
1166 
1167 	status = journal_load(journal);
1168 	if (status < 0) {
1169 		mlog_errno(status);
1170 		if (!igrab(inode))
1171 			BUG();
1172 		journal_destroy(journal);
1173 		goto done;
1174 	}
1175 
1176 	ocfs2_clear_journal_error(osb->sb, journal, slot_num);
1177 
1178 	/* wipe the journal */
1179 	mlog(0, "flushing the journal.\n");
1180 	journal_lock_updates(journal);
1181 	status = journal_flush(journal);
1182 	journal_unlock_updates(journal);
1183 	if (status < 0)
1184 		mlog_errno(status);
1185 
1186 	/* This will mark the node clean */
1187 	flags = le32_to_cpu(fe->id1.journal1.ij_flags);
1188 	flags &= ~OCFS2_JOURNAL_DIRTY_FL;
1189 	fe->id1.journal1.ij_flags = cpu_to_le32(flags);
1190 
1191 	status = ocfs2_write_block(osb, bh, inode);
1192 	if (status < 0)
1193 		mlog_errno(status);
1194 
1195 	if (!igrab(inode))
1196 		BUG();
1197 
1198 	journal_destroy(journal);
1199 
1200 done:
1201 	/* drop the lock on this nodes journal */
1202 	if (got_lock)
1203 		ocfs2_inode_unlock(inode, 1);
1204 
1205 	if (inode)
1206 		iput(inode);
1207 
1208 	if (bh)
1209 		brelse(bh);
1210 
1211 	mlog_exit(status);
1212 	return status;
1213 }
1214 
1215 /*
1216  * Do the most important parts of node recovery:
1217  *  - Replay it's journal
1218  *  - Stamp a clean local allocator file
1219  *  - Stamp a clean truncate log
1220  *  - Mark the node clean
1221  *
1222  * If this function completes without error, a node in OCFS2 can be
1223  * said to have been safely recovered. As a result, failure during the
1224  * second part of a nodes recovery process (local alloc recovery) is
1225  * far less concerning.
1226  */
1227 static int ocfs2_recover_node(struct ocfs2_super *osb,
1228 			      int node_num)
1229 {
1230 	int status = 0;
1231 	int slot_num;
1232 	struct ocfs2_dinode *la_copy = NULL;
1233 	struct ocfs2_dinode *tl_copy = NULL;
1234 
1235 	mlog_entry("(node_num=%d, osb->node_num = %d)\n",
1236 		   node_num, osb->node_num);
1237 
1238 	mlog(0, "checking node %d\n", node_num);
1239 
1240 	/* Should not ever be called to recover ourselves -- in that
1241 	 * case we should've called ocfs2_journal_load instead. */
1242 	BUG_ON(osb->node_num == node_num);
1243 
1244 	slot_num = ocfs2_node_num_to_slot(osb, node_num);
1245 	if (slot_num == -ENOENT) {
1246 		status = 0;
1247 		mlog(0, "no slot for this node, so no recovery required.\n");
1248 		goto done;
1249 	}
1250 
1251 	mlog(0, "node %d was using slot %d\n", node_num, slot_num);
1252 
1253 	status = ocfs2_replay_journal(osb, node_num, slot_num);
1254 	if (status < 0) {
1255 		mlog_errno(status);
1256 		goto done;
1257 	}
1258 
1259 	/* Stamp a clean local alloc file AFTER recovering the journal... */
1260 	status = ocfs2_begin_local_alloc_recovery(osb, slot_num, &la_copy);
1261 	if (status < 0) {
1262 		mlog_errno(status);
1263 		goto done;
1264 	}
1265 
1266 	/* An error from begin_truncate_log_recovery is not
1267 	 * serious enough to warrant halting the rest of
1268 	 * recovery. */
1269 	status = ocfs2_begin_truncate_log_recovery(osb, slot_num, &tl_copy);
1270 	if (status < 0)
1271 		mlog_errno(status);
1272 
1273 	/* Likewise, this would be a strange but ultimately not so
1274 	 * harmful place to get an error... */
1275 	status = ocfs2_clear_slot(osb, slot_num);
1276 	if (status < 0)
1277 		mlog_errno(status);
1278 
1279 	/* This will kfree the memory pointed to by la_copy and tl_copy */
1280 	ocfs2_queue_recovery_completion(osb->journal, slot_num, la_copy,
1281 					tl_copy);
1282 
1283 	status = 0;
1284 done:
1285 
1286 	mlog_exit(status);
1287 	return status;
1288 }
1289 
1290 /* Test node liveness by trylocking his journal. If we get the lock,
1291  * we drop it here. Return 0 if we got the lock, -EAGAIN if node is
1292  * still alive (we couldn't get the lock) and < 0 on error. */
1293 static int ocfs2_trylock_journal(struct ocfs2_super *osb,
1294 				 int slot_num)
1295 {
1296 	int status, flags;
1297 	struct inode *inode = NULL;
1298 
1299 	inode = ocfs2_get_system_file_inode(osb, JOURNAL_SYSTEM_INODE,
1300 					    slot_num);
1301 	if (inode == NULL) {
1302 		mlog(ML_ERROR, "access error\n");
1303 		status = -EACCES;
1304 		goto bail;
1305 	}
1306 	if (is_bad_inode(inode)) {
1307 		mlog(ML_ERROR, "access error (bad inode)\n");
1308 		iput(inode);
1309 		inode = NULL;
1310 		status = -EACCES;
1311 		goto bail;
1312 	}
1313 	SET_INODE_JOURNAL(inode);
1314 
1315 	flags = OCFS2_META_LOCK_RECOVERY | OCFS2_META_LOCK_NOQUEUE;
1316 	status = ocfs2_inode_lock_full(inode, NULL, 1, flags);
1317 	if (status < 0) {
1318 		if (status != -EAGAIN)
1319 			mlog_errno(status);
1320 		goto bail;
1321 	}
1322 
1323 	ocfs2_inode_unlock(inode, 1);
1324 bail:
1325 	if (inode)
1326 		iput(inode);
1327 
1328 	return status;
1329 }
1330 
1331 /* Call this underneath ocfs2_super_lock. It also assumes that the
1332  * slot info struct has been updated from disk. */
1333 int ocfs2_mark_dead_nodes(struct ocfs2_super *osb)
1334 {
1335 	unsigned int node_num;
1336 	int status, i;
1337 
1338 	/* This is called with the super block cluster lock, so we
1339 	 * know that the slot map can't change underneath us. */
1340 
1341 	spin_lock(&osb->osb_lock);
1342 	for (i = 0; i < osb->max_slots; i++) {
1343 		if (i == osb->slot_num)
1344 			continue;
1345 
1346 		status = ocfs2_slot_to_node_num_locked(osb, i, &node_num);
1347 		if (status == -ENOENT)
1348 			continue;
1349 
1350 		if (__ocfs2_recovery_map_test(osb, node_num))
1351 			continue;
1352 		spin_unlock(&osb->osb_lock);
1353 
1354 		/* Ok, we have a slot occupied by another node which
1355 		 * is not in the recovery map. We trylock his journal
1356 		 * file here to test if he's alive. */
1357 		status = ocfs2_trylock_journal(osb, i);
1358 		if (!status) {
1359 			/* Since we're called from mount, we know that
1360 			 * the recovery thread can't race us on
1361 			 * setting / checking the recovery bits. */
1362 			ocfs2_recovery_thread(osb, node_num);
1363 		} else if ((status < 0) && (status != -EAGAIN)) {
1364 			mlog_errno(status);
1365 			goto bail;
1366 		}
1367 
1368 		spin_lock(&osb->osb_lock);
1369 	}
1370 	spin_unlock(&osb->osb_lock);
1371 
1372 	status = 0;
1373 bail:
1374 	mlog_exit(status);
1375 	return status;
1376 }
1377 
1378 struct ocfs2_orphan_filldir_priv {
1379 	struct inode		*head;
1380 	struct ocfs2_super	*osb;
1381 };
1382 
1383 static int ocfs2_orphan_filldir(void *priv, const char *name, int name_len,
1384 				loff_t pos, u64 ino, unsigned type)
1385 {
1386 	struct ocfs2_orphan_filldir_priv *p = priv;
1387 	struct inode *iter;
1388 
1389 	if (name_len == 1 && !strncmp(".", name, 1))
1390 		return 0;
1391 	if (name_len == 2 && !strncmp("..", name, 2))
1392 		return 0;
1393 
1394 	/* Skip bad inodes so that recovery can continue */
1395 	iter = ocfs2_iget(p->osb, ino,
1396 			  OCFS2_FI_FLAG_ORPHAN_RECOVERY, 0);
1397 	if (IS_ERR(iter))
1398 		return 0;
1399 
1400 	mlog(0, "queue orphan %llu\n",
1401 	     (unsigned long long)OCFS2_I(iter)->ip_blkno);
1402 	/* No locking is required for the next_orphan queue as there
1403 	 * is only ever a single process doing orphan recovery. */
1404 	OCFS2_I(iter)->ip_next_orphan = p->head;
1405 	p->head = iter;
1406 
1407 	return 0;
1408 }
1409 
1410 static int ocfs2_queue_orphans(struct ocfs2_super *osb,
1411 			       int slot,
1412 			       struct inode **head)
1413 {
1414 	int status;
1415 	struct inode *orphan_dir_inode = NULL;
1416 	struct ocfs2_orphan_filldir_priv priv;
1417 	loff_t pos = 0;
1418 
1419 	priv.osb = osb;
1420 	priv.head = *head;
1421 
1422 	orphan_dir_inode = ocfs2_get_system_file_inode(osb,
1423 						       ORPHAN_DIR_SYSTEM_INODE,
1424 						       slot);
1425 	if  (!orphan_dir_inode) {
1426 		status = -ENOENT;
1427 		mlog_errno(status);
1428 		return status;
1429 	}
1430 
1431 	mutex_lock(&orphan_dir_inode->i_mutex);
1432 	status = ocfs2_inode_lock(orphan_dir_inode, NULL, 0);
1433 	if (status < 0) {
1434 		mlog_errno(status);
1435 		goto out;
1436 	}
1437 
1438 	status = ocfs2_dir_foreach(orphan_dir_inode, &pos, &priv,
1439 				   ocfs2_orphan_filldir);
1440 	if (status) {
1441 		mlog_errno(status);
1442 		goto out_cluster;
1443 	}
1444 
1445 	*head = priv.head;
1446 
1447 out_cluster:
1448 	ocfs2_inode_unlock(orphan_dir_inode, 0);
1449 out:
1450 	mutex_unlock(&orphan_dir_inode->i_mutex);
1451 	iput(orphan_dir_inode);
1452 	return status;
1453 }
1454 
1455 static int ocfs2_orphan_recovery_can_continue(struct ocfs2_super *osb,
1456 					      int slot)
1457 {
1458 	int ret;
1459 
1460 	spin_lock(&osb->osb_lock);
1461 	ret = !osb->osb_orphan_wipes[slot];
1462 	spin_unlock(&osb->osb_lock);
1463 	return ret;
1464 }
1465 
1466 static void ocfs2_mark_recovering_orphan_dir(struct ocfs2_super *osb,
1467 					     int slot)
1468 {
1469 	spin_lock(&osb->osb_lock);
1470 	/* Mark ourselves such that new processes in delete_inode()
1471 	 * know to quit early. */
1472 	ocfs2_node_map_set_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1473 	while (osb->osb_orphan_wipes[slot]) {
1474 		/* If any processes are already in the middle of an
1475 		 * orphan wipe on this dir, then we need to wait for
1476 		 * them. */
1477 		spin_unlock(&osb->osb_lock);
1478 		wait_event_interruptible(osb->osb_wipe_event,
1479 					 ocfs2_orphan_recovery_can_continue(osb, slot));
1480 		spin_lock(&osb->osb_lock);
1481 	}
1482 	spin_unlock(&osb->osb_lock);
1483 }
1484 
1485 static void ocfs2_clear_recovering_orphan_dir(struct ocfs2_super *osb,
1486 					      int slot)
1487 {
1488 	ocfs2_node_map_clear_bit(osb, &osb->osb_recovering_orphan_dirs, slot);
1489 }
1490 
1491 /*
1492  * Orphan recovery. Each mounted node has it's own orphan dir which we
1493  * must run during recovery. Our strategy here is to build a list of
1494  * the inodes in the orphan dir and iget/iput them. The VFS does
1495  * (most) of the rest of the work.
1496  *
1497  * Orphan recovery can happen at any time, not just mount so we have a
1498  * couple of extra considerations.
1499  *
1500  * - We grab as many inodes as we can under the orphan dir lock -
1501  *   doing iget() outside the orphan dir risks getting a reference on
1502  *   an invalid inode.
1503  * - We must be sure not to deadlock with other processes on the
1504  *   system wanting to run delete_inode(). This can happen when they go
1505  *   to lock the orphan dir and the orphan recovery process attempts to
1506  *   iget() inside the orphan dir lock. This can be avoided by
1507  *   advertising our state to ocfs2_delete_inode().
1508  */
1509 static int ocfs2_recover_orphans(struct ocfs2_super *osb,
1510 				 int slot)
1511 {
1512 	int ret = 0;
1513 	struct inode *inode = NULL;
1514 	struct inode *iter;
1515 	struct ocfs2_inode_info *oi;
1516 
1517 	mlog(0, "Recover inodes from orphan dir in slot %d\n", slot);
1518 
1519 	ocfs2_mark_recovering_orphan_dir(osb, slot);
1520 	ret = ocfs2_queue_orphans(osb, slot, &inode);
1521 	ocfs2_clear_recovering_orphan_dir(osb, slot);
1522 
1523 	/* Error here should be noted, but we want to continue with as
1524 	 * many queued inodes as we've got. */
1525 	if (ret)
1526 		mlog_errno(ret);
1527 
1528 	while (inode) {
1529 		oi = OCFS2_I(inode);
1530 		mlog(0, "iput orphan %llu\n", (unsigned long long)oi->ip_blkno);
1531 
1532 		iter = oi->ip_next_orphan;
1533 
1534 		spin_lock(&oi->ip_lock);
1535 		/* The remote delete code may have set these on the
1536 		 * assumption that the other node would wipe them
1537 		 * successfully.  If they are still in the node's
1538 		 * orphan dir, we need to reset that state. */
1539 		oi->ip_flags &= ~(OCFS2_INODE_DELETED|OCFS2_INODE_SKIP_DELETE);
1540 
1541 		/* Set the proper information to get us going into
1542 		 * ocfs2_delete_inode. */
1543 		oi->ip_flags |= OCFS2_INODE_MAYBE_ORPHANED;
1544 		spin_unlock(&oi->ip_lock);
1545 
1546 		iput(inode);
1547 
1548 		inode = iter;
1549 	}
1550 
1551 	return ret;
1552 }
1553 
1554 static int ocfs2_wait_on_mount(struct ocfs2_super *osb)
1555 {
1556 	/* This check is good because ocfs2 will wait on our recovery
1557 	 * thread before changing it to something other than MOUNTED
1558 	 * or DISABLED. */
1559 	wait_event(osb->osb_mount_event,
1560 		   atomic_read(&osb->vol_state) == VOLUME_MOUNTED ||
1561 		   atomic_read(&osb->vol_state) == VOLUME_DISABLED);
1562 
1563 	/* If there's an error on mount, then we may never get to the
1564 	 * MOUNTED flag, but this is set right before
1565 	 * dismount_volume() so we can trust it. */
1566 	if (atomic_read(&osb->vol_state) == VOLUME_DISABLED) {
1567 		mlog(0, "mount error, exiting!\n");
1568 		return -EBUSY;
1569 	}
1570 
1571 	return 0;
1572 }
1573 
1574 static int ocfs2_commit_thread(void *arg)
1575 {
1576 	int status;
1577 	struct ocfs2_super *osb = arg;
1578 	struct ocfs2_journal *journal = osb->journal;
1579 
1580 	/* we can trust j_num_trans here because _should_stop() is only set in
1581 	 * shutdown and nobody other than ourselves should be able to start
1582 	 * transactions.  committing on shutdown might take a few iterations
1583 	 * as final transactions put deleted inodes on the list */
1584 	while (!(kthread_should_stop() &&
1585 		 atomic_read(&journal->j_num_trans) == 0)) {
1586 
1587 		wait_event_interruptible(osb->checkpoint_event,
1588 					 atomic_read(&journal->j_num_trans)
1589 					 || kthread_should_stop());
1590 
1591 		status = ocfs2_commit_cache(osb);
1592 		if (status < 0)
1593 			mlog_errno(status);
1594 
1595 		if (kthread_should_stop() && atomic_read(&journal->j_num_trans)){
1596 			mlog(ML_KTHREAD,
1597 			     "commit_thread: %u transactions pending on "
1598 			     "shutdown\n",
1599 			     atomic_read(&journal->j_num_trans));
1600 		}
1601 	}
1602 
1603 	return 0;
1604 }
1605 
1606 /* Look for a dirty journal without taking any cluster locks. Used for
1607  * hard readonly access to determine whether the file system journals
1608  * require recovery. */
1609 int ocfs2_check_journals_nolocks(struct ocfs2_super *osb)
1610 {
1611 	int ret = 0;
1612 	unsigned int slot;
1613 	struct buffer_head *di_bh;
1614 	struct ocfs2_dinode *di;
1615 	struct inode *journal = NULL;
1616 
1617 	for(slot = 0; slot < osb->max_slots; slot++) {
1618 		journal = ocfs2_get_system_file_inode(osb,
1619 						      JOURNAL_SYSTEM_INODE,
1620 						      slot);
1621 		if (!journal || is_bad_inode(journal)) {
1622 			ret = -EACCES;
1623 			mlog_errno(ret);
1624 			goto out;
1625 		}
1626 
1627 		di_bh = NULL;
1628 		ret = ocfs2_read_block(osb, OCFS2_I(journal)->ip_blkno, &di_bh,
1629 				       0, journal);
1630 		if (ret < 0) {
1631 			mlog_errno(ret);
1632 			goto out;
1633 		}
1634 
1635 		di = (struct ocfs2_dinode *) di_bh->b_data;
1636 
1637 		if (le32_to_cpu(di->id1.journal1.ij_flags) &
1638 		    OCFS2_JOURNAL_DIRTY_FL)
1639 			ret = -EROFS;
1640 
1641 		brelse(di_bh);
1642 		if (ret)
1643 			break;
1644 	}
1645 
1646 out:
1647 	if (journal)
1648 		iput(journal);
1649 
1650 	return ret;
1651 }
1652