xref: /openbmc/linux/fs/jbd2/checkpoint.c (revision baf2c002)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * linux/fs/jbd2/checkpoint.c
4  *
5  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
6  *
7  * Copyright 1999 Red Hat Software --- All Rights Reserved
8  *
9  * Checkpoint routines for the generic filesystem journaling code.
10  * Part of the ext2fs journaling system.
11  *
12  * Checkpointing is the process of ensuring that a section of the log is
13  * committed fully to disk, so that that portion of the log can be
14  * reused.
15  */
16 
17 #include <linux/time.h>
18 #include <linux/fs.h>
19 #include <linux/jbd2.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/blkdev.h>
23 #include <trace/events/jbd2.h>
24 
25 /*
26  * Unlink a buffer from a transaction checkpoint list.
27  *
28  * Called with j_list_lock held.
29  */
30 static inline void __buffer_unlink_first(struct journal_head *jh)
31 {
32 	transaction_t *transaction = jh->b_cp_transaction;
33 
34 	jh->b_cpnext->b_cpprev = jh->b_cpprev;
35 	jh->b_cpprev->b_cpnext = jh->b_cpnext;
36 	if (transaction->t_checkpoint_list == jh) {
37 		transaction->t_checkpoint_list = jh->b_cpnext;
38 		if (transaction->t_checkpoint_list == jh)
39 			transaction->t_checkpoint_list = NULL;
40 	}
41 }
42 
43 /*
44  * Unlink a buffer from a transaction checkpoint(io) list.
45  *
46  * Called with j_list_lock held.
47  */
48 static inline void __buffer_unlink(struct journal_head *jh)
49 {
50 	transaction_t *transaction = jh->b_cp_transaction;
51 
52 	__buffer_unlink_first(jh);
53 	if (transaction->t_checkpoint_io_list == jh) {
54 		transaction->t_checkpoint_io_list = jh->b_cpnext;
55 		if (transaction->t_checkpoint_io_list == jh)
56 			transaction->t_checkpoint_io_list = NULL;
57 	}
58 }
59 
60 /*
61  * Move a buffer from the checkpoint list to the checkpoint io list
62  *
63  * Called with j_list_lock held
64  */
65 static inline void __buffer_relink_io(struct journal_head *jh)
66 {
67 	transaction_t *transaction = jh->b_cp_transaction;
68 
69 	__buffer_unlink_first(jh);
70 
71 	if (!transaction->t_checkpoint_io_list) {
72 		jh->b_cpnext = jh->b_cpprev = jh;
73 	} else {
74 		jh->b_cpnext = transaction->t_checkpoint_io_list;
75 		jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
76 		jh->b_cpprev->b_cpnext = jh;
77 		jh->b_cpnext->b_cpprev = jh;
78 	}
79 	transaction->t_checkpoint_io_list = jh;
80 }
81 
82 /*
83  * Check a checkpoint buffer could be release or not.
84  *
85  * Requires j_list_lock
86  */
87 static inline bool __cp_buffer_busy(struct journal_head *jh)
88 {
89 	struct buffer_head *bh = jh2bh(jh);
90 
91 	return (jh->b_transaction || buffer_locked(bh) || buffer_dirty(bh));
92 }
93 
94 /*
95  * __jbd2_log_wait_for_space: wait until there is space in the journal.
96  *
97  * Called under j-state_lock *only*.  It will be unlocked if we have to wait
98  * for a checkpoint to free up some space in the log.
99  */
100 void __jbd2_log_wait_for_space(journal_t *journal)
101 __acquires(&journal->j_state_lock)
102 __releases(&journal->j_state_lock)
103 {
104 	int nblocks, space_left;
105 	/* assert_spin_locked(&journal->j_state_lock); */
106 
107 	nblocks = journal->j_max_transaction_buffers;
108 	while (jbd2_log_space_left(journal) < nblocks) {
109 		write_unlock(&journal->j_state_lock);
110 		mutex_lock_io(&journal->j_checkpoint_mutex);
111 
112 		/*
113 		 * Test again, another process may have checkpointed while we
114 		 * were waiting for the checkpoint lock. If there are no
115 		 * transactions ready to be checkpointed, try to recover
116 		 * journal space by calling cleanup_journal_tail(), and if
117 		 * that doesn't work, by waiting for the currently committing
118 		 * transaction to complete.  If there is absolutely no way
119 		 * to make progress, this is either a BUG or corrupted
120 		 * filesystem, so abort the journal and leave a stack
121 		 * trace for forensic evidence.
122 		 */
123 		write_lock(&journal->j_state_lock);
124 		if (journal->j_flags & JBD2_ABORT) {
125 			mutex_unlock(&journal->j_checkpoint_mutex);
126 			return;
127 		}
128 		spin_lock(&journal->j_list_lock);
129 		space_left = jbd2_log_space_left(journal);
130 		if (space_left < nblocks) {
131 			int chkpt = journal->j_checkpoint_transactions != NULL;
132 			tid_t tid = 0;
133 
134 			if (journal->j_committing_transaction)
135 				tid = journal->j_committing_transaction->t_tid;
136 			spin_unlock(&journal->j_list_lock);
137 			write_unlock(&journal->j_state_lock);
138 			if (chkpt) {
139 				jbd2_log_do_checkpoint(journal);
140 			} else if (jbd2_cleanup_journal_tail(journal) == 0) {
141 				/* We were able to recover space; yay! */
142 				;
143 			} else if (tid) {
144 				/*
145 				 * jbd2_journal_commit_transaction() may want
146 				 * to take the checkpoint_mutex if JBD2_FLUSHED
147 				 * is set.  So we need to temporarily drop it.
148 				 */
149 				mutex_unlock(&journal->j_checkpoint_mutex);
150 				jbd2_log_wait_commit(journal, tid);
151 				write_lock(&journal->j_state_lock);
152 				continue;
153 			} else {
154 				printk(KERN_ERR "%s: needed %d blocks and "
155 				       "only had %d space available\n",
156 				       __func__, nblocks, space_left);
157 				printk(KERN_ERR "%s: no way to get more "
158 				       "journal space in %s\n", __func__,
159 				       journal->j_devname);
160 				WARN_ON(1);
161 				jbd2_journal_abort(journal, -EIO);
162 			}
163 			write_lock(&journal->j_state_lock);
164 		} else {
165 			spin_unlock(&journal->j_list_lock);
166 		}
167 		mutex_unlock(&journal->j_checkpoint_mutex);
168 	}
169 }
170 
171 static void
172 __flush_batch(journal_t *journal, int *batch_count)
173 {
174 	int i;
175 	struct blk_plug plug;
176 
177 	blk_start_plug(&plug);
178 	for (i = 0; i < *batch_count; i++)
179 		write_dirty_buffer(journal->j_chkpt_bhs[i], REQ_SYNC);
180 	blk_finish_plug(&plug);
181 
182 	for (i = 0; i < *batch_count; i++) {
183 		struct buffer_head *bh = journal->j_chkpt_bhs[i];
184 		BUFFER_TRACE(bh, "brelse");
185 		__brelse(bh);
186 	}
187 	*batch_count = 0;
188 }
189 
190 /*
191  * Perform an actual checkpoint. We take the first transaction on the
192  * list of transactions to be checkpointed and send all its buffers
193  * to disk. We submit larger chunks of data at once.
194  *
195  * The journal should be locked before calling this function.
196  * Called with j_checkpoint_mutex held.
197  */
198 int jbd2_log_do_checkpoint(journal_t *journal)
199 {
200 	struct journal_head	*jh;
201 	struct buffer_head	*bh;
202 	transaction_t		*transaction;
203 	tid_t			this_tid;
204 	int			result, batch_count = 0;
205 
206 	jbd2_debug(1, "Start checkpoint\n");
207 
208 	/*
209 	 * First thing: if there are any transactions in the log which
210 	 * don't need checkpointing, just eliminate them from the
211 	 * journal straight away.
212 	 */
213 	result = jbd2_cleanup_journal_tail(journal);
214 	trace_jbd2_checkpoint(journal, result);
215 	jbd2_debug(1, "cleanup_journal_tail returned %d\n", result);
216 	if (result <= 0)
217 		return result;
218 
219 	/*
220 	 * OK, we need to start writing disk blocks.  Take one transaction
221 	 * and write it.
222 	 */
223 	spin_lock(&journal->j_list_lock);
224 	if (!journal->j_checkpoint_transactions)
225 		goto out;
226 	transaction = journal->j_checkpoint_transactions;
227 	if (transaction->t_chp_stats.cs_chp_time == 0)
228 		transaction->t_chp_stats.cs_chp_time = jiffies;
229 	this_tid = transaction->t_tid;
230 restart:
231 	/*
232 	 * If someone cleaned up this transaction while we slept, we're
233 	 * done (maybe it's a new transaction, but it fell at the same
234 	 * address).
235 	 */
236 	if (journal->j_checkpoint_transactions != transaction ||
237 	    transaction->t_tid != this_tid)
238 		goto out;
239 
240 	/* checkpoint all of the transaction's buffers */
241 	while (transaction->t_checkpoint_list) {
242 		jh = transaction->t_checkpoint_list;
243 		bh = jh2bh(jh);
244 
245 		if (buffer_locked(bh)) {
246 			get_bh(bh);
247 			spin_unlock(&journal->j_list_lock);
248 			wait_on_buffer(bh);
249 			/* the journal_head may have gone by now */
250 			BUFFER_TRACE(bh, "brelse");
251 			__brelse(bh);
252 			goto retry;
253 		}
254 		if (jh->b_transaction != NULL) {
255 			transaction_t *t = jh->b_transaction;
256 			tid_t tid = t->t_tid;
257 
258 			transaction->t_chp_stats.cs_forced_to_close++;
259 			spin_unlock(&journal->j_list_lock);
260 			if (unlikely(journal->j_flags & JBD2_UNMOUNT))
261 				/*
262 				 * The journal thread is dead; so
263 				 * starting and waiting for a commit
264 				 * to finish will cause us to wait for
265 				 * a _very_ long time.
266 				 */
267 				printk(KERN_ERR
268 		"JBD2: %s: Waiting for Godot: block %llu\n",
269 		journal->j_devname, (unsigned long long) bh->b_blocknr);
270 
271 			if (batch_count)
272 				__flush_batch(journal, &batch_count);
273 			jbd2_log_start_commit(journal, tid);
274 			/*
275 			 * jbd2_journal_commit_transaction() may want
276 			 * to take the checkpoint_mutex if JBD2_FLUSHED
277 			 * is set, jbd2_update_log_tail() called by
278 			 * jbd2_journal_commit_transaction() may also take
279 			 * checkpoint_mutex.  So we need to temporarily
280 			 * drop it.
281 			 */
282 			mutex_unlock(&journal->j_checkpoint_mutex);
283 			jbd2_log_wait_commit(journal, tid);
284 			mutex_lock_io(&journal->j_checkpoint_mutex);
285 			spin_lock(&journal->j_list_lock);
286 			goto restart;
287 		}
288 		if (!buffer_dirty(bh)) {
289 			BUFFER_TRACE(bh, "remove from checkpoint");
290 			if (__jbd2_journal_remove_checkpoint(jh))
291 				/* The transaction was released; we're done */
292 				goto out;
293 			continue;
294 		}
295 		/*
296 		 * Important: we are about to write the buffer, and
297 		 * possibly block, while still holding the journal
298 		 * lock.  We cannot afford to let the transaction
299 		 * logic start messing around with this buffer before
300 		 * we write it to disk, as that would break
301 		 * recoverability.
302 		 */
303 		BUFFER_TRACE(bh, "queue");
304 		get_bh(bh);
305 		J_ASSERT_BH(bh, !buffer_jwrite(bh));
306 		journal->j_chkpt_bhs[batch_count++] = bh;
307 		__buffer_relink_io(jh);
308 		transaction->t_chp_stats.cs_written++;
309 		if ((batch_count == JBD2_NR_BATCH) ||
310 		    need_resched() ||
311 		    spin_needbreak(&journal->j_list_lock))
312 			goto unlock_and_flush;
313 	}
314 
315 	if (batch_count) {
316 		unlock_and_flush:
317 			spin_unlock(&journal->j_list_lock);
318 		retry:
319 			if (batch_count)
320 				__flush_batch(journal, &batch_count);
321 			spin_lock(&journal->j_list_lock);
322 			goto restart;
323 	}
324 
325 	/*
326 	 * Now we issued all of the transaction's buffers, let's deal
327 	 * with the buffers that are out for I/O.
328 	 */
329 restart2:
330 	/* Did somebody clean up the transaction in the meanwhile? */
331 	if (journal->j_checkpoint_transactions != transaction ||
332 	    transaction->t_tid != this_tid)
333 		goto out;
334 
335 	while (transaction->t_checkpoint_io_list) {
336 		jh = transaction->t_checkpoint_io_list;
337 		bh = jh2bh(jh);
338 		if (buffer_locked(bh)) {
339 			get_bh(bh);
340 			spin_unlock(&journal->j_list_lock);
341 			wait_on_buffer(bh);
342 			/* the journal_head may have gone by now */
343 			BUFFER_TRACE(bh, "brelse");
344 			__brelse(bh);
345 			spin_lock(&journal->j_list_lock);
346 			goto restart2;
347 		}
348 
349 		/*
350 		 * Now in whatever state the buffer currently is, we
351 		 * know that it has been written out and so we can
352 		 * drop it from the list
353 		 */
354 		if (__jbd2_journal_remove_checkpoint(jh))
355 			break;
356 	}
357 out:
358 	spin_unlock(&journal->j_list_lock);
359 	result = jbd2_cleanup_journal_tail(journal);
360 
361 	return (result < 0) ? result : 0;
362 }
363 
364 /*
365  * Check the list of checkpoint transactions for the journal to see if
366  * we have already got rid of any since the last update of the log tail
367  * in the journal superblock.  If so, we can instantly roll the
368  * superblock forward to remove those transactions from the log.
369  *
370  * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
371  *
372  * Called with the journal lock held.
373  *
374  * This is the only part of the journaling code which really needs to be
375  * aware of transaction aborts.  Checkpointing involves writing to the
376  * main filesystem area rather than to the journal, so it can proceed
377  * even in abort state, but we must not update the super block if
378  * checkpointing may have failed.  Otherwise, we would lose some metadata
379  * buffers which should be written-back to the filesystem.
380  */
381 
382 int jbd2_cleanup_journal_tail(journal_t *journal)
383 {
384 	tid_t		first_tid;
385 	unsigned long	blocknr;
386 
387 	if (is_journal_aborted(journal))
388 		return -EIO;
389 
390 	if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
391 		return 1;
392 	J_ASSERT(blocknr != 0);
393 
394 	/*
395 	 * We need to make sure that any blocks that were recently written out
396 	 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
397 	 * we drop the transactions from the journal. It's unlikely this will
398 	 * be necessary, especially with an appropriately sized journal, but we
399 	 * need this to guarantee correctness.  Fortunately
400 	 * jbd2_cleanup_journal_tail() doesn't get called all that often.
401 	 */
402 	if (journal->j_flags & JBD2_BARRIER)
403 		blkdev_issue_flush(journal->j_fs_dev);
404 
405 	return __jbd2_update_log_tail(journal, first_tid, blocknr);
406 }
407 
408 
409 /* Checkpoint list management */
410 
411 /*
412  * journal_clean_one_cp_list
413  *
414  * Find all the written-back checkpoint buffers in the given list and
415  * release them. If 'destroy' is set, clean all buffers unconditionally.
416  *
417  * Called with j_list_lock held.
418  * Returns 1 if we freed the transaction, 0 otherwise.
419  */
420 static int journal_clean_one_cp_list(struct journal_head *jh, bool destroy)
421 {
422 	struct journal_head *last_jh;
423 	struct journal_head *next_jh = jh;
424 
425 	if (!jh)
426 		return 0;
427 
428 	last_jh = jh->b_cpprev;
429 	do {
430 		jh = next_jh;
431 		next_jh = jh->b_cpnext;
432 
433 		if (!destroy && __cp_buffer_busy(jh))
434 			return 0;
435 
436 		if (__jbd2_journal_remove_checkpoint(jh))
437 			return 1;
438 		/*
439 		 * This function only frees up some memory
440 		 * if possible so we dont have an obligation
441 		 * to finish processing. Bail out if preemption
442 		 * requested:
443 		 */
444 		if (need_resched())
445 			return 0;
446 	} while (jh != last_jh);
447 
448 	return 0;
449 }
450 
451 /*
452  * journal_shrink_one_cp_list
453  *
454  * Find 'nr_to_scan' written-back checkpoint buffers in the given list
455  * and try to release them. If the whole transaction is released, set
456  * the 'released' parameter. Return the number of released checkpointed
457  * buffers.
458  *
459  * Called with j_list_lock held.
460  */
461 static unsigned long journal_shrink_one_cp_list(struct journal_head *jh,
462 						unsigned long *nr_to_scan,
463 						bool *released)
464 {
465 	struct journal_head *last_jh;
466 	struct journal_head *next_jh = jh;
467 	unsigned long nr_freed = 0;
468 	int ret;
469 
470 	if (!jh || *nr_to_scan == 0)
471 		return 0;
472 
473 	last_jh = jh->b_cpprev;
474 	do {
475 		jh = next_jh;
476 		next_jh = jh->b_cpnext;
477 
478 		(*nr_to_scan)--;
479 		if (__cp_buffer_busy(jh))
480 			continue;
481 
482 		nr_freed++;
483 		ret = __jbd2_journal_remove_checkpoint(jh);
484 		if (ret) {
485 			*released = true;
486 			break;
487 		}
488 
489 		if (need_resched())
490 			break;
491 	} while (jh != last_jh && *nr_to_scan);
492 
493 	return nr_freed;
494 }
495 
496 /*
497  * jbd2_journal_shrink_checkpoint_list
498  *
499  * Find 'nr_to_scan' written-back checkpoint buffers in the journal
500  * and try to release them. Return the number of released checkpointed
501  * buffers.
502  *
503  * Called with j_list_lock held.
504  */
505 unsigned long jbd2_journal_shrink_checkpoint_list(journal_t *journal,
506 						  unsigned long *nr_to_scan)
507 {
508 	transaction_t *transaction, *last_transaction, *next_transaction;
509 	bool released;
510 	tid_t first_tid = 0, last_tid = 0, next_tid = 0;
511 	tid_t tid = 0;
512 	unsigned long nr_freed = 0;
513 	unsigned long nr_scanned = *nr_to_scan;
514 
515 again:
516 	spin_lock(&journal->j_list_lock);
517 	if (!journal->j_checkpoint_transactions) {
518 		spin_unlock(&journal->j_list_lock);
519 		goto out;
520 	}
521 
522 	/*
523 	 * Get next shrink transaction, resume previous scan or start
524 	 * over again. If some others do checkpoint and drop transaction
525 	 * from the checkpoint list, we ignore saved j_shrink_transaction
526 	 * and start over unconditionally.
527 	 */
528 	if (journal->j_shrink_transaction)
529 		transaction = journal->j_shrink_transaction;
530 	else
531 		transaction = journal->j_checkpoint_transactions;
532 
533 	if (!first_tid)
534 		first_tid = transaction->t_tid;
535 	last_transaction = journal->j_checkpoint_transactions->t_cpprev;
536 	next_transaction = transaction;
537 	last_tid = last_transaction->t_tid;
538 	do {
539 		transaction = next_transaction;
540 		next_transaction = transaction->t_cpnext;
541 		tid = transaction->t_tid;
542 		released = false;
543 
544 		nr_freed += journal_shrink_one_cp_list(transaction->t_checkpoint_list,
545 						       nr_to_scan, &released);
546 		if (*nr_to_scan == 0)
547 			break;
548 		if (need_resched() || spin_needbreak(&journal->j_list_lock))
549 			break;
550 		if (released)
551 			continue;
552 
553 		nr_freed += journal_shrink_one_cp_list(transaction->t_checkpoint_io_list,
554 						       nr_to_scan, &released);
555 		if (*nr_to_scan == 0)
556 			break;
557 		if (need_resched() || spin_needbreak(&journal->j_list_lock))
558 			break;
559 	} while (transaction != last_transaction);
560 
561 	if (transaction != last_transaction) {
562 		journal->j_shrink_transaction = next_transaction;
563 		next_tid = next_transaction->t_tid;
564 	} else {
565 		journal->j_shrink_transaction = NULL;
566 		next_tid = 0;
567 	}
568 
569 	spin_unlock(&journal->j_list_lock);
570 	cond_resched();
571 
572 	if (*nr_to_scan && next_tid)
573 		goto again;
574 out:
575 	nr_scanned -= *nr_to_scan;
576 	trace_jbd2_shrink_checkpoint_list(journal, first_tid, tid, last_tid,
577 					  nr_freed, nr_scanned, next_tid);
578 
579 	return nr_freed;
580 }
581 
582 /*
583  * journal_clean_checkpoint_list
584  *
585  * Find all the written-back checkpoint buffers in the journal and release them.
586  * If 'destroy' is set, release all buffers unconditionally.
587  *
588  * Called with j_list_lock held.
589  */
590 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy)
591 {
592 	transaction_t *transaction, *last_transaction, *next_transaction;
593 	int ret;
594 
595 	transaction = journal->j_checkpoint_transactions;
596 	if (!transaction)
597 		return;
598 
599 	last_transaction = transaction->t_cpprev;
600 	next_transaction = transaction;
601 	do {
602 		transaction = next_transaction;
603 		next_transaction = transaction->t_cpnext;
604 		ret = journal_clean_one_cp_list(transaction->t_checkpoint_list,
605 						destroy);
606 		/*
607 		 * This function only frees up some memory if possible so we
608 		 * dont have an obligation to finish processing. Bail out if
609 		 * preemption requested:
610 		 */
611 		if (need_resched())
612 			return;
613 		if (ret)
614 			continue;
615 		/*
616 		 * It is essential that we are as careful as in the case of
617 		 * t_checkpoint_list with removing the buffer from the list as
618 		 * we can possibly see not yet submitted buffers on io_list
619 		 */
620 		ret = journal_clean_one_cp_list(transaction->
621 				t_checkpoint_io_list, destroy);
622 		if (need_resched())
623 			return;
624 		/*
625 		 * Stop scanning if we couldn't free the transaction. This
626 		 * avoids pointless scanning of transactions which still
627 		 * weren't checkpointed.
628 		 */
629 		if (!ret)
630 			return;
631 	} while (transaction != last_transaction);
632 }
633 
634 /*
635  * Remove buffers from all checkpoint lists as journal is aborted and we just
636  * need to free memory
637  */
638 void jbd2_journal_destroy_checkpoint(journal_t *journal)
639 {
640 	/*
641 	 * We loop because __jbd2_journal_clean_checkpoint_list() may abort
642 	 * early due to a need of rescheduling.
643 	 */
644 	while (1) {
645 		spin_lock(&journal->j_list_lock);
646 		if (!journal->j_checkpoint_transactions) {
647 			spin_unlock(&journal->j_list_lock);
648 			break;
649 		}
650 		__jbd2_journal_clean_checkpoint_list(journal, true);
651 		spin_unlock(&journal->j_list_lock);
652 		cond_resched();
653 	}
654 }
655 
656 /*
657  * journal_remove_checkpoint: called after a buffer has been committed
658  * to disk (either by being write-back flushed to disk, or being
659  * committed to the log).
660  *
661  * We cannot safely clean a transaction out of the log until all of the
662  * buffer updates committed in that transaction have safely been stored
663  * elsewhere on disk.  To achieve this, all of the buffers in a
664  * transaction need to be maintained on the transaction's checkpoint
665  * lists until they have been rewritten, at which point this function is
666  * called to remove the buffer from the existing transaction's
667  * checkpoint lists.
668  *
669  * The function returns 1 if it frees the transaction, 0 otherwise.
670  * The function can free jh and bh.
671  *
672  * This function is called with j_list_lock held.
673  */
674 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
675 {
676 	struct transaction_chp_stats_s *stats;
677 	transaction_t *transaction;
678 	journal_t *journal;
679 	struct buffer_head *bh = jh2bh(jh);
680 
681 	JBUFFER_TRACE(jh, "entry");
682 
683 	transaction = jh->b_cp_transaction;
684 	if (!transaction) {
685 		JBUFFER_TRACE(jh, "not on transaction");
686 		return 0;
687 	}
688 	journal = transaction->t_journal;
689 
690 	JBUFFER_TRACE(jh, "removing from transaction");
691 
692 	/*
693 	 * If we have failed to write the buffer out to disk, the filesystem
694 	 * may become inconsistent. We cannot abort the journal here since
695 	 * we hold j_list_lock and we have to be careful about races with
696 	 * jbd2_journal_destroy(). So mark the writeback IO error in the
697 	 * journal here and we abort the journal later from a better context.
698 	 */
699 	if (buffer_write_io_error(bh))
700 		set_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags);
701 
702 	__buffer_unlink(jh);
703 	jh->b_cp_transaction = NULL;
704 	percpu_counter_dec(&journal->j_checkpoint_jh_count);
705 	jbd2_journal_put_journal_head(jh);
706 
707 	/* Is this transaction empty? */
708 	if (transaction->t_checkpoint_list || transaction->t_checkpoint_io_list)
709 		return 0;
710 
711 	/*
712 	 * There is one special case to worry about: if we have just pulled the
713 	 * buffer off a running or committing transaction's checkpoing list,
714 	 * then even if the checkpoint list is empty, the transaction obviously
715 	 * cannot be dropped!
716 	 *
717 	 * The locking here around t_state is a bit sleazy.
718 	 * See the comment at the end of jbd2_journal_commit_transaction().
719 	 */
720 	if (transaction->t_state != T_FINISHED)
721 		return 0;
722 
723 	/*
724 	 * OK, that was the last buffer for the transaction, we can now
725 	 * safely remove this transaction from the log.
726 	 */
727 	stats = &transaction->t_chp_stats;
728 	if (stats->cs_chp_time)
729 		stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
730 						    jiffies);
731 	trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
732 				    transaction->t_tid, stats);
733 
734 	__jbd2_journal_drop_transaction(journal, transaction);
735 	jbd2_journal_free_transaction(transaction);
736 	return 1;
737 }
738 
739 /*
740  * journal_insert_checkpoint: put a committed buffer onto a checkpoint
741  * list so that we know when it is safe to clean the transaction out of
742  * the log.
743  *
744  * Called with the journal locked.
745  * Called with j_list_lock held.
746  */
747 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
748 			       transaction_t *transaction)
749 {
750 	JBUFFER_TRACE(jh, "entry");
751 	J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
752 	J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
753 
754 	/* Get reference for checkpointing transaction */
755 	jbd2_journal_grab_journal_head(jh2bh(jh));
756 	jh->b_cp_transaction = transaction;
757 
758 	if (!transaction->t_checkpoint_list) {
759 		jh->b_cpnext = jh->b_cpprev = jh;
760 	} else {
761 		jh->b_cpnext = transaction->t_checkpoint_list;
762 		jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
763 		jh->b_cpprev->b_cpnext = jh;
764 		jh->b_cpnext->b_cpprev = jh;
765 	}
766 	transaction->t_checkpoint_list = jh;
767 	percpu_counter_inc(&transaction->t_journal->j_checkpoint_jh_count);
768 }
769 
770 /*
771  * We've finished with this transaction structure: adios...
772  *
773  * The transaction must have no links except for the checkpoint by this
774  * point.
775  *
776  * Called with the journal locked.
777  * Called with j_list_lock held.
778  */
779 
780 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
781 {
782 	assert_spin_locked(&journal->j_list_lock);
783 
784 	journal->j_shrink_transaction = NULL;
785 	if (transaction->t_cpnext) {
786 		transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
787 		transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
788 		if (journal->j_checkpoint_transactions == transaction)
789 			journal->j_checkpoint_transactions =
790 				transaction->t_cpnext;
791 		if (journal->j_checkpoint_transactions == transaction)
792 			journal->j_checkpoint_transactions = NULL;
793 	}
794 
795 	J_ASSERT(transaction->t_state == T_FINISHED);
796 	J_ASSERT(transaction->t_buffers == NULL);
797 	J_ASSERT(transaction->t_forget == NULL);
798 	J_ASSERT(transaction->t_shadow_list == NULL);
799 	J_ASSERT(transaction->t_checkpoint_list == NULL);
800 	J_ASSERT(transaction->t_checkpoint_io_list == NULL);
801 	J_ASSERT(atomic_read(&transaction->t_updates) == 0);
802 	J_ASSERT(journal->j_committing_transaction != transaction);
803 	J_ASSERT(journal->j_running_transaction != transaction);
804 
805 	trace_jbd2_drop_transaction(journal, transaction);
806 
807 	jbd2_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
808 }
809