Loading...
1/*
2 * linux/fs/jbd2/checkpoint.c
3 *
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
5 *
6 * Copyright 1999 Red Hat Software --- All Rights Reserved
7 *
8 * This file is part of the Linux kernel and is made available under
9 * the terms of the GNU General Public License, version 2, or at your
10 * option, any later version, incorporated herein by reference.
11 *
12 * Checkpoint routines for the generic filesystem journaling code.
13 * Part of the ext2fs journaling system.
14 *
15 * Checkpointing is the process of ensuring that a section of the log is
16 * committed fully to disk, so that that portion of the log can be
17 * reused.
18 */
19
20#include <linux/time.h>
21#include <linux/fs.h>
22#include <linux/jbd2.h>
23#include <linux/errno.h>
24#include <linux/slab.h>
25#include <linux/blkdev.h>
26#include <trace/events/jbd2.h>
27
28/*
29 * Unlink a buffer from a transaction checkpoint list.
30 *
31 * Called with j_list_lock held.
32 */
33static inline void __buffer_unlink_first(struct journal_head *jh)
34{
35 transaction_t *transaction = jh->b_cp_transaction;
36
37 jh->b_cpnext->b_cpprev = jh->b_cpprev;
38 jh->b_cpprev->b_cpnext = jh->b_cpnext;
39 if (transaction->t_checkpoint_list == jh) {
40 transaction->t_checkpoint_list = jh->b_cpnext;
41 if (transaction->t_checkpoint_list == jh)
42 transaction->t_checkpoint_list = NULL;
43 }
44}
45
46/*
47 * Unlink a buffer from a transaction checkpoint(io) list.
48 *
49 * Called with j_list_lock held.
50 */
51static inline void __buffer_unlink(struct journal_head *jh)
52{
53 transaction_t *transaction = jh->b_cp_transaction;
54
55 __buffer_unlink_first(jh);
56 if (transaction->t_checkpoint_io_list == jh) {
57 transaction->t_checkpoint_io_list = jh->b_cpnext;
58 if (transaction->t_checkpoint_io_list == jh)
59 transaction->t_checkpoint_io_list = NULL;
60 }
61}
62
63/*
64 * Move a buffer from the checkpoint list to the checkpoint io list
65 *
66 * Called with j_list_lock held
67 */
68static inline void __buffer_relink_io(struct journal_head *jh)
69{
70 transaction_t *transaction = jh->b_cp_transaction;
71
72 __buffer_unlink_first(jh);
73
74 if (!transaction->t_checkpoint_io_list) {
75 jh->b_cpnext = jh->b_cpprev = jh;
76 } else {
77 jh->b_cpnext = transaction->t_checkpoint_io_list;
78 jh->b_cpprev = transaction->t_checkpoint_io_list->b_cpprev;
79 jh->b_cpprev->b_cpnext = jh;
80 jh->b_cpnext->b_cpprev = jh;
81 }
82 transaction->t_checkpoint_io_list = jh;
83}
84
85/*
86 * Try to release a checkpointed buffer from its transaction.
87 * Returns 1 if we released it and 2 if we also released the
88 * whole transaction.
89 *
90 * Requires j_list_lock
91 */
92static int __try_to_free_cp_buf(struct journal_head *jh)
93{
94 int ret = 0;
95 struct buffer_head *bh = jh2bh(jh);
96
97 if (jh->b_transaction == NULL && !buffer_locked(bh) &&
98 !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
99 /*
100 * Get our reference so that bh cannot be freed before
101 * we unlock it
102 */
103 get_bh(bh);
104 JBUFFER_TRACE(jh, "remove from checkpoint list");
105 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
106 BUFFER_TRACE(bh, "release");
107 __brelse(bh);
108 }
109 return ret;
110}
111
112/*
113 * __jbd2_log_wait_for_space: wait until there is space in the journal.
114 *
115 * Called under j-state_lock *only*. It will be unlocked if we have to wait
116 * for a checkpoint to free up some space in the log.
117 */
118void __jbd2_log_wait_for_space(journal_t *journal)
119{
120 int nblocks, space_left;
121 /* assert_spin_locked(&journal->j_state_lock); */
122
123 nblocks = jbd2_space_needed(journal);
124 while (jbd2_log_space_left(journal) < nblocks) {
125 if (journal->j_flags & JBD2_ABORT)
126 return;
127 write_unlock(&journal->j_state_lock);
128 mutex_lock(&journal->j_checkpoint_mutex);
129
130 /*
131 * Test again, another process may have checkpointed while we
132 * were waiting for the checkpoint lock. If there are no
133 * transactions ready to be checkpointed, try to recover
134 * journal space by calling cleanup_journal_tail(), and if
135 * that doesn't work, by waiting for the currently committing
136 * transaction to complete. If there is absolutely no way
137 * to make progress, this is either a BUG or corrupted
138 * filesystem, so abort the journal and leave a stack
139 * trace for forensic evidence.
140 */
141 write_lock(&journal->j_state_lock);
142 spin_lock(&journal->j_list_lock);
143 nblocks = jbd2_space_needed(journal);
144 space_left = jbd2_log_space_left(journal);
145 if (space_left < nblocks) {
146 int chkpt = journal->j_checkpoint_transactions != NULL;
147 tid_t tid = 0;
148
149 if (journal->j_committing_transaction)
150 tid = journal->j_committing_transaction->t_tid;
151 spin_unlock(&journal->j_list_lock);
152 write_unlock(&journal->j_state_lock);
153 if (chkpt) {
154 jbd2_log_do_checkpoint(journal);
155 } else if (jbd2_cleanup_journal_tail(journal) == 0) {
156 /* We were able to recover space; yay! */
157 ;
158 } else if (tid) {
159 /*
160 * jbd2_journal_commit_transaction() may want
161 * to take the checkpoint_mutex if JBD2_FLUSHED
162 * is set. So we need to temporarily drop it.
163 */
164 mutex_unlock(&journal->j_checkpoint_mutex);
165 jbd2_log_wait_commit(journal, tid);
166 write_lock(&journal->j_state_lock);
167 continue;
168 } else {
169 printk(KERN_ERR "%s: needed %d blocks and "
170 "only had %d space available\n",
171 __func__, nblocks, space_left);
172 printk(KERN_ERR "%s: no way to get more "
173 "journal space in %s\n", __func__,
174 journal->j_devname);
175 WARN_ON(1);
176 jbd2_journal_abort(journal, 0);
177 }
178 write_lock(&journal->j_state_lock);
179 } else {
180 spin_unlock(&journal->j_list_lock);
181 }
182 mutex_unlock(&journal->j_checkpoint_mutex);
183 }
184}
185
186/*
187 * Clean up transaction's list of buffers submitted for io.
188 * We wait for any pending IO to complete and remove any clean
189 * buffers. Note that we take the buffers in the opposite ordering
190 * from the one in which they were submitted for IO.
191 *
192 * Return 0 on success, and return <0 if some buffers have failed
193 * to be written out.
194 *
195 * Called with j_list_lock held.
196 */
197static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
198{
199 struct journal_head *jh;
200 struct buffer_head *bh;
201 tid_t this_tid;
202 int released = 0;
203 int ret = 0;
204
205 this_tid = transaction->t_tid;
206restart:
207 /* Did somebody clean up the transaction in the meanwhile? */
208 if (journal->j_checkpoint_transactions != transaction ||
209 transaction->t_tid != this_tid)
210 return ret;
211 while (!released && transaction->t_checkpoint_io_list) {
212 jh = transaction->t_checkpoint_io_list;
213 bh = jh2bh(jh);
214 get_bh(bh);
215 if (buffer_locked(bh)) {
216 spin_unlock(&journal->j_list_lock);
217 wait_on_buffer(bh);
218 /* the journal_head may have gone by now */
219 BUFFER_TRACE(bh, "brelse");
220 __brelse(bh);
221 spin_lock(&journal->j_list_lock);
222 goto restart;
223 }
224 if (unlikely(buffer_write_io_error(bh)))
225 ret = -EIO;
226
227 /*
228 * Now in whatever state the buffer currently is, we know that
229 * it has been written out and so we can drop it from the list
230 */
231 released = __jbd2_journal_remove_checkpoint(jh);
232 __brelse(bh);
233 }
234
235 return ret;
236}
237
238static void
239__flush_batch(journal_t *journal, int *batch_count)
240{
241 int i;
242 struct blk_plug plug;
243
244 blk_start_plug(&plug);
245 for (i = 0; i < *batch_count; i++)
246 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE_SYNC);
247 blk_finish_plug(&plug);
248
249 for (i = 0; i < *batch_count; i++) {
250 struct buffer_head *bh = journal->j_chkpt_bhs[i];
251 BUFFER_TRACE(bh, "brelse");
252 __brelse(bh);
253 }
254 *batch_count = 0;
255}
256
257/*
258 * Try to flush one buffer from the checkpoint list to disk.
259 *
260 * Return 1 if something happened which requires us to abort the current
261 * scan of the checkpoint list. Return <0 if the buffer has failed to
262 * be written out.
263 *
264 * Called with j_list_lock held and drops it if 1 is returned
265 */
266static int __process_buffer(journal_t *journal, struct journal_head *jh,
267 int *batch_count, transaction_t *transaction)
268{
269 struct buffer_head *bh = jh2bh(jh);
270 int ret = 0;
271
272 if (buffer_locked(bh)) {
273 get_bh(bh);
274 spin_unlock(&journal->j_list_lock);
275 wait_on_buffer(bh);
276 /* the journal_head may have gone by now */
277 BUFFER_TRACE(bh, "brelse");
278 __brelse(bh);
279 ret = 1;
280 } else if (jh->b_transaction != NULL) {
281 transaction_t *t = jh->b_transaction;
282 tid_t tid = t->t_tid;
283
284 transaction->t_chp_stats.cs_forced_to_close++;
285 spin_unlock(&journal->j_list_lock);
286 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
287 /*
288 * The journal thread is dead; so starting and
289 * waiting for a commit to finish will cause
290 * us to wait for a _very_ long time.
291 */
292 printk(KERN_ERR "JBD2: %s: "
293 "Waiting for Godot: block %llu\n",
294 journal->j_devname,
295 (unsigned long long) bh->b_blocknr);
296 jbd2_log_start_commit(journal, tid);
297 jbd2_log_wait_commit(journal, tid);
298 ret = 1;
299 } else if (!buffer_dirty(bh)) {
300 ret = 1;
301 if (unlikely(buffer_write_io_error(bh)))
302 ret = -EIO;
303 get_bh(bh);
304 BUFFER_TRACE(bh, "remove from checkpoint");
305 __jbd2_journal_remove_checkpoint(jh);
306 spin_unlock(&journal->j_list_lock);
307 __brelse(bh);
308 } else {
309 /*
310 * Important: we are about to write the buffer, and
311 * possibly block, while still holding the journal lock.
312 * We cannot afford to let the transaction logic start
313 * messing around with this buffer before we write it to
314 * disk, as that would break recoverability.
315 */
316 BUFFER_TRACE(bh, "queue");
317 get_bh(bh);
318 J_ASSERT_BH(bh, !buffer_jwrite(bh));
319 journal->j_chkpt_bhs[*batch_count] = bh;
320 __buffer_relink_io(jh);
321 transaction->t_chp_stats.cs_written++;
322 (*batch_count)++;
323 if (*batch_count == JBD2_NR_BATCH) {
324 spin_unlock(&journal->j_list_lock);
325 __flush_batch(journal, batch_count);
326 ret = 1;
327 }
328 }
329 return ret;
330}
331
332/*
333 * Perform an actual checkpoint. We take the first transaction on the
334 * list of transactions to be checkpointed and send all its buffers
335 * to disk. We submit larger chunks of data at once.
336 *
337 * The journal should be locked before calling this function.
338 * Called with j_checkpoint_mutex held.
339 */
340int jbd2_log_do_checkpoint(journal_t *journal)
341{
342 transaction_t *transaction;
343 tid_t this_tid;
344 int result;
345
346 jbd_debug(1, "Start checkpoint\n");
347
348 /*
349 * First thing: if there are any transactions in the log which
350 * don't need checkpointing, just eliminate them from the
351 * journal straight away.
352 */
353 result = jbd2_cleanup_journal_tail(journal);
354 trace_jbd2_checkpoint(journal, result);
355 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
356 if (result <= 0)
357 return result;
358
359 /*
360 * OK, we need to start writing disk blocks. Take one transaction
361 * and write it.
362 */
363 result = 0;
364 spin_lock(&journal->j_list_lock);
365 if (!journal->j_checkpoint_transactions)
366 goto out;
367 transaction = journal->j_checkpoint_transactions;
368 if (transaction->t_chp_stats.cs_chp_time == 0)
369 transaction->t_chp_stats.cs_chp_time = jiffies;
370 this_tid = transaction->t_tid;
371restart:
372 /*
373 * If someone cleaned up this transaction while we slept, we're
374 * done (maybe it's a new transaction, but it fell at the same
375 * address).
376 */
377 if (journal->j_checkpoint_transactions == transaction &&
378 transaction->t_tid == this_tid) {
379 int batch_count = 0;
380 struct journal_head *jh;
381 int retry = 0, err;
382
383 while (!retry && transaction->t_checkpoint_list) {
384 jh = transaction->t_checkpoint_list;
385 retry = __process_buffer(journal, jh, &batch_count,
386 transaction);
387 if (retry < 0 && !result)
388 result = retry;
389 if (!retry && (need_resched() ||
390 spin_needbreak(&journal->j_list_lock))) {
391 spin_unlock(&journal->j_list_lock);
392 retry = 1;
393 break;
394 }
395 }
396
397 if (batch_count) {
398 if (!retry) {
399 spin_unlock(&journal->j_list_lock);
400 retry = 1;
401 }
402 __flush_batch(journal, &batch_count);
403 }
404
405 if (retry) {
406 spin_lock(&journal->j_list_lock);
407 goto restart;
408 }
409 /*
410 * Now we have cleaned up the first transaction's checkpoint
411 * list. Let's clean up the second one
412 */
413 err = __wait_cp_io(journal, transaction);
414 if (!result)
415 result = err;
416 }
417out:
418 spin_unlock(&journal->j_list_lock);
419 if (result < 0)
420 jbd2_journal_abort(journal, result);
421 else
422 result = jbd2_cleanup_journal_tail(journal);
423
424 return (result < 0) ? result : 0;
425}
426
427/*
428 * Check the list of checkpoint transactions for the journal to see if
429 * we have already got rid of any since the last update of the log tail
430 * in the journal superblock. If so, we can instantly roll the
431 * superblock forward to remove those transactions from the log.
432 *
433 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
434 *
435 * Called with the journal lock held.
436 *
437 * This is the only part of the journaling code which really needs to be
438 * aware of transaction aborts. Checkpointing involves writing to the
439 * main filesystem area rather than to the journal, so it can proceed
440 * even in abort state, but we must not update the super block if
441 * checkpointing may have failed. Otherwise, we would lose some metadata
442 * buffers which should be written-back to the filesystem.
443 */
444
445int jbd2_cleanup_journal_tail(journal_t *journal)
446{
447 tid_t first_tid;
448 unsigned long blocknr;
449
450 if (is_journal_aborted(journal))
451 return 1;
452
453 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
454 return 1;
455 J_ASSERT(blocknr != 0);
456
457 /*
458 * We need to make sure that any blocks that were recently written out
459 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
460 * we drop the transactions from the journal. It's unlikely this will
461 * be necessary, especially with an appropriately sized journal, but we
462 * need this to guarantee correctness. Fortunately
463 * jbd2_cleanup_journal_tail() doesn't get called all that often.
464 */
465 if (journal->j_flags & JBD2_BARRIER)
466 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
467
468 __jbd2_update_log_tail(journal, first_tid, blocknr);
469 return 0;
470}
471
472
473/* Checkpoint list management */
474
475/*
476 * journal_clean_one_cp_list
477 *
478 * Find all the written-back checkpoint buffers in the given list and
479 * release them.
480 *
481 * Called with the journal locked.
482 * Called with j_list_lock held.
483 * Returns number of buffers reaped (for debug)
484 */
485
486static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
487{
488 struct journal_head *last_jh;
489 struct journal_head *next_jh = jh;
490 int ret, freed = 0;
491
492 *released = 0;
493 if (!jh)
494 return 0;
495
496 last_jh = jh->b_cpprev;
497 do {
498 jh = next_jh;
499 next_jh = jh->b_cpnext;
500 ret = __try_to_free_cp_buf(jh);
501 if (ret) {
502 freed++;
503 if (ret == 2) {
504 *released = 1;
505 return freed;
506 }
507 }
508 /*
509 * This function only frees up some memory
510 * if possible so we dont have an obligation
511 * to finish processing. Bail out if preemption
512 * requested:
513 */
514 if (need_resched())
515 return freed;
516 } while (jh != last_jh);
517
518 return freed;
519}
520
521/*
522 * journal_clean_checkpoint_list
523 *
524 * Find all the written-back checkpoint buffers in the journal and release them.
525 *
526 * Called with the journal locked.
527 * Called with j_list_lock held.
528 * Returns number of buffers reaped (for debug)
529 */
530
531int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
532{
533 transaction_t *transaction, *last_transaction, *next_transaction;
534 int ret = 0;
535 int released;
536
537 transaction = journal->j_checkpoint_transactions;
538 if (!transaction)
539 goto out;
540
541 last_transaction = transaction->t_cpprev;
542 next_transaction = transaction;
543 do {
544 transaction = next_transaction;
545 next_transaction = transaction->t_cpnext;
546 ret += journal_clean_one_cp_list(transaction->
547 t_checkpoint_list, &released);
548 /*
549 * This function only frees up some memory if possible so we
550 * dont have an obligation to finish processing. Bail out if
551 * preemption requested:
552 */
553 if (need_resched())
554 goto out;
555 if (released)
556 continue;
557 /*
558 * It is essential that we are as careful as in the case of
559 * t_checkpoint_list with removing the buffer from the list as
560 * we can possibly see not yet submitted buffers on io_list
561 */
562 ret += journal_clean_one_cp_list(transaction->
563 t_checkpoint_io_list, &released);
564 if (need_resched())
565 goto out;
566 } while (transaction != last_transaction);
567out:
568 return ret;
569}
570
571/*
572 * journal_remove_checkpoint: called after a buffer has been committed
573 * to disk (either by being write-back flushed to disk, or being
574 * committed to the log).
575 *
576 * We cannot safely clean a transaction out of the log until all of the
577 * buffer updates committed in that transaction have safely been stored
578 * elsewhere on disk. To achieve this, all of the buffers in a
579 * transaction need to be maintained on the transaction's checkpoint
580 * lists until they have been rewritten, at which point this function is
581 * called to remove the buffer from the existing transaction's
582 * checkpoint lists.
583 *
584 * The function returns 1 if it frees the transaction, 0 otherwise.
585 * The function can free jh and bh.
586 *
587 * This function is called with j_list_lock held.
588 */
589int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
590{
591 struct transaction_chp_stats_s *stats;
592 transaction_t *transaction;
593 journal_t *journal;
594 int ret = 0;
595
596 JBUFFER_TRACE(jh, "entry");
597
598 if ((transaction = jh->b_cp_transaction) == NULL) {
599 JBUFFER_TRACE(jh, "not on transaction");
600 goto out;
601 }
602 journal = transaction->t_journal;
603
604 JBUFFER_TRACE(jh, "removing from transaction");
605 __buffer_unlink(jh);
606 jh->b_cp_transaction = NULL;
607 jbd2_journal_put_journal_head(jh);
608
609 if (transaction->t_checkpoint_list != NULL ||
610 transaction->t_checkpoint_io_list != NULL)
611 goto out;
612
613 /*
614 * There is one special case to worry about: if we have just pulled the
615 * buffer off a running or committing transaction's checkpoing list,
616 * then even if the checkpoint list is empty, the transaction obviously
617 * cannot be dropped!
618 *
619 * The locking here around t_state is a bit sleazy.
620 * See the comment at the end of jbd2_journal_commit_transaction().
621 */
622 if (transaction->t_state != T_FINISHED)
623 goto out;
624
625 /* OK, that was the last buffer for the transaction: we can now
626 safely remove this transaction from the log */
627 stats = &transaction->t_chp_stats;
628 if (stats->cs_chp_time)
629 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
630 jiffies);
631 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
632 transaction->t_tid, stats);
633
634 __jbd2_journal_drop_transaction(journal, transaction);
635 jbd2_journal_free_transaction(transaction);
636 ret = 1;
637out:
638 return ret;
639}
640
641/*
642 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
643 * list so that we know when it is safe to clean the transaction out of
644 * the log.
645 *
646 * Called with the journal locked.
647 * Called with j_list_lock held.
648 */
649void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
650 transaction_t *transaction)
651{
652 JBUFFER_TRACE(jh, "entry");
653 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
654 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
655
656 /* Get reference for checkpointing transaction */
657 jbd2_journal_grab_journal_head(jh2bh(jh));
658 jh->b_cp_transaction = transaction;
659
660 if (!transaction->t_checkpoint_list) {
661 jh->b_cpnext = jh->b_cpprev = jh;
662 } else {
663 jh->b_cpnext = transaction->t_checkpoint_list;
664 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
665 jh->b_cpprev->b_cpnext = jh;
666 jh->b_cpnext->b_cpprev = jh;
667 }
668 transaction->t_checkpoint_list = jh;
669}
670
671/*
672 * We've finished with this transaction structure: adios...
673 *
674 * The transaction must have no links except for the checkpoint by this
675 * point.
676 *
677 * Called with the journal locked.
678 * Called with j_list_lock held.
679 */
680
681void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
682{
683 assert_spin_locked(&journal->j_list_lock);
684 if (transaction->t_cpnext) {
685 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
686 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
687 if (journal->j_checkpoint_transactions == transaction)
688 journal->j_checkpoint_transactions =
689 transaction->t_cpnext;
690 if (journal->j_checkpoint_transactions == transaction)
691 journal->j_checkpoint_transactions = NULL;
692 }
693
694 J_ASSERT(transaction->t_state == T_FINISHED);
695 J_ASSERT(transaction->t_buffers == NULL);
696 J_ASSERT(transaction->t_forget == NULL);
697 J_ASSERT(transaction->t_shadow_list == NULL);
698 J_ASSERT(transaction->t_checkpoint_list == NULL);
699 J_ASSERT(transaction->t_checkpoint_io_list == NULL);
700 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
701 J_ASSERT(journal->j_committing_transaction != transaction);
702 J_ASSERT(journal->j_running_transaction != transaction);
703
704 trace_jbd2_drop_transaction(journal, transaction);
705
706 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
707}
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 */
30static inline void __buffer_unlink(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 * __jbd2_log_wait_for_space: wait until there is space in the journal.
45 *
46 * Called under j-state_lock *only*. It will be unlocked if we have to wait
47 * for a checkpoint to free up some space in the log.
48 */
49void __jbd2_log_wait_for_space(journal_t *journal)
50__acquires(&journal->j_state_lock)
51__releases(&journal->j_state_lock)
52{
53 int nblocks, space_left;
54 /* assert_spin_locked(&journal->j_state_lock); */
55
56 nblocks = journal->j_max_transaction_buffers;
57 while (jbd2_log_space_left(journal) < nblocks) {
58 write_unlock(&journal->j_state_lock);
59 mutex_lock_io(&journal->j_checkpoint_mutex);
60
61 /*
62 * Test again, another process may have checkpointed while we
63 * were waiting for the checkpoint lock. If there are no
64 * transactions ready to be checkpointed, try to recover
65 * journal space by calling cleanup_journal_tail(), and if
66 * that doesn't work, by waiting for the currently committing
67 * transaction to complete. If there is absolutely no way
68 * to make progress, this is either a BUG or corrupted
69 * filesystem, so abort the journal and leave a stack
70 * trace for forensic evidence.
71 */
72 write_lock(&journal->j_state_lock);
73 if (journal->j_flags & JBD2_ABORT) {
74 mutex_unlock(&journal->j_checkpoint_mutex);
75 return;
76 }
77 spin_lock(&journal->j_list_lock);
78 space_left = jbd2_log_space_left(journal);
79 if (space_left < nblocks) {
80 int chkpt = journal->j_checkpoint_transactions != NULL;
81 tid_t tid = 0;
82 bool has_transaction = false;
83
84 if (journal->j_committing_transaction) {
85 tid = journal->j_committing_transaction->t_tid;
86 has_transaction = true;
87 }
88 spin_unlock(&journal->j_list_lock);
89 write_unlock(&journal->j_state_lock);
90 if (chkpt) {
91 jbd2_log_do_checkpoint(journal);
92 } else if (jbd2_cleanup_journal_tail(journal) <= 0) {
93 /*
94 * We were able to recover space or the
95 * journal was aborted due to an error.
96 */
97 ;
98 } else if (has_transaction) {
99 /*
100 * jbd2_journal_commit_transaction() may want
101 * to take the checkpoint_mutex if JBD2_FLUSHED
102 * is set. So we need to temporarily drop it.
103 */
104 mutex_unlock(&journal->j_checkpoint_mutex);
105 jbd2_log_wait_commit(journal, tid);
106 write_lock(&journal->j_state_lock);
107 continue;
108 } else {
109 printk(KERN_ERR "%s: needed %d blocks and "
110 "only had %d space available\n",
111 __func__, nblocks, space_left);
112 printk(KERN_ERR "%s: no way to get more "
113 "journal space in %s\n", __func__,
114 journal->j_devname);
115 WARN_ON(1);
116 jbd2_journal_abort(journal, -EIO);
117 }
118 write_lock(&journal->j_state_lock);
119 } else {
120 spin_unlock(&journal->j_list_lock);
121 }
122 mutex_unlock(&journal->j_checkpoint_mutex);
123 }
124}
125
126static void
127__flush_batch(journal_t *journal, int *batch_count)
128{
129 int i;
130 struct blk_plug plug;
131
132 blk_start_plug(&plug);
133 for (i = 0; i < *batch_count; i++)
134 write_dirty_buffer(journal->j_chkpt_bhs[i], REQ_SYNC);
135 blk_finish_plug(&plug);
136
137 for (i = 0; i < *batch_count; i++) {
138 struct buffer_head *bh = journal->j_chkpt_bhs[i];
139 BUFFER_TRACE(bh, "brelse");
140 __brelse(bh);
141 journal->j_chkpt_bhs[i] = NULL;
142 }
143 *batch_count = 0;
144}
145
146/*
147 * Perform an actual checkpoint. We take the first transaction on the
148 * list of transactions to be checkpointed and send all its buffers
149 * to disk. We submit larger chunks of data at once.
150 *
151 * The journal should be locked before calling this function.
152 * Called with j_checkpoint_mutex held.
153 */
154int jbd2_log_do_checkpoint(journal_t *journal)
155{
156 struct journal_head *jh;
157 struct buffer_head *bh;
158 transaction_t *transaction;
159 tid_t this_tid;
160 int result, batch_count = 0;
161
162 jbd2_debug(1, "Start checkpoint\n");
163
164 /*
165 * First thing: if there are any transactions in the log which
166 * don't need checkpointing, just eliminate them from the
167 * journal straight away.
168 */
169 result = jbd2_cleanup_journal_tail(journal);
170 trace_jbd2_checkpoint(journal, result);
171 jbd2_debug(1, "cleanup_journal_tail returned %d\n", result);
172 if (result <= 0)
173 return result;
174
175 /*
176 * OK, we need to start writing disk blocks. Take one transaction
177 * and write it.
178 */
179 spin_lock(&journal->j_list_lock);
180 if (!journal->j_checkpoint_transactions)
181 goto out;
182 transaction = journal->j_checkpoint_transactions;
183 if (transaction->t_chp_stats.cs_chp_time == 0)
184 transaction->t_chp_stats.cs_chp_time = jiffies;
185 this_tid = transaction->t_tid;
186restart:
187 /*
188 * If someone cleaned up this transaction while we slept, we're
189 * done (maybe it's a new transaction, but it fell at the same
190 * address).
191 */
192 if (journal->j_checkpoint_transactions != transaction ||
193 transaction->t_tid != this_tid)
194 goto out;
195
196 /* checkpoint all of the transaction's buffers */
197 while (transaction->t_checkpoint_list) {
198 jh = transaction->t_checkpoint_list;
199 bh = jh2bh(jh);
200
201 if (jh->b_transaction != NULL) {
202 transaction_t *t = jh->b_transaction;
203 tid_t tid = t->t_tid;
204
205 transaction->t_chp_stats.cs_forced_to_close++;
206 spin_unlock(&journal->j_list_lock);
207 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
208 /*
209 * The journal thread is dead; so
210 * starting and waiting for a commit
211 * to finish will cause us to wait for
212 * a _very_ long time.
213 */
214 printk(KERN_ERR
215 "JBD2: %s: Waiting for Godot: block %llu\n",
216 journal->j_devname, (unsigned long long) bh->b_blocknr);
217
218 if (batch_count)
219 __flush_batch(journal, &batch_count);
220 jbd2_log_start_commit(journal, tid);
221 /*
222 * jbd2_journal_commit_transaction() may want
223 * to take the checkpoint_mutex if JBD2_FLUSHED
224 * is set, jbd2_update_log_tail() called by
225 * jbd2_journal_commit_transaction() may also take
226 * checkpoint_mutex. So we need to temporarily
227 * drop it.
228 */
229 mutex_unlock(&journal->j_checkpoint_mutex);
230 jbd2_log_wait_commit(journal, tid);
231 mutex_lock_io(&journal->j_checkpoint_mutex);
232 spin_lock(&journal->j_list_lock);
233 goto restart;
234 }
235 if (!trylock_buffer(bh)) {
236 /*
237 * The buffer is locked, it may be writing back, or
238 * flushing out in the last couple of cycles, or
239 * re-adding into a new transaction, need to check
240 * it again until it's unlocked.
241 */
242 get_bh(bh);
243 spin_unlock(&journal->j_list_lock);
244 wait_on_buffer(bh);
245 /* the journal_head may have gone by now */
246 BUFFER_TRACE(bh, "brelse");
247 __brelse(bh);
248 goto retry;
249 } else if (!buffer_dirty(bh)) {
250 unlock_buffer(bh);
251 BUFFER_TRACE(bh, "remove from checkpoint");
252 /*
253 * If the transaction was released or the checkpoint
254 * list was empty, we're done.
255 */
256 if (__jbd2_journal_remove_checkpoint(jh) ||
257 !transaction->t_checkpoint_list)
258 goto out;
259 } else {
260 unlock_buffer(bh);
261 /*
262 * We are about to write the buffer, it could be
263 * raced by some other transaction shrink or buffer
264 * re-log logic once we release the j_list_lock,
265 * leave it on the checkpoint list and check status
266 * again to make sure it's clean.
267 */
268 BUFFER_TRACE(bh, "queue");
269 get_bh(bh);
270 J_ASSERT_BH(bh, !buffer_jwrite(bh));
271 journal->j_chkpt_bhs[batch_count++] = bh;
272 transaction->t_chp_stats.cs_written++;
273 transaction->t_checkpoint_list = jh->b_cpnext;
274 }
275
276 if ((batch_count == JBD2_NR_BATCH) ||
277 need_resched() || spin_needbreak(&journal->j_list_lock) ||
278 jh2bh(transaction->t_checkpoint_list) == journal->j_chkpt_bhs[0])
279 goto unlock_and_flush;
280 }
281
282 if (batch_count) {
283 unlock_and_flush:
284 spin_unlock(&journal->j_list_lock);
285 retry:
286 if (batch_count)
287 __flush_batch(journal, &batch_count);
288 spin_lock(&journal->j_list_lock);
289 goto restart;
290 }
291
292out:
293 spin_unlock(&journal->j_list_lock);
294 result = jbd2_cleanup_journal_tail(journal);
295
296 return (result < 0) ? result : 0;
297}
298
299/*
300 * Check the list of checkpoint transactions for the journal to see if
301 * we have already got rid of any since the last update of the log tail
302 * in the journal superblock. If so, we can instantly roll the
303 * superblock forward to remove those transactions from the log.
304 *
305 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
306 *
307 * Called with the journal lock held.
308 *
309 * This is the only part of the journaling code which really needs to be
310 * aware of transaction aborts. Checkpointing involves writing to the
311 * main filesystem area rather than to the journal, so it can proceed
312 * even in abort state, but we must not update the super block if
313 * checkpointing may have failed. Otherwise, we would lose some metadata
314 * buffers which should be written-back to the filesystem.
315 */
316
317int jbd2_cleanup_journal_tail(journal_t *journal)
318{
319 tid_t first_tid;
320 unsigned long blocknr;
321
322 if (is_journal_aborted(journal))
323 return -EIO;
324
325 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
326 return 1;
327 J_ASSERT(blocknr != 0);
328
329 /*
330 * We need to make sure that any blocks that were recently written out
331 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
332 * we drop the transactions from the journal. It's unlikely this will
333 * be necessary, especially with an appropriately sized journal, but we
334 * need this to guarantee correctness. Fortunately
335 * jbd2_cleanup_journal_tail() doesn't get called all that often.
336 */
337 if (journal->j_flags & JBD2_BARRIER)
338 blkdev_issue_flush(journal->j_fs_dev);
339
340 return __jbd2_update_log_tail(journal, first_tid, blocknr);
341}
342
343
344/* Checkpoint list management */
345
346/*
347 * journal_shrink_one_cp_list
348 *
349 * Find all the written-back checkpoint buffers in the given list
350 * and try to release them. If the whole transaction is released, set
351 * the 'released' parameter. Return the number of released checkpointed
352 * buffers.
353 *
354 * Called with j_list_lock held.
355 */
356static unsigned long journal_shrink_one_cp_list(struct journal_head *jh,
357 enum jbd2_shrink_type type,
358 bool *released)
359{
360 struct journal_head *last_jh;
361 struct journal_head *next_jh = jh;
362 unsigned long nr_freed = 0;
363 int ret;
364
365 *released = false;
366 if (!jh)
367 return 0;
368
369 last_jh = jh->b_cpprev;
370 do {
371 jh = next_jh;
372 next_jh = jh->b_cpnext;
373
374 if (type == JBD2_SHRINK_DESTROY) {
375 ret = __jbd2_journal_remove_checkpoint(jh);
376 } else {
377 ret = jbd2_journal_try_remove_checkpoint(jh);
378 if (ret < 0) {
379 if (type == JBD2_SHRINK_BUSY_SKIP)
380 continue;
381 break;
382 }
383 }
384
385 nr_freed++;
386 if (ret) {
387 *released = true;
388 break;
389 }
390
391 if (need_resched())
392 break;
393 } while (jh != last_jh);
394
395 return nr_freed;
396}
397
398/*
399 * jbd2_journal_shrink_checkpoint_list
400 *
401 * Find 'nr_to_scan' written-back checkpoint buffers in the journal
402 * and try to release them. Return the number of released checkpointed
403 * buffers.
404 *
405 * Called with j_list_lock held.
406 */
407unsigned long jbd2_journal_shrink_checkpoint_list(journal_t *journal,
408 unsigned long *nr_to_scan)
409{
410 transaction_t *transaction, *last_transaction, *next_transaction;
411 bool __maybe_unused released;
412 tid_t first_tid = 0, last_tid = 0, next_tid = 0;
413 tid_t tid = 0;
414 unsigned long nr_freed = 0;
415 unsigned long freed;
416 bool first_set = false;
417
418again:
419 spin_lock(&journal->j_list_lock);
420 if (!journal->j_checkpoint_transactions) {
421 spin_unlock(&journal->j_list_lock);
422 goto out;
423 }
424
425 /*
426 * Get next shrink transaction, resume previous scan or start
427 * over again. If some others do checkpoint and drop transaction
428 * from the checkpoint list, we ignore saved j_shrink_transaction
429 * and start over unconditionally.
430 */
431 if (journal->j_shrink_transaction)
432 transaction = journal->j_shrink_transaction;
433 else
434 transaction = journal->j_checkpoint_transactions;
435
436 if (!first_set) {
437 first_tid = transaction->t_tid;
438 first_set = true;
439 }
440 last_transaction = journal->j_checkpoint_transactions->t_cpprev;
441 next_transaction = transaction;
442 last_tid = last_transaction->t_tid;
443 do {
444 transaction = next_transaction;
445 next_transaction = transaction->t_cpnext;
446 tid = transaction->t_tid;
447
448 freed = journal_shrink_one_cp_list(transaction->t_checkpoint_list,
449 JBD2_SHRINK_BUSY_SKIP, &released);
450 nr_freed += freed;
451 (*nr_to_scan) -= min(*nr_to_scan, freed);
452 if (*nr_to_scan == 0)
453 break;
454 if (need_resched() || spin_needbreak(&journal->j_list_lock))
455 break;
456 } while (transaction != last_transaction);
457
458 if (transaction != last_transaction) {
459 journal->j_shrink_transaction = next_transaction;
460 next_tid = next_transaction->t_tid;
461 } else {
462 journal->j_shrink_transaction = NULL;
463 next_tid = 0;
464 }
465
466 spin_unlock(&journal->j_list_lock);
467 cond_resched();
468
469 if (*nr_to_scan && journal->j_shrink_transaction)
470 goto again;
471out:
472 trace_jbd2_shrink_checkpoint_list(journal, first_tid, tid, last_tid,
473 nr_freed, next_tid);
474
475 return nr_freed;
476}
477
478/*
479 * journal_clean_checkpoint_list
480 *
481 * Find all the written-back checkpoint buffers in the journal and release them.
482 * If 'type' is JBD2_SHRINK_DESTROY, release all buffers unconditionally. If
483 * 'type' is JBD2_SHRINK_BUSY_STOP, will stop release buffers if encounters a
484 * busy buffer. To avoid wasting CPU cycles scanning the buffer list in some
485 * cases, don't pass JBD2_SHRINK_BUSY_SKIP 'type' for this function.
486 *
487 * Called with j_list_lock held.
488 */
489void __jbd2_journal_clean_checkpoint_list(journal_t *journal,
490 enum jbd2_shrink_type type)
491{
492 transaction_t *transaction, *last_transaction, *next_transaction;
493 bool released;
494
495 WARN_ON_ONCE(type == JBD2_SHRINK_BUSY_SKIP);
496
497 transaction = journal->j_checkpoint_transactions;
498 if (!transaction)
499 return;
500
501 last_transaction = transaction->t_cpprev;
502 next_transaction = transaction;
503 do {
504 transaction = next_transaction;
505 next_transaction = transaction->t_cpnext;
506 journal_shrink_one_cp_list(transaction->t_checkpoint_list,
507 type, &released);
508 /*
509 * This function only frees up some memory if possible so we
510 * dont have an obligation to finish processing. Bail out if
511 * preemption requested:
512 */
513 if (need_resched())
514 return;
515 /*
516 * Stop scanning if we couldn't free the transaction. This
517 * avoids pointless scanning of transactions which still
518 * weren't checkpointed.
519 */
520 if (!released)
521 return;
522 } while (transaction != last_transaction);
523}
524
525/*
526 * Remove buffers from all checkpoint lists as journal is aborted and we just
527 * need to free memory
528 */
529void jbd2_journal_destroy_checkpoint(journal_t *journal)
530{
531 /*
532 * We loop because __jbd2_journal_clean_checkpoint_list() may abort
533 * early due to a need of rescheduling.
534 */
535 while (1) {
536 spin_lock(&journal->j_list_lock);
537 if (!journal->j_checkpoint_transactions) {
538 spin_unlock(&journal->j_list_lock);
539 break;
540 }
541 __jbd2_journal_clean_checkpoint_list(journal, JBD2_SHRINK_DESTROY);
542 spin_unlock(&journal->j_list_lock);
543 cond_resched();
544 }
545}
546
547/*
548 * journal_remove_checkpoint: called after a buffer has been committed
549 * to disk (either by being write-back flushed to disk, or being
550 * committed to the log).
551 *
552 * We cannot safely clean a transaction out of the log until all of the
553 * buffer updates committed in that transaction have safely been stored
554 * elsewhere on disk. To achieve this, all of the buffers in a
555 * transaction need to be maintained on the transaction's checkpoint
556 * lists until they have been rewritten, at which point this function is
557 * called to remove the buffer from the existing transaction's
558 * checkpoint lists.
559 *
560 * The function returns 1 if it frees the transaction, 0 otherwise.
561 * The function can free jh and bh.
562 *
563 * This function is called with j_list_lock held.
564 */
565int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
566{
567 struct transaction_chp_stats_s *stats;
568 transaction_t *transaction;
569 journal_t *journal;
570
571 JBUFFER_TRACE(jh, "entry");
572
573 transaction = jh->b_cp_transaction;
574 if (!transaction) {
575 JBUFFER_TRACE(jh, "not on transaction");
576 return 0;
577 }
578 journal = transaction->t_journal;
579
580 JBUFFER_TRACE(jh, "removing from transaction");
581
582 __buffer_unlink(jh);
583 jh->b_cp_transaction = NULL;
584 percpu_counter_dec(&journal->j_checkpoint_jh_count);
585 jbd2_journal_put_journal_head(jh);
586
587 /* Is this transaction empty? */
588 if (transaction->t_checkpoint_list)
589 return 0;
590
591 /*
592 * There is one special case to worry about: if we have just pulled the
593 * buffer off a running or committing transaction's checkpoing list,
594 * then even if the checkpoint list is empty, the transaction obviously
595 * cannot be dropped!
596 *
597 * The locking here around t_state is a bit sleazy.
598 * See the comment at the end of jbd2_journal_commit_transaction().
599 */
600 if (transaction->t_state != T_FINISHED)
601 return 0;
602
603 /*
604 * OK, that was the last buffer for the transaction, we can now
605 * safely remove this transaction from the log.
606 */
607 stats = &transaction->t_chp_stats;
608 if (stats->cs_chp_time)
609 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
610 jiffies);
611 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
612 transaction->t_tid, stats);
613
614 __jbd2_journal_drop_transaction(journal, transaction);
615 jbd2_journal_free_transaction(transaction);
616 return 1;
617}
618
619/*
620 * Check the checkpoint buffer and try to remove it from the checkpoint
621 * list if it's clean. Returns -EBUSY if it is not clean, returns 1 if
622 * it frees the transaction, 0 otherwise.
623 *
624 * This function is called with j_list_lock held.
625 */
626int jbd2_journal_try_remove_checkpoint(struct journal_head *jh)
627{
628 struct buffer_head *bh = jh2bh(jh);
629
630 if (jh->b_transaction)
631 return -EBUSY;
632 if (!trylock_buffer(bh))
633 return -EBUSY;
634 if (buffer_dirty(bh)) {
635 unlock_buffer(bh);
636 return -EBUSY;
637 }
638 unlock_buffer(bh);
639
640 /*
641 * Buffer is clean and the IO has finished (we held the buffer
642 * lock) so the checkpoint is done. We can safely remove the
643 * buffer from this transaction.
644 */
645 JBUFFER_TRACE(jh, "remove from checkpoint list");
646 return __jbd2_journal_remove_checkpoint(jh);
647}
648
649/*
650 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
651 * list so that we know when it is safe to clean the transaction out of
652 * the log.
653 *
654 * Called with the journal locked.
655 * Called with j_list_lock held.
656 */
657void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
658 transaction_t *transaction)
659{
660 JBUFFER_TRACE(jh, "entry");
661 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
662 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
663
664 /* Get reference for checkpointing transaction */
665 jbd2_journal_grab_journal_head(jh2bh(jh));
666 jh->b_cp_transaction = transaction;
667
668 if (!transaction->t_checkpoint_list) {
669 jh->b_cpnext = jh->b_cpprev = jh;
670 } else {
671 jh->b_cpnext = transaction->t_checkpoint_list;
672 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
673 jh->b_cpprev->b_cpnext = jh;
674 jh->b_cpnext->b_cpprev = jh;
675 }
676 transaction->t_checkpoint_list = jh;
677 percpu_counter_inc(&transaction->t_journal->j_checkpoint_jh_count);
678}
679
680/*
681 * We've finished with this transaction structure: adios...
682 *
683 * The transaction must have no links except for the checkpoint by this
684 * point.
685 *
686 * Called with the journal locked.
687 * Called with j_list_lock held.
688 */
689
690void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
691{
692 assert_spin_locked(&journal->j_list_lock);
693
694 journal->j_shrink_transaction = NULL;
695 if (transaction->t_cpnext) {
696 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
697 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
698 if (journal->j_checkpoint_transactions == transaction)
699 journal->j_checkpoint_transactions =
700 transaction->t_cpnext;
701 if (journal->j_checkpoint_transactions == transaction)
702 journal->j_checkpoint_transactions = NULL;
703 }
704
705 J_ASSERT(transaction->t_state == T_FINISHED);
706 J_ASSERT(transaction->t_buffers == NULL);
707 J_ASSERT(transaction->t_forget == NULL);
708 J_ASSERT(transaction->t_shadow_list == NULL);
709 J_ASSERT(transaction->t_checkpoint_list == NULL);
710 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
711 J_ASSERT(journal->j_committing_transaction != transaction);
712 J_ASSERT(journal->j_running_transaction != transaction);
713
714 trace_jbd2_drop_transaction(journal, transaction);
715
716 jbd2_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
717}