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 * Try to release a checkpointed buffer from its transaction.
84 * Returns 1 if we released it and 2 if we also released the
85 * whole transaction.
86 *
87 * Requires j_list_lock
88 */
89 static int __try_to_free_cp_buf(struct journal_head *jh)
90 {
91 int ret = 0;
92 struct buffer_head *bh = jh2bh(jh);
93
94 if (jh->b_transaction == NULL && !buffer_locked(bh) &&
95 !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
96 JBUFFER_TRACE(jh, "remove from checkpoint list");
97 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
98 }
99 return ret;
100 }
101
102 /*
103 * __jbd2_log_wait_for_space: wait until there is space in the journal.
104 *
105 * Called under j-state_lock *only*. It will be unlocked if we have to wait
106 * for a checkpoint to free up some space in the log.
107 */
108 void __jbd2_log_wait_for_space(journal_t *journal)
109 {
110 int nblocks, space_left;
111 /* assert_spin_locked(&journal->j_state_lock); */
112
113 nblocks = jbd2_space_needed(journal);
114 while (jbd2_log_space_left(journal) < nblocks) {
115 write_unlock(&journal->j_state_lock);
116 mutex_lock_io(&journal->j_checkpoint_mutex);
117
118 /*
119 * Test again, another process may have checkpointed while we
120 * were waiting for the checkpoint lock. If there are no
121 * transactions ready to be checkpointed, try to recover
122 * journal space by calling cleanup_journal_tail(), and if
123 * that doesn't work, by waiting for the currently committing
124 * transaction to complete. If there is absolutely no way
125 * to make progress, this is either a BUG or corrupted
126 * filesystem, so abort the journal and leave a stack
127 * trace for forensic evidence.
128 */
129 write_lock(&journal->j_state_lock);
130 if (journal->j_flags & JBD2_ABORT) {
131 mutex_unlock(&journal->j_checkpoint_mutex);
132 return;
133 }
134 spin_lock(&journal->j_list_lock);
135 space_left = jbd2_log_space_left(journal);
136 if (space_left < nblocks) {
137 int chkpt = journal->j_checkpoint_transactions != NULL;
138 tid_t tid = 0;
139
140 if (journal->j_committing_transaction)
141 tid = journal->j_committing_transaction->t_tid;
142 spin_unlock(&journal->j_list_lock);
143 write_unlock(&journal->j_state_lock);
144 if (chkpt) {
145 jbd2_log_do_checkpoint(journal);
146 } else if (jbd2_cleanup_journal_tail(journal) == 0) {
147 /* We were able to recover space; yay! */
148 ;
149 } else if (tid) {
150 /*
151 * jbd2_journal_commit_transaction() may want
152 * to take the checkpoint_mutex if JBD2_FLUSHED
153 * is set. So we need to temporarily drop it.
154 */
155 mutex_unlock(&journal->j_checkpoint_mutex);
156 jbd2_log_wait_commit(journal, tid);
157 write_lock(&journal->j_state_lock);
158 continue;
159 } else {
160 printk(KERN_ERR "%s: needed %d blocks and "
161 "only had %d space available\n",
162 __func__, nblocks, space_left);
163 printk(KERN_ERR "%s: no way to get more "
164 "journal space in %s\n", __func__,
165 journal->j_devname);
166 WARN_ON(1);
167 jbd2_journal_abort(journal, -EIO);
168 }
169 write_lock(&journal->j_state_lock);
170 } else {
171 spin_unlock(&journal->j_list_lock);
172 }
173 mutex_unlock(&journal->j_checkpoint_mutex);
174 }
175 }
176
177 static void
178 __flush_batch(journal_t *journal, int *batch_count)
179 {
180 int i;
181 struct blk_plug plug;
182
183 blk_start_plug(&plug);
184 for (i = 0; i < *batch_count; i++)
185 write_dirty_buffer(journal->j_chkpt_bhs[i], REQ_SYNC);
186 blk_finish_plug(&plug);
187
188 for (i = 0; i < *batch_count; i++) {
189 struct buffer_head *bh = journal->j_chkpt_bhs[i];
190 BUFFER_TRACE(bh, "brelse");
191 __brelse(bh);
192 }
193 *batch_count = 0;
194 }
195
196 /*
197 * Perform an actual checkpoint. We take the first transaction on the
198 * list of transactions to be checkpointed and send all its buffers
199 * to disk. We submit larger chunks of data at once.
200 *
201 * The journal should be locked before calling this function.
202 * Called with j_checkpoint_mutex held.
203 */
204 int jbd2_log_do_checkpoint(journal_t *journal)
205 {
206 struct journal_head *jh;
207 struct buffer_head *bh;
208 transaction_t *transaction;
209 tid_t this_tid;
210 int result, batch_count = 0;
211
212 jbd_debug(1, "Start checkpoint\n");
213
214 /*
215 * First thing: if there are any transactions in the log which
216 * don't need checkpointing, just eliminate them from the
217 * journal straight away.
218 */
219 result = jbd2_cleanup_journal_tail(journal);
220 trace_jbd2_checkpoint(journal, result);
221 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
222 if (result <= 0)
223 return result;
224
225 /*
226 * OK, we need to start writing disk blocks. Take one transaction
227 * and write it.
228 */
229 result = 0;
230 spin_lock(&journal->j_list_lock);
231 if (!journal->j_checkpoint_transactions)
232 goto out;
233 transaction = journal->j_checkpoint_transactions;
234 if (transaction->t_chp_stats.cs_chp_time == 0)
235 transaction->t_chp_stats.cs_chp_time = jiffies;
236 this_tid = transaction->t_tid;
237 restart:
238 /*
239 * If someone cleaned up this transaction while we slept, we're
240 * done (maybe it's a new transaction, but it fell at the same
241 * address).
242 */
243 if (journal->j_checkpoint_transactions != transaction ||
244 transaction->t_tid != this_tid)
245 goto out;
246
247 /* checkpoint all of the transaction's buffers */
248 while (transaction->t_checkpoint_list) {
249 jh = transaction->t_checkpoint_list;
250 bh = jh2bh(jh);
251
252 if (buffer_locked(bh)) {
253 get_bh(bh);
254 spin_unlock(&journal->j_list_lock);
255 wait_on_buffer(bh);
256 /* the journal_head may have gone by now */
257 BUFFER_TRACE(bh, "brelse");
258 __brelse(bh);
259 goto retry;
260 }
261 if (jh->b_transaction != NULL) {
262 transaction_t *t = jh->b_transaction;
263 tid_t tid = t->t_tid;
264
265 transaction->t_chp_stats.cs_forced_to_close++;
266 spin_unlock(&journal->j_list_lock);
267 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
268 /*
269 * The journal thread is dead; so
270 * starting and waiting for a commit
271 * to finish will cause us to wait for
272 * a _very_ long time.
273 */
274 printk(KERN_ERR
275 "JBD2: %s: Waiting for Godot: block %llu\n",
276 journal->j_devname, (unsigned long long) bh->b_blocknr);
277
278 if (batch_count)
279 __flush_batch(journal, &batch_count);
280 jbd2_log_start_commit(journal, tid);
281 /*
282 * jbd2_journal_commit_transaction() may want
283 * to take the checkpoint_mutex if JBD2_FLUSHED
284 * is set, jbd2_update_log_tail() called by
285 * jbd2_journal_commit_transaction() may also take
286 * checkpoint_mutex. So we need to temporarily
287 * drop it.
288 */
289 mutex_unlock(&journal->j_checkpoint_mutex);
290 jbd2_log_wait_commit(journal, tid);
291 mutex_lock_io(&journal->j_checkpoint_mutex);
292 spin_lock(&journal->j_list_lock);
293 goto restart;
294 }
295 if (!buffer_dirty(bh)) {
296 if (unlikely(buffer_write_io_error(bh)) && !result)
297 result = -EIO;
298 BUFFER_TRACE(bh, "remove from checkpoint");
299 if (__jbd2_journal_remove_checkpoint(jh))
300 /* The transaction was released; we're done */
301 goto out;
302 continue;
303 }
304 /*
305 * Important: we are about to write the buffer, and
306 * possibly block, while still holding the journal
307 * lock. We cannot afford to let the transaction
308 * logic start messing around with this buffer before
309 * we write it to disk, as that would break
310 * recoverability.
311 */
312 BUFFER_TRACE(bh, "queue");
313 get_bh(bh);
314 J_ASSERT_BH(bh, !buffer_jwrite(bh));
315 journal->j_chkpt_bhs[batch_count++] = bh;
316 __buffer_relink_io(jh);
317 transaction->t_chp_stats.cs_written++;
318 if ((batch_count == JBD2_NR_BATCH) ||
319 need_resched() ||
320 spin_needbreak(&journal->j_list_lock))
321 goto unlock_and_flush;
322 }
323
324 if (batch_count) {
325 unlock_and_flush:
326 spin_unlock(&journal->j_list_lock);
327 retry:
328 if (batch_count)
329 __flush_batch(journal, &batch_count);
330 spin_lock(&journal->j_list_lock);
331 goto restart;
332 }
333
334 /*
335 * Now we issued all of the transaction's buffers, let's deal
336 * with the buffers that are out for I/O.
337 */
338 restart2:
339 /* Did somebody clean up the transaction in the meanwhile? */
340 if (journal->j_checkpoint_transactions != transaction ||
341 transaction->t_tid != this_tid)
342 goto out;
343
344 while (transaction->t_checkpoint_io_list) {
345 jh = transaction->t_checkpoint_io_list;
346 bh = jh2bh(jh);
347 if (buffer_locked(bh)) {
348 get_bh(bh);
349 spin_unlock(&journal->j_list_lock);
350 wait_on_buffer(bh);
351 /* the journal_head may have gone by now */
352 BUFFER_TRACE(bh, "brelse");
353 __brelse(bh);
354 spin_lock(&journal->j_list_lock);
355 goto restart2;
356 }
357 if (unlikely(buffer_write_io_error(bh)) && !result)
358 result = -EIO;
359
360 /*
361 * Now in whatever state the buffer currently is, we
362 * know that it has been written out and so we can
363 * drop it from the list
364 */
365 if (__jbd2_journal_remove_checkpoint(jh))
366 break;
367 }
368 out:
369 spin_unlock(&journal->j_list_lock);
370 if (result < 0)
371 jbd2_journal_abort(journal, result);
372 else
373 result = jbd2_cleanup_journal_tail(journal);
374
375 return (result < 0) ? result : 0;
376 }
377
378 /*
379 * Check the list of checkpoint transactions for the journal to see if
380 * we have already got rid of any since the last update of the log tail
381 * in the journal superblock. If so, we can instantly roll the
382 * superblock forward to remove those transactions from the log.
383 *
384 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
385 *
386 * Called with the journal lock held.
387 *
388 * This is the only part of the journaling code which really needs to be
389 * aware of transaction aborts. Checkpointing involves writing to the
390 * main filesystem area rather than to the journal, so it can proceed
391 * even in abort state, but we must not update the super block if
392 * checkpointing may have failed. Otherwise, we would lose some metadata
393 * buffers which should be written-back to the filesystem.
394 */
395
396 int jbd2_cleanup_journal_tail(journal_t *journal)
397 {
398 tid_t first_tid;
399 unsigned long blocknr;
400
401 if (is_journal_aborted(journal))
402 return -EIO;
403
404 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
405 return 1;
406 J_ASSERT(blocknr != 0);
407
408 /*
409 * We need to make sure that any blocks that were recently written out
410 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
411 * we drop the transactions from the journal. It's unlikely this will
412 * be necessary, especially with an appropriately sized journal, but we
413 * need this to guarantee correctness. Fortunately
414 * jbd2_cleanup_journal_tail() doesn't get called all that often.
415 */
416 if (journal->j_flags & JBD2_BARRIER)
417 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
418
419 return __jbd2_update_log_tail(journal, first_tid, blocknr);
420 }
421
422
423 /* Checkpoint list management */
424
425 /*
426 * journal_clean_one_cp_list
427 *
428 * Find all the written-back checkpoint buffers in the given list and
429 * release them. If 'destroy' is set, clean all buffers unconditionally.
430 *
431 * Called with j_list_lock held.
432 * Returns 1 if we freed the transaction, 0 otherwise.
433 */
434 static int journal_clean_one_cp_list(struct journal_head *jh, bool destroy)
435 {
436 struct journal_head *last_jh;
437 struct journal_head *next_jh = jh;
438 int ret;
439
440 if (!jh)
441 return 0;
442
443 last_jh = jh->b_cpprev;
444 do {
445 jh = next_jh;
446 next_jh = jh->b_cpnext;
447 if (!destroy)
448 ret = __try_to_free_cp_buf(jh);
449 else
450 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
451 if (!ret)
452 return 0;
453 if (ret == 2)
454 return 1;
455 /*
456 * This function only frees up some memory
457 * if possible so we dont have an obligation
458 * to finish processing. Bail out if preemption
459 * requested:
460 */
461 if (need_resched())
462 return 0;
463 } while (jh != last_jh);
464
465 return 0;
466 }
467
468 /*
469 * journal_clean_checkpoint_list
470 *
471 * Find all the written-back checkpoint buffers in the journal and release them.
472 * If 'destroy' is set, release all buffers unconditionally.
473 *
474 * Called with j_list_lock held.
475 */
476 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy)
477 {
478 transaction_t *transaction, *last_transaction, *next_transaction;
479 int ret;
480
481 transaction = journal->j_checkpoint_transactions;
482 if (!transaction)
483 return;
484
485 last_transaction = transaction->t_cpprev;
486 next_transaction = transaction;
487 do {
488 transaction = next_transaction;
489 next_transaction = transaction->t_cpnext;
490 ret = journal_clean_one_cp_list(transaction->t_checkpoint_list,
491 destroy);
492 /*
493 * This function only frees up some memory if possible so we
494 * dont have an obligation to finish processing. Bail out if
495 * preemption requested:
496 */
497 if (need_resched())
498 return;
499 if (ret)
500 continue;
501 /*
502 * It is essential that we are as careful as in the case of
503 * t_checkpoint_list with removing the buffer from the list as
504 * we can possibly see not yet submitted buffers on io_list
505 */
506 ret = journal_clean_one_cp_list(transaction->
507 t_checkpoint_io_list, destroy);
508 if (need_resched())
509 return;
510 /*
511 * Stop scanning if we couldn't free the transaction. This
512 * avoids pointless scanning of transactions which still
513 * weren't checkpointed.
514 */
515 if (!ret)
516 return;
517 } while (transaction != last_transaction);
518 }
519
520 /*
521 * Remove buffers from all checkpoint lists as journal is aborted and we just
522 * need to free memory
523 */
524 void jbd2_journal_destroy_checkpoint(journal_t *journal)
525 {
526 /*
527 * We loop because __jbd2_journal_clean_checkpoint_list() may abort
528 * early due to a need of rescheduling.
529 */
530 while (1) {
531 spin_lock(&journal->j_list_lock);
532 if (!journal->j_checkpoint_transactions) {
533 spin_unlock(&journal->j_list_lock);
534 break;
535 }
536 __jbd2_journal_clean_checkpoint_list(journal, true);
537 spin_unlock(&journal->j_list_lock);
538 cond_resched();
539 }
540 }
541
542 /*
543 * journal_remove_checkpoint: called after a buffer has been committed
544 * to disk (either by being write-back flushed to disk, or being
545 * committed to the log).
546 *
547 * We cannot safely clean a transaction out of the log until all of the
548 * buffer updates committed in that transaction have safely been stored
549 * elsewhere on disk. To achieve this, all of the buffers in a
550 * transaction need to be maintained on the transaction's checkpoint
551 * lists until they have been rewritten, at which point this function is
552 * called to remove the buffer from the existing transaction's
553 * checkpoint lists.
554 *
555 * The function returns 1 if it frees the transaction, 0 otherwise.
556 * The function can free jh and bh.
557 *
558 * This function is called with j_list_lock held.
559 */
560 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
561 {
562 struct transaction_chp_stats_s *stats;
563 transaction_t *transaction;
564 journal_t *journal;
565 int ret = 0;
566
567 JBUFFER_TRACE(jh, "entry");
568
569 if ((transaction = jh->b_cp_transaction) == NULL) {
570 JBUFFER_TRACE(jh, "not on transaction");
571 goto out;
572 }
573 journal = transaction->t_journal;
574
575 JBUFFER_TRACE(jh, "removing from transaction");
576 __buffer_unlink(jh);
577 jh->b_cp_transaction = NULL;
578 jbd2_journal_put_journal_head(jh);
579
580 if (transaction->t_checkpoint_list != NULL ||
581 transaction->t_checkpoint_io_list != NULL)
582 goto out;
583
584 /*
585 * There is one special case to worry about: if we have just pulled the
586 * buffer off a running or committing transaction's checkpoing list,
587 * then even if the checkpoint list is empty, the transaction obviously
588 * cannot be dropped!
589 *
590 * The locking here around t_state is a bit sleazy.
591 * See the comment at the end of jbd2_journal_commit_transaction().
592 */
593 if (transaction->t_state != T_FINISHED)
594 goto out;
595
596 /* OK, that was the last buffer for the transaction: we can now
597 safely remove this transaction from the log */
598 stats = &transaction->t_chp_stats;
599 if (stats->cs_chp_time)
600 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
601 jiffies);
602 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
603 transaction->t_tid, stats);
604
605 __jbd2_journal_drop_transaction(journal, transaction);
606 jbd2_journal_free_transaction(transaction);
607 ret = 1;
608 out:
609 return ret;
610 }
611
612 /*
613 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
614 * list so that we know when it is safe to clean the transaction out of
615 * the log.
616 *
617 * Called with the journal locked.
618 * Called with j_list_lock held.
619 */
620 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
621 transaction_t *transaction)
622 {
623 JBUFFER_TRACE(jh, "entry");
624 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
625 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
626
627 /* Get reference for checkpointing transaction */
628 jbd2_journal_grab_journal_head(jh2bh(jh));
629 jh->b_cp_transaction = transaction;
630
631 if (!transaction->t_checkpoint_list) {
632 jh->b_cpnext = jh->b_cpprev = jh;
633 } else {
634 jh->b_cpnext = transaction->t_checkpoint_list;
635 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
636 jh->b_cpprev->b_cpnext = jh;
637 jh->b_cpnext->b_cpprev = jh;
638 }
639 transaction->t_checkpoint_list = jh;
640 }
641
642 /*
643 * We've finished with this transaction structure: adios...
644 *
645 * The transaction must have no links except for the checkpoint by this
646 * point.
647 *
648 * Called with the journal locked.
649 * Called with j_list_lock held.
650 */
651
652 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
653 {
654 assert_spin_locked(&journal->j_list_lock);
655 if (transaction->t_cpnext) {
656 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
657 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
658 if (journal->j_checkpoint_transactions == transaction)
659 journal->j_checkpoint_transactions =
660 transaction->t_cpnext;
661 if (journal->j_checkpoint_transactions == transaction)
662 journal->j_checkpoint_transactions = NULL;
663 }
664
665 J_ASSERT(transaction->t_state == T_FINISHED);
666 J_ASSERT(transaction->t_buffers == NULL);
667 J_ASSERT(transaction->t_forget == NULL);
668 J_ASSERT(transaction->t_shadow_list == NULL);
669 J_ASSERT(transaction->t_checkpoint_list == NULL);
670 J_ASSERT(transaction->t_checkpoint_io_list == NULL);
671 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
672 J_ASSERT(journal->j_committing_transaction != transaction);
673 J_ASSERT(journal->j_running_transaction != transaction);
674
675 trace_jbd2_drop_transaction(journal, transaction);
676
677 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
678 }