1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/checkpoint.c
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
7 * Copyright 1999 Red Hat Software --- All Rights Reserved
9 * Checkpoint routines for the generic filesystem journaling code.
10 * Part of the ext2fs journaling system.
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
17 #include <linux/time.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>
26 * Unlink a buffer from a transaction checkpoint list.
28 * Called with j_list_lock held.
30 static inline void __buffer_unlink(struct journal_head *jh)
32 transaction_t *transaction = jh->b_cp_transaction;
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;
44 * Check a checkpoint buffer could be release or not.
46 * Requires j_list_lock
48 static inline bool __cp_buffer_busy(struct journal_head *jh)
50 struct buffer_head *bh = jh2bh(jh);
52 return (jh->b_transaction || buffer_locked(bh) || buffer_dirty(bh));
56 * __jbd2_log_wait_for_space: wait until there is space in the journal.
58 * Called under j-state_lock *only*. It will be unlocked if we have to wait
59 * for a checkpoint to free up some space in the log.
61 void __jbd2_log_wait_for_space(journal_t *journal)
62 __acquires(&journal->j_state_lock)
63 __releases(&journal->j_state_lock)
65 int nblocks, space_left;
66 /* assert_spin_locked(&journal->j_state_lock); */
68 nblocks = journal->j_max_transaction_buffers;
69 while (jbd2_log_space_left(journal) < nblocks) {
70 write_unlock(&journal->j_state_lock);
71 mutex_lock_io(&journal->j_checkpoint_mutex);
74 * Test again, another process may have checkpointed while we
75 * were waiting for the checkpoint lock. If there are no
76 * transactions ready to be checkpointed, try to recover
77 * journal space by calling cleanup_journal_tail(), and if
78 * that doesn't work, by waiting for the currently committing
79 * transaction to complete. If there is absolutely no way
80 * to make progress, this is either a BUG or corrupted
81 * filesystem, so abort the journal and leave a stack
82 * trace for forensic evidence.
84 write_lock(&journal->j_state_lock);
85 if (journal->j_flags & JBD2_ABORT) {
86 mutex_unlock(&journal->j_checkpoint_mutex);
89 spin_lock(&journal->j_list_lock);
90 space_left = jbd2_log_space_left(journal);
91 if (space_left < nblocks) {
92 int chkpt = journal->j_checkpoint_transactions != NULL;
95 if (journal->j_committing_transaction)
96 tid = journal->j_committing_transaction->t_tid;
97 spin_unlock(&journal->j_list_lock);
98 write_unlock(&journal->j_state_lock);
100 jbd2_log_do_checkpoint(journal);
101 } else if (jbd2_cleanup_journal_tail(journal) == 0) {
102 /* We were able to recover space; yay! */
106 * jbd2_journal_commit_transaction() may want
107 * to take the checkpoint_mutex if JBD2_FLUSHED
108 * is set. So we need to temporarily drop it.
110 mutex_unlock(&journal->j_checkpoint_mutex);
111 jbd2_log_wait_commit(journal, tid);
112 write_lock(&journal->j_state_lock);
115 printk(KERN_ERR "%s: needed %d blocks and "
116 "only had %d space available\n",
117 __func__, nblocks, space_left);
118 printk(KERN_ERR "%s: no way to get more "
119 "journal space in %s\n", __func__,
122 jbd2_journal_abort(journal, -EIO);
124 write_lock(&journal->j_state_lock);
126 spin_unlock(&journal->j_list_lock);
128 mutex_unlock(&journal->j_checkpoint_mutex);
133 __flush_batch(journal_t *journal, int *batch_count)
136 struct blk_plug plug;
138 blk_start_plug(&plug);
139 for (i = 0; i < *batch_count; i++)
140 write_dirty_buffer(journal->j_chkpt_bhs[i], REQ_SYNC);
141 blk_finish_plug(&plug);
143 for (i = 0; i < *batch_count; i++) {
144 struct buffer_head *bh = journal->j_chkpt_bhs[i];
145 BUFFER_TRACE(bh, "brelse");
147 journal->j_chkpt_bhs[i] = NULL;
153 * Perform an actual checkpoint. We take the first transaction on the
154 * list of transactions to be checkpointed and send all its buffers
155 * to disk. We submit larger chunks of data at once.
157 * The journal should be locked before calling this function.
158 * Called with j_checkpoint_mutex held.
160 int jbd2_log_do_checkpoint(journal_t *journal)
162 struct journal_head *jh;
163 struct buffer_head *bh;
164 transaction_t *transaction;
166 int result, batch_count = 0;
168 jbd2_debug(1, "Start checkpoint\n");
171 * First thing: if there are any transactions in the log which
172 * don't need checkpointing, just eliminate them from the
173 * journal straight away.
175 result = jbd2_cleanup_journal_tail(journal);
176 trace_jbd2_checkpoint(journal, result);
177 jbd2_debug(1, "cleanup_journal_tail returned %d\n", result);
182 * OK, we need to start writing disk blocks. Take one transaction
185 spin_lock(&journal->j_list_lock);
186 if (!journal->j_checkpoint_transactions)
188 transaction = journal->j_checkpoint_transactions;
189 if (transaction->t_chp_stats.cs_chp_time == 0)
190 transaction->t_chp_stats.cs_chp_time = jiffies;
191 this_tid = transaction->t_tid;
194 * If someone cleaned up this transaction while we slept, we're
195 * done (maybe it's a new transaction, but it fell at the same
198 if (journal->j_checkpoint_transactions != transaction ||
199 transaction->t_tid != this_tid)
202 /* checkpoint all of the transaction's buffers */
203 while (transaction->t_checkpoint_list) {
204 jh = transaction->t_checkpoint_list;
208 * The buffer may be writing back, or flushing out in the
209 * last couple of cycles, or re-adding into a new transaction,
210 * need to check it again until it's unlocked.
212 if (buffer_locked(bh)) {
214 spin_unlock(&journal->j_list_lock);
216 /* the journal_head may have gone by now */
217 BUFFER_TRACE(bh, "brelse");
221 if (jh->b_transaction != NULL) {
222 transaction_t *t = jh->b_transaction;
223 tid_t tid = t->t_tid;
225 transaction->t_chp_stats.cs_forced_to_close++;
226 spin_unlock(&journal->j_list_lock);
227 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
229 * The journal thread is dead; so
230 * starting and waiting for a commit
231 * to finish will cause us to wait for
232 * a _very_ long time.
235 "JBD2: %s: Waiting for Godot: block %llu\n",
236 journal->j_devname, (unsigned long long) bh->b_blocknr);
239 __flush_batch(journal, &batch_count);
240 jbd2_log_start_commit(journal, tid);
242 * jbd2_journal_commit_transaction() may want
243 * to take the checkpoint_mutex if JBD2_FLUSHED
244 * is set, jbd2_update_log_tail() called by
245 * jbd2_journal_commit_transaction() may also take
246 * checkpoint_mutex. So we need to temporarily
249 mutex_unlock(&journal->j_checkpoint_mutex);
250 jbd2_log_wait_commit(journal, tid);
251 mutex_lock_io(&journal->j_checkpoint_mutex);
252 spin_lock(&journal->j_list_lock);
255 if (!buffer_dirty(bh)) {
256 BUFFER_TRACE(bh, "remove from checkpoint");
258 * If the transaction was released or the checkpoint
259 * list was empty, we're done.
261 if (__jbd2_journal_remove_checkpoint(jh) ||
262 !transaction->t_checkpoint_list)
266 * We are about to write the buffer, it could be
267 * raced by some other transaction shrink or buffer
268 * re-log logic once we release the j_list_lock,
269 * leave it on the checkpoint list and check status
270 * again to make sure it's clean.
272 BUFFER_TRACE(bh, "queue");
274 J_ASSERT_BH(bh, !buffer_jwrite(bh));
275 journal->j_chkpt_bhs[batch_count++] = bh;
276 transaction->t_chp_stats.cs_written++;
277 transaction->t_checkpoint_list = jh->b_cpnext;
280 if ((batch_count == JBD2_NR_BATCH) ||
281 need_resched() || spin_needbreak(&journal->j_list_lock) ||
282 jh2bh(transaction->t_checkpoint_list) == journal->j_chkpt_bhs[0])
283 goto unlock_and_flush;
288 spin_unlock(&journal->j_list_lock);
291 __flush_batch(journal, &batch_count);
292 spin_lock(&journal->j_list_lock);
297 spin_unlock(&journal->j_list_lock);
298 result = jbd2_cleanup_journal_tail(journal);
300 return (result < 0) ? result : 0;
304 * Check the list of checkpoint transactions for the journal to see if
305 * we have already got rid of any since the last update of the log tail
306 * in the journal superblock. If so, we can instantly roll the
307 * superblock forward to remove those transactions from the log.
309 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
311 * Called with the journal lock held.
313 * This is the only part of the journaling code which really needs to be
314 * aware of transaction aborts. Checkpointing involves writing to the
315 * main filesystem area rather than to the journal, so it can proceed
316 * even in abort state, but we must not update the super block if
317 * checkpointing may have failed. Otherwise, we would lose some metadata
318 * buffers which should be written-back to the filesystem.
321 int jbd2_cleanup_journal_tail(journal_t *journal)
324 unsigned long blocknr;
326 if (is_journal_aborted(journal))
329 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
331 J_ASSERT(blocknr != 0);
334 * We need to make sure that any blocks that were recently written out
335 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
336 * we drop the transactions from the journal. It's unlikely this will
337 * be necessary, especially with an appropriately sized journal, but we
338 * need this to guarantee correctness. Fortunately
339 * jbd2_cleanup_journal_tail() doesn't get called all that often.
341 if (journal->j_flags & JBD2_BARRIER)
342 blkdev_issue_flush(journal->j_fs_dev);
344 return __jbd2_update_log_tail(journal, first_tid, blocknr);
348 /* Checkpoint list management */
351 * journal_clean_one_cp_list
353 * Find all the written-back checkpoint buffers in the given list and
354 * release them. If 'destroy' is set, clean all buffers unconditionally.
356 * Called with j_list_lock held.
357 * Returns 1 if we freed the transaction, 0 otherwise.
359 static int journal_clean_one_cp_list(struct journal_head *jh, bool destroy)
361 struct journal_head *last_jh;
362 struct journal_head *next_jh = jh;
367 last_jh = jh->b_cpprev;
370 next_jh = jh->b_cpnext;
372 if (!destroy && __cp_buffer_busy(jh))
375 if (__jbd2_journal_remove_checkpoint(jh))
378 * This function only frees up some memory
379 * if possible so we dont have an obligation
380 * to finish processing. Bail out if preemption
385 } while (jh != last_jh);
391 * journal_shrink_one_cp_list
393 * Find 'nr_to_scan' written-back checkpoint buffers in the given list
394 * and try to release them. If the whole transaction is released, set
395 * the 'released' parameter. Return the number of released checkpointed
398 * Called with j_list_lock held.
400 static unsigned long journal_shrink_one_cp_list(struct journal_head *jh,
401 unsigned long *nr_to_scan,
404 struct journal_head *last_jh;
405 struct journal_head *next_jh = jh;
406 unsigned long nr_freed = 0;
409 if (!jh || *nr_to_scan == 0)
412 last_jh = jh->b_cpprev;
415 next_jh = jh->b_cpnext;
418 if (__cp_buffer_busy(jh))
422 ret = __jbd2_journal_remove_checkpoint(jh);
430 } while (jh != last_jh && *nr_to_scan);
436 * jbd2_journal_shrink_checkpoint_list
438 * Find 'nr_to_scan' written-back checkpoint buffers in the journal
439 * and try to release them. Return the number of released checkpointed
442 * Called with j_list_lock held.
444 unsigned long jbd2_journal_shrink_checkpoint_list(journal_t *journal,
445 unsigned long *nr_to_scan)
447 transaction_t *transaction, *last_transaction, *next_transaction;
449 tid_t first_tid = 0, last_tid = 0, next_tid = 0;
451 unsigned long nr_freed = 0;
452 unsigned long nr_scanned = *nr_to_scan;
455 spin_lock(&journal->j_list_lock);
456 if (!journal->j_checkpoint_transactions) {
457 spin_unlock(&journal->j_list_lock);
462 * Get next shrink transaction, resume previous scan or start
463 * over again. If some others do checkpoint and drop transaction
464 * from the checkpoint list, we ignore saved j_shrink_transaction
465 * and start over unconditionally.
467 if (journal->j_shrink_transaction)
468 transaction = journal->j_shrink_transaction;
470 transaction = journal->j_checkpoint_transactions;
473 first_tid = transaction->t_tid;
474 last_transaction = journal->j_checkpoint_transactions->t_cpprev;
475 next_transaction = transaction;
476 last_tid = last_transaction->t_tid;
478 transaction = next_transaction;
479 next_transaction = transaction->t_cpnext;
480 tid = transaction->t_tid;
483 nr_freed += journal_shrink_one_cp_list(transaction->t_checkpoint_list,
484 nr_to_scan, &released);
485 if (*nr_to_scan == 0)
487 if (need_resched() || spin_needbreak(&journal->j_list_lock))
489 } while (transaction != last_transaction);
491 if (transaction != last_transaction) {
492 journal->j_shrink_transaction = next_transaction;
493 next_tid = next_transaction->t_tid;
495 journal->j_shrink_transaction = NULL;
499 spin_unlock(&journal->j_list_lock);
502 if (*nr_to_scan && next_tid)
505 nr_scanned -= *nr_to_scan;
506 trace_jbd2_shrink_checkpoint_list(journal, first_tid, tid, last_tid,
507 nr_freed, nr_scanned, next_tid);
513 * journal_clean_checkpoint_list
515 * Find all the written-back checkpoint buffers in the journal and release them.
516 * If 'destroy' is set, release all buffers unconditionally.
518 * Called with j_list_lock held.
520 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy)
522 transaction_t *transaction, *last_transaction, *next_transaction;
525 transaction = journal->j_checkpoint_transactions;
529 last_transaction = transaction->t_cpprev;
530 next_transaction = transaction;
532 transaction = next_transaction;
533 next_transaction = transaction->t_cpnext;
534 ret = journal_clean_one_cp_list(transaction->t_checkpoint_list,
537 * This function only frees up some memory if possible so we
538 * dont have an obligation to finish processing. Bail out if
539 * preemption requested:
544 * Stop scanning if we couldn't free the transaction. This
545 * avoids pointless scanning of transactions which still
546 * weren't checkpointed.
550 } while (transaction != last_transaction);
554 * Remove buffers from all checkpoint lists as journal is aborted and we just
555 * need to free memory
557 void jbd2_journal_destroy_checkpoint(journal_t *journal)
560 * We loop because __jbd2_journal_clean_checkpoint_list() may abort
561 * early due to a need of rescheduling.
564 spin_lock(&journal->j_list_lock);
565 if (!journal->j_checkpoint_transactions) {
566 spin_unlock(&journal->j_list_lock);
569 __jbd2_journal_clean_checkpoint_list(journal, true);
570 spin_unlock(&journal->j_list_lock);
576 * journal_remove_checkpoint: called after a buffer has been committed
577 * to disk (either by being write-back flushed to disk, or being
578 * committed to the log).
580 * We cannot safely clean a transaction out of the log until all of the
581 * buffer updates committed in that transaction have safely been stored
582 * elsewhere on disk. To achieve this, all of the buffers in a
583 * transaction need to be maintained on the transaction's checkpoint
584 * lists until they have been rewritten, at which point this function is
585 * called to remove the buffer from the existing transaction's
588 * The function returns 1 if it frees the transaction, 0 otherwise.
589 * The function can free jh and bh.
591 * This function is called with j_list_lock held.
593 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
595 struct transaction_chp_stats_s *stats;
596 transaction_t *transaction;
598 struct buffer_head *bh = jh2bh(jh);
600 JBUFFER_TRACE(jh, "entry");
602 transaction = jh->b_cp_transaction;
604 JBUFFER_TRACE(jh, "not on transaction");
607 journal = transaction->t_journal;
609 JBUFFER_TRACE(jh, "removing from transaction");
612 * If we have failed to write the buffer out to disk, the filesystem
613 * may become inconsistent. We cannot abort the journal here since
614 * we hold j_list_lock and we have to be careful about races with
615 * jbd2_journal_destroy(). So mark the writeback IO error in the
616 * journal here and we abort the journal later from a better context.
618 if (buffer_write_io_error(bh))
619 set_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags);
622 jh->b_cp_transaction = NULL;
623 percpu_counter_dec(&journal->j_checkpoint_jh_count);
624 jbd2_journal_put_journal_head(jh);
626 /* Is this transaction empty? */
627 if (transaction->t_checkpoint_list)
631 * There is one special case to worry about: if we have just pulled the
632 * buffer off a running or committing transaction's checkpoing list,
633 * then even if the checkpoint list is empty, the transaction obviously
636 * The locking here around t_state is a bit sleazy.
637 * See the comment at the end of jbd2_journal_commit_transaction().
639 if (transaction->t_state != T_FINISHED)
643 * OK, that was the last buffer for the transaction, we can now
644 * safely remove this transaction from the log.
646 stats = &transaction->t_chp_stats;
647 if (stats->cs_chp_time)
648 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
650 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
651 transaction->t_tid, stats);
653 __jbd2_journal_drop_transaction(journal, transaction);
654 jbd2_journal_free_transaction(transaction);
659 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
660 * list so that we know when it is safe to clean the transaction out of
663 * Called with the journal locked.
664 * Called with j_list_lock held.
666 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
667 transaction_t *transaction)
669 JBUFFER_TRACE(jh, "entry");
670 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
671 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
673 /* Get reference for checkpointing transaction */
674 jbd2_journal_grab_journal_head(jh2bh(jh));
675 jh->b_cp_transaction = transaction;
677 if (!transaction->t_checkpoint_list) {
678 jh->b_cpnext = jh->b_cpprev = jh;
680 jh->b_cpnext = transaction->t_checkpoint_list;
681 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
682 jh->b_cpprev->b_cpnext = jh;
683 jh->b_cpnext->b_cpprev = jh;
685 transaction->t_checkpoint_list = jh;
686 percpu_counter_inc(&transaction->t_journal->j_checkpoint_jh_count);
690 * We've finished with this transaction structure: adios...
692 * The transaction must have no links except for the checkpoint by this
695 * Called with the journal locked.
696 * Called with j_list_lock held.
699 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
701 assert_spin_locked(&journal->j_list_lock);
703 journal->j_shrink_transaction = NULL;
704 if (transaction->t_cpnext) {
705 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
706 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
707 if (journal->j_checkpoint_transactions == transaction)
708 journal->j_checkpoint_transactions =
709 transaction->t_cpnext;
710 if (journal->j_checkpoint_transactions == transaction)
711 journal->j_checkpoint_transactions = NULL;
714 J_ASSERT(transaction->t_state == T_FINISHED);
715 J_ASSERT(transaction->t_buffers == NULL);
716 J_ASSERT(transaction->t_forget == NULL);
717 J_ASSERT(transaction->t_shadow_list == NULL);
718 J_ASSERT(transaction->t_checkpoint_list == NULL);
719 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
720 J_ASSERT(journal->j_committing_transaction != transaction);
721 J_ASSERT(journal->j_running_transaction != transaction);
723 trace_jbd2_drop_transaction(journal, transaction);
725 jbd2_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);