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;
207 if (jh->b_transaction != NULL) {
208 transaction_t *t = jh->b_transaction;
209 tid_t tid = t->t_tid;
211 transaction->t_chp_stats.cs_forced_to_close++;
212 spin_unlock(&journal->j_list_lock);
213 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
215 * The journal thread is dead; so
216 * starting and waiting for a commit
217 * to finish will cause us to wait for
218 * a _very_ long time.
221 "JBD2: %s: Waiting for Godot: block %llu\n",
222 journal->j_devname, (unsigned long long) bh->b_blocknr);
225 __flush_batch(journal, &batch_count);
226 jbd2_log_start_commit(journal, tid);
228 * jbd2_journal_commit_transaction() may want
229 * to take the checkpoint_mutex if JBD2_FLUSHED
230 * is set, jbd2_update_log_tail() called by
231 * jbd2_journal_commit_transaction() may also take
232 * checkpoint_mutex. So we need to temporarily
235 mutex_unlock(&journal->j_checkpoint_mutex);
236 jbd2_log_wait_commit(journal, tid);
237 mutex_lock_io(&journal->j_checkpoint_mutex);
238 spin_lock(&journal->j_list_lock);
241 if (!trylock_buffer(bh)) {
243 * The buffer is locked, it may be writing back, or
244 * flushing out in the last couple of cycles, or
245 * re-adding into a new transaction, need to check
246 * it again until it's unlocked.
249 spin_unlock(&journal->j_list_lock);
251 /* the journal_head may have gone by now */
252 BUFFER_TRACE(bh, "brelse");
255 } else if (!buffer_dirty(bh)) {
257 BUFFER_TRACE(bh, "remove from checkpoint");
259 * If the transaction was released or the checkpoint
260 * list was empty, we're done.
262 if (__jbd2_journal_remove_checkpoint(jh) ||
263 !transaction->t_checkpoint_list)
268 * We are about to write the buffer, it could be
269 * raced by some other transaction shrink or buffer
270 * re-log logic once we release the j_list_lock,
271 * leave it on the checkpoint list and check status
272 * again to make sure it's clean.
274 BUFFER_TRACE(bh, "queue");
276 J_ASSERT_BH(bh, !buffer_jwrite(bh));
277 journal->j_chkpt_bhs[batch_count++] = bh;
278 transaction->t_chp_stats.cs_written++;
279 transaction->t_checkpoint_list = jh->b_cpnext;
282 if ((batch_count == JBD2_NR_BATCH) ||
283 need_resched() || spin_needbreak(&journal->j_list_lock) ||
284 jh2bh(transaction->t_checkpoint_list) == journal->j_chkpt_bhs[0])
285 goto unlock_and_flush;
290 spin_unlock(&journal->j_list_lock);
293 __flush_batch(journal, &batch_count);
294 spin_lock(&journal->j_list_lock);
299 spin_unlock(&journal->j_list_lock);
300 result = jbd2_cleanup_journal_tail(journal);
302 return (result < 0) ? result : 0;
306 * Check the list of checkpoint transactions for the journal to see if
307 * we have already got rid of any since the last update of the log tail
308 * in the journal superblock. If so, we can instantly roll the
309 * superblock forward to remove those transactions from the log.
311 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
313 * Called with the journal lock held.
315 * This is the only part of the journaling code which really needs to be
316 * aware of transaction aborts. Checkpointing involves writing to the
317 * main filesystem area rather than to the journal, so it can proceed
318 * even in abort state, but we must not update the super block if
319 * checkpointing may have failed. Otherwise, we would lose some metadata
320 * buffers which should be written-back to the filesystem.
323 int jbd2_cleanup_journal_tail(journal_t *journal)
326 unsigned long blocknr;
328 if (is_journal_aborted(journal))
331 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
333 J_ASSERT(blocknr != 0);
336 * We need to make sure that any blocks that were recently written out
337 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
338 * we drop the transactions from the journal. It's unlikely this will
339 * be necessary, especially with an appropriately sized journal, but we
340 * need this to guarantee correctness. Fortunately
341 * jbd2_cleanup_journal_tail() doesn't get called all that often.
343 if (journal->j_flags & JBD2_BARRIER)
344 blkdev_issue_flush(journal->j_fs_dev);
346 return __jbd2_update_log_tail(journal, first_tid, blocknr);
350 /* Checkpoint list management */
352 enum shrink_type {SHRINK_DESTROY, SHRINK_BUSY_STOP, SHRINK_BUSY_SKIP};
355 * journal_shrink_one_cp_list
357 * Find all the written-back checkpoint buffers in the given list
358 * and try to release them. If the whole transaction is released, set
359 * the 'released' parameter. Return the number of released checkpointed
362 * Called with j_list_lock held.
364 static unsigned long journal_shrink_one_cp_list(struct journal_head *jh,
365 enum shrink_type type,
368 struct journal_head *last_jh;
369 struct journal_head *next_jh = jh;
370 unsigned long nr_freed = 0;
377 last_jh = jh->b_cpprev;
380 next_jh = jh->b_cpnext;
382 if (type == SHRINK_DESTROY) {
383 ret = __jbd2_journal_remove_checkpoint(jh);
385 ret = jbd2_journal_try_remove_checkpoint(jh);
387 if (type == SHRINK_BUSY_SKIP)
401 } while (jh != last_jh);
407 * jbd2_journal_shrink_checkpoint_list
409 * Find 'nr_to_scan' written-back checkpoint buffers in the journal
410 * and try to release them. Return the number of released checkpointed
413 * Called with j_list_lock held.
415 unsigned long jbd2_journal_shrink_checkpoint_list(journal_t *journal,
416 unsigned long *nr_to_scan)
418 transaction_t *transaction, *last_transaction, *next_transaction;
419 bool __maybe_unused released;
420 tid_t first_tid = 0, last_tid = 0, next_tid = 0;
422 unsigned long nr_freed = 0;
426 spin_lock(&journal->j_list_lock);
427 if (!journal->j_checkpoint_transactions) {
428 spin_unlock(&journal->j_list_lock);
433 * Get next shrink transaction, resume previous scan or start
434 * over again. If some others do checkpoint and drop transaction
435 * from the checkpoint list, we ignore saved j_shrink_transaction
436 * and start over unconditionally.
438 if (journal->j_shrink_transaction)
439 transaction = journal->j_shrink_transaction;
441 transaction = journal->j_checkpoint_transactions;
444 first_tid = transaction->t_tid;
445 last_transaction = journal->j_checkpoint_transactions->t_cpprev;
446 next_transaction = transaction;
447 last_tid = last_transaction->t_tid;
449 transaction = next_transaction;
450 next_transaction = transaction->t_cpnext;
451 tid = transaction->t_tid;
453 freed = journal_shrink_one_cp_list(transaction->t_checkpoint_list,
454 SHRINK_BUSY_SKIP, &released);
456 (*nr_to_scan) -= min(*nr_to_scan, freed);
457 if (*nr_to_scan == 0)
459 if (need_resched() || spin_needbreak(&journal->j_list_lock))
461 } while (transaction != last_transaction);
463 if (transaction != last_transaction) {
464 journal->j_shrink_transaction = next_transaction;
465 next_tid = next_transaction->t_tid;
467 journal->j_shrink_transaction = NULL;
471 spin_unlock(&journal->j_list_lock);
474 if (*nr_to_scan && next_tid)
477 trace_jbd2_shrink_checkpoint_list(journal, first_tid, tid, last_tid,
484 * journal_clean_checkpoint_list
486 * Find all the written-back checkpoint buffers in the journal and release them.
487 * If 'destroy' is set, release all buffers unconditionally.
489 * Called with j_list_lock held.
491 void __jbd2_journal_clean_checkpoint_list(journal_t *journal, bool destroy)
493 transaction_t *transaction, *last_transaction, *next_transaction;
494 enum shrink_type type;
497 transaction = journal->j_checkpoint_transactions;
501 type = destroy ? SHRINK_DESTROY : SHRINK_BUSY_STOP;
502 last_transaction = transaction->t_cpprev;
503 next_transaction = transaction;
505 transaction = next_transaction;
506 next_transaction = transaction->t_cpnext;
507 journal_shrink_one_cp_list(transaction->t_checkpoint_list,
510 * This function only frees up some memory if possible so we
511 * dont have an obligation to finish processing. Bail out if
512 * preemption requested:
517 * Stop scanning if we couldn't free the transaction. This
518 * avoids pointless scanning of transactions which still
519 * weren't checkpointed.
523 } while (transaction != last_transaction);
527 * Remove buffers from all checkpoint lists as journal is aborted and we just
528 * need to free memory
530 void jbd2_journal_destroy_checkpoint(journal_t *journal)
533 * We loop because __jbd2_journal_clean_checkpoint_list() may abort
534 * early due to a need of rescheduling.
537 spin_lock(&journal->j_list_lock);
538 if (!journal->j_checkpoint_transactions) {
539 spin_unlock(&journal->j_list_lock);
542 __jbd2_journal_clean_checkpoint_list(journal, true);
543 spin_unlock(&journal->j_list_lock);
549 * journal_remove_checkpoint: called after a buffer has been committed
550 * to disk (either by being write-back flushed to disk, or being
551 * committed to the log).
553 * We cannot safely clean a transaction out of the log until all of the
554 * buffer updates committed in that transaction have safely been stored
555 * elsewhere on disk. To achieve this, all of the buffers in a
556 * transaction need to be maintained on the transaction's checkpoint
557 * lists until they have been rewritten, at which point this function is
558 * called to remove the buffer from the existing transaction's
561 * The function returns 1 if it frees the transaction, 0 otherwise.
562 * The function can free jh and bh.
564 * This function is called with j_list_lock held.
566 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
568 struct transaction_chp_stats_s *stats;
569 transaction_t *transaction;
571 struct buffer_head *bh = jh2bh(jh);
573 JBUFFER_TRACE(jh, "entry");
575 transaction = jh->b_cp_transaction;
577 JBUFFER_TRACE(jh, "not on transaction");
580 journal = transaction->t_journal;
582 JBUFFER_TRACE(jh, "removing from transaction");
585 * If we have failed to write the buffer out to disk, the filesystem
586 * may become inconsistent. We cannot abort the journal here since
587 * we hold j_list_lock and we have to be careful about races with
588 * jbd2_journal_destroy(). So mark the writeback IO error in the
589 * journal here and we abort the journal later from a better context.
591 if (buffer_write_io_error(bh))
592 set_bit(JBD2_CHECKPOINT_IO_ERROR, &journal->j_atomic_flags);
595 jh->b_cp_transaction = NULL;
596 percpu_counter_dec(&journal->j_checkpoint_jh_count);
597 jbd2_journal_put_journal_head(jh);
599 /* Is this transaction empty? */
600 if (transaction->t_checkpoint_list)
604 * There is one special case to worry about: if we have just pulled the
605 * buffer off a running or committing transaction's checkpoing list,
606 * then even if the checkpoint list is empty, the transaction obviously
609 * The locking here around t_state is a bit sleazy.
610 * See the comment at the end of jbd2_journal_commit_transaction().
612 if (transaction->t_state != T_FINISHED)
616 * OK, that was the last buffer for the transaction, we can now
617 * safely remove this transaction from the log.
619 stats = &transaction->t_chp_stats;
620 if (stats->cs_chp_time)
621 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
623 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
624 transaction->t_tid, stats);
626 __jbd2_journal_drop_transaction(journal, transaction);
627 jbd2_journal_free_transaction(transaction);
632 * Check the checkpoint buffer and try to remove it from the checkpoint
633 * list if it's clean. Returns -EBUSY if it is not clean, returns 1 if
634 * it frees the transaction, 0 otherwise.
636 * This function is called with j_list_lock held.
638 int jbd2_journal_try_remove_checkpoint(struct journal_head *jh)
640 struct buffer_head *bh = jh2bh(jh);
642 if (jh->b_transaction)
644 if (!trylock_buffer(bh))
646 if (buffer_dirty(bh)) {
653 * Buffer is clean and the IO has finished (we held the buffer
654 * lock) so the checkpoint is done. We can safely remove the
655 * buffer from this transaction.
657 JBUFFER_TRACE(jh, "remove from checkpoint list");
658 return __jbd2_journal_remove_checkpoint(jh);
662 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
663 * list so that we know when it is safe to clean the transaction out of
666 * Called with the journal locked.
667 * Called with j_list_lock held.
669 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
670 transaction_t *transaction)
672 JBUFFER_TRACE(jh, "entry");
673 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
674 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
676 /* Get reference for checkpointing transaction */
677 jbd2_journal_grab_journal_head(jh2bh(jh));
678 jh->b_cp_transaction = transaction;
680 if (!transaction->t_checkpoint_list) {
681 jh->b_cpnext = jh->b_cpprev = jh;
683 jh->b_cpnext = transaction->t_checkpoint_list;
684 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
685 jh->b_cpprev->b_cpnext = jh;
686 jh->b_cpnext->b_cpprev = jh;
688 transaction->t_checkpoint_list = jh;
689 percpu_counter_inc(&transaction->t_journal->j_checkpoint_jh_count);
693 * We've finished with this transaction structure: adios...
695 * The transaction must have no links except for the checkpoint by this
698 * Called with the journal locked.
699 * Called with j_list_lock held.
702 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
704 assert_spin_locked(&journal->j_list_lock);
706 journal->j_shrink_transaction = NULL;
707 if (transaction->t_cpnext) {
708 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
709 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
710 if (journal->j_checkpoint_transactions == transaction)
711 journal->j_checkpoint_transactions =
712 transaction->t_cpnext;
713 if (journal->j_checkpoint_transactions == transaction)
714 journal->j_checkpoint_transactions = NULL;
717 J_ASSERT(transaction->t_state == T_FINISHED);
718 J_ASSERT(transaction->t_buffers == NULL);
719 J_ASSERT(transaction->t_forget == NULL);
720 J_ASSERT(transaction->t_shadow_list == NULL);
721 J_ASSERT(transaction->t_checkpoint_list == NULL);
722 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
723 J_ASSERT(journal->j_committing_transaction != transaction);
724 J_ASSERT(journal->j_running_transaction != transaction);
726 trace_jbd2_drop_transaction(journal, transaction);
728 jbd2_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);