2 * linux/fs/jbd2/checkpoint.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
6 * Copyright 1999 Red Hat Software --- All Rights Reserved
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.
12 * Checkpoint routines for the generic filesystem journaling code.
13 * Part of the ext2fs journaling system.
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
20 #include <linux/time.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>
29 * Unlink a buffer from a transaction checkpoint list.
31 * Called with j_list_lock held.
33 static inline void __buffer_unlink_first(struct journal_head *jh)
35 transaction_t *transaction = jh->b_cp_transaction;
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;
47 * Unlink a buffer from a transaction checkpoint(io) list.
49 * Called with j_list_lock held.
51 static inline void __buffer_unlink(struct journal_head *jh)
53 transaction_t *transaction = jh->b_cp_transaction;
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;
64 * Move a buffer from the checkpoint list to the checkpoint io list
66 * Called with j_list_lock held
68 static inline void __buffer_relink_io(struct journal_head *jh)
70 transaction_t *transaction = jh->b_cp_transaction;
72 __buffer_unlink_first(jh);
74 if (!transaction->t_checkpoint_io_list) {
75 jh->b_cpnext = jh->b_cpprev = jh;
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;
82 transaction->t_checkpoint_io_list = jh;
86 * Try to release a checkpointed buffer from its transaction.
87 * Returns 1 if we released it and 2 if we also released the
90 * Requires j_list_lock
92 static int __try_to_free_cp_buf(struct journal_head *jh)
95 struct buffer_head *bh = jh2bh(jh);
97 if (jh->b_transaction == NULL && !buffer_locked(bh) &&
98 !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
100 * Get our reference so that bh cannot be freed before
104 JBUFFER_TRACE(jh, "remove from checkpoint list");
105 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
106 BUFFER_TRACE(bh, "release");
113 * __jbd2_log_wait_for_space: wait until there is space in the journal.
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.
118 void __jbd2_log_wait_for_space(journal_t *journal)
120 int nblocks, space_left;
121 /* assert_spin_locked(&journal->j_state_lock); */
123 nblocks = jbd2_space_needed(journal);
124 while (jbd2_log_space_left(journal) < nblocks) {
125 if (journal->j_flags & JBD2_ABORT)
127 write_unlock(&journal->j_state_lock);
128 mutex_lock(&journal->j_checkpoint_mutex);
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.
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;
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);
154 jbd2_log_do_checkpoint(journal);
155 } else if (jbd2_cleanup_journal_tail(journal) == 0) {
156 /* We were able to recover space; yay! */
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.
164 mutex_unlock(&journal->j_checkpoint_mutex);
165 jbd2_log_wait_commit(journal, tid);
166 write_lock(&journal->j_state_lock);
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__,
176 jbd2_journal_abort(journal, 0);
178 write_lock(&journal->j_state_lock);
180 spin_unlock(&journal->j_list_lock);
182 mutex_unlock(&journal->j_checkpoint_mutex);
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.
192 * Return 0 on success, and return <0 if some buffers have failed
195 * Called with j_list_lock held.
197 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
199 struct journal_head *jh;
200 struct buffer_head *bh;
205 this_tid = transaction->t_tid;
207 /* Did somebody clean up the transaction in the meanwhile? */
208 if (journal->j_checkpoint_transactions != transaction ||
209 transaction->t_tid != this_tid)
211 while (!released && transaction->t_checkpoint_io_list) {
212 jh = transaction->t_checkpoint_io_list;
215 if (buffer_locked(bh)) {
216 spin_unlock(&journal->j_list_lock);
218 /* the journal_head may have gone by now */
219 BUFFER_TRACE(bh, "brelse");
221 spin_lock(&journal->j_list_lock);
224 if (unlikely(buffer_write_io_error(bh)))
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
231 released = __jbd2_journal_remove_checkpoint(jh);
239 __flush_batch(journal_t *journal, int *batch_count)
242 struct blk_plug plug;
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);
249 for (i = 0; i < *batch_count; i++) {
250 struct buffer_head *bh = journal->j_chkpt_bhs[i];
251 BUFFER_TRACE(bh, "brelse");
258 * Try to flush one buffer from the checkpoint list to disk.
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
264 * Called with j_list_lock held and drops it if 1 is returned
266 static int __process_buffer(journal_t *journal, struct journal_head *jh,
267 int *batch_count, transaction_t *transaction)
269 struct buffer_head *bh = jh2bh(jh);
272 if (buffer_locked(bh)) {
274 spin_unlock(&journal->j_list_lock);
276 /* the journal_head may have gone by now */
277 BUFFER_TRACE(bh, "brelse");
280 } else if (jh->b_transaction != NULL) {
281 transaction_t *t = jh->b_transaction;
282 tid_t tid = t->t_tid;
284 transaction->t_chp_stats.cs_forced_to_close++;
285 spin_unlock(&journal->j_list_lock);
286 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
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.
292 printk(KERN_ERR "JBD2: %s: "
293 "Waiting for Godot: block %llu\n",
295 (unsigned long long) bh->b_blocknr);
296 jbd2_log_start_commit(journal, tid);
297 jbd2_log_wait_commit(journal, tid);
299 } else if (!buffer_dirty(bh)) {
301 if (unlikely(buffer_write_io_error(bh)))
304 BUFFER_TRACE(bh, "remove from checkpoint");
305 __jbd2_journal_remove_checkpoint(jh);
306 spin_unlock(&journal->j_list_lock);
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.
316 BUFFER_TRACE(bh, "queue");
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++;
323 if (*batch_count == JBD2_NR_BATCH) {
324 spin_unlock(&journal->j_list_lock);
325 __flush_batch(journal, batch_count);
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.
337 * The journal should be locked before calling this function.
338 * Called with j_checkpoint_mutex held.
340 int jbd2_log_do_checkpoint(journal_t *journal)
342 transaction_t *transaction;
346 jbd_debug(1, "Start checkpoint\n");
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.
353 result = jbd2_cleanup_journal_tail(journal);
354 trace_jbd2_checkpoint(journal, result);
355 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
360 * OK, we need to start writing disk blocks. Take one transaction
364 spin_lock(&journal->j_list_lock);
365 if (!journal->j_checkpoint_transactions)
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;
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
377 if (journal->j_checkpoint_transactions == transaction &&
378 transaction->t_tid == this_tid) {
380 struct journal_head *jh;
383 while (!retry && transaction->t_checkpoint_list) {
384 jh = transaction->t_checkpoint_list;
385 retry = __process_buffer(journal, jh, &batch_count,
387 if (retry < 0 && !result)
389 if (!retry && (need_resched() ||
390 spin_needbreak(&journal->j_list_lock))) {
391 spin_unlock(&journal->j_list_lock);
399 spin_unlock(&journal->j_list_lock);
402 __flush_batch(journal, &batch_count);
406 spin_lock(&journal->j_list_lock);
410 * Now we have cleaned up the first transaction's checkpoint
411 * list. Let's clean up the second one
413 err = __wait_cp_io(journal, transaction);
418 spin_unlock(&journal->j_list_lock);
420 jbd2_journal_abort(journal, result);
422 result = jbd2_cleanup_journal_tail(journal);
424 return (result < 0) ? result : 0;
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.
433 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
435 * Called with the journal lock held.
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.
445 int jbd2_cleanup_journal_tail(journal_t *journal)
448 unsigned long blocknr;
450 if (is_journal_aborted(journal))
453 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
455 J_ASSERT(blocknr != 0);
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.
465 if (journal->j_flags & JBD2_BARRIER)
466 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
468 __jbd2_update_log_tail(journal, first_tid, blocknr);
473 /* Checkpoint list management */
476 * journal_clean_one_cp_list
478 * Find all the written-back checkpoint buffers in the given list and
481 * Called with the journal locked.
482 * Called with j_list_lock held.
483 * Returns number of buffers reaped (for debug)
486 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
488 struct journal_head *last_jh;
489 struct journal_head *next_jh = jh;
496 last_jh = jh->b_cpprev;
499 next_jh = jh->b_cpnext;
500 ret = __try_to_free_cp_buf(jh);
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
516 } while (jh != last_jh);
522 * journal_clean_checkpoint_list
524 * Find all the written-back checkpoint buffers in the journal and release them.
526 * Called with the journal locked.
527 * Called with j_list_lock held.
528 * Returns number of buffers reaped (for debug)
531 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
533 transaction_t *transaction, *last_transaction, *next_transaction;
537 transaction = journal->j_checkpoint_transactions;
541 last_transaction = transaction->t_cpprev;
542 next_transaction = transaction;
544 transaction = next_transaction;
545 next_transaction = transaction->t_cpnext;
546 ret += journal_clean_one_cp_list(transaction->
547 t_checkpoint_list, &released);
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:
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
562 ret += journal_clean_one_cp_list(transaction->
563 t_checkpoint_io_list, &released);
566 } while (transaction != last_transaction);
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).
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
584 * The function returns 1 if it frees the transaction, 0 otherwise.
585 * The function can free jh and bh.
587 * This function is called with j_list_lock held.
589 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
591 struct transaction_chp_stats_s *stats;
592 transaction_t *transaction;
596 JBUFFER_TRACE(jh, "entry");
598 if ((transaction = jh->b_cp_transaction) == NULL) {
599 JBUFFER_TRACE(jh, "not on transaction");
602 journal = transaction->t_journal;
604 JBUFFER_TRACE(jh, "removing from transaction");
606 jh->b_cp_transaction = NULL;
607 jbd2_journal_put_journal_head(jh);
609 if (transaction->t_checkpoint_list != NULL ||
610 transaction->t_checkpoint_io_list != NULL)
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
619 * The locking here around t_state is a bit sleazy.
620 * See the comment at the end of jbd2_journal_commit_transaction().
622 if (transaction->t_state != T_FINISHED)
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,
631 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
632 transaction->t_tid, stats);
634 __jbd2_journal_drop_transaction(journal, transaction);
635 jbd2_journal_free_transaction(transaction);
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
646 * Called with the journal locked.
647 * Called with j_list_lock held.
649 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
650 transaction_t *transaction)
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);
656 /* Get reference for checkpointing transaction */
657 jbd2_journal_grab_journal_head(jh2bh(jh));
658 jh->b_cp_transaction = transaction;
660 if (!transaction->t_checkpoint_list) {
661 jh->b_cpnext = jh->b_cpprev = jh;
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;
668 transaction->t_checkpoint_list = jh;
672 * We've finished with this transaction structure: adios...
674 * The transaction must have no links except for the checkpoint by this
677 * Called with the journal locked.
678 * Called with j_list_lock held.
681 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
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;
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);
704 trace_jbd2_drop_transaction(journal, transaction);
706 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);