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
91 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
93 static int __try_to_free_cp_buf(struct journal_head *jh)
96 struct buffer_head *bh = jh2bh(jh);
98 if (jh->b_jlist == BJ_None && !buffer_locked(bh) &&
99 !buffer_dirty(bh) && !buffer_write_io_error(bh)) {
101 * Get our reference so that bh cannot be freed before
105 JBUFFER_TRACE(jh, "remove from checkpoint list");
106 ret = __jbd2_journal_remove_checkpoint(jh) + 1;
107 jbd_unlock_bh_state(bh);
108 BUFFER_TRACE(bh, "release");
111 jbd_unlock_bh_state(bh);
117 * __jbd2_log_wait_for_space: wait until there is space in the journal.
119 * Called under j-state_lock *only*. It will be unlocked if we have to wait
120 * for a checkpoint to free up some space in the log.
122 void __jbd2_log_wait_for_space(journal_t *journal)
124 int nblocks, space_left;
125 /* assert_spin_locked(&journal->j_state_lock); */
127 nblocks = jbd_space_needed(journal);
128 while (__jbd2_log_space_left(journal) < nblocks) {
129 if (journal->j_flags & JBD2_ABORT)
131 write_unlock(&journal->j_state_lock);
132 mutex_lock(&journal->j_checkpoint_mutex);
135 * Test again, another process may have checkpointed while we
136 * were waiting for the checkpoint lock. If there are no
137 * transactions ready to be checkpointed, try to recover
138 * journal space by calling cleanup_journal_tail(), and if
139 * that doesn't work, by waiting for the currently committing
140 * transaction to complete. If there is absolutely no way
141 * to make progress, this is either a BUG or corrupted
142 * filesystem, so abort the journal and leave a stack
143 * trace for forensic evidence.
145 write_lock(&journal->j_state_lock);
146 spin_lock(&journal->j_list_lock);
147 nblocks = jbd_space_needed(journal);
148 space_left = __jbd2_log_space_left(journal);
149 if (space_left < nblocks) {
150 int chkpt = journal->j_checkpoint_transactions != NULL;
153 if (journal->j_committing_transaction)
154 tid = journal->j_committing_transaction->t_tid;
155 spin_unlock(&journal->j_list_lock);
156 write_unlock(&journal->j_state_lock);
158 jbd2_log_do_checkpoint(journal);
159 } else if (jbd2_cleanup_journal_tail(journal) == 0) {
160 /* We were able to recover space; yay! */
163 jbd2_log_wait_commit(journal, tid);
165 printk(KERN_ERR "%s: needed %d blocks and "
166 "only had %d space available\n",
167 __func__, nblocks, space_left);
168 printk(KERN_ERR "%s: no way to get more "
169 "journal space in %s\n", __func__,
172 jbd2_journal_abort(journal, 0);
174 write_lock(&journal->j_state_lock);
176 spin_unlock(&journal->j_list_lock);
178 mutex_unlock(&journal->j_checkpoint_mutex);
183 * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
184 * The caller must restart a list walk. Wait for someone else to run
185 * jbd_unlock_bh_state().
187 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
188 __releases(journal->j_list_lock)
191 spin_unlock(&journal->j_list_lock);
192 jbd_lock_bh_state(bh);
193 jbd_unlock_bh_state(bh);
198 * Clean up transaction's list of buffers submitted for io.
199 * We wait for any pending IO to complete and remove any clean
200 * buffers. Note that we take the buffers in the opposite ordering
201 * from the one in which they were submitted for IO.
203 * Return 0 on success, and return <0 if some buffers have failed
206 * Called with j_list_lock held.
208 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
210 struct journal_head *jh;
211 struct buffer_head *bh;
216 this_tid = transaction->t_tid;
218 /* Did somebody clean up the transaction in the meanwhile? */
219 if (journal->j_checkpoint_transactions != transaction ||
220 transaction->t_tid != this_tid)
222 while (!released && transaction->t_checkpoint_io_list) {
223 jh = transaction->t_checkpoint_io_list;
225 if (!jbd_trylock_bh_state(bh)) {
226 jbd_sync_bh(journal, bh);
227 spin_lock(&journal->j_list_lock);
231 if (buffer_locked(bh)) {
232 spin_unlock(&journal->j_list_lock);
233 jbd_unlock_bh_state(bh);
235 /* the journal_head may have gone by now */
236 BUFFER_TRACE(bh, "brelse");
238 spin_lock(&journal->j_list_lock);
241 if (unlikely(buffer_write_io_error(bh)))
245 * Now in whatever state the buffer currently is, we know that
246 * it has been written out and so we can drop it from the list
248 released = __jbd2_journal_remove_checkpoint(jh);
249 jbd_unlock_bh_state(bh);
257 __flush_batch(journal_t *journal, int *batch_count)
260 struct blk_plug plug;
262 blk_start_plug(&plug);
263 for (i = 0; i < *batch_count; i++)
264 write_dirty_buffer(journal->j_chkpt_bhs[i], WRITE_SYNC);
265 blk_finish_plug(&plug);
267 for (i = 0; i < *batch_count; i++) {
268 struct buffer_head *bh = journal->j_chkpt_bhs[i];
269 clear_buffer_jwrite(bh);
270 BUFFER_TRACE(bh, "brelse");
277 * Try to flush one buffer from the checkpoint list to disk.
279 * Return 1 if something happened which requires us to abort the current
280 * scan of the checkpoint list. Return <0 if the buffer has failed to
283 * Called with j_list_lock held and drops it if 1 is returned
284 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
286 static int __process_buffer(journal_t *journal, struct journal_head *jh,
287 int *batch_count, transaction_t *transaction)
289 struct buffer_head *bh = jh2bh(jh);
292 if (buffer_locked(bh)) {
294 spin_unlock(&journal->j_list_lock);
295 jbd_unlock_bh_state(bh);
297 /* the journal_head may have gone by now */
298 BUFFER_TRACE(bh, "brelse");
301 } else if (jh->b_transaction != NULL) {
302 transaction_t *t = jh->b_transaction;
303 tid_t tid = t->t_tid;
305 transaction->t_chp_stats.cs_forced_to_close++;
306 spin_unlock(&journal->j_list_lock);
307 jbd_unlock_bh_state(bh);
308 if (unlikely(journal->j_flags & JBD2_UNMOUNT))
310 * The journal thread is dead; so starting and
311 * waiting for a commit to finish will cause
312 * us to wait for a _very_ long time.
314 printk(KERN_ERR "JBD2: %s: "
315 "Waiting for Godot: block %llu\n",
317 (unsigned long long) bh->b_blocknr);
318 jbd2_log_start_commit(journal, tid);
319 jbd2_log_wait_commit(journal, tid);
321 } else if (!buffer_dirty(bh)) {
323 if (unlikely(buffer_write_io_error(bh)))
326 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
327 BUFFER_TRACE(bh, "remove from checkpoint");
328 __jbd2_journal_remove_checkpoint(jh);
329 spin_unlock(&journal->j_list_lock);
330 jbd_unlock_bh_state(bh);
334 * Important: we are about to write the buffer, and
335 * possibly block, while still holding the journal lock.
336 * We cannot afford to let the transaction logic start
337 * messing around with this buffer before we write it to
338 * disk, as that would break recoverability.
340 BUFFER_TRACE(bh, "queue");
342 J_ASSERT_BH(bh, !buffer_jwrite(bh));
343 set_buffer_jwrite(bh);
344 journal->j_chkpt_bhs[*batch_count] = bh;
345 __buffer_relink_io(jh);
346 jbd_unlock_bh_state(bh);
347 transaction->t_chp_stats.cs_written++;
349 if (*batch_count == JBD2_NR_BATCH) {
350 spin_unlock(&journal->j_list_lock);
351 __flush_batch(journal, batch_count);
359 * Perform an actual checkpoint. We take the first transaction on the
360 * list of transactions to be checkpointed and send all its buffers
361 * to disk. We submit larger chunks of data at once.
363 * The journal should be locked before calling this function.
364 * Called with j_checkpoint_mutex held.
366 int jbd2_log_do_checkpoint(journal_t *journal)
368 transaction_t *transaction;
372 jbd_debug(1, "Start checkpoint\n");
375 * First thing: if there are any transactions in the log which
376 * don't need checkpointing, just eliminate them from the
377 * journal straight away.
379 result = jbd2_cleanup_journal_tail(journal);
380 trace_jbd2_checkpoint(journal, result);
381 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
386 * OK, we need to start writing disk blocks. Take one transaction
390 spin_lock(&journal->j_list_lock);
391 if (!journal->j_checkpoint_transactions)
393 transaction = journal->j_checkpoint_transactions;
394 if (transaction->t_chp_stats.cs_chp_time == 0)
395 transaction->t_chp_stats.cs_chp_time = jiffies;
396 this_tid = transaction->t_tid;
399 * If someone cleaned up this transaction while we slept, we're
400 * done (maybe it's a new transaction, but it fell at the same
403 if (journal->j_checkpoint_transactions == transaction &&
404 transaction->t_tid == this_tid) {
406 struct journal_head *jh;
409 while (!retry && transaction->t_checkpoint_list) {
410 struct buffer_head *bh;
412 jh = transaction->t_checkpoint_list;
414 if (!jbd_trylock_bh_state(bh)) {
415 jbd_sync_bh(journal, bh);
419 retry = __process_buffer(journal, jh, &batch_count,
421 if (retry < 0 && !result)
423 if (!retry && (need_resched() ||
424 spin_needbreak(&journal->j_list_lock))) {
425 spin_unlock(&journal->j_list_lock);
433 spin_unlock(&journal->j_list_lock);
436 __flush_batch(journal, &batch_count);
440 spin_lock(&journal->j_list_lock);
444 * Now we have cleaned up the first transaction's checkpoint
445 * list. Let's clean up the second one
447 err = __wait_cp_io(journal, transaction);
452 spin_unlock(&journal->j_list_lock);
454 jbd2_journal_abort(journal, result);
456 result = jbd2_cleanup_journal_tail(journal);
458 return (result < 0) ? result : 0;
462 * Check the list of checkpoint transactions for the journal to see if
463 * we have already got rid of any since the last update of the log tail
464 * in the journal superblock. If so, we can instantly roll the
465 * superblock forward to remove those transactions from the log.
467 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
469 * Called with the journal lock held.
471 * This is the only part of the journaling code which really needs to be
472 * aware of transaction aborts. Checkpointing involves writing to the
473 * main filesystem area rather than to the journal, so it can proceed
474 * even in abort state, but we must not update the super block if
475 * checkpointing may have failed. Otherwise, we would lose some metadata
476 * buffers which should be written-back to the filesystem.
479 int jbd2_cleanup_journal_tail(journal_t *journal)
482 unsigned long blocknr;
484 if (is_journal_aborted(journal))
487 if (!jbd2_journal_get_log_tail(journal, &first_tid, &blocknr))
489 J_ASSERT(blocknr != 0);
492 * We need to make sure that any blocks that were recently written out
493 * --- perhaps by jbd2_log_do_checkpoint() --- are flushed out before
494 * we drop the transactions from the journal. It's unlikely this will
495 * be necessary, especially with an appropriately sized journal, but we
496 * need this to guarantee correctness. Fortunately
497 * jbd2_cleanup_journal_tail() doesn't get called all that often.
499 if (journal->j_flags & JBD2_BARRIER)
500 blkdev_issue_flush(journal->j_fs_dev, GFP_NOFS, NULL);
502 __jbd2_update_log_tail(journal, first_tid, blocknr);
507 /* Checkpoint list management */
510 * journal_clean_one_cp_list
512 * Find all the written-back checkpoint buffers in the given list and
515 * Called with the journal locked.
516 * Called with j_list_lock held.
517 * Returns number of bufers reaped (for debug)
520 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
522 struct journal_head *last_jh;
523 struct journal_head *next_jh = jh;
530 last_jh = jh->b_cpprev;
533 next_jh = jh->b_cpnext;
534 /* Use trylock because of the ranking */
535 if (jbd_trylock_bh_state(jh2bh(jh))) {
536 ret = __try_to_free_cp_buf(jh);
546 * This function only frees up some memory
547 * if possible so we dont have an obligation
548 * to finish processing. Bail out if preemption
553 } while (jh != last_jh);
559 * journal_clean_checkpoint_list
561 * Find all the written-back checkpoint buffers in the journal and release them.
563 * Called with the journal locked.
564 * Called with j_list_lock held.
565 * Returns number of buffers reaped (for debug)
568 int __jbd2_journal_clean_checkpoint_list(journal_t *journal)
570 transaction_t *transaction, *last_transaction, *next_transaction;
574 transaction = journal->j_checkpoint_transactions;
578 last_transaction = transaction->t_cpprev;
579 next_transaction = transaction;
581 transaction = next_transaction;
582 next_transaction = transaction->t_cpnext;
583 ret += journal_clean_one_cp_list(transaction->
584 t_checkpoint_list, &released);
586 * This function only frees up some memory if possible so we
587 * dont have an obligation to finish processing. Bail out if
588 * preemption requested:
595 * It is essential that we are as careful as in the case of
596 * t_checkpoint_list with removing the buffer from the list as
597 * we can possibly see not yet submitted buffers on io_list
599 ret += journal_clean_one_cp_list(transaction->
600 t_checkpoint_io_list, &released);
603 } while (transaction != last_transaction);
609 * journal_remove_checkpoint: called after a buffer has been committed
610 * to disk (either by being write-back flushed to disk, or being
611 * committed to the log).
613 * We cannot safely clean a transaction out of the log until all of the
614 * buffer updates committed in that transaction have safely been stored
615 * elsewhere on disk. To achieve this, all of the buffers in a
616 * transaction need to be maintained on the transaction's checkpoint
617 * lists until they have been rewritten, at which point this function is
618 * called to remove the buffer from the existing transaction's
621 * The function returns 1 if it frees the transaction, 0 otherwise.
622 * The function can free jh and bh.
624 * This function is called with j_list_lock held.
625 * This function is called with jbd_lock_bh_state(jh2bh(jh))
628 int __jbd2_journal_remove_checkpoint(struct journal_head *jh)
630 struct transaction_chp_stats_s *stats;
631 transaction_t *transaction;
635 JBUFFER_TRACE(jh, "entry");
637 if ((transaction = jh->b_cp_transaction) == NULL) {
638 JBUFFER_TRACE(jh, "not on transaction");
641 journal = transaction->t_journal;
643 JBUFFER_TRACE(jh, "removing from transaction");
645 jh->b_cp_transaction = NULL;
646 jbd2_journal_put_journal_head(jh);
648 if (transaction->t_checkpoint_list != NULL ||
649 transaction->t_checkpoint_io_list != NULL)
653 * There is one special case to worry about: if we have just pulled the
654 * buffer off a running or committing transaction's checkpoing list,
655 * then even if the checkpoint list is empty, the transaction obviously
658 * The locking here around t_state is a bit sleazy.
659 * See the comment at the end of jbd2_journal_commit_transaction().
661 if (transaction->t_state != T_FINISHED)
664 /* OK, that was the last buffer for the transaction: we can now
665 safely remove this transaction from the log */
666 stats = &transaction->t_chp_stats;
667 if (stats->cs_chp_time)
668 stats->cs_chp_time = jbd2_time_diff(stats->cs_chp_time,
670 trace_jbd2_checkpoint_stats(journal->j_fs_dev->bd_dev,
671 transaction->t_tid, stats);
673 __jbd2_journal_drop_transaction(journal, transaction);
676 /* Just in case anybody was waiting for more transactions to be
678 wake_up(&journal->j_wait_logspace);
685 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
686 * list so that we know when it is safe to clean the transaction out of
689 * Called with the journal locked.
690 * Called with j_list_lock held.
692 void __jbd2_journal_insert_checkpoint(struct journal_head *jh,
693 transaction_t *transaction)
695 JBUFFER_TRACE(jh, "entry");
696 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
697 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
699 /* Get reference for checkpointing transaction */
700 jbd2_journal_grab_journal_head(jh2bh(jh));
701 jh->b_cp_transaction = transaction;
703 if (!transaction->t_checkpoint_list) {
704 jh->b_cpnext = jh->b_cpprev = jh;
706 jh->b_cpnext = transaction->t_checkpoint_list;
707 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
708 jh->b_cpprev->b_cpnext = jh;
709 jh->b_cpnext->b_cpprev = jh;
711 transaction->t_checkpoint_list = jh;
715 * We've finished with this transaction structure: adios...
717 * The transaction must have no links except for the checkpoint by this
720 * Called with the journal locked.
721 * Called with j_list_lock held.
724 void __jbd2_journal_drop_transaction(journal_t *journal, transaction_t *transaction)
726 assert_spin_locked(&journal->j_list_lock);
727 if (transaction->t_cpnext) {
728 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
729 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
730 if (journal->j_checkpoint_transactions == transaction)
731 journal->j_checkpoint_transactions =
732 transaction->t_cpnext;
733 if (journal->j_checkpoint_transactions == transaction)
734 journal->j_checkpoint_transactions = NULL;
737 J_ASSERT(transaction->t_state == T_FINISHED);
738 J_ASSERT(transaction->t_buffers == NULL);
739 J_ASSERT(transaction->t_forget == NULL);
740 J_ASSERT(transaction->t_iobuf_list == NULL);
741 J_ASSERT(transaction->t_shadow_list == NULL);
742 J_ASSERT(transaction->t_log_list == NULL);
743 J_ASSERT(transaction->t_checkpoint_list == NULL);
744 J_ASSERT(transaction->t_checkpoint_io_list == NULL);
745 J_ASSERT(atomic_read(&transaction->t_updates) == 0);
746 J_ASSERT(journal->j_committing_transaction != transaction);
747 J_ASSERT(journal->j_running_transaction != transaction);
749 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);