2 * linux/fs/jbd/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/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/blkdev.h>
26 #include <trace/events/jbd.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 = __journal_remove_checkpoint(jh) + 1;
107 jbd_unlock_bh_state(bh);
108 BUFFER_TRACE(bh, "release");
111 jbd_unlock_bh_state(bh);
117 * __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 __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 (__log_space_left(journal) < nblocks) {
129 if (journal->j_flags & JFS_ABORT)
131 spin_unlock(&journal->j_state_lock);
134 mutex_lock(&journal->j_checkpoint_mutex);
137 * Test again, another process may have checkpointed while we
138 * were waiting for the checkpoint lock. If there are no
139 * transactions ready to be checkpointed, try to recover
140 * journal space by calling cleanup_journal_tail(), and if
141 * that doesn't work, by waiting for the currently committing
142 * transaction to complete. If there is absolutely no way
143 * to make progress, this is either a BUG or corrupted
144 * filesystem, so abort the journal and leave a stack
145 * trace for forensic evidence.
147 spin_lock(&journal->j_state_lock);
148 spin_lock(&journal->j_list_lock);
149 nblocks = jbd_space_needed(journal);
150 space_left = __log_space_left(journal);
151 if (space_left < nblocks) {
152 int chkpt = journal->j_checkpoint_transactions != NULL;
155 if (journal->j_committing_transaction)
156 tid = journal->j_committing_transaction->t_tid;
157 spin_unlock(&journal->j_list_lock);
158 spin_unlock(&journal->j_state_lock);
160 log_do_checkpoint(journal);
161 } else if (cleanup_journal_tail(journal) == 0) {
162 /* We were able to recover space; yay! */
165 log_wait_commit(journal, tid);
167 printk(KERN_ERR "%s: needed %d blocks and "
168 "only had %d space available\n",
169 __func__, nblocks, space_left);
170 printk(KERN_ERR "%s: no way to get more "
171 "journal space\n", __func__);
173 journal_abort(journal, 0);
175 spin_lock(&journal->j_state_lock);
177 spin_unlock(&journal->j_list_lock);
179 mutex_unlock(&journal->j_checkpoint_mutex);
184 * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
185 * The caller must restart a list walk. Wait for someone else to run
186 * jbd_unlock_bh_state().
188 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
189 __releases(journal->j_list_lock)
192 spin_unlock(&journal->j_list_lock);
193 jbd_lock_bh_state(bh);
194 jbd_unlock_bh_state(bh);
199 * Clean up transaction's list of buffers submitted for io.
200 * We wait for any pending IO to complete and remove any clean
201 * buffers. Note that we take the buffers in the opposite ordering
202 * from the one in which they were submitted for IO.
204 * Return 0 on success, and return <0 if some buffers have failed
207 * Called with j_list_lock held.
209 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
211 struct journal_head *jh;
212 struct buffer_head *bh;
217 this_tid = transaction->t_tid;
219 /* Did somebody clean up the transaction in the meanwhile? */
220 if (journal->j_checkpoint_transactions != transaction ||
221 transaction->t_tid != this_tid)
223 while (!released && transaction->t_checkpoint_io_list) {
224 jh = transaction->t_checkpoint_io_list;
226 if (!jbd_trylock_bh_state(bh)) {
227 jbd_sync_bh(journal, bh);
228 spin_lock(&journal->j_list_lock);
232 if (buffer_locked(bh)) {
233 spin_unlock(&journal->j_list_lock);
234 jbd_unlock_bh_state(bh);
236 /* the journal_head may have gone by now */
237 BUFFER_TRACE(bh, "brelse");
239 spin_lock(&journal->j_list_lock);
242 if (unlikely(buffer_write_io_error(bh)))
246 * Now in whatever state the buffer currently is, we know that
247 * it has been written out and so we can drop it from the list
249 released = __journal_remove_checkpoint(jh);
250 jbd_unlock_bh_state(bh);
260 __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
263 struct blk_plug plug;
265 blk_start_plug(&plug);
266 for (i = 0; i < *batch_count; i++)
267 write_dirty_buffer(bhs[i], WRITE_SYNC);
268 blk_finish_plug(&plug);
270 for (i = 0; i < *batch_count; i++) {
271 struct buffer_head *bh = bhs[i];
272 clear_buffer_jwrite(bh);
273 BUFFER_TRACE(bh, "brelse");
280 * Try to flush one buffer from the checkpoint list to disk.
282 * Return 1 if something happened which requires us to abort the current
283 * scan of the checkpoint list. Return <0 if the buffer has failed to
286 * Called with j_list_lock held and drops it if 1 is returned
287 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
289 static int __process_buffer(journal_t *journal, struct journal_head *jh,
290 struct buffer_head **bhs, int *batch_count)
292 struct buffer_head *bh = jh2bh(jh);
295 if (buffer_locked(bh)) {
297 spin_unlock(&journal->j_list_lock);
298 jbd_unlock_bh_state(bh);
300 /* the journal_head may have gone by now */
301 BUFFER_TRACE(bh, "brelse");
304 } else if (jh->b_transaction != NULL) {
305 transaction_t *t = jh->b_transaction;
306 tid_t tid = t->t_tid;
308 spin_unlock(&journal->j_list_lock);
309 jbd_unlock_bh_state(bh);
310 log_start_commit(journal, tid);
311 log_wait_commit(journal, tid);
313 } else if (!buffer_dirty(bh)) {
315 if (unlikely(buffer_write_io_error(bh)))
318 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
319 BUFFER_TRACE(bh, "remove from checkpoint");
320 __journal_remove_checkpoint(jh);
321 spin_unlock(&journal->j_list_lock);
322 jbd_unlock_bh_state(bh);
326 * Important: we are about to write the buffer, and
327 * possibly block, while still holding the journal lock.
328 * We cannot afford to let the transaction logic start
329 * messing around with this buffer before we write it to
330 * disk, as that would break recoverability.
332 BUFFER_TRACE(bh, "queue");
334 J_ASSERT_BH(bh, !buffer_jwrite(bh));
335 set_buffer_jwrite(bh);
336 bhs[*batch_count] = bh;
337 __buffer_relink_io(jh);
338 jbd_unlock_bh_state(bh);
340 if (*batch_count == NR_BATCH) {
341 spin_unlock(&journal->j_list_lock);
342 __flush_batch(journal, bhs, batch_count);
350 * Perform an actual checkpoint. We take the first transaction on the
351 * list of transactions to be checkpointed and send all its buffers
352 * to disk. We submit larger chunks of data at once.
354 * The journal should be locked before calling this function.
355 * Called with j_checkpoint_mutex held.
357 int log_do_checkpoint(journal_t *journal)
359 transaction_t *transaction;
363 jbd_debug(1, "Start checkpoint\n");
366 * First thing: if there are any transactions in the log which
367 * don't need checkpointing, just eliminate them from the
368 * journal straight away.
370 result = cleanup_journal_tail(journal);
371 trace_jbd_checkpoint(journal, result);
372 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
377 * OK, we need to start writing disk blocks. Take one transaction
381 spin_lock(&journal->j_list_lock);
382 if (!journal->j_checkpoint_transactions)
384 transaction = journal->j_checkpoint_transactions;
385 this_tid = transaction->t_tid;
388 * If someone cleaned up this transaction while we slept, we're
389 * done (maybe it's a new transaction, but it fell at the same
392 if (journal->j_checkpoint_transactions == transaction &&
393 transaction->t_tid == this_tid) {
395 struct buffer_head *bhs[NR_BATCH];
396 struct journal_head *jh;
399 while (!retry && transaction->t_checkpoint_list) {
400 struct buffer_head *bh;
402 jh = transaction->t_checkpoint_list;
404 if (!jbd_trylock_bh_state(bh)) {
405 jbd_sync_bh(journal, bh);
409 retry = __process_buffer(journal, jh, bhs,&batch_count);
410 if (retry < 0 && !result)
412 if (!retry && (need_resched() ||
413 spin_needbreak(&journal->j_list_lock))) {
414 spin_unlock(&journal->j_list_lock);
422 spin_unlock(&journal->j_list_lock);
425 __flush_batch(journal, bhs, &batch_count);
429 spin_lock(&journal->j_list_lock);
433 * Now we have cleaned up the first transaction's checkpoint
434 * list. Let's clean up the second one
436 err = __wait_cp_io(journal, transaction);
441 spin_unlock(&journal->j_list_lock);
443 journal_abort(journal, result);
445 result = cleanup_journal_tail(journal);
447 return (result < 0) ? result : 0;
451 * Check the list of checkpoint transactions for the journal to see if
452 * we have already got rid of any since the last update of the log tail
453 * in the journal superblock. If so, we can instantly roll the
454 * superblock forward to remove those transactions from the log.
456 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
458 * This is the only part of the journaling code which really needs to be
459 * aware of transaction aborts. Checkpointing involves writing to the
460 * main filesystem area rather than to the journal, so it can proceed
461 * even in abort state, but we must not update the super block if
462 * checkpointing may have failed. Otherwise, we would lose some metadata
463 * buffers which should be written-back to the filesystem.
466 int cleanup_journal_tail(journal_t *journal)
468 transaction_t * transaction;
470 unsigned int blocknr, freed;
472 if (is_journal_aborted(journal))
476 * OK, work out the oldest transaction remaining in the log, and
477 * the log block it starts at.
479 * If the log is now empty, we need to work out which is the
480 * next transaction ID we will write, and where it will
483 spin_lock(&journal->j_state_lock);
484 spin_lock(&journal->j_list_lock);
485 transaction = journal->j_checkpoint_transactions;
487 first_tid = transaction->t_tid;
488 blocknr = transaction->t_log_start;
489 } else if ((transaction = journal->j_committing_transaction) != NULL) {
490 first_tid = transaction->t_tid;
491 blocknr = transaction->t_log_start;
492 } else if ((transaction = journal->j_running_transaction) != NULL) {
493 first_tid = transaction->t_tid;
494 blocknr = journal->j_head;
496 first_tid = journal->j_transaction_sequence;
497 blocknr = journal->j_head;
499 spin_unlock(&journal->j_list_lock);
500 J_ASSERT(blocknr != 0);
502 /* If the oldest pinned transaction is at the tail of the log
503 already then there's not much we can do right now. */
504 if (journal->j_tail_sequence == first_tid) {
505 spin_unlock(&journal->j_state_lock);
508 spin_unlock(&journal->j_state_lock);
511 * We need to make sure that any blocks that were recently written out
512 * --- perhaps by log_do_checkpoint() --- are flushed out before we
513 * drop the transactions from the journal. Similarly we need to be sure
514 * superblock makes it to disk before next transaction starts reusing
515 * freed space (otherwise we could replay some blocks of the new
516 * transaction thinking they belong to the old one). So we use
517 * WRITE_FLUSH_FUA. It's unlikely this will be necessary, especially
518 * with an appropriately sized journal, but we need this to guarantee
519 * correctness. Fortunately cleanup_journal_tail() doesn't get called
522 journal_update_sb_log_tail(journal, first_tid, blocknr,
525 spin_lock(&journal->j_state_lock);
526 /* OK, update the superblock to recover the freed space.
527 * Physical blocks come first: have we wrapped beyond the end of
529 freed = blocknr - journal->j_tail;
530 if (blocknr < journal->j_tail)
531 freed = freed + journal->j_last - journal->j_first;
533 trace_jbd_cleanup_journal_tail(journal, first_tid, blocknr, freed);
535 "Cleaning journal tail from %d to %d (offset %u), "
537 journal->j_tail_sequence, first_tid, blocknr, freed);
539 journal->j_free += freed;
540 journal->j_tail_sequence = first_tid;
541 journal->j_tail = blocknr;
542 spin_unlock(&journal->j_state_lock);
547 /* Checkpoint list management */
550 * journal_clean_one_cp_list
552 * Find all the written-back checkpoint buffers in the given list and release
555 * Called with j_list_lock held.
556 * Returns number of buffers reaped (for debug)
559 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
561 struct journal_head *last_jh;
562 struct journal_head *next_jh = jh;
569 last_jh = jh->b_cpprev;
572 next_jh = jh->b_cpnext;
573 /* Use trylock because of the ranking */
574 if (jbd_trylock_bh_state(jh2bh(jh))) {
575 ret = __try_to_free_cp_buf(jh);
585 * This function only frees up some memory
586 * if possible so we dont have an obligation
587 * to finish processing. Bail out if preemption
592 } while (jh != last_jh);
598 * journal_clean_checkpoint_list
600 * Find all the written-back checkpoint buffers in the journal and release them.
602 * Called with the journal locked.
603 * Called with j_list_lock held.
604 * Returns number of buffers reaped (for debug)
607 int __journal_clean_checkpoint_list(journal_t *journal)
609 transaction_t *transaction, *last_transaction, *next_transaction;
613 transaction = journal->j_checkpoint_transactions;
617 last_transaction = transaction->t_cpprev;
618 next_transaction = transaction;
620 transaction = next_transaction;
621 next_transaction = transaction->t_cpnext;
622 ret += journal_clean_one_cp_list(transaction->
623 t_checkpoint_list, &released);
625 * This function only frees up some memory if possible so we
626 * dont have an obligation to finish processing. Bail out if
627 * preemption requested:
634 * It is essential that we are as careful as in the case of
635 * t_checkpoint_list with removing the buffer from the list as
636 * we can possibly see not yet submitted buffers on io_list
638 ret += journal_clean_one_cp_list(transaction->
639 t_checkpoint_io_list, &released);
642 } while (transaction != last_transaction);
648 * journal_remove_checkpoint: called after a buffer has been committed
649 * to disk (either by being write-back flushed to disk, or being
650 * committed to the log).
652 * We cannot safely clean a transaction out of the log until all of the
653 * buffer updates committed in that transaction have safely been stored
654 * elsewhere on disk. To achieve this, all of the buffers in a
655 * transaction need to be maintained on the transaction's checkpoint
656 * lists until they have been rewritten, at which point this function is
657 * called to remove the buffer from the existing transaction's
660 * The function returns 1 if it frees the transaction, 0 otherwise.
661 * The function can free jh and bh.
663 * This function is called with j_list_lock held.
664 * This function is called with jbd_lock_bh_state(jh2bh(jh))
667 int __journal_remove_checkpoint(struct journal_head *jh)
669 transaction_t *transaction;
673 JBUFFER_TRACE(jh, "entry");
675 if ((transaction = jh->b_cp_transaction) == NULL) {
676 JBUFFER_TRACE(jh, "not on transaction");
679 journal = transaction->t_journal;
681 JBUFFER_TRACE(jh, "removing from transaction");
683 jh->b_cp_transaction = NULL;
684 journal_put_journal_head(jh);
686 if (transaction->t_checkpoint_list != NULL ||
687 transaction->t_checkpoint_io_list != NULL)
691 * There is one special case to worry about: if we have just pulled the
692 * buffer off a running or committing transaction's checkpoing list,
693 * then even if the checkpoint list is empty, the transaction obviously
696 * The locking here around t_state is a bit sleazy.
697 * See the comment at the end of journal_commit_transaction().
699 if (transaction->t_state != T_FINISHED)
702 /* OK, that was the last buffer for the transaction: we can now
703 safely remove this transaction from the log */
705 __journal_drop_transaction(journal, transaction);
707 /* Just in case anybody was waiting for more transactions to be
709 wake_up(&journal->j_wait_logspace);
716 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
717 * list so that we know when it is safe to clean the transaction out of
720 * Called with the journal locked.
721 * Called with j_list_lock held.
723 void __journal_insert_checkpoint(struct journal_head *jh,
724 transaction_t *transaction)
726 JBUFFER_TRACE(jh, "entry");
727 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
728 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
730 /* Get reference for checkpointing transaction */
731 journal_grab_journal_head(jh2bh(jh));
732 jh->b_cp_transaction = transaction;
734 if (!transaction->t_checkpoint_list) {
735 jh->b_cpnext = jh->b_cpprev = jh;
737 jh->b_cpnext = transaction->t_checkpoint_list;
738 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
739 jh->b_cpprev->b_cpnext = jh;
740 jh->b_cpnext->b_cpprev = jh;
742 transaction->t_checkpoint_list = jh;
746 * We've finished with this transaction structure: adios...
748 * The transaction must have no links except for the checkpoint by this
751 * Called with the journal locked.
752 * Called with j_list_lock held.
755 void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
757 assert_spin_locked(&journal->j_list_lock);
758 if (transaction->t_cpnext) {
759 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
760 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
761 if (journal->j_checkpoint_transactions == transaction)
762 journal->j_checkpoint_transactions =
763 transaction->t_cpnext;
764 if (journal->j_checkpoint_transactions == transaction)
765 journal->j_checkpoint_transactions = NULL;
768 J_ASSERT(transaction->t_state == T_FINISHED);
769 J_ASSERT(transaction->t_buffers == NULL);
770 J_ASSERT(transaction->t_sync_datalist == NULL);
771 J_ASSERT(transaction->t_forget == NULL);
772 J_ASSERT(transaction->t_iobuf_list == NULL);
773 J_ASSERT(transaction->t_shadow_list == NULL);
774 J_ASSERT(transaction->t_log_list == NULL);
775 J_ASSERT(transaction->t_checkpoint_list == NULL);
776 J_ASSERT(transaction->t_checkpoint_io_list == NULL);
777 J_ASSERT(transaction->t_updates == 0);
778 J_ASSERT(journal->j_committing_transaction != transaction);
779 J_ASSERT(journal->j_running_transaction != transaction);
781 trace_jbd_drop_transaction(journal, transaction);
782 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);