Merge branch 'linux-next' of git://git.kernel.org/pub/scm/linux/kernel/git/konrad...
[zen-stable.git] / fs / jbd / checkpoint.c
blobf94fc48ff3a0c2676156c4a3545f011f9e340a7e
1 /*
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
17 * reused.
20 #include <linux/time.h>
21 #include <linux/fs.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;
76 } else {
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
88 * whole transaction.
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)
95 int ret = 0;
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
102 * we unlock it
104 get_bh(bh);
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");
109 __brelse(bh);
110 } else {
111 jbd_unlock_bh_state(bh);
113 return ret;
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)
130 return;
131 spin_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 spin_lock(&journal->j_state_lock);
146 spin_lock(&journal->j_list_lock);
147 nblocks = jbd_space_needed(journal);
148 space_left = __log_space_left(journal);
149 if (space_left < nblocks) {
150 int chkpt = journal->j_checkpoint_transactions != NULL;
151 tid_t tid = 0;
153 if (journal->j_committing_transaction)
154 tid = journal->j_committing_transaction->t_tid;
155 spin_unlock(&journal->j_list_lock);
156 spin_unlock(&journal->j_state_lock);
157 if (chkpt) {
158 log_do_checkpoint(journal);
159 } else if (cleanup_journal_tail(journal) == 0) {
160 /* We were able to recover space; yay! */
162 } else if (tid) {
163 log_wait_commit(journal, tid);
164 } else {
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\n", __func__);
170 WARN_ON(1);
171 journal_abort(journal, 0);
173 spin_lock(&journal->j_state_lock);
174 } else {
175 spin_unlock(&journal->j_list_lock);
177 mutex_unlock(&journal->j_checkpoint_mutex);
182 * We were unable to perform jbd_trylock_bh_state() inside j_list_lock.
183 * The caller must restart a list walk. Wait for someone else to run
184 * jbd_unlock_bh_state().
186 static void jbd_sync_bh(journal_t *journal, struct buffer_head *bh)
187 __releases(journal->j_list_lock)
189 get_bh(bh);
190 spin_unlock(&journal->j_list_lock);
191 jbd_lock_bh_state(bh);
192 jbd_unlock_bh_state(bh);
193 put_bh(bh);
197 * Clean up transaction's list of buffers submitted for io.
198 * We wait for any pending IO to complete and remove any clean
199 * buffers. Note that we take the buffers in the opposite ordering
200 * from the one in which they were submitted for IO.
202 * Return 0 on success, and return <0 if some buffers have failed
203 * to be written out.
205 * Called with j_list_lock held.
207 static int __wait_cp_io(journal_t *journal, transaction_t *transaction)
209 struct journal_head *jh;
210 struct buffer_head *bh;
211 tid_t this_tid;
212 int released = 0;
213 int ret = 0;
215 this_tid = transaction->t_tid;
216 restart:
217 /* Did somebody clean up the transaction in the meanwhile? */
218 if (journal->j_checkpoint_transactions != transaction ||
219 transaction->t_tid != this_tid)
220 return ret;
221 while (!released && transaction->t_checkpoint_io_list) {
222 jh = transaction->t_checkpoint_io_list;
223 bh = jh2bh(jh);
224 if (!jbd_trylock_bh_state(bh)) {
225 jbd_sync_bh(journal, bh);
226 spin_lock(&journal->j_list_lock);
227 goto restart;
229 get_bh(bh);
230 if (buffer_locked(bh)) {
231 spin_unlock(&journal->j_list_lock);
232 jbd_unlock_bh_state(bh);
233 wait_on_buffer(bh);
234 /* the journal_head may have gone by now */
235 BUFFER_TRACE(bh, "brelse");
236 __brelse(bh);
237 spin_lock(&journal->j_list_lock);
238 goto restart;
240 if (unlikely(buffer_write_io_error(bh)))
241 ret = -EIO;
244 * Now in whatever state the buffer currently is, we know that
245 * it has been written out and so we can drop it from the list
247 released = __journal_remove_checkpoint(jh);
248 jbd_unlock_bh_state(bh);
249 __brelse(bh);
252 return ret;
255 #define NR_BATCH 64
257 static void
258 __flush_batch(journal_t *journal, struct buffer_head **bhs, int *batch_count)
260 int i;
261 struct blk_plug plug;
263 blk_start_plug(&plug);
264 for (i = 0; i < *batch_count; i++)
265 write_dirty_buffer(bhs[i], WRITE_SYNC);
266 blk_finish_plug(&plug);
268 for (i = 0; i < *batch_count; i++) {
269 struct buffer_head *bh = bhs[i];
270 clear_buffer_jwrite(bh);
271 BUFFER_TRACE(bh, "brelse");
272 __brelse(bh);
274 *batch_count = 0;
278 * Try to flush one buffer from the checkpoint list to disk.
280 * Return 1 if something happened which requires us to abort the current
281 * scan of the checkpoint list. Return <0 if the buffer has failed to
282 * be written out.
284 * Called with j_list_lock held and drops it if 1 is returned
285 * Called under jbd_lock_bh_state(jh2bh(jh)), and drops it
287 static int __process_buffer(journal_t *journal, struct journal_head *jh,
288 struct buffer_head **bhs, int *batch_count)
290 struct buffer_head *bh = jh2bh(jh);
291 int ret = 0;
293 if (buffer_locked(bh)) {
294 get_bh(bh);
295 spin_unlock(&journal->j_list_lock);
296 jbd_unlock_bh_state(bh);
297 wait_on_buffer(bh);
298 /* the journal_head may have gone by now */
299 BUFFER_TRACE(bh, "brelse");
300 __brelse(bh);
301 ret = 1;
302 } else if (jh->b_transaction != NULL) {
303 transaction_t *t = jh->b_transaction;
304 tid_t tid = t->t_tid;
306 spin_unlock(&journal->j_list_lock);
307 jbd_unlock_bh_state(bh);
308 log_start_commit(journal, tid);
309 log_wait_commit(journal, tid);
310 ret = 1;
311 } else if (!buffer_dirty(bh)) {
312 ret = 1;
313 if (unlikely(buffer_write_io_error(bh)))
314 ret = -EIO;
315 get_bh(bh);
316 J_ASSERT_JH(jh, !buffer_jbddirty(bh));
317 BUFFER_TRACE(bh, "remove from checkpoint");
318 __journal_remove_checkpoint(jh);
319 spin_unlock(&journal->j_list_lock);
320 jbd_unlock_bh_state(bh);
321 __brelse(bh);
322 } else {
324 * Important: we are about to write the buffer, and
325 * possibly block, while still holding the journal lock.
326 * We cannot afford to let the transaction logic start
327 * messing around with this buffer before we write it to
328 * disk, as that would break recoverability.
330 BUFFER_TRACE(bh, "queue");
331 get_bh(bh);
332 J_ASSERT_BH(bh, !buffer_jwrite(bh));
333 set_buffer_jwrite(bh);
334 bhs[*batch_count] = bh;
335 __buffer_relink_io(jh);
336 jbd_unlock_bh_state(bh);
337 (*batch_count)++;
338 if (*batch_count == NR_BATCH) {
339 spin_unlock(&journal->j_list_lock);
340 __flush_batch(journal, bhs, batch_count);
341 ret = 1;
344 return ret;
348 * Perform an actual checkpoint. We take the first transaction on the
349 * list of transactions to be checkpointed and send all its buffers
350 * to disk. We submit larger chunks of data at once.
352 * The journal should be locked before calling this function.
353 * Called with j_checkpoint_mutex held.
355 int log_do_checkpoint(journal_t *journal)
357 transaction_t *transaction;
358 tid_t this_tid;
359 int result;
361 jbd_debug(1, "Start checkpoint\n");
364 * First thing: if there are any transactions in the log which
365 * don't need checkpointing, just eliminate them from the
366 * journal straight away.
368 result = cleanup_journal_tail(journal);
369 trace_jbd_checkpoint(journal, result);
370 jbd_debug(1, "cleanup_journal_tail returned %d\n", result);
371 if (result <= 0)
372 return result;
375 * OK, we need to start writing disk blocks. Take one transaction
376 * and write it.
378 result = 0;
379 spin_lock(&journal->j_list_lock);
380 if (!journal->j_checkpoint_transactions)
381 goto out;
382 transaction = journal->j_checkpoint_transactions;
383 this_tid = transaction->t_tid;
384 restart:
386 * If someone cleaned up this transaction while we slept, we're
387 * done (maybe it's a new transaction, but it fell at the same
388 * address).
390 if (journal->j_checkpoint_transactions == transaction &&
391 transaction->t_tid == this_tid) {
392 int batch_count = 0;
393 struct buffer_head *bhs[NR_BATCH];
394 struct journal_head *jh;
395 int retry = 0, err;
397 while (!retry && transaction->t_checkpoint_list) {
398 struct buffer_head *bh;
400 jh = transaction->t_checkpoint_list;
401 bh = jh2bh(jh);
402 if (!jbd_trylock_bh_state(bh)) {
403 jbd_sync_bh(journal, bh);
404 retry = 1;
405 break;
407 retry = __process_buffer(journal, jh, bhs,&batch_count);
408 if (retry < 0 && !result)
409 result = retry;
410 if (!retry && (need_resched() ||
411 spin_needbreak(&journal->j_list_lock))) {
412 spin_unlock(&journal->j_list_lock);
413 retry = 1;
414 break;
418 if (batch_count) {
419 if (!retry) {
420 spin_unlock(&journal->j_list_lock);
421 retry = 1;
423 __flush_batch(journal, bhs, &batch_count);
426 if (retry) {
427 spin_lock(&journal->j_list_lock);
428 goto restart;
431 * Now we have cleaned up the first transaction's checkpoint
432 * list. Let's clean up the second one
434 err = __wait_cp_io(journal, transaction);
435 if (!result)
436 result = err;
438 out:
439 spin_unlock(&journal->j_list_lock);
440 if (result < 0)
441 journal_abort(journal, result);
442 else
443 result = cleanup_journal_tail(journal);
445 return (result < 0) ? result : 0;
449 * Check the list of checkpoint transactions for the journal to see if
450 * we have already got rid of any since the last update of the log tail
451 * in the journal superblock. If so, we can instantly roll the
452 * superblock forward to remove those transactions from the log.
454 * Return <0 on error, 0 on success, 1 if there was nothing to clean up.
456 * Called with the journal lock held.
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;
469 tid_t first_tid;
470 unsigned int blocknr, freed;
472 if (is_journal_aborted(journal))
473 return 1;
475 /* OK, work out the oldest transaction remaining in the log, and
476 * the log block it starts at.
478 * If the log is now empty, we need to work out which is the
479 * next transaction ID we will write, and where it will
480 * start. */
482 spin_lock(&journal->j_state_lock);
483 spin_lock(&journal->j_list_lock);
484 transaction = journal->j_checkpoint_transactions;
485 if (transaction) {
486 first_tid = transaction->t_tid;
487 blocknr = transaction->t_log_start;
488 } else if ((transaction = journal->j_committing_transaction) != NULL) {
489 first_tid = transaction->t_tid;
490 blocknr = transaction->t_log_start;
491 } else if ((transaction = journal->j_running_transaction) != NULL) {
492 first_tid = transaction->t_tid;
493 blocknr = journal->j_head;
494 } else {
495 first_tid = journal->j_transaction_sequence;
496 blocknr = journal->j_head;
498 spin_unlock(&journal->j_list_lock);
499 J_ASSERT(blocknr != 0);
501 /* If the oldest pinned transaction is at the tail of the log
502 already then there's not much we can do right now. */
503 if (journal->j_tail_sequence == first_tid) {
504 spin_unlock(&journal->j_state_lock);
505 return 1;
508 /* OK, update the superblock to recover the freed space.
509 * Physical blocks come first: have we wrapped beyond the end of
510 * the log? */
511 freed = blocknr - journal->j_tail;
512 if (blocknr < journal->j_tail)
513 freed = freed + journal->j_last - journal->j_first;
515 trace_jbd_cleanup_journal_tail(journal, first_tid, blocknr, freed);
516 jbd_debug(1,
517 "Cleaning journal tail from %d to %d (offset %u), "
518 "freeing %u\n",
519 journal->j_tail_sequence, first_tid, blocknr, freed);
521 journal->j_free += freed;
522 journal->j_tail_sequence = first_tid;
523 journal->j_tail = blocknr;
524 spin_unlock(&journal->j_state_lock);
525 if (!(journal->j_flags & JFS_ABORT))
526 journal_update_superblock(journal, 1);
527 return 0;
531 /* Checkpoint list management */
534 * journal_clean_one_cp_list
536 * Find all the written-back checkpoint buffers in the given list and release
537 * them.
539 * Called with j_list_lock held.
540 * Returns number of bufers reaped (for debug)
543 static int journal_clean_one_cp_list(struct journal_head *jh, int *released)
545 struct journal_head *last_jh;
546 struct journal_head *next_jh = jh;
547 int ret, freed = 0;
549 *released = 0;
550 if (!jh)
551 return 0;
553 last_jh = jh->b_cpprev;
554 do {
555 jh = next_jh;
556 next_jh = jh->b_cpnext;
557 /* Use trylock because of the ranking */
558 if (jbd_trylock_bh_state(jh2bh(jh))) {
559 ret = __try_to_free_cp_buf(jh);
560 if (ret) {
561 freed++;
562 if (ret == 2) {
563 *released = 1;
564 return freed;
569 * This function only frees up some memory
570 * if possible so we dont have an obligation
571 * to finish processing. Bail out if preemption
572 * requested:
574 if (need_resched())
575 return freed;
576 } while (jh != last_jh);
578 return freed;
582 * journal_clean_checkpoint_list
584 * Find all the written-back checkpoint buffers in the journal and release them.
586 * Called with the journal locked.
587 * Called with j_list_lock held.
588 * Returns number of buffers reaped (for debug)
591 int __journal_clean_checkpoint_list(journal_t *journal)
593 transaction_t *transaction, *last_transaction, *next_transaction;
594 int ret = 0;
595 int released;
597 transaction = journal->j_checkpoint_transactions;
598 if (!transaction)
599 goto out;
601 last_transaction = transaction->t_cpprev;
602 next_transaction = transaction;
603 do {
604 transaction = next_transaction;
605 next_transaction = transaction->t_cpnext;
606 ret += journal_clean_one_cp_list(transaction->
607 t_checkpoint_list, &released);
609 * This function only frees up some memory if possible so we
610 * dont have an obligation to finish processing. Bail out if
611 * preemption requested:
613 if (need_resched())
614 goto out;
615 if (released)
616 continue;
618 * It is essential that we are as careful as in the case of
619 * t_checkpoint_list with removing the buffer from the list as
620 * we can possibly see not yet submitted buffers on io_list
622 ret += journal_clean_one_cp_list(transaction->
623 t_checkpoint_io_list, &released);
624 if (need_resched())
625 goto out;
626 } while (transaction != last_transaction);
627 out:
628 return ret;
632 * journal_remove_checkpoint: called after a buffer has been committed
633 * to disk (either by being write-back flushed to disk, or being
634 * committed to the log).
636 * We cannot safely clean a transaction out of the log until all of the
637 * buffer updates committed in that transaction have safely been stored
638 * elsewhere on disk. To achieve this, all of the buffers in a
639 * transaction need to be maintained on the transaction's checkpoint
640 * lists until they have been rewritten, at which point this function is
641 * called to remove the buffer from the existing transaction's
642 * checkpoint lists.
644 * The function returns 1 if it frees the transaction, 0 otherwise.
645 * The function can free jh and bh.
647 * This function is called with j_list_lock held.
648 * This function is called with jbd_lock_bh_state(jh2bh(jh))
651 int __journal_remove_checkpoint(struct journal_head *jh)
653 transaction_t *transaction;
654 journal_t *journal;
655 int ret = 0;
657 JBUFFER_TRACE(jh, "entry");
659 if ((transaction = jh->b_cp_transaction) == NULL) {
660 JBUFFER_TRACE(jh, "not on transaction");
661 goto out;
663 journal = transaction->t_journal;
665 JBUFFER_TRACE(jh, "removing from transaction");
666 __buffer_unlink(jh);
667 jh->b_cp_transaction = NULL;
668 journal_put_journal_head(jh);
670 if (transaction->t_checkpoint_list != NULL ||
671 transaction->t_checkpoint_io_list != NULL)
672 goto out;
675 * There is one special case to worry about: if we have just pulled the
676 * buffer off a running or committing transaction's checkpoing list,
677 * then even if the checkpoint list is empty, the transaction obviously
678 * cannot be dropped!
680 * The locking here around t_state is a bit sleazy.
681 * See the comment at the end of journal_commit_transaction().
683 if (transaction->t_state != T_FINISHED)
684 goto out;
686 /* OK, that was the last buffer for the transaction: we can now
687 safely remove this transaction from the log */
689 __journal_drop_transaction(journal, transaction);
691 /* Just in case anybody was waiting for more transactions to be
692 checkpointed... */
693 wake_up(&journal->j_wait_logspace);
694 ret = 1;
695 out:
696 return ret;
700 * journal_insert_checkpoint: put a committed buffer onto a checkpoint
701 * list so that we know when it is safe to clean the transaction out of
702 * the log.
704 * Called with the journal locked.
705 * Called with j_list_lock held.
707 void __journal_insert_checkpoint(struct journal_head *jh,
708 transaction_t *transaction)
710 JBUFFER_TRACE(jh, "entry");
711 J_ASSERT_JH(jh, buffer_dirty(jh2bh(jh)) || buffer_jbddirty(jh2bh(jh)));
712 J_ASSERT_JH(jh, jh->b_cp_transaction == NULL);
714 /* Get reference for checkpointing transaction */
715 journal_grab_journal_head(jh2bh(jh));
716 jh->b_cp_transaction = transaction;
718 if (!transaction->t_checkpoint_list) {
719 jh->b_cpnext = jh->b_cpprev = jh;
720 } else {
721 jh->b_cpnext = transaction->t_checkpoint_list;
722 jh->b_cpprev = transaction->t_checkpoint_list->b_cpprev;
723 jh->b_cpprev->b_cpnext = jh;
724 jh->b_cpnext->b_cpprev = jh;
726 transaction->t_checkpoint_list = jh;
730 * We've finished with this transaction structure: adios...
732 * The transaction must have no links except for the checkpoint by this
733 * point.
735 * Called with the journal locked.
736 * Called with j_list_lock held.
739 void __journal_drop_transaction(journal_t *journal, transaction_t *transaction)
741 assert_spin_locked(&journal->j_list_lock);
742 if (transaction->t_cpnext) {
743 transaction->t_cpnext->t_cpprev = transaction->t_cpprev;
744 transaction->t_cpprev->t_cpnext = transaction->t_cpnext;
745 if (journal->j_checkpoint_transactions == transaction)
746 journal->j_checkpoint_transactions =
747 transaction->t_cpnext;
748 if (journal->j_checkpoint_transactions == transaction)
749 journal->j_checkpoint_transactions = NULL;
752 J_ASSERT(transaction->t_state == T_FINISHED);
753 J_ASSERT(transaction->t_buffers == NULL);
754 J_ASSERT(transaction->t_sync_datalist == NULL);
755 J_ASSERT(transaction->t_forget == NULL);
756 J_ASSERT(transaction->t_iobuf_list == NULL);
757 J_ASSERT(transaction->t_shadow_list == NULL);
758 J_ASSERT(transaction->t_log_list == NULL);
759 J_ASSERT(transaction->t_checkpoint_list == NULL);
760 J_ASSERT(transaction->t_checkpoint_io_list == NULL);
761 J_ASSERT(transaction->t_updates == 0);
762 J_ASSERT(journal->j_committing_transaction != transaction);
763 J_ASSERT(journal->j_running_transaction != transaction);
765 trace_jbd_drop_transaction(journal, transaction);
766 jbd_debug(1, "Dropping transaction %d, all done\n", transaction->t_tid);
767 kfree(transaction);