Btrfs: wait for ordered extents if we didn't reclaim enough
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / fs / jbd / transaction.c
blob7e59c6e66f9b79a9fd57d3d450e51ee3de79b722
1 /*
2 * linux/fs/jbd/transaction.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6 * Copyright 1998 Red Hat corp --- 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 * Generic filesystem transaction handling code; part of the ext2fs
13 * journaling system.
15 * This file manages transactions (compound commits managed by the
16 * journaling code) and handles (individual atomic operations by the
17 * filesystem).
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/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29 #include <linux/backing-dev.h>
31 static void __journal_temp_unlink_buffer(struct journal_head *jh);
34 * get_transaction: obtain a new transaction_t object.
36 * Simply allocate and initialise a new transaction. Create it in
37 * RUNNING state and add it to the current journal (which should not
38 * have an existing running transaction: we only make a new transaction
39 * once we have started to commit the old one).
41 * Preconditions:
42 * The journal MUST be locked. We don't perform atomic mallocs on the
43 * new transaction and we can't block without protecting against other
44 * processes trying to touch the journal while it is in transition.
46 * Called under j_state_lock
49 static transaction_t *
50 get_transaction(journal_t *journal, transaction_t *transaction)
52 transaction->t_journal = journal;
53 transaction->t_state = T_RUNNING;
54 transaction->t_start_time = ktime_get();
55 transaction->t_tid = journal->j_transaction_sequence++;
56 transaction->t_expires = jiffies + journal->j_commit_interval;
57 spin_lock_init(&transaction->t_handle_lock);
59 /* Set up the commit timer for the new transaction. */
60 journal->j_commit_timer.expires =
61 round_jiffies_up(transaction->t_expires);
62 add_timer(&journal->j_commit_timer);
64 J_ASSERT(journal->j_running_transaction == NULL);
65 journal->j_running_transaction = transaction;
67 return transaction;
71 * Handle management.
73 * A handle_t is an object which represents a single atomic update to a
74 * filesystem, and which tracks all of the modifications which form part
75 * of that one update.
79 * start_this_handle: Given a handle, deal with any locking or stalling
80 * needed to make sure that there is enough journal space for the handle
81 * to begin. Attach the handle to a transaction and set up the
82 * transaction's buffer credits.
85 static int start_this_handle(journal_t *journal, handle_t *handle)
87 transaction_t *transaction;
88 int needed;
89 int nblocks = handle->h_buffer_credits;
90 transaction_t *new_transaction = NULL;
91 int ret = 0;
93 if (nblocks > journal->j_max_transaction_buffers) {
94 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
95 current->comm, nblocks,
96 journal->j_max_transaction_buffers);
97 ret = -ENOSPC;
98 goto out;
101 alloc_transaction:
102 if (!journal->j_running_transaction) {
103 new_transaction = kzalloc(sizeof(*new_transaction), GFP_NOFS);
104 if (!new_transaction) {
105 congestion_wait(BLK_RW_ASYNC, HZ/50);
106 goto alloc_transaction;
110 jbd_debug(3, "New handle %p going live.\n", handle);
112 repeat:
115 * We need to hold j_state_lock until t_updates has been incremented,
116 * for proper journal barrier handling
118 spin_lock(&journal->j_state_lock);
119 repeat_locked:
120 if (is_journal_aborted(journal) ||
121 (journal->j_errno != 0 && !(journal->j_flags & JFS_ACK_ERR))) {
122 spin_unlock(&journal->j_state_lock);
123 ret = -EROFS;
124 goto out;
127 /* Wait on the journal's transaction barrier if necessary */
128 if (journal->j_barrier_count) {
129 spin_unlock(&journal->j_state_lock);
130 wait_event(journal->j_wait_transaction_locked,
131 journal->j_barrier_count == 0);
132 goto repeat;
135 if (!journal->j_running_transaction) {
136 if (!new_transaction) {
137 spin_unlock(&journal->j_state_lock);
138 goto alloc_transaction;
140 get_transaction(journal, new_transaction);
141 new_transaction = NULL;
144 transaction = journal->j_running_transaction;
147 * If the current transaction is locked down for commit, wait for the
148 * lock to be released.
150 if (transaction->t_state == T_LOCKED) {
151 DEFINE_WAIT(wait);
153 prepare_to_wait(&journal->j_wait_transaction_locked,
154 &wait, TASK_UNINTERRUPTIBLE);
155 spin_unlock(&journal->j_state_lock);
156 schedule();
157 finish_wait(&journal->j_wait_transaction_locked, &wait);
158 goto repeat;
162 * If there is not enough space left in the log to write all potential
163 * buffers requested by this operation, we need to stall pending a log
164 * checkpoint to free some more log space.
166 spin_lock(&transaction->t_handle_lock);
167 needed = transaction->t_outstanding_credits + nblocks;
169 if (needed > journal->j_max_transaction_buffers) {
171 * If the current transaction is already too large, then start
172 * to commit it: we can then go back and attach this handle to
173 * a new transaction.
175 DEFINE_WAIT(wait);
177 jbd_debug(2, "Handle %p starting new commit...\n", handle);
178 spin_unlock(&transaction->t_handle_lock);
179 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
180 TASK_UNINTERRUPTIBLE);
181 __log_start_commit(journal, transaction->t_tid);
182 spin_unlock(&journal->j_state_lock);
183 schedule();
184 finish_wait(&journal->j_wait_transaction_locked, &wait);
185 goto repeat;
189 * The commit code assumes that it can get enough log space
190 * without forcing a checkpoint. This is *critical* for
191 * correctness: a checkpoint of a buffer which is also
192 * associated with a committing transaction creates a deadlock,
193 * so commit simply cannot force through checkpoints.
195 * We must therefore ensure the necessary space in the journal
196 * *before* starting to dirty potentially checkpointed buffers
197 * in the new transaction.
199 * The worst part is, any transaction currently committing can
200 * reduce the free space arbitrarily. Be careful to account for
201 * those buffers when checkpointing.
205 * @@@ AKPM: This seems rather over-defensive. We're giving commit
206 * a _lot_ of headroom: 1/4 of the journal plus the size of
207 * the committing transaction. Really, we only need to give it
208 * committing_transaction->t_outstanding_credits plus "enough" for
209 * the log control blocks.
210 * Also, this test is inconsistent with the matching one in
211 * journal_extend().
213 if (__log_space_left(journal) < jbd_space_needed(journal)) {
214 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
215 spin_unlock(&transaction->t_handle_lock);
216 __log_wait_for_space(journal);
217 goto repeat_locked;
220 /* OK, account for the buffers that this operation expects to
221 * use and add the handle to the running transaction. */
223 handle->h_transaction = transaction;
224 transaction->t_outstanding_credits += nblocks;
225 transaction->t_updates++;
226 transaction->t_handle_count++;
227 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
228 handle, nblocks, transaction->t_outstanding_credits,
229 __log_space_left(journal));
230 spin_unlock(&transaction->t_handle_lock);
231 spin_unlock(&journal->j_state_lock);
233 lock_map_acquire(&handle->h_lockdep_map);
234 out:
235 if (unlikely(new_transaction)) /* It's usually NULL */
236 kfree(new_transaction);
237 return ret;
240 static struct lock_class_key jbd_handle_key;
242 /* Allocate a new handle. This should probably be in a slab... */
243 static handle_t *new_handle(int nblocks)
245 handle_t *handle = jbd_alloc_handle(GFP_NOFS);
246 if (!handle)
247 return NULL;
248 memset(handle, 0, sizeof(*handle));
249 handle->h_buffer_credits = nblocks;
250 handle->h_ref = 1;
252 lockdep_init_map(&handle->h_lockdep_map, "jbd_handle", &jbd_handle_key, 0);
254 return handle;
258 * handle_t *journal_start() - Obtain a new handle.
259 * @journal: Journal to start transaction on.
260 * @nblocks: number of block buffer we might modify
262 * We make sure that the transaction can guarantee at least nblocks of
263 * modified buffers in the log. We block until the log can guarantee
264 * that much space.
266 * This function is visible to journal users (like ext3fs), so is not
267 * called with the journal already locked.
269 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
270 * on failure.
272 handle_t *journal_start(journal_t *journal, int nblocks)
274 handle_t *handle = journal_current_handle();
275 int err;
277 if (!journal)
278 return ERR_PTR(-EROFS);
280 if (handle) {
281 J_ASSERT(handle->h_transaction->t_journal == journal);
282 handle->h_ref++;
283 return handle;
286 handle = new_handle(nblocks);
287 if (!handle)
288 return ERR_PTR(-ENOMEM);
290 current->journal_info = handle;
292 err = start_this_handle(journal, handle);
293 if (err < 0) {
294 jbd_free_handle(handle);
295 current->journal_info = NULL;
296 handle = ERR_PTR(err);
298 return handle;
302 * int journal_extend() - extend buffer credits.
303 * @handle: handle to 'extend'
304 * @nblocks: nr blocks to try to extend by.
306 * Some transactions, such as large extends and truncates, can be done
307 * atomically all at once or in several stages. The operation requests
308 * a credit for a number of buffer modications in advance, but can
309 * extend its credit if it needs more.
311 * journal_extend tries to give the running handle more buffer credits.
312 * It does not guarantee that allocation - this is a best-effort only.
313 * The calling process MUST be able to deal cleanly with a failure to
314 * extend here.
316 * Return 0 on success, non-zero on failure.
318 * return code < 0 implies an error
319 * return code > 0 implies normal transaction-full status.
321 int journal_extend(handle_t *handle, int nblocks)
323 transaction_t *transaction = handle->h_transaction;
324 journal_t *journal = transaction->t_journal;
325 int result;
326 int wanted;
328 result = -EIO;
329 if (is_handle_aborted(handle))
330 goto out;
332 result = 1;
334 spin_lock(&journal->j_state_lock);
336 /* Don't extend a locked-down transaction! */
337 if (handle->h_transaction->t_state != T_RUNNING) {
338 jbd_debug(3, "denied handle %p %d blocks: "
339 "transaction not running\n", handle, nblocks);
340 goto error_out;
343 spin_lock(&transaction->t_handle_lock);
344 wanted = transaction->t_outstanding_credits + nblocks;
346 if (wanted > journal->j_max_transaction_buffers) {
347 jbd_debug(3, "denied handle %p %d blocks: "
348 "transaction too large\n", handle, nblocks);
349 goto unlock;
352 if (wanted > __log_space_left(journal)) {
353 jbd_debug(3, "denied handle %p %d blocks: "
354 "insufficient log space\n", handle, nblocks);
355 goto unlock;
358 handle->h_buffer_credits += nblocks;
359 transaction->t_outstanding_credits += nblocks;
360 result = 0;
362 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
363 unlock:
364 spin_unlock(&transaction->t_handle_lock);
365 error_out:
366 spin_unlock(&journal->j_state_lock);
367 out:
368 return result;
373 * int journal_restart() - restart a handle.
374 * @handle: handle to restart
375 * @nblocks: nr credits requested
377 * Restart a handle for a multi-transaction filesystem
378 * operation.
380 * If the journal_extend() call above fails to grant new buffer credits
381 * to a running handle, a call to journal_restart will commit the
382 * handle's transaction so far and reattach the handle to a new
383 * transaction capabable of guaranteeing the requested number of
384 * credits.
387 int journal_restart(handle_t *handle, int nblocks)
389 transaction_t *transaction = handle->h_transaction;
390 journal_t *journal = transaction->t_journal;
391 int ret;
393 /* If we've had an abort of any type, don't even think about
394 * actually doing the restart! */
395 if (is_handle_aborted(handle))
396 return 0;
399 * First unlink the handle from its current transaction, and start the
400 * commit on that.
402 J_ASSERT(transaction->t_updates > 0);
403 J_ASSERT(journal_current_handle() == handle);
405 spin_lock(&journal->j_state_lock);
406 spin_lock(&transaction->t_handle_lock);
407 transaction->t_outstanding_credits -= handle->h_buffer_credits;
408 transaction->t_updates--;
410 if (!transaction->t_updates)
411 wake_up(&journal->j_wait_updates);
412 spin_unlock(&transaction->t_handle_lock);
414 jbd_debug(2, "restarting handle %p\n", handle);
415 __log_start_commit(journal, transaction->t_tid);
416 spin_unlock(&journal->j_state_lock);
418 lock_map_release(&handle->h_lockdep_map);
419 handle->h_buffer_credits = nblocks;
420 ret = start_this_handle(journal, handle);
421 return ret;
426 * void journal_lock_updates () - establish a transaction barrier.
427 * @journal: Journal to establish a barrier on.
429 * This locks out any further updates from being started, and blocks
430 * until all existing updates have completed, returning only once the
431 * journal is in a quiescent state with no updates running.
433 * The journal lock should not be held on entry.
435 void journal_lock_updates(journal_t *journal)
437 DEFINE_WAIT(wait);
439 spin_lock(&journal->j_state_lock);
440 ++journal->j_barrier_count;
442 /* Wait until there are no running updates */
443 while (1) {
444 transaction_t *transaction = journal->j_running_transaction;
446 if (!transaction)
447 break;
449 spin_lock(&transaction->t_handle_lock);
450 if (!transaction->t_updates) {
451 spin_unlock(&transaction->t_handle_lock);
452 break;
454 prepare_to_wait(&journal->j_wait_updates, &wait,
455 TASK_UNINTERRUPTIBLE);
456 spin_unlock(&transaction->t_handle_lock);
457 spin_unlock(&journal->j_state_lock);
458 schedule();
459 finish_wait(&journal->j_wait_updates, &wait);
460 spin_lock(&journal->j_state_lock);
462 spin_unlock(&journal->j_state_lock);
465 * We have now established a barrier against other normal updates, but
466 * we also need to barrier against other journal_lock_updates() calls
467 * to make sure that we serialise special journal-locked operations
468 * too.
470 mutex_lock(&journal->j_barrier);
474 * void journal_unlock_updates (journal_t* journal) - release barrier
475 * @journal: Journal to release the barrier on.
477 * Release a transaction barrier obtained with journal_lock_updates().
479 * Should be called without the journal lock held.
481 void journal_unlock_updates (journal_t *journal)
483 J_ASSERT(journal->j_barrier_count != 0);
485 mutex_unlock(&journal->j_barrier);
486 spin_lock(&journal->j_state_lock);
487 --journal->j_barrier_count;
488 spin_unlock(&journal->j_state_lock);
489 wake_up(&journal->j_wait_transaction_locked);
492 static void warn_dirty_buffer(struct buffer_head *bh)
494 char b[BDEVNAME_SIZE];
496 printk(KERN_WARNING
497 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
498 "There's a risk of filesystem corruption in case of system "
499 "crash.\n",
500 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
504 * If the buffer is already part of the current transaction, then there
505 * is nothing we need to do. If it is already part of a prior
506 * transaction which we are still committing to disk, then we need to
507 * make sure that we do not overwrite the old copy: we do copy-out to
508 * preserve the copy going to disk. We also account the buffer against
509 * the handle's metadata buffer credits (unless the buffer is already
510 * part of the transaction, that is).
513 static int
514 do_get_write_access(handle_t *handle, struct journal_head *jh,
515 int force_copy)
517 struct buffer_head *bh;
518 transaction_t *transaction;
519 journal_t *journal;
520 int error;
521 char *frozen_buffer = NULL;
522 int need_copy = 0;
524 if (is_handle_aborted(handle))
525 return -EROFS;
527 transaction = handle->h_transaction;
528 journal = transaction->t_journal;
530 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
532 JBUFFER_TRACE(jh, "entry");
533 repeat:
534 bh = jh2bh(jh);
536 /* @@@ Need to check for errors here at some point. */
538 lock_buffer(bh);
539 jbd_lock_bh_state(bh);
541 /* We now hold the buffer lock so it is safe to query the buffer
542 * state. Is the buffer dirty?
544 * If so, there are two possibilities. The buffer may be
545 * non-journaled, and undergoing a quite legitimate writeback.
546 * Otherwise, it is journaled, and we don't expect dirty buffers
547 * in that state (the buffers should be marked JBD_Dirty
548 * instead.) So either the IO is being done under our own
549 * control and this is a bug, or it's a third party IO such as
550 * dump(8) (which may leave the buffer scheduled for read ---
551 * ie. locked but not dirty) or tune2fs (which may actually have
552 * the buffer dirtied, ugh.) */
554 if (buffer_dirty(bh)) {
556 * First question: is this buffer already part of the current
557 * transaction or the existing committing transaction?
559 if (jh->b_transaction) {
560 J_ASSERT_JH(jh,
561 jh->b_transaction == transaction ||
562 jh->b_transaction ==
563 journal->j_committing_transaction);
564 if (jh->b_next_transaction)
565 J_ASSERT_JH(jh, jh->b_next_transaction ==
566 transaction);
567 warn_dirty_buffer(bh);
570 * In any case we need to clean the dirty flag and we must
571 * do it under the buffer lock to be sure we don't race
572 * with running write-out.
574 JBUFFER_TRACE(jh, "Journalling dirty buffer");
575 clear_buffer_dirty(bh);
576 set_buffer_jbddirty(bh);
579 unlock_buffer(bh);
581 error = -EROFS;
582 if (is_handle_aborted(handle)) {
583 jbd_unlock_bh_state(bh);
584 goto out;
586 error = 0;
589 * The buffer is already part of this transaction if b_transaction or
590 * b_next_transaction points to it
592 if (jh->b_transaction == transaction ||
593 jh->b_next_transaction == transaction)
594 goto done;
597 * this is the first time this transaction is touching this buffer,
598 * reset the modified flag
600 jh->b_modified = 0;
603 * If there is already a copy-out version of this buffer, then we don't
604 * need to make another one
606 if (jh->b_frozen_data) {
607 JBUFFER_TRACE(jh, "has frozen data");
608 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
609 jh->b_next_transaction = transaction;
610 goto done;
613 /* Is there data here we need to preserve? */
615 if (jh->b_transaction && jh->b_transaction != transaction) {
616 JBUFFER_TRACE(jh, "owned by older transaction");
617 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
618 J_ASSERT_JH(jh, jh->b_transaction ==
619 journal->j_committing_transaction);
621 /* There is one case we have to be very careful about.
622 * If the committing transaction is currently writing
623 * this buffer out to disk and has NOT made a copy-out,
624 * then we cannot modify the buffer contents at all
625 * right now. The essence of copy-out is that it is the
626 * extra copy, not the primary copy, which gets
627 * journaled. If the primary copy is already going to
628 * disk then we cannot do copy-out here. */
630 if (jh->b_jlist == BJ_Shadow) {
631 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
632 wait_queue_head_t *wqh;
634 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
636 JBUFFER_TRACE(jh, "on shadow: sleep");
637 jbd_unlock_bh_state(bh);
638 /* commit wakes up all shadow buffers after IO */
639 for ( ; ; ) {
640 prepare_to_wait(wqh, &wait.wait,
641 TASK_UNINTERRUPTIBLE);
642 if (jh->b_jlist != BJ_Shadow)
643 break;
644 schedule();
646 finish_wait(wqh, &wait.wait);
647 goto repeat;
650 /* Only do the copy if the currently-owning transaction
651 * still needs it. If it is on the Forget list, the
652 * committing transaction is past that stage. The
653 * buffer had better remain locked during the kmalloc,
654 * but that should be true --- we hold the journal lock
655 * still and the buffer is already on the BUF_JOURNAL
656 * list so won't be flushed.
658 * Subtle point, though: if this is a get_undo_access,
659 * then we will be relying on the frozen_data to contain
660 * the new value of the committed_data record after the
661 * transaction, so we HAVE to force the frozen_data copy
662 * in that case. */
664 if (jh->b_jlist != BJ_Forget || force_copy) {
665 JBUFFER_TRACE(jh, "generate frozen data");
666 if (!frozen_buffer) {
667 JBUFFER_TRACE(jh, "allocate memory for buffer");
668 jbd_unlock_bh_state(bh);
669 frozen_buffer =
670 jbd_alloc(jh2bh(jh)->b_size,
671 GFP_NOFS);
672 if (!frozen_buffer) {
673 printk(KERN_EMERG
674 "%s: OOM for frozen_buffer\n",
675 __func__);
676 JBUFFER_TRACE(jh, "oom!");
677 error = -ENOMEM;
678 jbd_lock_bh_state(bh);
679 goto done;
681 goto repeat;
683 jh->b_frozen_data = frozen_buffer;
684 frozen_buffer = NULL;
685 need_copy = 1;
687 jh->b_next_transaction = transaction;
692 * Finally, if the buffer is not journaled right now, we need to make
693 * sure it doesn't get written to disk before the caller actually
694 * commits the new data
696 if (!jh->b_transaction) {
697 JBUFFER_TRACE(jh, "no transaction");
698 J_ASSERT_JH(jh, !jh->b_next_transaction);
699 JBUFFER_TRACE(jh, "file as BJ_Reserved");
700 spin_lock(&journal->j_list_lock);
701 __journal_file_buffer(jh, transaction, BJ_Reserved);
702 spin_unlock(&journal->j_list_lock);
705 done:
706 if (need_copy) {
707 struct page *page;
708 int offset;
709 char *source;
711 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
712 "Possible IO failure.\n");
713 page = jh2bh(jh)->b_page;
714 offset = offset_in_page(jh2bh(jh)->b_data);
715 source = kmap_atomic(page, KM_USER0);
716 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
717 kunmap_atomic(source, KM_USER0);
719 jbd_unlock_bh_state(bh);
722 * If we are about to journal a buffer, then any revoke pending on it is
723 * no longer valid
725 journal_cancel_revoke(handle, jh);
727 out:
728 if (unlikely(frozen_buffer)) /* It's usually NULL */
729 jbd_free(frozen_buffer, bh->b_size);
731 JBUFFER_TRACE(jh, "exit");
732 return error;
736 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
737 * @handle: transaction to add buffer modifications to
738 * @bh: bh to be used for metadata writes
740 * Returns an error code or 0 on success.
742 * In full data journalling mode the buffer may be of type BJ_AsyncData,
743 * because we're write()ing a buffer which is also part of a shared mapping.
746 int journal_get_write_access(handle_t *handle, struct buffer_head *bh)
748 struct journal_head *jh = journal_add_journal_head(bh);
749 int rc;
751 /* We do not want to get caught playing with fields which the
752 * log thread also manipulates. Make sure that the buffer
753 * completes any outstanding IO before proceeding. */
754 rc = do_get_write_access(handle, jh, 0);
755 journal_put_journal_head(jh);
756 return rc;
761 * When the user wants to journal a newly created buffer_head
762 * (ie. getblk() returned a new buffer and we are going to populate it
763 * manually rather than reading off disk), then we need to keep the
764 * buffer_head locked until it has been completely filled with new
765 * data. In this case, we should be able to make the assertion that
766 * the bh is not already part of an existing transaction.
768 * The buffer should already be locked by the caller by this point.
769 * There is no lock ranking violation: it was a newly created,
770 * unlocked buffer beforehand. */
773 * int journal_get_create_access () - notify intent to use newly created bh
774 * @handle: transaction to new buffer to
775 * @bh: new buffer.
777 * Call this if you create a new bh.
779 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
781 transaction_t *transaction = handle->h_transaction;
782 journal_t *journal = transaction->t_journal;
783 struct journal_head *jh = journal_add_journal_head(bh);
784 int err;
786 jbd_debug(5, "journal_head %p\n", jh);
787 err = -EROFS;
788 if (is_handle_aborted(handle))
789 goto out;
790 err = 0;
792 JBUFFER_TRACE(jh, "entry");
794 * The buffer may already belong to this transaction due to pre-zeroing
795 * in the filesystem's new_block code. It may also be on the previous,
796 * committing transaction's lists, but it HAS to be in Forget state in
797 * that case: the transaction must have deleted the buffer for it to be
798 * reused here.
800 jbd_lock_bh_state(bh);
801 spin_lock(&journal->j_list_lock);
802 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
803 jh->b_transaction == NULL ||
804 (jh->b_transaction == journal->j_committing_transaction &&
805 jh->b_jlist == BJ_Forget)));
807 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
808 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
810 if (jh->b_transaction == NULL) {
812 * Previous journal_forget() could have left the buffer
813 * with jbddirty bit set because it was being committed. When
814 * the commit finished, we've filed the buffer for
815 * checkpointing and marked it dirty. Now we are reallocating
816 * the buffer so the transaction freeing it must have
817 * committed and so it's safe to clear the dirty bit.
819 clear_buffer_dirty(jh2bh(jh));
821 /* first access by this transaction */
822 jh->b_modified = 0;
824 JBUFFER_TRACE(jh, "file as BJ_Reserved");
825 __journal_file_buffer(jh, transaction, BJ_Reserved);
826 } else if (jh->b_transaction == journal->j_committing_transaction) {
827 /* first access by this transaction */
828 jh->b_modified = 0;
830 JBUFFER_TRACE(jh, "set next transaction");
831 jh->b_next_transaction = transaction;
833 spin_unlock(&journal->j_list_lock);
834 jbd_unlock_bh_state(bh);
837 * akpm: I added this. ext3_alloc_branch can pick up new indirect
838 * blocks which contain freed but then revoked metadata. We need
839 * to cancel the revoke in case we end up freeing it yet again
840 * and the reallocating as data - this would cause a second revoke,
841 * which hits an assertion error.
843 JBUFFER_TRACE(jh, "cancelling revoke");
844 journal_cancel_revoke(handle, jh);
845 out:
846 journal_put_journal_head(jh);
847 return err;
851 * int journal_get_undo_access() - Notify intent to modify metadata with non-rewindable consequences
852 * @handle: transaction
853 * @bh: buffer to undo
855 * Sometimes there is a need to distinguish between metadata which has
856 * been committed to disk and that which has not. The ext3fs code uses
857 * this for freeing and allocating space, we have to make sure that we
858 * do not reuse freed space until the deallocation has been committed,
859 * since if we overwrote that space we would make the delete
860 * un-rewindable in case of a crash.
862 * To deal with that, journal_get_undo_access requests write access to a
863 * buffer for parts of non-rewindable operations such as delete
864 * operations on the bitmaps. The journaling code must keep a copy of
865 * the buffer's contents prior to the undo_access call until such time
866 * as we know that the buffer has definitely been committed to disk.
868 * We never need to know which transaction the committed data is part
869 * of, buffers touched here are guaranteed to be dirtied later and so
870 * will be committed to a new transaction in due course, at which point
871 * we can discard the old committed data pointer.
873 * Returns error number or 0 on success.
875 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
877 int err;
878 struct journal_head *jh = journal_add_journal_head(bh);
879 char *committed_data = NULL;
881 JBUFFER_TRACE(jh, "entry");
884 * Do this first --- it can drop the journal lock, so we want to
885 * make sure that obtaining the committed_data is done
886 * atomically wrt. completion of any outstanding commits.
888 err = do_get_write_access(handle, jh, 1);
889 if (err)
890 goto out;
892 repeat:
893 if (!jh->b_committed_data) {
894 committed_data = jbd_alloc(jh2bh(jh)->b_size, GFP_NOFS);
895 if (!committed_data) {
896 printk(KERN_EMERG "%s: No memory for committed data\n",
897 __func__);
898 err = -ENOMEM;
899 goto out;
903 jbd_lock_bh_state(bh);
904 if (!jh->b_committed_data) {
905 /* Copy out the current buffer contents into the
906 * preserved, committed copy. */
907 JBUFFER_TRACE(jh, "generate b_committed data");
908 if (!committed_data) {
909 jbd_unlock_bh_state(bh);
910 goto repeat;
913 jh->b_committed_data = committed_data;
914 committed_data = NULL;
915 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
917 jbd_unlock_bh_state(bh);
918 out:
919 journal_put_journal_head(jh);
920 if (unlikely(committed_data))
921 jbd_free(committed_data, bh->b_size);
922 return err;
926 * int journal_dirty_data() - mark a buffer as containing dirty data to be flushed
927 * @handle: transaction
928 * @bh: bufferhead to mark
930 * Description:
931 * Mark a buffer as containing dirty data which needs to be flushed before
932 * we can commit the current transaction.
934 * The buffer is placed on the transaction's data list and is marked as
935 * belonging to the transaction.
937 * Returns error number or 0 on success.
939 * journal_dirty_data() can be called via page_launder->ext3_writepage
940 * by kswapd.
942 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
944 journal_t *journal = handle->h_transaction->t_journal;
945 int need_brelse = 0;
946 struct journal_head *jh;
947 int ret = 0;
949 if (is_handle_aborted(handle))
950 return ret;
952 jh = journal_add_journal_head(bh);
953 JBUFFER_TRACE(jh, "entry");
956 * The buffer could *already* be dirty. Writeout can start
957 * at any time.
959 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
962 * What if the buffer is already part of a running transaction?
964 * There are two cases:
965 * 1) It is part of the current running transaction. Refile it,
966 * just in case we have allocated it as metadata, deallocated
967 * it, then reallocated it as data.
968 * 2) It is part of the previous, still-committing transaction.
969 * If all we want to do is to guarantee that the buffer will be
970 * written to disk before this new transaction commits, then
971 * being sure that the *previous* transaction has this same
972 * property is sufficient for us! Just leave it on its old
973 * transaction.
975 * In case (2), the buffer must not already exist as metadata
976 * --- that would violate write ordering (a transaction is free
977 * to write its data at any point, even before the previous
978 * committing transaction has committed). The caller must
979 * never, ever allow this to happen: there's nothing we can do
980 * about it in this layer.
982 jbd_lock_bh_state(bh);
983 spin_lock(&journal->j_list_lock);
985 /* Now that we have bh_state locked, are we really still mapped? */
986 if (!buffer_mapped(bh)) {
987 JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
988 goto no_journal;
991 if (jh->b_transaction) {
992 JBUFFER_TRACE(jh, "has transaction");
993 if (jh->b_transaction != handle->h_transaction) {
994 JBUFFER_TRACE(jh, "belongs to older transaction");
995 J_ASSERT_JH(jh, jh->b_transaction ==
996 journal->j_committing_transaction);
998 /* @@@ IS THIS TRUE ? */
1000 * Not any more. Scenario: someone does a write()
1001 * in data=journal mode. The buffer's transaction has
1002 * moved into commit. Then someone does another
1003 * write() to the file. We do the frozen data copyout
1004 * and set b_next_transaction to point to j_running_t.
1005 * And while we're in that state, someone does a
1006 * writepage() in an attempt to pageout the same area
1007 * of the file via a shared mapping. At present that
1008 * calls journal_dirty_data(), and we get right here.
1009 * It may be too late to journal the data. Simply
1010 * falling through to the next test will suffice: the
1011 * data will be dirty and wil be checkpointed. The
1012 * ordering comments in the next comment block still
1013 * apply.
1015 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1018 * If we're journalling data, and this buffer was
1019 * subject to a write(), it could be metadata, forget
1020 * or shadow against the committing transaction. Now,
1021 * someone has dirtied the same darn page via a mapping
1022 * and it is being writepage()'d.
1023 * We *could* just steal the page from commit, with some
1024 * fancy locking there. Instead, we just skip it -
1025 * don't tie the page's buffers to the new transaction
1026 * at all.
1027 * Implication: if we crash before the writepage() data
1028 * is written into the filesystem, recovery will replay
1029 * the write() data.
1031 if (jh->b_jlist != BJ_None &&
1032 jh->b_jlist != BJ_SyncData &&
1033 jh->b_jlist != BJ_Locked) {
1034 JBUFFER_TRACE(jh, "Not stealing");
1035 goto no_journal;
1039 * This buffer may be undergoing writeout in commit. We
1040 * can't return from here and let the caller dirty it
1041 * again because that can cause the write-out loop in
1042 * commit to never terminate.
1044 if (buffer_dirty(bh)) {
1045 get_bh(bh);
1046 spin_unlock(&journal->j_list_lock);
1047 jbd_unlock_bh_state(bh);
1048 need_brelse = 1;
1049 sync_dirty_buffer(bh);
1050 jbd_lock_bh_state(bh);
1051 spin_lock(&journal->j_list_lock);
1052 /* Since we dropped the lock... */
1053 if (!buffer_mapped(bh)) {
1054 JBUFFER_TRACE(jh, "buffer got unmapped");
1055 goto no_journal;
1057 /* The buffer may become locked again at any
1058 time if it is redirtied */
1062 * We cannot remove the buffer with io error from the
1063 * committing transaction, because otherwise it would
1064 * miss the error and the commit would not abort.
1066 if (unlikely(!buffer_uptodate(bh))) {
1067 ret = -EIO;
1068 goto no_journal;
1070 /* We might have slept so buffer could be refiled now */
1071 if (jh->b_transaction != NULL &&
1072 jh->b_transaction != handle->h_transaction) {
1073 JBUFFER_TRACE(jh, "unfile from commit");
1074 __journal_temp_unlink_buffer(jh);
1075 /* It still points to the committing
1076 * transaction; move it to this one so
1077 * that the refile assert checks are
1078 * happy. */
1079 jh->b_transaction = handle->h_transaction;
1081 /* The buffer will be refiled below */
1085 * Special case --- the buffer might actually have been
1086 * allocated and then immediately deallocated in the previous,
1087 * committing transaction, so might still be left on that
1088 * transaction's metadata lists.
1090 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1091 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1092 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1093 JBUFFER_TRACE(jh, "file as data");
1094 __journal_file_buffer(jh, handle->h_transaction,
1095 BJ_SyncData);
1097 } else {
1098 JBUFFER_TRACE(jh, "not on a transaction");
1099 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1101 no_journal:
1102 spin_unlock(&journal->j_list_lock);
1103 jbd_unlock_bh_state(bh);
1104 if (need_brelse) {
1105 BUFFER_TRACE(bh, "brelse");
1106 __brelse(bh);
1108 JBUFFER_TRACE(jh, "exit");
1109 journal_put_journal_head(jh);
1110 return ret;
1114 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1115 * @handle: transaction to add buffer to.
1116 * @bh: buffer to mark
1118 * Mark dirty metadata which needs to be journaled as part of the current
1119 * transaction.
1121 * The buffer is placed on the transaction's metadata list and is marked
1122 * as belonging to the transaction.
1124 * Returns error number or 0 on success.
1126 * Special care needs to be taken if the buffer already belongs to the
1127 * current committing transaction (in which case we should have frozen
1128 * data present for that commit). In that case, we don't relink the
1129 * buffer: that only gets done when the old transaction finally
1130 * completes its commit.
1132 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1134 transaction_t *transaction = handle->h_transaction;
1135 journal_t *journal = transaction->t_journal;
1136 struct journal_head *jh = bh2jh(bh);
1138 jbd_debug(5, "journal_head %p\n", jh);
1139 JBUFFER_TRACE(jh, "entry");
1140 if (is_handle_aborted(handle))
1141 goto out;
1143 jbd_lock_bh_state(bh);
1145 if (jh->b_modified == 0) {
1147 * This buffer's got modified and becoming part
1148 * of the transaction. This needs to be done
1149 * once a transaction -bzzz
1151 jh->b_modified = 1;
1152 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1153 handle->h_buffer_credits--;
1157 * fastpath, to avoid expensive locking. If this buffer is already
1158 * on the running transaction's metadata list there is nothing to do.
1159 * Nobody can take it off again because there is a handle open.
1160 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1161 * result in this test being false, so we go in and take the locks.
1163 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1164 JBUFFER_TRACE(jh, "fastpath");
1165 J_ASSERT_JH(jh, jh->b_transaction ==
1166 journal->j_running_transaction);
1167 goto out_unlock_bh;
1170 set_buffer_jbddirty(bh);
1173 * Metadata already on the current transaction list doesn't
1174 * need to be filed. Metadata on another transaction's list must
1175 * be committing, and will be refiled once the commit completes:
1176 * leave it alone for now.
1178 if (jh->b_transaction != transaction) {
1179 JBUFFER_TRACE(jh, "already on other transaction");
1180 J_ASSERT_JH(jh, jh->b_transaction ==
1181 journal->j_committing_transaction);
1182 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1183 /* And this case is illegal: we can't reuse another
1184 * transaction's data buffer, ever. */
1185 goto out_unlock_bh;
1188 /* That test should have eliminated the following case: */
1189 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1191 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1192 spin_lock(&journal->j_list_lock);
1193 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1194 spin_unlock(&journal->j_list_lock);
1195 out_unlock_bh:
1196 jbd_unlock_bh_state(bh);
1197 out:
1198 JBUFFER_TRACE(jh, "exit");
1199 return 0;
1203 * journal_release_buffer: undo a get_write_access without any buffer
1204 * updates, if the update decided in the end that it didn't need access.
1207 void
1208 journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1210 BUFFER_TRACE(bh, "entry");
1214 * void journal_forget() - bforget() for potentially-journaled buffers.
1215 * @handle: transaction handle
1216 * @bh: bh to 'forget'
1218 * We can only do the bforget if there are no commits pending against the
1219 * buffer. If the buffer is dirty in the current running transaction we
1220 * can safely unlink it.
1222 * bh may not be a journalled buffer at all - it may be a non-JBD
1223 * buffer which came off the hashtable. Check for this.
1225 * Decrements bh->b_count by one.
1227 * Allow this call even if the handle has aborted --- it may be part of
1228 * the caller's cleanup after an abort.
1230 int journal_forget (handle_t *handle, struct buffer_head *bh)
1232 transaction_t *transaction = handle->h_transaction;
1233 journal_t *journal = transaction->t_journal;
1234 struct journal_head *jh;
1235 int drop_reserve = 0;
1236 int err = 0;
1237 int was_modified = 0;
1239 BUFFER_TRACE(bh, "entry");
1241 jbd_lock_bh_state(bh);
1242 spin_lock(&journal->j_list_lock);
1244 if (!buffer_jbd(bh))
1245 goto not_jbd;
1246 jh = bh2jh(bh);
1248 /* Critical error: attempting to delete a bitmap buffer, maybe?
1249 * Don't do any jbd operations, and return an error. */
1250 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1251 "inconsistent data on disk")) {
1252 err = -EIO;
1253 goto not_jbd;
1256 /* keep track of wether or not this transaction modified us */
1257 was_modified = jh->b_modified;
1260 * The buffer's going from the transaction, we must drop
1261 * all references -bzzz
1263 jh->b_modified = 0;
1265 if (jh->b_transaction == handle->h_transaction) {
1266 J_ASSERT_JH(jh, !jh->b_frozen_data);
1268 /* If we are forgetting a buffer which is already part
1269 * of this transaction, then we can just drop it from
1270 * the transaction immediately. */
1271 clear_buffer_dirty(bh);
1272 clear_buffer_jbddirty(bh);
1274 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1277 * we only want to drop a reference if this transaction
1278 * modified the buffer
1280 if (was_modified)
1281 drop_reserve = 1;
1284 * We are no longer going to journal this buffer.
1285 * However, the commit of this transaction is still
1286 * important to the buffer: the delete that we are now
1287 * processing might obsolete an old log entry, so by
1288 * committing, we can satisfy the buffer's checkpoint.
1290 * So, if we have a checkpoint on the buffer, we should
1291 * now refile the buffer on our BJ_Forget list so that
1292 * we know to remove the checkpoint after we commit.
1295 if (jh->b_cp_transaction) {
1296 __journal_temp_unlink_buffer(jh);
1297 __journal_file_buffer(jh, transaction, BJ_Forget);
1298 } else {
1299 __journal_unfile_buffer(jh);
1300 if (!buffer_jbd(bh)) {
1301 spin_unlock(&journal->j_list_lock);
1302 jbd_unlock_bh_state(bh);
1303 __bforget(bh);
1304 goto drop;
1307 } else if (jh->b_transaction) {
1308 J_ASSERT_JH(jh, (jh->b_transaction ==
1309 journal->j_committing_transaction));
1310 /* However, if the buffer is still owned by a prior
1311 * (committing) transaction, we can't drop it yet... */
1312 JBUFFER_TRACE(jh, "belongs to older transaction");
1313 /* ... but we CAN drop it from the new transaction if we
1314 * have also modified it since the original commit. */
1316 if (jh->b_next_transaction) {
1317 J_ASSERT(jh->b_next_transaction == transaction);
1318 jh->b_next_transaction = NULL;
1321 * only drop a reference if this transaction modified
1322 * the buffer
1324 if (was_modified)
1325 drop_reserve = 1;
1329 not_jbd:
1330 spin_unlock(&journal->j_list_lock);
1331 jbd_unlock_bh_state(bh);
1332 __brelse(bh);
1333 drop:
1334 if (drop_reserve) {
1335 /* no need to reserve log space for this block -bzzz */
1336 handle->h_buffer_credits++;
1338 return err;
1342 * int journal_stop() - complete a transaction
1343 * @handle: tranaction to complete.
1345 * All done for a particular handle.
1347 * There is not much action needed here. We just return any remaining
1348 * buffer credits to the transaction and remove the handle. The only
1349 * complication is that we need to start a commit operation if the
1350 * filesystem is marked for synchronous update.
1352 * journal_stop itself will not usually return an error, but it may
1353 * do so in unusual circumstances. In particular, expect it to
1354 * return -EIO if a journal_abort has been executed since the
1355 * transaction began.
1357 int journal_stop(handle_t *handle)
1359 transaction_t *transaction = handle->h_transaction;
1360 journal_t *journal = transaction->t_journal;
1361 int err;
1362 pid_t pid;
1364 J_ASSERT(journal_current_handle() == handle);
1366 if (is_handle_aborted(handle))
1367 err = -EIO;
1368 else {
1369 J_ASSERT(transaction->t_updates > 0);
1370 err = 0;
1373 if (--handle->h_ref > 0) {
1374 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1375 handle->h_ref);
1376 return err;
1379 jbd_debug(4, "Handle %p going down\n", handle);
1382 * Implement synchronous transaction batching. If the handle
1383 * was synchronous, don't force a commit immediately. Let's
1384 * yield and let another thread piggyback onto this transaction.
1385 * Keep doing that while new threads continue to arrive.
1386 * It doesn't cost much - we're about to run a commit and sleep
1387 * on IO anyway. Speeds up many-threaded, many-dir operations
1388 * by 30x or more...
1390 * We try and optimize the sleep time against what the underlying disk
1391 * can do, instead of having a static sleep time. This is useful for
1392 * the case where our storage is so fast that it is more optimal to go
1393 * ahead and force a flush and wait for the transaction to be committed
1394 * than it is to wait for an arbitrary amount of time for new writers to
1395 * join the transaction. We achieve this by measuring how long it takes
1396 * to commit a transaction, and compare it with how long this
1397 * transaction has been running, and if run time < commit time then we
1398 * sleep for the delta and commit. This greatly helps super fast disks
1399 * that would see slowdowns as more threads started doing fsyncs.
1401 * But don't do this if this process was the most recent one to
1402 * perform a synchronous write. We do this to detect the case where a
1403 * single process is doing a stream of sync writes. No point in waiting
1404 * for joiners in that case.
1406 pid = current->pid;
1407 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1408 u64 commit_time, trans_time;
1410 journal->j_last_sync_writer = pid;
1412 spin_lock(&journal->j_state_lock);
1413 commit_time = journal->j_average_commit_time;
1414 spin_unlock(&journal->j_state_lock);
1416 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1417 transaction->t_start_time));
1419 commit_time = min_t(u64, commit_time,
1420 1000*jiffies_to_usecs(1));
1422 if (trans_time < commit_time) {
1423 ktime_t expires = ktime_add_ns(ktime_get(),
1424 commit_time);
1425 set_current_state(TASK_UNINTERRUPTIBLE);
1426 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1430 if (handle->h_sync)
1431 transaction->t_synchronous_commit = 1;
1432 current->journal_info = NULL;
1433 spin_lock(&journal->j_state_lock);
1434 spin_lock(&transaction->t_handle_lock);
1435 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1436 transaction->t_updates--;
1437 if (!transaction->t_updates) {
1438 wake_up(&journal->j_wait_updates);
1439 if (journal->j_barrier_count)
1440 wake_up(&journal->j_wait_transaction_locked);
1444 * If the handle is marked SYNC, we need to set another commit
1445 * going! We also want to force a commit if the current
1446 * transaction is occupying too much of the log, or if the
1447 * transaction is too old now.
1449 if (handle->h_sync ||
1450 transaction->t_outstanding_credits >
1451 journal->j_max_transaction_buffers ||
1452 time_after_eq(jiffies, transaction->t_expires)) {
1453 /* Do this even for aborted journals: an abort still
1454 * completes the commit thread, it just doesn't write
1455 * anything to disk. */
1456 tid_t tid = transaction->t_tid;
1458 spin_unlock(&transaction->t_handle_lock);
1459 jbd_debug(2, "transaction too old, requesting commit for "
1460 "handle %p\n", handle);
1461 /* This is non-blocking */
1462 __log_start_commit(journal, transaction->t_tid);
1463 spin_unlock(&journal->j_state_lock);
1466 * Special case: JFS_SYNC synchronous updates require us
1467 * to wait for the commit to complete.
1469 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1470 err = log_wait_commit(journal, tid);
1471 } else {
1472 spin_unlock(&transaction->t_handle_lock);
1473 spin_unlock(&journal->j_state_lock);
1476 lock_map_release(&handle->h_lockdep_map);
1478 jbd_free_handle(handle);
1479 return err;
1483 * int journal_force_commit() - force any uncommitted transactions
1484 * @journal: journal to force
1486 * For synchronous operations: force any uncommitted transactions
1487 * to disk. May seem kludgy, but it reuses all the handle batching
1488 * code in a very simple manner.
1490 int journal_force_commit(journal_t *journal)
1492 handle_t *handle;
1493 int ret;
1495 handle = journal_start(journal, 1);
1496 if (IS_ERR(handle)) {
1497 ret = PTR_ERR(handle);
1498 } else {
1499 handle->h_sync = 1;
1500 ret = journal_stop(handle);
1502 return ret;
1507 * List management code snippets: various functions for manipulating the
1508 * transaction buffer lists.
1513 * Append a buffer to a transaction list, given the transaction's list head
1514 * pointer.
1516 * j_list_lock is held.
1518 * jbd_lock_bh_state(jh2bh(jh)) is held.
1521 static inline void
1522 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1524 if (!*list) {
1525 jh->b_tnext = jh->b_tprev = jh;
1526 *list = jh;
1527 } else {
1528 /* Insert at the tail of the list to preserve order */
1529 struct journal_head *first = *list, *last = first->b_tprev;
1530 jh->b_tprev = last;
1531 jh->b_tnext = first;
1532 last->b_tnext = first->b_tprev = jh;
1537 * Remove a buffer from a transaction list, given the transaction's list
1538 * head pointer.
1540 * Called with j_list_lock held, and the journal may not be locked.
1542 * jbd_lock_bh_state(jh2bh(jh)) is held.
1545 static inline void
1546 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1548 if (*list == jh) {
1549 *list = jh->b_tnext;
1550 if (*list == jh)
1551 *list = NULL;
1553 jh->b_tprev->b_tnext = jh->b_tnext;
1554 jh->b_tnext->b_tprev = jh->b_tprev;
1558 * Remove a buffer from the appropriate transaction list.
1560 * Note that this function can *change* the value of
1561 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1562 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1563 * is holding onto a copy of one of thee pointers, it could go bad.
1564 * Generally the caller needs to re-read the pointer from the transaction_t.
1566 * Called under j_list_lock. The journal may not be locked.
1568 static void __journal_temp_unlink_buffer(struct journal_head *jh)
1570 struct journal_head **list = NULL;
1571 transaction_t *transaction;
1572 struct buffer_head *bh = jh2bh(jh);
1574 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1575 transaction = jh->b_transaction;
1576 if (transaction)
1577 assert_spin_locked(&transaction->t_journal->j_list_lock);
1579 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1580 if (jh->b_jlist != BJ_None)
1581 J_ASSERT_JH(jh, transaction != NULL);
1583 switch (jh->b_jlist) {
1584 case BJ_None:
1585 return;
1586 case BJ_SyncData:
1587 list = &transaction->t_sync_datalist;
1588 break;
1589 case BJ_Metadata:
1590 transaction->t_nr_buffers--;
1591 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1592 list = &transaction->t_buffers;
1593 break;
1594 case BJ_Forget:
1595 list = &transaction->t_forget;
1596 break;
1597 case BJ_IO:
1598 list = &transaction->t_iobuf_list;
1599 break;
1600 case BJ_Shadow:
1601 list = &transaction->t_shadow_list;
1602 break;
1603 case BJ_LogCtl:
1604 list = &transaction->t_log_list;
1605 break;
1606 case BJ_Reserved:
1607 list = &transaction->t_reserved_list;
1608 break;
1609 case BJ_Locked:
1610 list = &transaction->t_locked_list;
1611 break;
1614 __blist_del_buffer(list, jh);
1615 jh->b_jlist = BJ_None;
1616 if (test_clear_buffer_jbddirty(bh))
1617 mark_buffer_dirty(bh); /* Expose it to the VM */
1621 * Remove buffer from all transactions.
1623 * Called with bh_state lock and j_list_lock
1625 * jh and bh may be already freed when this function returns.
1627 void __journal_unfile_buffer(struct journal_head *jh)
1629 __journal_temp_unlink_buffer(jh);
1630 jh->b_transaction = NULL;
1631 journal_put_journal_head(jh);
1634 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1636 struct buffer_head *bh = jh2bh(jh);
1638 /* Get reference so that buffer cannot be freed before we unlock it */
1639 get_bh(bh);
1640 jbd_lock_bh_state(bh);
1641 spin_lock(&journal->j_list_lock);
1642 __journal_unfile_buffer(jh);
1643 spin_unlock(&journal->j_list_lock);
1644 jbd_unlock_bh_state(bh);
1645 __brelse(bh);
1649 * Called from journal_try_to_free_buffers().
1651 * Called under jbd_lock_bh_state(bh)
1653 static void
1654 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1656 struct journal_head *jh;
1658 jh = bh2jh(bh);
1660 if (buffer_locked(bh) || buffer_dirty(bh))
1661 goto out;
1663 if (jh->b_next_transaction != NULL)
1664 goto out;
1666 spin_lock(&journal->j_list_lock);
1667 if (jh->b_transaction != NULL && jh->b_cp_transaction == NULL) {
1668 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1669 /* A written-back ordered data buffer */
1670 JBUFFER_TRACE(jh, "release data");
1671 __journal_unfile_buffer(jh);
1673 } else if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1674 /* written-back checkpointed metadata buffer */
1675 if (jh->b_jlist == BJ_None) {
1676 JBUFFER_TRACE(jh, "remove from checkpoint list");
1677 __journal_remove_checkpoint(jh);
1680 spin_unlock(&journal->j_list_lock);
1681 out:
1682 return;
1686 * int journal_try_to_free_buffers() - try to free page buffers.
1687 * @journal: journal for operation
1688 * @page: to try and free
1689 * @gfp_mask: we use the mask to detect how hard should we try to release
1690 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1691 * release the buffers.
1694 * For all the buffers on this page,
1695 * if they are fully written out ordered data, move them onto BUF_CLEAN
1696 * so try_to_free_buffers() can reap them.
1698 * This function returns non-zero if we wish try_to_free_buffers()
1699 * to be called. We do this if the page is releasable by try_to_free_buffers().
1700 * We also do it if the page has locked or dirty buffers and the caller wants
1701 * us to perform sync or async writeout.
1703 * This complicates JBD locking somewhat. We aren't protected by the
1704 * BKL here. We wish to remove the buffer from its committing or
1705 * running transaction's ->t_datalist via __journal_unfile_buffer.
1707 * This may *change* the value of transaction_t->t_datalist, so anyone
1708 * who looks at t_datalist needs to lock against this function.
1710 * Even worse, someone may be doing a journal_dirty_data on this
1711 * buffer. So we need to lock against that. journal_dirty_data()
1712 * will come out of the lock with the buffer dirty, which makes it
1713 * ineligible for release here.
1715 * Who else is affected by this? hmm... Really the only contender
1716 * is do_get_write_access() - it could be looking at the buffer while
1717 * journal_try_to_free_buffer() is changing its state. But that
1718 * cannot happen because we never reallocate freed data as metadata
1719 * while the data is part of a transaction. Yes?
1721 * Return 0 on failure, 1 on success
1723 int journal_try_to_free_buffers(journal_t *journal,
1724 struct page *page, gfp_t gfp_mask)
1726 struct buffer_head *head;
1727 struct buffer_head *bh;
1728 int ret = 0;
1730 J_ASSERT(PageLocked(page));
1732 head = page_buffers(page);
1733 bh = head;
1734 do {
1735 struct journal_head *jh;
1738 * We take our own ref against the journal_head here to avoid
1739 * having to add tons of locking around each instance of
1740 * journal_put_journal_head().
1742 jh = journal_grab_journal_head(bh);
1743 if (!jh)
1744 continue;
1746 jbd_lock_bh_state(bh);
1747 __journal_try_to_free_buffer(journal, bh);
1748 journal_put_journal_head(jh);
1749 jbd_unlock_bh_state(bh);
1750 if (buffer_jbd(bh))
1751 goto busy;
1752 } while ((bh = bh->b_this_page) != head);
1754 ret = try_to_free_buffers(page);
1756 busy:
1757 return ret;
1761 * This buffer is no longer needed. If it is on an older transaction's
1762 * checkpoint list we need to record it on this transaction's forget list
1763 * to pin this buffer (and hence its checkpointing transaction) down until
1764 * this transaction commits. If the buffer isn't on a checkpoint list, we
1765 * release it.
1766 * Returns non-zero if JBD no longer has an interest in the buffer.
1768 * Called under j_list_lock.
1770 * Called under jbd_lock_bh_state(bh).
1772 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1774 int may_free = 1;
1775 struct buffer_head *bh = jh2bh(jh);
1777 if (jh->b_cp_transaction) {
1778 JBUFFER_TRACE(jh, "on running+cp transaction");
1779 __journal_temp_unlink_buffer(jh);
1781 * We don't want to write the buffer anymore, clear the
1782 * bit so that we don't confuse checks in
1783 * __journal_file_buffer
1785 clear_buffer_dirty(bh);
1786 __journal_file_buffer(jh, transaction, BJ_Forget);
1787 may_free = 0;
1788 } else {
1789 JBUFFER_TRACE(jh, "on running transaction");
1790 __journal_unfile_buffer(jh);
1792 return may_free;
1796 * journal_invalidatepage
1798 * This code is tricky. It has a number of cases to deal with.
1800 * There are two invariants which this code relies on:
1802 * i_size must be updated on disk before we start calling invalidatepage on the
1803 * data.
1805 * This is done in ext3 by defining an ext3_setattr method which
1806 * updates i_size before truncate gets going. By maintaining this
1807 * invariant, we can be sure that it is safe to throw away any buffers
1808 * attached to the current transaction: once the transaction commits,
1809 * we know that the data will not be needed.
1811 * Note however that we can *not* throw away data belonging to the
1812 * previous, committing transaction!
1814 * Any disk blocks which *are* part of the previous, committing
1815 * transaction (and which therefore cannot be discarded immediately) are
1816 * not going to be reused in the new running transaction
1818 * The bitmap committed_data images guarantee this: any block which is
1819 * allocated in one transaction and removed in the next will be marked
1820 * as in-use in the committed_data bitmap, so cannot be reused until
1821 * the next transaction to delete the block commits. This means that
1822 * leaving committing buffers dirty is quite safe: the disk blocks
1823 * cannot be reallocated to a different file and so buffer aliasing is
1824 * not possible.
1827 * The above applies mainly to ordered data mode. In writeback mode we
1828 * don't make guarantees about the order in which data hits disk --- in
1829 * particular we don't guarantee that new dirty data is flushed before
1830 * transaction commit --- so it is always safe just to discard data
1831 * immediately in that mode. --sct
1835 * The journal_unmap_buffer helper function returns zero if the buffer
1836 * concerned remains pinned as an anonymous buffer belonging to an older
1837 * transaction.
1839 * We're outside-transaction here. Either or both of j_running_transaction
1840 * and j_committing_transaction may be NULL.
1842 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1844 transaction_t *transaction;
1845 struct journal_head *jh;
1846 int may_free = 1;
1847 int ret;
1849 BUFFER_TRACE(bh, "entry");
1852 * It is safe to proceed here without the j_list_lock because the
1853 * buffers cannot be stolen by try_to_free_buffers as long as we are
1854 * holding the page lock. --sct
1857 if (!buffer_jbd(bh))
1858 goto zap_buffer_unlocked;
1860 spin_lock(&journal->j_state_lock);
1861 jbd_lock_bh_state(bh);
1862 spin_lock(&journal->j_list_lock);
1864 jh = journal_grab_journal_head(bh);
1865 if (!jh)
1866 goto zap_buffer_no_jh;
1869 * We cannot remove the buffer from checkpoint lists until the
1870 * transaction adding inode to orphan list (let's call it T)
1871 * is committed. Otherwise if the transaction changing the
1872 * buffer would be cleaned from the journal before T is
1873 * committed, a crash will cause that the correct contents of
1874 * the buffer will be lost. On the other hand we have to
1875 * clear the buffer dirty bit at latest at the moment when the
1876 * transaction marking the buffer as freed in the filesystem
1877 * structures is committed because from that moment on the
1878 * buffer can be reallocated and used by a different page.
1879 * Since the block hasn't been freed yet but the inode has
1880 * already been added to orphan list, it is safe for us to add
1881 * the buffer to BJ_Forget list of the newest transaction.
1883 transaction = jh->b_transaction;
1884 if (transaction == NULL) {
1885 /* First case: not on any transaction. If it
1886 * has no checkpoint link, then we can zap it:
1887 * it's a writeback-mode buffer so we don't care
1888 * if it hits disk safely. */
1889 if (!jh->b_cp_transaction) {
1890 JBUFFER_TRACE(jh, "not on any transaction: zap");
1891 goto zap_buffer;
1894 if (!buffer_dirty(bh)) {
1895 /* bdflush has written it. We can drop it now */
1896 goto zap_buffer;
1899 /* OK, it must be in the journal but still not
1900 * written fully to disk: it's metadata or
1901 * journaled data... */
1903 if (journal->j_running_transaction) {
1904 /* ... and once the current transaction has
1905 * committed, the buffer won't be needed any
1906 * longer. */
1907 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1908 ret = __dispose_buffer(jh,
1909 journal->j_running_transaction);
1910 journal_put_journal_head(jh);
1911 spin_unlock(&journal->j_list_lock);
1912 jbd_unlock_bh_state(bh);
1913 spin_unlock(&journal->j_state_lock);
1914 return ret;
1915 } else {
1916 /* There is no currently-running transaction. So the
1917 * orphan record which we wrote for this file must have
1918 * passed into commit. We must attach this buffer to
1919 * the committing transaction, if it exists. */
1920 if (journal->j_committing_transaction) {
1921 JBUFFER_TRACE(jh, "give to committing trans");
1922 ret = __dispose_buffer(jh,
1923 journal->j_committing_transaction);
1924 journal_put_journal_head(jh);
1925 spin_unlock(&journal->j_list_lock);
1926 jbd_unlock_bh_state(bh);
1927 spin_unlock(&journal->j_state_lock);
1928 return ret;
1929 } else {
1930 /* The orphan record's transaction has
1931 * committed. We can cleanse this buffer */
1932 clear_buffer_jbddirty(bh);
1933 goto zap_buffer;
1936 } else if (transaction == journal->j_committing_transaction) {
1937 JBUFFER_TRACE(jh, "on committing transaction");
1938 if (jh->b_jlist == BJ_Locked) {
1940 * The buffer is on the committing transaction's locked
1941 * list. We have the buffer locked, so I/O has
1942 * completed. So we can nail the buffer now.
1944 may_free = __dispose_buffer(jh, transaction);
1945 goto zap_buffer;
1948 * The buffer is committing, we simply cannot touch
1949 * it. So we just set j_next_transaction to the
1950 * running transaction (if there is one) and mark
1951 * buffer as freed so that commit code knows it should
1952 * clear dirty bits when it is done with the buffer.
1954 set_buffer_freed(bh);
1955 if (journal->j_running_transaction && buffer_jbddirty(bh))
1956 jh->b_next_transaction = journal->j_running_transaction;
1957 journal_put_journal_head(jh);
1958 spin_unlock(&journal->j_list_lock);
1959 jbd_unlock_bh_state(bh);
1960 spin_unlock(&journal->j_state_lock);
1961 return 0;
1962 } else {
1963 /* Good, the buffer belongs to the running transaction.
1964 * We are writing our own transaction's data, not any
1965 * previous one's, so it is safe to throw it away
1966 * (remember that we expect the filesystem to have set
1967 * i_size already for this truncate so recovery will not
1968 * expose the disk blocks we are discarding here.) */
1969 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1970 JBUFFER_TRACE(jh, "on running transaction");
1971 may_free = __dispose_buffer(jh, transaction);
1974 zap_buffer:
1975 journal_put_journal_head(jh);
1976 zap_buffer_no_jh:
1977 spin_unlock(&journal->j_list_lock);
1978 jbd_unlock_bh_state(bh);
1979 spin_unlock(&journal->j_state_lock);
1980 zap_buffer_unlocked:
1981 clear_buffer_dirty(bh);
1982 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1983 clear_buffer_mapped(bh);
1984 clear_buffer_req(bh);
1985 clear_buffer_new(bh);
1986 bh->b_bdev = NULL;
1987 return may_free;
1991 * void journal_invalidatepage() - invalidate a journal page
1992 * @journal: journal to use for flush
1993 * @page: page to flush
1994 * @offset: length of page to invalidate.
1996 * Reap page buffers containing data after offset in page.
1998 void journal_invalidatepage(journal_t *journal,
1999 struct page *page,
2000 unsigned long offset)
2002 struct buffer_head *head, *bh, *next;
2003 unsigned int curr_off = 0;
2004 int may_free = 1;
2006 if (!PageLocked(page))
2007 BUG();
2008 if (!page_has_buffers(page))
2009 return;
2011 /* We will potentially be playing with lists other than just the
2012 * data lists (especially for journaled data mode), so be
2013 * cautious in our locking. */
2015 head = bh = page_buffers(page);
2016 do {
2017 unsigned int next_off = curr_off + bh->b_size;
2018 next = bh->b_this_page;
2020 if (offset <= curr_off) {
2021 /* This block is wholly outside the truncation point */
2022 lock_buffer(bh);
2023 may_free &= journal_unmap_buffer(journal, bh);
2024 unlock_buffer(bh);
2026 curr_off = next_off;
2027 bh = next;
2029 } while (bh != head);
2031 if (!offset) {
2032 if (may_free && try_to_free_buffers(page))
2033 J_ASSERT(!page_has_buffers(page));
2038 * File a buffer on the given transaction list.
2040 void __journal_file_buffer(struct journal_head *jh,
2041 transaction_t *transaction, int jlist)
2043 struct journal_head **list = NULL;
2044 int was_dirty = 0;
2045 struct buffer_head *bh = jh2bh(jh);
2047 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2048 assert_spin_locked(&transaction->t_journal->j_list_lock);
2050 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2051 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2052 jh->b_transaction == NULL);
2054 if (jh->b_transaction && jh->b_jlist == jlist)
2055 return;
2057 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2058 jlist == BJ_Shadow || jlist == BJ_Forget) {
2060 * For metadata buffers, we track dirty bit in buffer_jbddirty
2061 * instead of buffer_dirty. We should not see a dirty bit set
2062 * here because we clear it in do_get_write_access but e.g.
2063 * tune2fs can modify the sb and set the dirty bit at any time
2064 * so we try to gracefully handle that.
2066 if (buffer_dirty(bh))
2067 warn_dirty_buffer(bh);
2068 if (test_clear_buffer_dirty(bh) ||
2069 test_clear_buffer_jbddirty(bh))
2070 was_dirty = 1;
2073 if (jh->b_transaction)
2074 __journal_temp_unlink_buffer(jh);
2075 else
2076 journal_grab_journal_head(bh);
2077 jh->b_transaction = transaction;
2079 switch (jlist) {
2080 case BJ_None:
2081 J_ASSERT_JH(jh, !jh->b_committed_data);
2082 J_ASSERT_JH(jh, !jh->b_frozen_data);
2083 return;
2084 case BJ_SyncData:
2085 list = &transaction->t_sync_datalist;
2086 break;
2087 case BJ_Metadata:
2088 transaction->t_nr_buffers++;
2089 list = &transaction->t_buffers;
2090 break;
2091 case BJ_Forget:
2092 list = &transaction->t_forget;
2093 break;
2094 case BJ_IO:
2095 list = &transaction->t_iobuf_list;
2096 break;
2097 case BJ_Shadow:
2098 list = &transaction->t_shadow_list;
2099 break;
2100 case BJ_LogCtl:
2101 list = &transaction->t_log_list;
2102 break;
2103 case BJ_Reserved:
2104 list = &transaction->t_reserved_list;
2105 break;
2106 case BJ_Locked:
2107 list = &transaction->t_locked_list;
2108 break;
2111 __blist_add_buffer(list, jh);
2112 jh->b_jlist = jlist;
2114 if (was_dirty)
2115 set_buffer_jbddirty(bh);
2118 void journal_file_buffer(struct journal_head *jh,
2119 transaction_t *transaction, int jlist)
2121 jbd_lock_bh_state(jh2bh(jh));
2122 spin_lock(&transaction->t_journal->j_list_lock);
2123 __journal_file_buffer(jh, transaction, jlist);
2124 spin_unlock(&transaction->t_journal->j_list_lock);
2125 jbd_unlock_bh_state(jh2bh(jh));
2129 * Remove a buffer from its current buffer list in preparation for
2130 * dropping it from its current transaction entirely. If the buffer has
2131 * already started to be used by a subsequent transaction, refile the
2132 * buffer on that transaction's metadata list.
2134 * Called under j_list_lock
2135 * Called under jbd_lock_bh_state(jh2bh(jh))
2137 * jh and bh may be already free when this function returns
2139 void __journal_refile_buffer(struct journal_head *jh)
2141 int was_dirty, jlist;
2142 struct buffer_head *bh = jh2bh(jh);
2144 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2145 if (jh->b_transaction)
2146 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2148 /* If the buffer is now unused, just drop it. */
2149 if (jh->b_next_transaction == NULL) {
2150 __journal_unfile_buffer(jh);
2151 return;
2155 * It has been modified by a later transaction: add it to the new
2156 * transaction's metadata list.
2159 was_dirty = test_clear_buffer_jbddirty(bh);
2160 __journal_temp_unlink_buffer(jh);
2162 * We set b_transaction here because b_next_transaction will inherit
2163 * our jh reference and thus __journal_file_buffer() must not take a
2164 * new one.
2166 jh->b_transaction = jh->b_next_transaction;
2167 jh->b_next_transaction = NULL;
2168 if (buffer_freed(bh))
2169 jlist = BJ_Forget;
2170 else if (jh->b_modified)
2171 jlist = BJ_Metadata;
2172 else
2173 jlist = BJ_Reserved;
2174 __journal_file_buffer(jh, jh->b_transaction, jlist);
2175 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2177 if (was_dirty)
2178 set_buffer_jbddirty(bh);
2182 * __journal_refile_buffer() with necessary locking added. We take our bh
2183 * reference so that we can safely unlock bh.
2185 * The jh and bh may be freed by this call.
2187 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2189 struct buffer_head *bh = jh2bh(jh);
2191 /* Get reference so that buffer cannot be freed before we unlock it */
2192 get_bh(bh);
2193 jbd_lock_bh_state(bh);
2194 spin_lock(&journal->j_list_lock);
2195 __journal_refile_buffer(jh);
2196 jbd_unlock_bh_state(bh);
2197 spin_unlock(&journal->j_list_lock);
2198 __brelse(bh);