2 * linux/fs/jbd2/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
15 * This file manages transactions (compound commits managed by the
16 * journaling code) and handles (individual atomic operations by the
20 #include <linux/time.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29 #include <linux/backing-dev.h>
30 #include <linux/module.h>
32 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
35 * jbd2_get_transaction: obtain a new transaction_t object.
37 * Simply allocate and initialise a new transaction. Create it in
38 * RUNNING state and add it to the current journal (which should not
39 * have an existing running transaction: we only make a new transaction
40 * once we have started to commit the old one).
43 * The journal MUST be locked. We don't perform atomic mallocs on the
44 * new transaction and we can't block without protecting against other
45 * processes trying to touch the journal while it is in transition.
49 static transaction_t *
50 jbd2_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);
58 atomic_set(&transaction->t_updates, 0);
59 atomic_set(&transaction->t_outstanding_credits, 0);
60 INIT_LIST_HEAD(&transaction->t_inode_list);
61 INIT_LIST_HEAD(&transaction->t_private_list);
63 /* Set up the commit timer for the new transaction. */
64 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
65 add_timer(&journal->j_commit_timer);
67 J_ASSERT(journal->j_running_transaction == NULL);
68 journal->j_running_transaction = transaction;
69 transaction->t_max_wait = 0;
70 transaction->t_start = jiffies;
78 * A handle_t is an object which represents a single atomic update to a
79 * filesystem, and which tracks all of the modifications which form part
84 * start_this_handle: Given a handle, deal with any locking or stalling
85 * needed to make sure that there is enough journal space for the handle
86 * to begin. Attach the handle to a transaction and set up the
87 * transaction's buffer credits.
90 static int start_this_handle(journal_t *journal, handle_t *handle,
93 transaction_t *transaction;
95 int nblocks = handle->h_buffer_credits;
96 transaction_t *new_transaction = NULL;
97 unsigned long ts = jiffies;
99 if (nblocks > journal->j_max_transaction_buffers) {
100 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
101 current->comm, nblocks,
102 journal->j_max_transaction_buffers);
107 if (!journal->j_running_transaction) {
108 new_transaction = kzalloc(sizeof(*new_transaction), gfp_mask);
109 if (!new_transaction) {
111 * If __GFP_FS is not present, then we may be
112 * being called from inside the fs writeback
113 * layer, so we MUST NOT fail. Since
114 * __GFP_NOFAIL is going away, we will arrange
115 * to retry the allocation ourselves.
117 if ((gfp_mask & __GFP_FS) == 0) {
118 congestion_wait(BLK_RW_ASYNC, HZ/50);
119 goto alloc_transaction;
125 jbd_debug(3, "New handle %p going live.\n", handle);
128 * We need to hold j_state_lock until t_updates has been incremented,
129 * for proper journal barrier handling
132 read_lock(&journal->j_state_lock);
133 if (is_journal_aborted(journal) ||
134 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
135 read_unlock(&journal->j_state_lock);
136 kfree(new_transaction);
140 /* Wait on the journal's transaction barrier if necessary */
141 if (journal->j_barrier_count) {
142 read_unlock(&journal->j_state_lock);
143 wait_event(journal->j_wait_transaction_locked,
144 journal->j_barrier_count == 0);
148 if (!journal->j_running_transaction) {
149 read_unlock(&journal->j_state_lock);
150 if (!new_transaction)
151 goto alloc_transaction;
152 write_lock(&journal->j_state_lock);
153 if (!journal->j_running_transaction) {
154 jbd2_get_transaction(journal, new_transaction);
155 new_transaction = NULL;
157 write_unlock(&journal->j_state_lock);
161 transaction = journal->j_running_transaction;
164 * If the current transaction is locked down for commit, wait for the
165 * lock to be released.
167 if (transaction->t_state == T_LOCKED) {
170 prepare_to_wait(&journal->j_wait_transaction_locked,
171 &wait, TASK_UNINTERRUPTIBLE);
172 read_unlock(&journal->j_state_lock);
174 finish_wait(&journal->j_wait_transaction_locked, &wait);
179 * If there is not enough space left in the log to write all potential
180 * buffers requested by this operation, we need to stall pending a log
181 * checkpoint to free some more log space.
183 spin_lock(&transaction->t_handle_lock);
184 needed = atomic_read(&transaction->t_outstanding_credits) + nblocks;
186 if (needed > journal->j_max_transaction_buffers) {
188 * If the current transaction is already too large, then start
189 * to commit it: we can then go back and attach this handle to
194 jbd_debug(2, "Handle %p starting new commit...\n", handle);
195 spin_unlock(&transaction->t_handle_lock);
196 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
197 TASK_UNINTERRUPTIBLE);
198 __jbd2_log_start_commit(journal, transaction->t_tid);
199 read_unlock(&journal->j_state_lock);
201 finish_wait(&journal->j_wait_transaction_locked, &wait);
206 * The commit code assumes that it can get enough log space
207 * without forcing a checkpoint. This is *critical* for
208 * correctness: a checkpoint of a buffer which is also
209 * associated with a committing transaction creates a deadlock,
210 * so commit simply cannot force through checkpoints.
212 * We must therefore ensure the necessary space in the journal
213 * *before* starting to dirty potentially checkpointed buffers
214 * in the new transaction.
216 * The worst part is, any transaction currently committing can
217 * reduce the free space arbitrarily. Be careful to account for
218 * those buffers when checkpointing.
222 * @@@ AKPM: This seems rather over-defensive. We're giving commit
223 * a _lot_ of headroom: 1/4 of the journal plus the size of
224 * the committing transaction. Really, we only need to give it
225 * committing_transaction->t_outstanding_credits plus "enough" for
226 * the log control blocks.
227 * Also, this test is inconsitent with the matching one in
228 * jbd2_journal_extend().
230 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
231 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
232 spin_unlock(&transaction->t_handle_lock);
233 read_unlock(&journal->j_state_lock);
234 write_lock(&journal->j_state_lock);
235 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
236 __jbd2_log_wait_for_space(journal);
237 write_unlock(&journal->j_state_lock);
241 /* OK, account for the buffers that this operation expects to
242 * use and add the handle to the running transaction. */
244 if (time_after(transaction->t_start, ts)) {
245 ts = jbd2_time_diff(ts, transaction->t_start);
246 if (ts > transaction->t_max_wait)
247 transaction->t_max_wait = ts;
250 handle->h_transaction = transaction;
251 atomic_add(nblocks, &transaction->t_outstanding_credits);
252 atomic_inc(&transaction->t_updates);
253 transaction->t_handle_count++;
254 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
256 atomic_read(&transaction->t_outstanding_credits),
257 __jbd2_log_space_left(journal));
258 spin_unlock(&transaction->t_handle_lock);
259 read_unlock(&journal->j_state_lock);
261 lock_map_acquire(&handle->h_lockdep_map);
262 kfree(new_transaction);
266 static struct lock_class_key jbd2_handle_key;
268 /* Allocate a new handle. This should probably be in a slab... */
269 static handle_t *new_handle(int nblocks)
271 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
274 memset(handle, 0, sizeof(*handle));
275 handle->h_buffer_credits = nblocks;
278 lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
279 &jbd2_handle_key, 0);
285 * handle_t *jbd2_journal_start() - Obtain a new handle.
286 * @journal: Journal to start transaction on.
287 * @nblocks: number of block buffer we might modify
289 * We make sure that the transaction can guarantee at least nblocks of
290 * modified buffers in the log. We block until the log can guarantee
293 * This function is visible to journal users (like ext3fs), so is not
294 * called with the journal already locked.
296 * Return a pointer to a newly allocated handle, or NULL on failure
298 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int gfp_mask)
300 handle_t *handle = journal_current_handle();
304 return ERR_PTR(-EROFS);
307 J_ASSERT(handle->h_transaction->t_journal == journal);
312 handle = new_handle(nblocks);
314 return ERR_PTR(-ENOMEM);
316 current->journal_info = handle;
318 err = start_this_handle(journal, handle, gfp_mask);
320 jbd2_free_handle(handle);
321 current->journal_info = NULL;
322 handle = ERR_PTR(err);
328 EXPORT_SYMBOL(jbd2__journal_start);
331 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
333 return jbd2__journal_start(journal, nblocks, GFP_NOFS);
335 EXPORT_SYMBOL(jbd2_journal_start);
339 * int jbd2_journal_extend() - extend buffer credits.
340 * @handle: handle to 'extend'
341 * @nblocks: nr blocks to try to extend by.
343 * Some transactions, such as large extends and truncates, can be done
344 * atomically all at once or in several stages. The operation requests
345 * a credit for a number of buffer modications in advance, but can
346 * extend its credit if it needs more.
348 * jbd2_journal_extend tries to give the running handle more buffer credits.
349 * It does not guarantee that allocation - this is a best-effort only.
350 * The calling process MUST be able to deal cleanly with a failure to
353 * Return 0 on success, non-zero on failure.
355 * return code < 0 implies an error
356 * return code > 0 implies normal transaction-full status.
358 int jbd2_journal_extend(handle_t *handle, int nblocks)
360 transaction_t *transaction = handle->h_transaction;
361 journal_t *journal = transaction->t_journal;
366 if (is_handle_aborted(handle))
371 read_lock(&journal->j_state_lock);
373 /* Don't extend a locked-down transaction! */
374 if (handle->h_transaction->t_state != T_RUNNING) {
375 jbd_debug(3, "denied handle %p %d blocks: "
376 "transaction not running\n", handle, nblocks);
380 spin_lock(&transaction->t_handle_lock);
381 wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
383 if (wanted > journal->j_max_transaction_buffers) {
384 jbd_debug(3, "denied handle %p %d blocks: "
385 "transaction too large\n", handle, nblocks);
389 if (wanted > __jbd2_log_space_left(journal)) {
390 jbd_debug(3, "denied handle %p %d blocks: "
391 "insufficient log space\n", handle, nblocks);
395 handle->h_buffer_credits += nblocks;
396 atomic_add(nblocks, &transaction->t_outstanding_credits);
399 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
401 spin_unlock(&transaction->t_handle_lock);
403 read_unlock(&journal->j_state_lock);
410 * int jbd2_journal_restart() - restart a handle .
411 * @handle: handle to restart
412 * @nblocks: nr credits requested
414 * Restart a handle for a multi-transaction filesystem
417 * If the jbd2_journal_extend() call above fails to grant new buffer credits
418 * to a running handle, a call to jbd2_journal_restart will commit the
419 * handle's transaction so far and reattach the handle to a new
420 * transaction capabable of guaranteeing the requested number of
423 int jbd2__journal_restart(handle_t *handle, int nblocks, int gfp_mask)
425 transaction_t *transaction = handle->h_transaction;
426 journal_t *journal = transaction->t_journal;
429 /* If we've had an abort of any type, don't even think about
430 * actually doing the restart! */
431 if (is_handle_aborted(handle))
435 * First unlink the handle from its current transaction, and start the
438 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
439 J_ASSERT(journal_current_handle() == handle);
441 read_lock(&journal->j_state_lock);
442 spin_lock(&transaction->t_handle_lock);
443 atomic_sub(handle->h_buffer_credits,
444 &transaction->t_outstanding_credits);
445 if (atomic_dec_and_test(&transaction->t_updates))
446 wake_up(&journal->j_wait_updates);
447 spin_unlock(&transaction->t_handle_lock);
449 jbd_debug(2, "restarting handle %p\n", handle);
450 __jbd2_log_start_commit(journal, transaction->t_tid);
451 read_unlock(&journal->j_state_lock);
453 lock_map_release(&handle->h_lockdep_map);
454 handle->h_buffer_credits = nblocks;
455 ret = start_this_handle(journal, handle, gfp_mask);
458 EXPORT_SYMBOL(jbd2__journal_restart);
461 int jbd2_journal_restart(handle_t *handle, int nblocks)
463 return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
465 EXPORT_SYMBOL(jbd2_journal_restart);
468 * void jbd2_journal_lock_updates () - establish a transaction barrier.
469 * @journal: Journal to establish a barrier on.
471 * This locks out any further updates from being started, and blocks
472 * until all existing updates have completed, returning only once the
473 * journal is in a quiescent state with no updates running.
475 * The journal lock should not be held on entry.
477 void jbd2_journal_lock_updates(journal_t *journal)
481 write_lock(&journal->j_state_lock);
482 ++journal->j_barrier_count;
484 /* Wait until there are no running updates */
486 transaction_t *transaction = journal->j_running_transaction;
491 spin_lock(&transaction->t_handle_lock);
492 if (!atomic_read(&transaction->t_updates)) {
493 spin_unlock(&transaction->t_handle_lock);
496 prepare_to_wait(&journal->j_wait_updates, &wait,
497 TASK_UNINTERRUPTIBLE);
498 spin_unlock(&transaction->t_handle_lock);
499 write_unlock(&journal->j_state_lock);
501 finish_wait(&journal->j_wait_updates, &wait);
502 write_lock(&journal->j_state_lock);
504 write_unlock(&journal->j_state_lock);
507 * We have now established a barrier against other normal updates, but
508 * we also need to barrier against other jbd2_journal_lock_updates() calls
509 * to make sure that we serialise special journal-locked operations
512 mutex_lock(&journal->j_barrier);
516 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
517 * @journal: Journal to release the barrier on.
519 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
521 * Should be called without the journal lock held.
523 void jbd2_journal_unlock_updates (journal_t *journal)
525 J_ASSERT(journal->j_barrier_count != 0);
527 mutex_unlock(&journal->j_barrier);
528 write_lock(&journal->j_state_lock);
529 --journal->j_barrier_count;
530 write_unlock(&journal->j_state_lock);
531 wake_up(&journal->j_wait_transaction_locked);
534 static void warn_dirty_buffer(struct buffer_head *bh)
536 char b[BDEVNAME_SIZE];
539 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
540 "There's a risk of filesystem corruption in case of system "
542 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
546 * If the buffer is already part of the current transaction, then there
547 * is nothing we need to do. If it is already part of a prior
548 * transaction which we are still committing to disk, then we need to
549 * make sure that we do not overwrite the old copy: we do copy-out to
550 * preserve the copy going to disk. We also account the buffer against
551 * the handle's metadata buffer credits (unless the buffer is already
552 * part of the transaction, that is).
556 do_get_write_access(handle_t *handle, struct journal_head *jh,
559 struct buffer_head *bh;
560 transaction_t *transaction;
563 char *frozen_buffer = NULL;
566 if (is_handle_aborted(handle))
569 transaction = handle->h_transaction;
570 journal = transaction->t_journal;
572 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
574 JBUFFER_TRACE(jh, "entry");
578 /* @@@ Need to check for errors here at some point. */
581 jbd_lock_bh_state(bh);
583 /* We now hold the buffer lock so it is safe to query the buffer
584 * state. Is the buffer dirty?
586 * If so, there are two possibilities. The buffer may be
587 * non-journaled, and undergoing a quite legitimate writeback.
588 * Otherwise, it is journaled, and we don't expect dirty buffers
589 * in that state (the buffers should be marked JBD_Dirty
590 * instead.) So either the IO is being done under our own
591 * control and this is a bug, or it's a third party IO such as
592 * dump(8) (which may leave the buffer scheduled for read ---
593 * ie. locked but not dirty) or tune2fs (which may actually have
594 * the buffer dirtied, ugh.) */
596 if (buffer_dirty(bh)) {
598 * First question: is this buffer already part of the current
599 * transaction or the existing committing transaction?
601 if (jh->b_transaction) {
603 jh->b_transaction == transaction ||
605 journal->j_committing_transaction);
606 if (jh->b_next_transaction)
607 J_ASSERT_JH(jh, jh->b_next_transaction ==
609 warn_dirty_buffer(bh);
612 * In any case we need to clean the dirty flag and we must
613 * do it under the buffer lock to be sure we don't race
614 * with running write-out.
616 JBUFFER_TRACE(jh, "Journalling dirty buffer");
617 clear_buffer_dirty(bh);
618 set_buffer_jbddirty(bh);
624 if (is_handle_aborted(handle)) {
625 jbd_unlock_bh_state(bh);
631 * The buffer is already part of this transaction if b_transaction or
632 * b_next_transaction points to it
634 if (jh->b_transaction == transaction ||
635 jh->b_next_transaction == transaction)
639 * this is the first time this transaction is touching this buffer,
640 * reset the modified flag
645 * If there is already a copy-out version of this buffer, then we don't
646 * need to make another one
648 if (jh->b_frozen_data) {
649 JBUFFER_TRACE(jh, "has frozen data");
650 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
651 jh->b_next_transaction = transaction;
655 /* Is there data here we need to preserve? */
657 if (jh->b_transaction && jh->b_transaction != transaction) {
658 JBUFFER_TRACE(jh, "owned by older transaction");
659 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
660 J_ASSERT_JH(jh, jh->b_transaction ==
661 journal->j_committing_transaction);
663 /* There is one case we have to be very careful about.
664 * If the committing transaction is currently writing
665 * this buffer out to disk and has NOT made a copy-out,
666 * then we cannot modify the buffer contents at all
667 * right now. The essence of copy-out is that it is the
668 * extra copy, not the primary copy, which gets
669 * journaled. If the primary copy is already going to
670 * disk then we cannot do copy-out here. */
672 if (jh->b_jlist == BJ_Shadow) {
673 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
674 wait_queue_head_t *wqh;
676 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
678 JBUFFER_TRACE(jh, "on shadow: sleep");
679 jbd_unlock_bh_state(bh);
680 /* commit wakes up all shadow buffers after IO */
682 prepare_to_wait(wqh, &wait.wait,
683 TASK_UNINTERRUPTIBLE);
684 if (jh->b_jlist != BJ_Shadow)
688 finish_wait(wqh, &wait.wait);
692 /* Only do the copy if the currently-owning transaction
693 * still needs it. If it is on the Forget list, the
694 * committing transaction is past that stage. The
695 * buffer had better remain locked during the kmalloc,
696 * but that should be true --- we hold the journal lock
697 * still and the buffer is already on the BUF_JOURNAL
698 * list so won't be flushed.
700 * Subtle point, though: if this is a get_undo_access,
701 * then we will be relying on the frozen_data to contain
702 * the new value of the committed_data record after the
703 * transaction, so we HAVE to force the frozen_data copy
706 if (jh->b_jlist != BJ_Forget || force_copy) {
707 JBUFFER_TRACE(jh, "generate frozen data");
708 if (!frozen_buffer) {
709 JBUFFER_TRACE(jh, "allocate memory for buffer");
710 jbd_unlock_bh_state(bh);
712 jbd2_alloc(jh2bh(jh)->b_size,
714 if (!frozen_buffer) {
716 "%s: OOM for frozen_buffer\n",
718 JBUFFER_TRACE(jh, "oom!");
720 jbd_lock_bh_state(bh);
725 jh->b_frozen_data = frozen_buffer;
726 frozen_buffer = NULL;
729 jh->b_next_transaction = transaction;
734 * Finally, if the buffer is not journaled right now, we need to make
735 * sure it doesn't get written to disk before the caller actually
736 * commits the new data
738 if (!jh->b_transaction) {
739 JBUFFER_TRACE(jh, "no transaction");
740 J_ASSERT_JH(jh, !jh->b_next_transaction);
741 jh->b_transaction = transaction;
742 JBUFFER_TRACE(jh, "file as BJ_Reserved");
743 spin_lock(&journal->j_list_lock);
744 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
745 spin_unlock(&journal->j_list_lock);
754 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
755 "Possible IO failure.\n");
756 page = jh2bh(jh)->b_page;
757 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
758 source = kmap_atomic(page, KM_USER0);
759 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
760 kunmap_atomic(source, KM_USER0);
763 * Now that the frozen data is saved off, we need to store
764 * any matching triggers.
766 jh->b_frozen_triggers = jh->b_triggers;
768 jbd_unlock_bh_state(bh);
771 * If we are about to journal a buffer, then any revoke pending on it is
774 jbd2_journal_cancel_revoke(handle, jh);
777 if (unlikely(frozen_buffer)) /* It's usually NULL */
778 jbd2_free(frozen_buffer, bh->b_size);
780 JBUFFER_TRACE(jh, "exit");
785 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
786 * @handle: transaction to add buffer modifications to
787 * @bh: bh to be used for metadata writes
788 * @credits: variable that will receive credits for the buffer
790 * Returns an error code or 0 on success.
792 * In full data journalling mode the buffer may be of type BJ_AsyncData,
793 * because we're write()ing a buffer which is also part of a shared mapping.
796 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
798 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
801 /* We do not want to get caught playing with fields which the
802 * log thread also manipulates. Make sure that the buffer
803 * completes any outstanding IO before proceeding. */
804 rc = do_get_write_access(handle, jh, 0);
805 jbd2_journal_put_journal_head(jh);
811 * When the user wants to journal a newly created buffer_head
812 * (ie. getblk() returned a new buffer and we are going to populate it
813 * manually rather than reading off disk), then we need to keep the
814 * buffer_head locked until it has been completely filled with new
815 * data. In this case, we should be able to make the assertion that
816 * the bh is not already part of an existing transaction.
818 * The buffer should already be locked by the caller by this point.
819 * There is no lock ranking violation: it was a newly created,
820 * unlocked buffer beforehand. */
823 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
824 * @handle: transaction to new buffer to
827 * Call this if you create a new bh.
829 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
831 transaction_t *transaction = handle->h_transaction;
832 journal_t *journal = transaction->t_journal;
833 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
836 jbd_debug(5, "journal_head %p\n", jh);
838 if (is_handle_aborted(handle))
842 JBUFFER_TRACE(jh, "entry");
844 * The buffer may already belong to this transaction due to pre-zeroing
845 * in the filesystem's new_block code. It may also be on the previous,
846 * committing transaction's lists, but it HAS to be in Forget state in
847 * that case: the transaction must have deleted the buffer for it to be
850 jbd_lock_bh_state(bh);
851 spin_lock(&journal->j_list_lock);
852 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
853 jh->b_transaction == NULL ||
854 (jh->b_transaction == journal->j_committing_transaction &&
855 jh->b_jlist == BJ_Forget)));
857 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
858 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
860 if (jh->b_transaction == NULL) {
862 * Previous jbd2_journal_forget() could have left the buffer
863 * with jbddirty bit set because it was being committed. When
864 * the commit finished, we've filed the buffer for
865 * checkpointing and marked it dirty. Now we are reallocating
866 * the buffer so the transaction freeing it must have
867 * committed and so it's safe to clear the dirty bit.
869 clear_buffer_dirty(jh2bh(jh));
870 jh->b_transaction = transaction;
872 /* first access by this transaction */
875 JBUFFER_TRACE(jh, "file as BJ_Reserved");
876 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
877 } else if (jh->b_transaction == journal->j_committing_transaction) {
878 /* first access by this transaction */
881 JBUFFER_TRACE(jh, "set next transaction");
882 jh->b_next_transaction = transaction;
884 spin_unlock(&journal->j_list_lock);
885 jbd_unlock_bh_state(bh);
888 * akpm: I added this. ext3_alloc_branch can pick up new indirect
889 * blocks which contain freed but then revoked metadata. We need
890 * to cancel the revoke in case we end up freeing it yet again
891 * and the reallocating as data - this would cause a second revoke,
892 * which hits an assertion error.
894 JBUFFER_TRACE(jh, "cancelling revoke");
895 jbd2_journal_cancel_revoke(handle, jh);
896 jbd2_journal_put_journal_head(jh);
902 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
903 * non-rewindable consequences
904 * @handle: transaction
905 * @bh: buffer to undo
906 * @credits: store the number of taken credits here (if not NULL)
908 * Sometimes there is a need to distinguish between metadata which has
909 * been committed to disk and that which has not. The ext3fs code uses
910 * this for freeing and allocating space, we have to make sure that we
911 * do not reuse freed space until the deallocation has been committed,
912 * since if we overwrote that space we would make the delete
913 * un-rewindable in case of a crash.
915 * To deal with that, jbd2_journal_get_undo_access requests write access to a
916 * buffer for parts of non-rewindable operations such as delete
917 * operations on the bitmaps. The journaling code must keep a copy of
918 * the buffer's contents prior to the undo_access call until such time
919 * as we know that the buffer has definitely been committed to disk.
921 * We never need to know which transaction the committed data is part
922 * of, buffers touched here are guaranteed to be dirtied later and so
923 * will be committed to a new transaction in due course, at which point
924 * we can discard the old committed data pointer.
926 * Returns error number or 0 on success.
928 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
931 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
932 char *committed_data = NULL;
934 JBUFFER_TRACE(jh, "entry");
937 * Do this first --- it can drop the journal lock, so we want to
938 * make sure that obtaining the committed_data is done
939 * atomically wrt. completion of any outstanding commits.
941 err = do_get_write_access(handle, jh, 1);
946 if (!jh->b_committed_data) {
947 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
948 if (!committed_data) {
949 printk(KERN_EMERG "%s: No memory for committed data\n",
956 jbd_lock_bh_state(bh);
957 if (!jh->b_committed_data) {
958 /* Copy out the current buffer contents into the
959 * preserved, committed copy. */
960 JBUFFER_TRACE(jh, "generate b_committed data");
961 if (!committed_data) {
962 jbd_unlock_bh_state(bh);
966 jh->b_committed_data = committed_data;
967 committed_data = NULL;
968 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
970 jbd_unlock_bh_state(bh);
972 jbd2_journal_put_journal_head(jh);
973 if (unlikely(committed_data))
974 jbd2_free(committed_data, bh->b_size);
979 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
980 * @bh: buffer to trigger on
981 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
983 * Set any triggers on this journal_head. This is always safe, because
984 * triggers for a committing buffer will be saved off, and triggers for
985 * a running transaction will match the buffer in that transaction.
987 * Call with NULL to clear the triggers.
989 void jbd2_journal_set_triggers(struct buffer_head *bh,
990 struct jbd2_buffer_trigger_type *type)
992 struct journal_head *jh = bh2jh(bh);
994 jh->b_triggers = type;
997 void jbd2_buffer_commit_trigger(struct journal_head *jh, void *mapped_data,
998 struct jbd2_buffer_trigger_type *triggers)
1000 struct buffer_head *bh = jh2bh(jh);
1002 if (!triggers || !triggers->t_commit)
1005 triggers->t_commit(triggers, bh, mapped_data, bh->b_size);
1008 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1009 struct jbd2_buffer_trigger_type *triggers)
1011 if (!triggers || !triggers->t_abort)
1014 triggers->t_abort(triggers, jh2bh(jh));
1020 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1021 * @handle: transaction to add buffer to.
1022 * @bh: buffer to mark
1024 * mark dirty metadata which needs to be journaled as part of the current
1027 * The buffer is placed on the transaction's metadata list and is marked
1028 * as belonging to the transaction.
1030 * Returns error number or 0 on success.
1032 * Special care needs to be taken if the buffer already belongs to the
1033 * current committing transaction (in which case we should have frozen
1034 * data present for that commit). In that case, we don't relink the
1035 * buffer: that only gets done when the old transaction finally
1036 * completes its commit.
1038 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1040 transaction_t *transaction = handle->h_transaction;
1041 journal_t *journal = transaction->t_journal;
1042 struct journal_head *jh = bh2jh(bh);
1044 jbd_debug(5, "journal_head %p\n", jh);
1045 JBUFFER_TRACE(jh, "entry");
1046 if (is_handle_aborted(handle))
1049 jbd_lock_bh_state(bh);
1051 if (jh->b_modified == 0) {
1053 * This buffer's got modified and becoming part
1054 * of the transaction. This needs to be done
1055 * once a transaction -bzzz
1058 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1059 handle->h_buffer_credits--;
1063 * fastpath, to avoid expensive locking. If this buffer is already
1064 * on the running transaction's metadata list there is nothing to do.
1065 * Nobody can take it off again because there is a handle open.
1066 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1067 * result in this test being false, so we go in and take the locks.
1069 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1070 JBUFFER_TRACE(jh, "fastpath");
1071 J_ASSERT_JH(jh, jh->b_transaction ==
1072 journal->j_running_transaction);
1076 set_buffer_jbddirty(bh);
1079 * Metadata already on the current transaction list doesn't
1080 * need to be filed. Metadata on another transaction's list must
1081 * be committing, and will be refiled once the commit completes:
1082 * leave it alone for now.
1084 if (jh->b_transaction != transaction) {
1085 JBUFFER_TRACE(jh, "already on other transaction");
1086 J_ASSERT_JH(jh, jh->b_transaction ==
1087 journal->j_committing_transaction);
1088 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1089 /* And this case is illegal: we can't reuse another
1090 * transaction's data buffer, ever. */
1094 /* That test should have eliminated the following case: */
1095 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1097 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1098 spin_lock(&journal->j_list_lock);
1099 __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1100 spin_unlock(&journal->j_list_lock);
1102 jbd_unlock_bh_state(bh);
1104 JBUFFER_TRACE(jh, "exit");
1109 * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1110 * updates, if the update decided in the end that it didn't need access.
1114 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1116 BUFFER_TRACE(bh, "entry");
1120 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1121 * @handle: transaction handle
1122 * @bh: bh to 'forget'
1124 * We can only do the bforget if there are no commits pending against the
1125 * buffer. If the buffer is dirty in the current running transaction we
1126 * can safely unlink it.
1128 * bh may not be a journalled buffer at all - it may be a non-JBD
1129 * buffer which came off the hashtable. Check for this.
1131 * Decrements bh->b_count by one.
1133 * Allow this call even if the handle has aborted --- it may be part of
1134 * the caller's cleanup after an abort.
1136 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1138 transaction_t *transaction = handle->h_transaction;
1139 journal_t *journal = transaction->t_journal;
1140 struct journal_head *jh;
1141 int drop_reserve = 0;
1143 int was_modified = 0;
1145 BUFFER_TRACE(bh, "entry");
1147 jbd_lock_bh_state(bh);
1148 spin_lock(&journal->j_list_lock);
1150 if (!buffer_jbd(bh))
1154 /* Critical error: attempting to delete a bitmap buffer, maybe?
1155 * Don't do any jbd operations, and return an error. */
1156 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1157 "inconsistent data on disk")) {
1162 /* keep track of wether or not this transaction modified us */
1163 was_modified = jh->b_modified;
1166 * The buffer's going from the transaction, we must drop
1167 * all references -bzzz
1171 if (jh->b_transaction == handle->h_transaction) {
1172 J_ASSERT_JH(jh, !jh->b_frozen_data);
1174 /* If we are forgetting a buffer which is already part
1175 * of this transaction, then we can just drop it from
1176 * the transaction immediately. */
1177 clear_buffer_dirty(bh);
1178 clear_buffer_jbddirty(bh);
1180 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1183 * we only want to drop a reference if this transaction
1184 * modified the buffer
1190 * We are no longer going to journal this buffer.
1191 * However, the commit of this transaction is still
1192 * important to the buffer: the delete that we are now
1193 * processing might obsolete an old log entry, so by
1194 * committing, we can satisfy the buffer's checkpoint.
1196 * So, if we have a checkpoint on the buffer, we should
1197 * now refile the buffer on our BJ_Forget list so that
1198 * we know to remove the checkpoint after we commit.
1201 if (jh->b_cp_transaction) {
1202 __jbd2_journal_temp_unlink_buffer(jh);
1203 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1205 __jbd2_journal_unfile_buffer(jh);
1206 jbd2_journal_remove_journal_head(bh);
1208 if (!buffer_jbd(bh)) {
1209 spin_unlock(&journal->j_list_lock);
1210 jbd_unlock_bh_state(bh);
1215 } else if (jh->b_transaction) {
1216 J_ASSERT_JH(jh, (jh->b_transaction ==
1217 journal->j_committing_transaction));
1218 /* However, if the buffer is still owned by a prior
1219 * (committing) transaction, we can't drop it yet... */
1220 JBUFFER_TRACE(jh, "belongs to older transaction");
1221 /* ... but we CAN drop it from the new transaction if we
1222 * have also modified it since the original commit. */
1224 if (jh->b_next_transaction) {
1225 J_ASSERT(jh->b_next_transaction == transaction);
1226 jh->b_next_transaction = NULL;
1229 * only drop a reference if this transaction modified
1238 spin_unlock(&journal->j_list_lock);
1239 jbd_unlock_bh_state(bh);
1243 /* no need to reserve log space for this block -bzzz */
1244 handle->h_buffer_credits++;
1250 * int jbd2_journal_stop() - complete a transaction
1251 * @handle: tranaction to complete.
1253 * All done for a particular handle.
1255 * There is not much action needed here. We just return any remaining
1256 * buffer credits to the transaction and remove the handle. The only
1257 * complication is that we need to start a commit operation if the
1258 * filesystem is marked for synchronous update.
1260 * jbd2_journal_stop itself will not usually return an error, but it may
1261 * do so in unusual circumstances. In particular, expect it to
1262 * return -EIO if a jbd2_journal_abort has been executed since the
1263 * transaction began.
1265 int jbd2_journal_stop(handle_t *handle)
1267 transaction_t *transaction = handle->h_transaction;
1268 journal_t *journal = transaction->t_journal;
1269 int err, wait_for_commit = 0;
1273 J_ASSERT(journal_current_handle() == handle);
1275 if (is_handle_aborted(handle))
1278 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1282 if (--handle->h_ref > 0) {
1283 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1288 jbd_debug(4, "Handle %p going down\n", handle);
1291 * Implement synchronous transaction batching. If the handle
1292 * was synchronous, don't force a commit immediately. Let's
1293 * yield and let another thread piggyback onto this
1294 * transaction. Keep doing that while new threads continue to
1295 * arrive. It doesn't cost much - we're about to run a commit
1296 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1297 * operations by 30x or more...
1299 * We try and optimize the sleep time against what the
1300 * underlying disk can do, instead of having a static sleep
1301 * time. This is useful for the case where our storage is so
1302 * fast that it is more optimal to go ahead and force a flush
1303 * and wait for the transaction to be committed than it is to
1304 * wait for an arbitrary amount of time for new writers to
1305 * join the transaction. We achieve this by measuring how
1306 * long it takes to commit a transaction, and compare it with
1307 * how long this transaction has been running, and if run time
1308 * < commit time then we sleep for the delta and commit. This
1309 * greatly helps super fast disks that would see slowdowns as
1310 * more threads started doing fsyncs.
1312 * But don't do this if this process was the most recent one
1313 * to perform a synchronous write. We do this to detect the
1314 * case where a single process is doing a stream of sync
1315 * writes. No point in waiting for joiners in that case.
1318 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1319 u64 commit_time, trans_time;
1321 journal->j_last_sync_writer = pid;
1323 read_lock(&journal->j_state_lock);
1324 commit_time = journal->j_average_commit_time;
1325 read_unlock(&journal->j_state_lock);
1327 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1328 transaction->t_start_time));
1330 commit_time = max_t(u64, commit_time,
1331 1000*journal->j_min_batch_time);
1332 commit_time = min_t(u64, commit_time,
1333 1000*journal->j_max_batch_time);
1335 if (trans_time < commit_time) {
1336 ktime_t expires = ktime_add_ns(ktime_get(),
1338 set_current_state(TASK_UNINTERRUPTIBLE);
1339 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1344 transaction->t_synchronous_commit = 1;
1345 current->journal_info = NULL;
1346 atomic_sub(handle->h_buffer_credits,
1347 &transaction->t_outstanding_credits);
1350 * If the handle is marked SYNC, we need to set another commit
1351 * going! We also want to force a commit if the current
1352 * transaction is occupying too much of the log, or if the
1353 * transaction is too old now.
1355 if (handle->h_sync ||
1356 (atomic_read(&transaction->t_outstanding_credits) >
1357 journal->j_max_transaction_buffers) ||
1358 time_after_eq(jiffies, transaction->t_expires)) {
1359 /* Do this even for aborted journals: an abort still
1360 * completes the commit thread, it just doesn't write
1361 * anything to disk. */
1363 jbd_debug(2, "transaction too old, requesting commit for "
1364 "handle %p\n", handle);
1365 /* This is non-blocking */
1366 jbd2_log_start_commit(journal, transaction->t_tid);
1369 * Special case: JBD2_SYNC synchronous updates require us
1370 * to wait for the commit to complete.
1372 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1373 wait_for_commit = 1;
1377 * Once we drop t_updates, if it goes to zero the transaction
1378 * could start commiting on us and eventually disappear. So
1379 * once we do this, we must not dereference transaction
1382 tid = transaction->t_tid;
1383 if (atomic_dec_and_test(&transaction->t_updates)) {
1384 wake_up(&journal->j_wait_updates);
1385 if (journal->j_barrier_count)
1386 wake_up(&journal->j_wait_transaction_locked);
1389 if (wait_for_commit)
1390 err = jbd2_log_wait_commit(journal, tid);
1392 lock_map_release(&handle->h_lockdep_map);
1394 jbd2_free_handle(handle);
1399 * int jbd2_journal_force_commit() - force any uncommitted transactions
1400 * @journal: journal to force
1402 * For synchronous operations: force any uncommitted transactions
1403 * to disk. May seem kludgy, but it reuses all the handle batching
1404 * code in a very simple manner.
1406 int jbd2_journal_force_commit(journal_t *journal)
1411 handle = jbd2_journal_start(journal, 1);
1412 if (IS_ERR(handle)) {
1413 ret = PTR_ERR(handle);
1416 ret = jbd2_journal_stop(handle);
1423 * List management code snippets: various functions for manipulating the
1424 * transaction buffer lists.
1429 * Append a buffer to a transaction list, given the transaction's list head
1432 * j_list_lock is held.
1434 * jbd_lock_bh_state(jh2bh(jh)) is held.
1438 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1441 jh->b_tnext = jh->b_tprev = jh;
1444 /* Insert at the tail of the list to preserve order */
1445 struct journal_head *first = *list, *last = first->b_tprev;
1447 jh->b_tnext = first;
1448 last->b_tnext = first->b_tprev = jh;
1453 * Remove a buffer from a transaction list, given the transaction's list
1456 * Called with j_list_lock held, and the journal may not be locked.
1458 * jbd_lock_bh_state(jh2bh(jh)) is held.
1462 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1465 *list = jh->b_tnext;
1469 jh->b_tprev->b_tnext = jh->b_tnext;
1470 jh->b_tnext->b_tprev = jh->b_tprev;
1474 * Remove a buffer from the appropriate transaction list.
1476 * Note that this function can *change* the value of
1477 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1478 * t_log_list or t_reserved_list. If the caller is holding onto a copy of one
1479 * of these pointers, it could go bad. Generally the caller needs to re-read
1480 * the pointer from the transaction_t.
1482 * Called under j_list_lock. The journal may not be locked.
1484 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1486 struct journal_head **list = NULL;
1487 transaction_t *transaction;
1488 struct buffer_head *bh = jh2bh(jh);
1490 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1491 transaction = jh->b_transaction;
1493 assert_spin_locked(&transaction->t_journal->j_list_lock);
1495 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1496 if (jh->b_jlist != BJ_None)
1497 J_ASSERT_JH(jh, transaction != NULL);
1499 switch (jh->b_jlist) {
1503 transaction->t_nr_buffers--;
1504 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1505 list = &transaction->t_buffers;
1508 list = &transaction->t_forget;
1511 list = &transaction->t_iobuf_list;
1514 list = &transaction->t_shadow_list;
1517 list = &transaction->t_log_list;
1520 list = &transaction->t_reserved_list;
1524 __blist_del_buffer(list, jh);
1525 jh->b_jlist = BJ_None;
1526 if (test_clear_buffer_jbddirty(bh))
1527 mark_buffer_dirty(bh); /* Expose it to the VM */
1530 void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1532 __jbd2_journal_temp_unlink_buffer(jh);
1533 jh->b_transaction = NULL;
1536 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1538 jbd_lock_bh_state(jh2bh(jh));
1539 spin_lock(&journal->j_list_lock);
1540 __jbd2_journal_unfile_buffer(jh);
1541 spin_unlock(&journal->j_list_lock);
1542 jbd_unlock_bh_state(jh2bh(jh));
1546 * Called from jbd2_journal_try_to_free_buffers().
1548 * Called under jbd_lock_bh_state(bh)
1551 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1553 struct journal_head *jh;
1557 if (buffer_locked(bh) || buffer_dirty(bh))
1560 if (jh->b_next_transaction != NULL)
1563 spin_lock(&journal->j_list_lock);
1564 if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1565 /* written-back checkpointed metadata buffer */
1566 if (jh->b_jlist == BJ_None) {
1567 JBUFFER_TRACE(jh, "remove from checkpoint list");
1568 __jbd2_journal_remove_checkpoint(jh);
1569 jbd2_journal_remove_journal_head(bh);
1573 spin_unlock(&journal->j_list_lock);
1579 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1580 * @journal: journal for operation
1581 * @page: to try and free
1582 * @gfp_mask: we use the mask to detect how hard should we try to release
1583 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1584 * release the buffers.
1587 * For all the buffers on this page,
1588 * if they are fully written out ordered data, move them onto BUF_CLEAN
1589 * so try_to_free_buffers() can reap them.
1591 * This function returns non-zero if we wish try_to_free_buffers()
1592 * to be called. We do this if the page is releasable by try_to_free_buffers().
1593 * We also do it if the page has locked or dirty buffers and the caller wants
1594 * us to perform sync or async writeout.
1596 * This complicates JBD locking somewhat. We aren't protected by the
1597 * BKL here. We wish to remove the buffer from its committing or
1598 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1600 * This may *change* the value of transaction_t->t_datalist, so anyone
1601 * who looks at t_datalist needs to lock against this function.
1603 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1604 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
1605 * will come out of the lock with the buffer dirty, which makes it
1606 * ineligible for release here.
1608 * Who else is affected by this? hmm... Really the only contender
1609 * is do_get_write_access() - it could be looking at the buffer while
1610 * journal_try_to_free_buffer() is changing its state. But that
1611 * cannot happen because we never reallocate freed data as metadata
1612 * while the data is part of a transaction. Yes?
1614 * Return 0 on failure, 1 on success
1616 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1617 struct page *page, gfp_t gfp_mask)
1619 struct buffer_head *head;
1620 struct buffer_head *bh;
1623 J_ASSERT(PageLocked(page));
1625 head = page_buffers(page);
1628 struct journal_head *jh;
1631 * We take our own ref against the journal_head here to avoid
1632 * having to add tons of locking around each instance of
1633 * jbd2_journal_remove_journal_head() and
1634 * jbd2_journal_put_journal_head().
1636 jh = jbd2_journal_grab_journal_head(bh);
1640 jbd_lock_bh_state(bh);
1641 __journal_try_to_free_buffer(journal, bh);
1642 jbd2_journal_put_journal_head(jh);
1643 jbd_unlock_bh_state(bh);
1646 } while ((bh = bh->b_this_page) != head);
1648 ret = try_to_free_buffers(page);
1655 * This buffer is no longer needed. If it is on an older transaction's
1656 * checkpoint list we need to record it on this transaction's forget list
1657 * to pin this buffer (and hence its checkpointing transaction) down until
1658 * this transaction commits. If the buffer isn't on a checkpoint list, we
1660 * Returns non-zero if JBD no longer has an interest in the buffer.
1662 * Called under j_list_lock.
1664 * Called under jbd_lock_bh_state(bh).
1666 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1669 struct buffer_head *bh = jh2bh(jh);
1671 __jbd2_journal_unfile_buffer(jh);
1673 if (jh->b_cp_transaction) {
1674 JBUFFER_TRACE(jh, "on running+cp transaction");
1676 * We don't want to write the buffer anymore, clear the
1677 * bit so that we don't confuse checks in
1678 * __journal_file_buffer
1680 clear_buffer_dirty(bh);
1681 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1684 JBUFFER_TRACE(jh, "on running transaction");
1685 jbd2_journal_remove_journal_head(bh);
1692 * jbd2_journal_invalidatepage
1694 * This code is tricky. It has a number of cases to deal with.
1696 * There are two invariants which this code relies on:
1698 * i_size must be updated on disk before we start calling invalidatepage on the
1701 * This is done in ext3 by defining an ext3_setattr method which
1702 * updates i_size before truncate gets going. By maintaining this
1703 * invariant, we can be sure that it is safe to throw away any buffers
1704 * attached to the current transaction: once the transaction commits,
1705 * we know that the data will not be needed.
1707 * Note however that we can *not* throw away data belonging to the
1708 * previous, committing transaction!
1710 * Any disk blocks which *are* part of the previous, committing
1711 * transaction (and which therefore cannot be discarded immediately) are
1712 * not going to be reused in the new running transaction
1714 * The bitmap committed_data images guarantee this: any block which is
1715 * allocated in one transaction and removed in the next will be marked
1716 * as in-use in the committed_data bitmap, so cannot be reused until
1717 * the next transaction to delete the block commits. This means that
1718 * leaving committing buffers dirty is quite safe: the disk blocks
1719 * cannot be reallocated to a different file and so buffer aliasing is
1723 * The above applies mainly to ordered data mode. In writeback mode we
1724 * don't make guarantees about the order in which data hits disk --- in
1725 * particular we don't guarantee that new dirty data is flushed before
1726 * transaction commit --- so it is always safe just to discard data
1727 * immediately in that mode. --sct
1731 * The journal_unmap_buffer helper function returns zero if the buffer
1732 * concerned remains pinned as an anonymous buffer belonging to an older
1735 * We're outside-transaction here. Either or both of j_running_transaction
1736 * and j_committing_transaction may be NULL.
1738 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1740 transaction_t *transaction;
1741 struct journal_head *jh;
1745 BUFFER_TRACE(bh, "entry");
1748 * It is safe to proceed here without the j_list_lock because the
1749 * buffers cannot be stolen by try_to_free_buffers as long as we are
1750 * holding the page lock. --sct
1753 if (!buffer_jbd(bh))
1754 goto zap_buffer_unlocked;
1756 /* OK, we have data buffer in journaled mode */
1757 write_lock(&journal->j_state_lock);
1758 jbd_lock_bh_state(bh);
1759 spin_lock(&journal->j_list_lock);
1761 jh = jbd2_journal_grab_journal_head(bh);
1763 goto zap_buffer_no_jh;
1766 * We cannot remove the buffer from checkpoint lists until the
1767 * transaction adding inode to orphan list (let's call it T)
1768 * is committed. Otherwise if the transaction changing the
1769 * buffer would be cleaned from the journal before T is
1770 * committed, a crash will cause that the correct contents of
1771 * the buffer will be lost. On the other hand we have to
1772 * clear the buffer dirty bit at latest at the moment when the
1773 * transaction marking the buffer as freed in the filesystem
1774 * structures is committed because from that moment on the
1775 * buffer can be reallocated and used by a different page.
1776 * Since the block hasn't been freed yet but the inode has
1777 * already been added to orphan list, it is safe for us to add
1778 * the buffer to BJ_Forget list of the newest transaction.
1780 transaction = jh->b_transaction;
1781 if (transaction == NULL) {
1782 /* First case: not on any transaction. If it
1783 * has no checkpoint link, then we can zap it:
1784 * it's a writeback-mode buffer so we don't care
1785 * if it hits disk safely. */
1786 if (!jh->b_cp_transaction) {
1787 JBUFFER_TRACE(jh, "not on any transaction: zap");
1791 if (!buffer_dirty(bh)) {
1792 /* bdflush has written it. We can drop it now */
1796 /* OK, it must be in the journal but still not
1797 * written fully to disk: it's metadata or
1798 * journaled data... */
1800 if (journal->j_running_transaction) {
1801 /* ... and once the current transaction has
1802 * committed, the buffer won't be needed any
1804 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1805 ret = __dispose_buffer(jh,
1806 journal->j_running_transaction);
1807 jbd2_journal_put_journal_head(jh);
1808 spin_unlock(&journal->j_list_lock);
1809 jbd_unlock_bh_state(bh);
1810 write_unlock(&journal->j_state_lock);
1813 /* There is no currently-running transaction. So the
1814 * orphan record which we wrote for this file must have
1815 * passed into commit. We must attach this buffer to
1816 * the committing transaction, if it exists. */
1817 if (journal->j_committing_transaction) {
1818 JBUFFER_TRACE(jh, "give to committing trans");
1819 ret = __dispose_buffer(jh,
1820 journal->j_committing_transaction);
1821 jbd2_journal_put_journal_head(jh);
1822 spin_unlock(&journal->j_list_lock);
1823 jbd_unlock_bh_state(bh);
1824 write_unlock(&journal->j_state_lock);
1827 /* The orphan record's transaction has
1828 * committed. We can cleanse this buffer */
1829 clear_buffer_jbddirty(bh);
1833 } else if (transaction == journal->j_committing_transaction) {
1834 JBUFFER_TRACE(jh, "on committing transaction");
1836 * The buffer is committing, we simply cannot touch
1837 * it. So we just set j_next_transaction to the
1838 * running transaction (if there is one) and mark
1839 * buffer as freed so that commit code knows it should
1840 * clear dirty bits when it is done with the buffer.
1842 set_buffer_freed(bh);
1843 if (journal->j_running_transaction && buffer_jbddirty(bh))
1844 jh->b_next_transaction = journal->j_running_transaction;
1845 jbd2_journal_put_journal_head(jh);
1846 spin_unlock(&journal->j_list_lock);
1847 jbd_unlock_bh_state(bh);
1848 write_unlock(&journal->j_state_lock);
1851 /* Good, the buffer belongs to the running transaction.
1852 * We are writing our own transaction's data, not any
1853 * previous one's, so it is safe to throw it away
1854 * (remember that we expect the filesystem to have set
1855 * i_size already for this truncate so recovery will not
1856 * expose the disk blocks we are discarding here.) */
1857 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1858 JBUFFER_TRACE(jh, "on running transaction");
1859 may_free = __dispose_buffer(jh, transaction);
1863 jbd2_journal_put_journal_head(jh);
1865 spin_unlock(&journal->j_list_lock);
1866 jbd_unlock_bh_state(bh);
1867 write_unlock(&journal->j_state_lock);
1868 zap_buffer_unlocked:
1869 clear_buffer_dirty(bh);
1870 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1871 clear_buffer_mapped(bh);
1872 clear_buffer_req(bh);
1873 clear_buffer_new(bh);
1879 * void jbd2_journal_invalidatepage()
1880 * @journal: journal to use for flush...
1881 * @page: page to flush
1882 * @offset: length of page to invalidate.
1884 * Reap page buffers containing data after offset in page.
1887 void jbd2_journal_invalidatepage(journal_t *journal,
1889 unsigned long offset)
1891 struct buffer_head *head, *bh, *next;
1892 unsigned int curr_off = 0;
1895 if (!PageLocked(page))
1897 if (!page_has_buffers(page))
1900 /* We will potentially be playing with lists other than just the
1901 * data lists (especially for journaled data mode), so be
1902 * cautious in our locking. */
1904 head = bh = page_buffers(page);
1906 unsigned int next_off = curr_off + bh->b_size;
1907 next = bh->b_this_page;
1909 if (offset <= curr_off) {
1910 /* This block is wholly outside the truncation point */
1912 may_free &= journal_unmap_buffer(journal, bh);
1915 curr_off = next_off;
1918 } while (bh != head);
1921 if (may_free && try_to_free_buffers(page))
1922 J_ASSERT(!page_has_buffers(page));
1927 * File a buffer on the given transaction list.
1929 void __jbd2_journal_file_buffer(struct journal_head *jh,
1930 transaction_t *transaction, int jlist)
1932 struct journal_head **list = NULL;
1934 struct buffer_head *bh = jh2bh(jh);
1936 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1937 assert_spin_locked(&transaction->t_journal->j_list_lock);
1939 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1940 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1941 jh->b_transaction == NULL);
1943 if (jh->b_transaction && jh->b_jlist == jlist)
1946 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1947 jlist == BJ_Shadow || jlist == BJ_Forget) {
1949 * For metadata buffers, we track dirty bit in buffer_jbddirty
1950 * instead of buffer_dirty. We should not see a dirty bit set
1951 * here because we clear it in do_get_write_access but e.g.
1952 * tune2fs can modify the sb and set the dirty bit at any time
1953 * so we try to gracefully handle that.
1955 if (buffer_dirty(bh))
1956 warn_dirty_buffer(bh);
1957 if (test_clear_buffer_dirty(bh) ||
1958 test_clear_buffer_jbddirty(bh))
1962 if (jh->b_transaction)
1963 __jbd2_journal_temp_unlink_buffer(jh);
1964 jh->b_transaction = transaction;
1968 J_ASSERT_JH(jh, !jh->b_committed_data);
1969 J_ASSERT_JH(jh, !jh->b_frozen_data);
1972 transaction->t_nr_buffers++;
1973 list = &transaction->t_buffers;
1976 list = &transaction->t_forget;
1979 list = &transaction->t_iobuf_list;
1982 list = &transaction->t_shadow_list;
1985 list = &transaction->t_log_list;
1988 list = &transaction->t_reserved_list;
1992 __blist_add_buffer(list, jh);
1993 jh->b_jlist = jlist;
1996 set_buffer_jbddirty(bh);
1999 void jbd2_journal_file_buffer(struct journal_head *jh,
2000 transaction_t *transaction, int jlist)
2002 jbd_lock_bh_state(jh2bh(jh));
2003 spin_lock(&transaction->t_journal->j_list_lock);
2004 __jbd2_journal_file_buffer(jh, transaction, jlist);
2005 spin_unlock(&transaction->t_journal->j_list_lock);
2006 jbd_unlock_bh_state(jh2bh(jh));
2010 * Remove a buffer from its current buffer list in preparation for
2011 * dropping it from its current transaction entirely. If the buffer has
2012 * already started to be used by a subsequent transaction, refile the
2013 * buffer on that transaction's metadata list.
2015 * Called under journal->j_list_lock
2017 * Called under jbd_lock_bh_state(jh2bh(jh))
2019 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2021 int was_dirty, jlist;
2022 struct buffer_head *bh = jh2bh(jh);
2024 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2025 if (jh->b_transaction)
2026 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2028 /* If the buffer is now unused, just drop it. */
2029 if (jh->b_next_transaction == NULL) {
2030 __jbd2_journal_unfile_buffer(jh);
2035 * It has been modified by a later transaction: add it to the new
2036 * transaction's metadata list.
2039 was_dirty = test_clear_buffer_jbddirty(bh);
2040 __jbd2_journal_temp_unlink_buffer(jh);
2041 jh->b_transaction = jh->b_next_transaction;
2042 jh->b_next_transaction = NULL;
2043 if (buffer_freed(bh))
2045 else if (jh->b_modified)
2046 jlist = BJ_Metadata;
2048 jlist = BJ_Reserved;
2049 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2050 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2053 set_buffer_jbddirty(bh);
2057 * For the unlocked version of this call, also make sure that any
2058 * hanging journal_head is cleaned up if necessary.
2060 * __jbd2_journal_refile_buffer is usually called as part of a single locked
2061 * operation on a buffer_head, in which the caller is probably going to
2062 * be hooking the journal_head onto other lists. In that case it is up
2063 * to the caller to remove the journal_head if necessary. For the
2064 * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2065 * doing anything else to the buffer so we need to do the cleanup
2066 * ourselves to avoid a jh leak.
2068 * *** The journal_head may be freed by this call! ***
2070 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2072 struct buffer_head *bh = jh2bh(jh);
2074 jbd_lock_bh_state(bh);
2075 spin_lock(&journal->j_list_lock);
2077 __jbd2_journal_refile_buffer(jh);
2078 jbd_unlock_bh_state(bh);
2079 jbd2_journal_remove_journal_head(bh);
2081 spin_unlock(&journal->j_list_lock);
2086 * File inode in the inode list of the handle's transaction
2088 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2090 transaction_t *transaction = handle->h_transaction;
2091 journal_t *journal = transaction->t_journal;
2093 if (is_handle_aborted(handle))
2096 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2097 transaction->t_tid);
2100 * First check whether inode isn't already on the transaction's
2101 * lists without taking the lock. Note that this check is safe
2102 * without the lock as we cannot race with somebody removing inode
2103 * from the transaction. The reason is that we remove inode from the
2104 * transaction only in journal_release_jbd_inode() and when we commit
2105 * the transaction. We are guarded from the first case by holding
2106 * a reference to the inode. We are safe against the second case
2107 * because if jinode->i_transaction == transaction, commit code
2108 * cannot touch the transaction because we hold reference to it,
2109 * and if jinode->i_next_transaction == transaction, commit code
2110 * will only file the inode where we want it.
2112 if (jinode->i_transaction == transaction ||
2113 jinode->i_next_transaction == transaction)
2116 spin_lock(&journal->j_list_lock);
2118 if (jinode->i_transaction == transaction ||
2119 jinode->i_next_transaction == transaction)
2122 /* On some different transaction's list - should be
2123 * the committing one */
2124 if (jinode->i_transaction) {
2125 J_ASSERT(jinode->i_next_transaction == NULL);
2126 J_ASSERT(jinode->i_transaction ==
2127 journal->j_committing_transaction);
2128 jinode->i_next_transaction = transaction;
2131 /* Not on any transaction list... */
2132 J_ASSERT(!jinode->i_next_transaction);
2133 jinode->i_transaction = transaction;
2134 list_add(&jinode->i_list, &transaction->t_inode_list);
2136 spin_unlock(&journal->j_list_lock);
2142 * File truncate and transaction commit interact with each other in a
2143 * non-trivial way. If a transaction writing data block A is
2144 * committing, we cannot discard the data by truncate until we have
2145 * written them. Otherwise if we crashed after the transaction with
2146 * write has committed but before the transaction with truncate has
2147 * committed, we could see stale data in block A. This function is a
2148 * helper to solve this problem. It starts writeout of the truncated
2149 * part in case it is in the committing transaction.
2151 * Filesystem code must call this function when inode is journaled in
2152 * ordered mode before truncation happens and after the inode has been
2153 * placed on orphan list with the new inode size. The second condition
2154 * avoids the race that someone writes new data and we start
2155 * committing the transaction after this function has been called but
2156 * before a transaction for truncate is started (and furthermore it
2157 * allows us to optimize the case where the addition to orphan list
2158 * happens in the same transaction as write --- we don't have to write
2159 * any data in such case).
2161 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2162 struct jbd2_inode *jinode,
2165 transaction_t *inode_trans, *commit_trans;
2168 /* This is a quick check to avoid locking if not necessary */
2169 if (!jinode->i_transaction)
2171 /* Locks are here just to force reading of recent values, it is
2172 * enough that the transaction was not committing before we started
2173 * a transaction adding the inode to orphan list */
2174 read_lock(&journal->j_state_lock);
2175 commit_trans = journal->j_committing_transaction;
2176 read_unlock(&journal->j_state_lock);
2177 spin_lock(&journal->j_list_lock);
2178 inode_trans = jinode->i_transaction;
2179 spin_unlock(&journal->j_list_lock);
2180 if (inode_trans == commit_trans) {
2181 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2182 new_size, LLONG_MAX);
2184 jbd2_journal_abort(journal, ret);