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 atomic_set(&transaction
->t_handle_count
, 0);
61 INIT_LIST_HEAD(&transaction
->t_inode_list
);
62 INIT_LIST_HEAD(&transaction
->t_private_list
);
64 /* Set up the commit timer for the new transaction. */
65 journal
->j_commit_timer
.expires
= round_jiffies_up(transaction
->t_expires
);
66 add_timer(&journal
->j_commit_timer
);
68 J_ASSERT(journal
->j_running_transaction
== NULL
);
69 journal
->j_running_transaction
= transaction
;
70 transaction
->t_max_wait
= 0;
71 transaction
->t_start
= jiffies
;
79 * A handle_t is an object which represents a single atomic update to a
80 * filesystem, and which tracks all of the modifications which form part
85 * Update transiaction's maximum wait time, if debugging is enabled.
87 * In order for t_max_wait to be reliable, it must be protected by a
88 * lock. But doing so will mean that start_this_handle() can not be
89 * run in parallel on SMP systems, which limits our scalability. So
90 * unless debugging is enabled, we no longer update t_max_wait, which
91 * means that maximum wait time reported by the jbd2_run_stats
92 * tracepoint will always be zero.
94 static inline void update_t_max_wait(transaction_t
*transaction
)
96 #ifdef CONFIG_JBD2_DEBUG
97 unsigned long ts
= jiffies
;
99 if (jbd2_journal_enable_debug
&&
100 time_after(transaction
->t_start
, ts
)) {
101 ts
= jbd2_time_diff(ts
, transaction
->t_start
);
102 spin_lock(&transaction
->t_handle_lock
);
103 if (ts
> transaction
->t_max_wait
)
104 transaction
->t_max_wait
= ts
;
105 spin_unlock(&transaction
->t_handle_lock
);
111 * start_this_handle: Given a handle, deal with any locking or stalling
112 * needed to make sure that there is enough journal space for the handle
113 * to begin. Attach the handle to a transaction and set up the
114 * transaction's buffer credits.
117 static int start_this_handle(journal_t
*journal
, handle_t
*handle
,
120 transaction_t
*transaction
;
122 int nblocks
= handle
->h_buffer_credits
;
123 transaction_t
*new_transaction
= NULL
;
125 if (nblocks
> journal
->j_max_transaction_buffers
) {
126 printk(KERN_ERR
"JBD: %s wants too many credits (%d > %d)\n",
127 current
->comm
, nblocks
,
128 journal
->j_max_transaction_buffers
);
133 if (!journal
->j_running_transaction
) {
134 new_transaction
= kzalloc(sizeof(*new_transaction
), gfp_mask
);
135 if (!new_transaction
) {
137 * If __GFP_FS is not present, then we may be
138 * being called from inside the fs writeback
139 * layer, so we MUST NOT fail. Since
140 * __GFP_NOFAIL is going away, we will arrange
141 * to retry the allocation ourselves.
143 if ((gfp_mask
& __GFP_FS
) == 0) {
144 congestion_wait(BLK_RW_ASYNC
, HZ
/50);
145 goto alloc_transaction
;
151 jbd_debug(3, "New handle %p going live.\n", handle
);
154 * We need to hold j_state_lock until t_updates has been incremented,
155 * for proper journal barrier handling
158 read_lock(&journal
->j_state_lock
);
159 BUG_ON(journal
->j_flags
& JBD2_UNMOUNT
);
160 if (is_journal_aborted(journal
) ||
161 (journal
->j_errno
!= 0 && !(journal
->j_flags
& JBD2_ACK_ERR
))) {
162 read_unlock(&journal
->j_state_lock
);
163 kfree(new_transaction
);
167 /* Wait on the journal's transaction barrier if necessary */
168 if (journal
->j_barrier_count
) {
169 read_unlock(&journal
->j_state_lock
);
170 wait_event(journal
->j_wait_transaction_locked
,
171 journal
->j_barrier_count
== 0);
175 if (!journal
->j_running_transaction
) {
176 read_unlock(&journal
->j_state_lock
);
177 if (!new_transaction
)
178 goto alloc_transaction
;
179 write_lock(&journal
->j_state_lock
);
180 if (!journal
->j_running_transaction
) {
181 jbd2_get_transaction(journal
, new_transaction
);
182 new_transaction
= NULL
;
184 write_unlock(&journal
->j_state_lock
);
188 transaction
= journal
->j_running_transaction
;
191 * If the current transaction is locked down for commit, wait for the
192 * lock to be released.
194 if (transaction
->t_state
== T_LOCKED
) {
197 prepare_to_wait(&journal
->j_wait_transaction_locked
,
198 &wait
, TASK_UNINTERRUPTIBLE
);
199 read_unlock(&journal
->j_state_lock
);
201 finish_wait(&journal
->j_wait_transaction_locked
, &wait
);
206 * If there is not enough space left in the log to write all potential
207 * buffers requested by this operation, we need to stall pending a log
208 * checkpoint to free some more log space.
210 needed
= atomic_add_return(nblocks
,
211 &transaction
->t_outstanding_credits
);
213 if (needed
> journal
->j_max_transaction_buffers
) {
215 * If the current transaction is already too large, then start
216 * to commit it: we can then go back and attach this handle to
221 jbd_debug(2, "Handle %p starting new commit...\n", handle
);
222 atomic_sub(nblocks
, &transaction
->t_outstanding_credits
);
223 prepare_to_wait(&journal
->j_wait_transaction_locked
, &wait
,
224 TASK_UNINTERRUPTIBLE
);
225 __jbd2_log_start_commit(journal
, transaction
->t_tid
);
226 read_unlock(&journal
->j_state_lock
);
228 finish_wait(&journal
->j_wait_transaction_locked
, &wait
);
233 * The commit code assumes that it can get enough log space
234 * without forcing a checkpoint. This is *critical* for
235 * correctness: a checkpoint of a buffer which is also
236 * associated with a committing transaction creates a deadlock,
237 * so commit simply cannot force through checkpoints.
239 * We must therefore ensure the necessary space in the journal
240 * *before* starting to dirty potentially checkpointed buffers
241 * in the new transaction.
243 * The worst part is, any transaction currently committing can
244 * reduce the free space arbitrarily. Be careful to account for
245 * those buffers when checkpointing.
249 * @@@ AKPM: This seems rather over-defensive. We're giving commit
250 * a _lot_ of headroom: 1/4 of the journal plus the size of
251 * the committing transaction. Really, we only need to give it
252 * committing_transaction->t_outstanding_credits plus "enough" for
253 * the log control blocks.
254 * Also, this test is inconsitent with the matching one in
255 * jbd2_journal_extend().
257 if (__jbd2_log_space_left(journal
) < jbd_space_needed(journal
)) {
258 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle
);
259 atomic_sub(nblocks
, &transaction
->t_outstanding_credits
);
260 read_unlock(&journal
->j_state_lock
);
261 write_lock(&journal
->j_state_lock
);
262 if (__jbd2_log_space_left(journal
) < jbd_space_needed(journal
))
263 __jbd2_log_wait_for_space(journal
);
264 write_unlock(&journal
->j_state_lock
);
268 /* OK, account for the buffers that this operation expects to
269 * use and add the handle to the running transaction.
271 update_t_max_wait(transaction
);
272 handle
->h_transaction
= transaction
;
273 atomic_inc(&transaction
->t_updates
);
274 atomic_inc(&transaction
->t_handle_count
);
275 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
277 atomic_read(&transaction
->t_outstanding_credits
),
278 __jbd2_log_space_left(journal
));
279 read_unlock(&journal
->j_state_lock
);
281 lock_map_acquire(&handle
->h_lockdep_map
);
282 kfree(new_transaction
);
286 static struct lock_class_key jbd2_handle_key
;
288 /* Allocate a new handle. This should probably be in a slab... */
289 static handle_t
*new_handle(int nblocks
)
291 handle_t
*handle
= jbd2_alloc_handle(GFP_NOFS
);
294 memset(handle
, 0, sizeof(*handle
));
295 handle
->h_buffer_credits
= nblocks
;
298 lockdep_init_map(&handle
->h_lockdep_map
, "jbd2_handle",
299 &jbd2_handle_key
, 0);
305 * handle_t *jbd2_journal_start() - Obtain a new handle.
306 * @journal: Journal to start transaction on.
307 * @nblocks: number of block buffer we might modify
309 * We make sure that the transaction can guarantee at least nblocks of
310 * modified buffers in the log. We block until the log can guarantee
313 * This function is visible to journal users (like ext3fs), so is not
314 * called with the journal already locked.
316 * Return a pointer to a newly allocated handle, or NULL on failure
318 handle_t
*jbd2__journal_start(journal_t
*journal
, int nblocks
, int gfp_mask
)
320 handle_t
*handle
= journal_current_handle();
324 return ERR_PTR(-EROFS
);
327 J_ASSERT(handle
->h_transaction
->t_journal
== journal
);
332 handle
= new_handle(nblocks
);
334 return ERR_PTR(-ENOMEM
);
336 current
->journal_info
= handle
;
338 err
= start_this_handle(journal
, handle
, gfp_mask
);
340 jbd2_free_handle(handle
);
341 current
->journal_info
= NULL
;
342 handle
= ERR_PTR(err
);
348 EXPORT_SYMBOL(jbd2__journal_start
);
351 handle_t
*jbd2_journal_start(journal_t
*journal
, int nblocks
)
353 return jbd2__journal_start(journal
, nblocks
, GFP_NOFS
);
355 EXPORT_SYMBOL(jbd2_journal_start
);
359 * int jbd2_journal_extend() - extend buffer credits.
360 * @handle: handle to 'extend'
361 * @nblocks: nr blocks to try to extend by.
363 * Some transactions, such as large extends and truncates, can be done
364 * atomically all at once or in several stages. The operation requests
365 * a credit for a number of buffer modications in advance, but can
366 * extend its credit if it needs more.
368 * jbd2_journal_extend tries to give the running handle more buffer credits.
369 * It does not guarantee that allocation - this is a best-effort only.
370 * The calling process MUST be able to deal cleanly with a failure to
373 * Return 0 on success, non-zero on failure.
375 * return code < 0 implies an error
376 * return code > 0 implies normal transaction-full status.
378 int jbd2_journal_extend(handle_t
*handle
, int nblocks
)
380 transaction_t
*transaction
= handle
->h_transaction
;
381 journal_t
*journal
= transaction
->t_journal
;
386 if (is_handle_aborted(handle
))
391 read_lock(&journal
->j_state_lock
);
393 /* Don't extend a locked-down transaction! */
394 if (handle
->h_transaction
->t_state
!= T_RUNNING
) {
395 jbd_debug(3, "denied handle %p %d blocks: "
396 "transaction not running\n", handle
, nblocks
);
400 spin_lock(&transaction
->t_handle_lock
);
401 wanted
= atomic_read(&transaction
->t_outstanding_credits
) + nblocks
;
403 if (wanted
> journal
->j_max_transaction_buffers
) {
404 jbd_debug(3, "denied handle %p %d blocks: "
405 "transaction too large\n", handle
, nblocks
);
409 if (wanted
> __jbd2_log_space_left(journal
)) {
410 jbd_debug(3, "denied handle %p %d blocks: "
411 "insufficient log space\n", handle
, nblocks
);
415 handle
->h_buffer_credits
+= nblocks
;
416 atomic_add(nblocks
, &transaction
->t_outstanding_credits
);
419 jbd_debug(3, "extended handle %p by %d\n", handle
, nblocks
);
421 spin_unlock(&transaction
->t_handle_lock
);
423 read_unlock(&journal
->j_state_lock
);
430 * int jbd2_journal_restart() - restart a handle .
431 * @handle: handle to restart
432 * @nblocks: nr credits requested
434 * Restart a handle for a multi-transaction filesystem
437 * If the jbd2_journal_extend() call above fails to grant new buffer credits
438 * to a running handle, a call to jbd2_journal_restart will commit the
439 * handle's transaction so far and reattach the handle to a new
440 * transaction capabable of guaranteeing the requested number of
443 int jbd2__journal_restart(handle_t
*handle
, int nblocks
, int gfp_mask
)
445 transaction_t
*transaction
= handle
->h_transaction
;
446 journal_t
*journal
= transaction
->t_journal
;
449 /* If we've had an abort of any type, don't even think about
450 * actually doing the restart! */
451 if (is_handle_aborted(handle
))
455 * First unlink the handle from its current transaction, and start the
458 J_ASSERT(atomic_read(&transaction
->t_updates
) > 0);
459 J_ASSERT(journal_current_handle() == handle
);
461 read_lock(&journal
->j_state_lock
);
462 spin_lock(&transaction
->t_handle_lock
);
463 atomic_sub(handle
->h_buffer_credits
,
464 &transaction
->t_outstanding_credits
);
465 if (atomic_dec_and_test(&transaction
->t_updates
))
466 wake_up(&journal
->j_wait_updates
);
467 spin_unlock(&transaction
->t_handle_lock
);
469 jbd_debug(2, "restarting handle %p\n", handle
);
470 __jbd2_log_start_commit(journal
, transaction
->t_tid
);
471 read_unlock(&journal
->j_state_lock
);
473 lock_map_release(&handle
->h_lockdep_map
);
474 handle
->h_buffer_credits
= nblocks
;
475 ret
= start_this_handle(journal
, handle
, gfp_mask
);
478 EXPORT_SYMBOL(jbd2__journal_restart
);
481 int jbd2_journal_restart(handle_t
*handle
, int nblocks
)
483 return jbd2__journal_restart(handle
, nblocks
, GFP_NOFS
);
485 EXPORT_SYMBOL(jbd2_journal_restart
);
488 * void jbd2_journal_lock_updates () - establish a transaction barrier.
489 * @journal: Journal to establish a barrier on.
491 * This locks out any further updates from being started, and blocks
492 * until all existing updates have completed, returning only once the
493 * journal is in a quiescent state with no updates running.
495 * The journal lock should not be held on entry.
497 void jbd2_journal_lock_updates(journal_t
*journal
)
501 write_lock(&journal
->j_state_lock
);
502 ++journal
->j_barrier_count
;
504 /* Wait until there are no running updates */
506 transaction_t
*transaction
= journal
->j_running_transaction
;
511 spin_lock(&transaction
->t_handle_lock
);
512 if (!atomic_read(&transaction
->t_updates
)) {
513 spin_unlock(&transaction
->t_handle_lock
);
516 prepare_to_wait(&journal
->j_wait_updates
, &wait
,
517 TASK_UNINTERRUPTIBLE
);
518 spin_unlock(&transaction
->t_handle_lock
);
519 write_unlock(&journal
->j_state_lock
);
521 finish_wait(&journal
->j_wait_updates
, &wait
);
522 write_lock(&journal
->j_state_lock
);
524 write_unlock(&journal
->j_state_lock
);
527 * We have now established a barrier against other normal updates, but
528 * we also need to barrier against other jbd2_journal_lock_updates() calls
529 * to make sure that we serialise special journal-locked operations
532 mutex_lock(&journal
->j_barrier
);
536 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
537 * @journal: Journal to release the barrier on.
539 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
541 * Should be called without the journal lock held.
543 void jbd2_journal_unlock_updates (journal_t
*journal
)
545 J_ASSERT(journal
->j_barrier_count
!= 0);
547 mutex_unlock(&journal
->j_barrier
);
548 write_lock(&journal
->j_state_lock
);
549 --journal
->j_barrier_count
;
550 write_unlock(&journal
->j_state_lock
);
551 wake_up(&journal
->j_wait_transaction_locked
);
554 static void warn_dirty_buffer(struct buffer_head
*bh
)
556 char b
[BDEVNAME_SIZE
];
559 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
560 "There's a risk of filesystem corruption in case of system "
562 bdevname(bh
->b_bdev
, b
), (unsigned long long)bh
->b_blocknr
);
566 * If the buffer is already part of the current transaction, then there
567 * is nothing we need to do. If it is already part of a prior
568 * transaction which we are still committing to disk, then we need to
569 * make sure that we do not overwrite the old copy: we do copy-out to
570 * preserve the copy going to disk. We also account the buffer against
571 * the handle's metadata buffer credits (unless the buffer is already
572 * part of the transaction, that is).
576 do_get_write_access(handle_t
*handle
, struct journal_head
*jh
,
579 struct buffer_head
*bh
;
580 transaction_t
*transaction
;
583 char *frozen_buffer
= NULL
;
586 if (is_handle_aborted(handle
))
589 transaction
= handle
->h_transaction
;
590 journal
= transaction
->t_journal
;
592 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh
, force_copy
);
594 JBUFFER_TRACE(jh
, "entry");
598 /* @@@ Need to check for errors here at some point. */
601 jbd_lock_bh_state(bh
);
603 /* We now hold the buffer lock so it is safe to query the buffer
604 * state. Is the buffer dirty?
606 * If so, there are two possibilities. The buffer may be
607 * non-journaled, and undergoing a quite legitimate writeback.
608 * Otherwise, it is journaled, and we don't expect dirty buffers
609 * in that state (the buffers should be marked JBD_Dirty
610 * instead.) So either the IO is being done under our own
611 * control and this is a bug, or it's a third party IO such as
612 * dump(8) (which may leave the buffer scheduled for read ---
613 * ie. locked but not dirty) or tune2fs (which may actually have
614 * the buffer dirtied, ugh.) */
616 if (buffer_dirty(bh
)) {
618 * First question: is this buffer already part of the current
619 * transaction or the existing committing transaction?
621 if (jh
->b_transaction
) {
623 jh
->b_transaction
== transaction
||
625 journal
->j_committing_transaction
);
626 if (jh
->b_next_transaction
)
627 J_ASSERT_JH(jh
, jh
->b_next_transaction
==
629 warn_dirty_buffer(bh
);
632 * In any case we need to clean the dirty flag and we must
633 * do it under the buffer lock to be sure we don't race
634 * with running write-out.
636 JBUFFER_TRACE(jh
, "Journalling dirty buffer");
637 clear_buffer_dirty(bh
);
638 set_buffer_jbddirty(bh
);
644 if (is_handle_aborted(handle
)) {
645 jbd_unlock_bh_state(bh
);
651 * The buffer is already part of this transaction if b_transaction or
652 * b_next_transaction points to it
654 if (jh
->b_transaction
== transaction
||
655 jh
->b_next_transaction
== transaction
)
659 * this is the first time this transaction is touching this buffer,
660 * reset the modified flag
665 * If there is already a copy-out version of this buffer, then we don't
666 * need to make another one
668 if (jh
->b_frozen_data
) {
669 JBUFFER_TRACE(jh
, "has frozen data");
670 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
671 jh
->b_next_transaction
= transaction
;
675 /* Is there data here we need to preserve? */
677 if (jh
->b_transaction
&& jh
->b_transaction
!= transaction
) {
678 JBUFFER_TRACE(jh
, "owned by older transaction");
679 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
680 J_ASSERT_JH(jh
, jh
->b_transaction
==
681 journal
->j_committing_transaction
);
683 /* There is one case we have to be very careful about.
684 * If the committing transaction is currently writing
685 * this buffer out to disk and has NOT made a copy-out,
686 * then we cannot modify the buffer contents at all
687 * right now. The essence of copy-out is that it is the
688 * extra copy, not the primary copy, which gets
689 * journaled. If the primary copy is already going to
690 * disk then we cannot do copy-out here. */
692 if (jh
->b_jlist
== BJ_Shadow
) {
693 DEFINE_WAIT_BIT(wait
, &bh
->b_state
, BH_Unshadow
);
694 wait_queue_head_t
*wqh
;
696 wqh
= bit_waitqueue(&bh
->b_state
, BH_Unshadow
);
698 JBUFFER_TRACE(jh
, "on shadow: sleep");
699 jbd_unlock_bh_state(bh
);
700 /* commit wakes up all shadow buffers after IO */
702 prepare_to_wait(wqh
, &wait
.wait
,
703 TASK_UNINTERRUPTIBLE
);
704 if (jh
->b_jlist
!= BJ_Shadow
)
708 finish_wait(wqh
, &wait
.wait
);
712 /* Only do the copy if the currently-owning transaction
713 * still needs it. If it is on the Forget list, the
714 * committing transaction is past that stage. The
715 * buffer had better remain locked during the kmalloc,
716 * but that should be true --- we hold the journal lock
717 * still and the buffer is already on the BUF_JOURNAL
718 * list so won't be flushed.
720 * Subtle point, though: if this is a get_undo_access,
721 * then we will be relying on the frozen_data to contain
722 * the new value of the committed_data record after the
723 * transaction, so we HAVE to force the frozen_data copy
726 if (jh
->b_jlist
!= BJ_Forget
|| force_copy
) {
727 JBUFFER_TRACE(jh
, "generate frozen data");
728 if (!frozen_buffer
) {
729 JBUFFER_TRACE(jh
, "allocate memory for buffer");
730 jbd_unlock_bh_state(bh
);
732 jbd2_alloc(jh2bh(jh
)->b_size
,
734 if (!frozen_buffer
) {
736 "%s: OOM for frozen_buffer\n",
738 JBUFFER_TRACE(jh
, "oom!");
740 jbd_lock_bh_state(bh
);
745 jh
->b_frozen_data
= frozen_buffer
;
746 frozen_buffer
= NULL
;
749 jh
->b_next_transaction
= transaction
;
754 * Finally, if the buffer is not journaled right now, we need to make
755 * sure it doesn't get written to disk before the caller actually
756 * commits the new data
758 if (!jh
->b_transaction
) {
759 JBUFFER_TRACE(jh
, "no transaction");
760 J_ASSERT_JH(jh
, !jh
->b_next_transaction
);
761 jh
->b_transaction
= transaction
;
762 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
763 spin_lock(&journal
->j_list_lock
);
764 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Reserved
);
765 spin_unlock(&journal
->j_list_lock
);
774 J_EXPECT_JH(jh
, buffer_uptodate(jh2bh(jh
)),
775 "Possible IO failure.\n");
776 page
= jh2bh(jh
)->b_page
;
777 offset
= ((unsigned long) jh2bh(jh
)->b_data
) & ~PAGE_MASK
;
778 source
= kmap_atomic(page
, KM_USER0
);
779 /* Fire data frozen trigger just before we copy the data */
780 jbd2_buffer_frozen_trigger(jh
, source
+ offset
,
782 memcpy(jh
->b_frozen_data
, source
+offset
, jh2bh(jh
)->b_size
);
783 kunmap_atomic(source
, KM_USER0
);
786 * Now that the frozen data is saved off, we need to store
787 * any matching triggers.
789 jh
->b_frozen_triggers
= jh
->b_triggers
;
791 jbd_unlock_bh_state(bh
);
794 * If we are about to journal a buffer, then any revoke pending on it is
797 jbd2_journal_cancel_revoke(handle
, jh
);
800 if (unlikely(frozen_buffer
)) /* It's usually NULL */
801 jbd2_free(frozen_buffer
, bh
->b_size
);
803 JBUFFER_TRACE(jh
, "exit");
808 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
809 * @handle: transaction to add buffer modifications to
810 * @bh: bh to be used for metadata writes
811 * @credits: variable that will receive credits for the buffer
813 * Returns an error code or 0 on success.
815 * In full data journalling mode the buffer may be of type BJ_AsyncData,
816 * because we're write()ing a buffer which is also part of a shared mapping.
819 int jbd2_journal_get_write_access(handle_t
*handle
, struct buffer_head
*bh
)
821 struct journal_head
*jh
= jbd2_journal_add_journal_head(bh
);
824 /* We do not want to get caught playing with fields which the
825 * log thread also manipulates. Make sure that the buffer
826 * completes any outstanding IO before proceeding. */
827 rc
= do_get_write_access(handle
, jh
, 0);
828 jbd2_journal_put_journal_head(jh
);
834 * When the user wants to journal a newly created buffer_head
835 * (ie. getblk() returned a new buffer and we are going to populate it
836 * manually rather than reading off disk), then we need to keep the
837 * buffer_head locked until it has been completely filled with new
838 * data. In this case, we should be able to make the assertion that
839 * the bh is not already part of an existing transaction.
841 * The buffer should already be locked by the caller by this point.
842 * There is no lock ranking violation: it was a newly created,
843 * unlocked buffer beforehand. */
846 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
847 * @handle: transaction to new buffer to
850 * Call this if you create a new bh.
852 int jbd2_journal_get_create_access(handle_t
*handle
, struct buffer_head
*bh
)
854 transaction_t
*transaction
= handle
->h_transaction
;
855 journal_t
*journal
= transaction
->t_journal
;
856 struct journal_head
*jh
= jbd2_journal_add_journal_head(bh
);
859 jbd_debug(5, "journal_head %p\n", jh
);
861 if (is_handle_aborted(handle
))
865 JBUFFER_TRACE(jh
, "entry");
867 * The buffer may already belong to this transaction due to pre-zeroing
868 * in the filesystem's new_block code. It may also be on the previous,
869 * committing transaction's lists, but it HAS to be in Forget state in
870 * that case: the transaction must have deleted the buffer for it to be
873 jbd_lock_bh_state(bh
);
874 spin_lock(&journal
->j_list_lock
);
875 J_ASSERT_JH(jh
, (jh
->b_transaction
== transaction
||
876 jh
->b_transaction
== NULL
||
877 (jh
->b_transaction
== journal
->j_committing_transaction
&&
878 jh
->b_jlist
== BJ_Forget
)));
880 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
881 J_ASSERT_JH(jh
, buffer_locked(jh2bh(jh
)));
883 if (jh
->b_transaction
== NULL
) {
885 * Previous jbd2_journal_forget() could have left the buffer
886 * with jbddirty bit set because it was being committed. When
887 * the commit finished, we've filed the buffer for
888 * checkpointing and marked it dirty. Now we are reallocating
889 * the buffer so the transaction freeing it must have
890 * committed and so it's safe to clear the dirty bit.
892 clear_buffer_dirty(jh2bh(jh
));
893 jh
->b_transaction
= transaction
;
895 /* first access by this transaction */
898 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
899 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Reserved
);
900 } else if (jh
->b_transaction
== journal
->j_committing_transaction
) {
901 /* first access by this transaction */
904 JBUFFER_TRACE(jh
, "set next transaction");
905 jh
->b_next_transaction
= transaction
;
907 spin_unlock(&journal
->j_list_lock
);
908 jbd_unlock_bh_state(bh
);
911 * akpm: I added this. ext3_alloc_branch can pick up new indirect
912 * blocks which contain freed but then revoked metadata. We need
913 * to cancel the revoke in case we end up freeing it yet again
914 * and the reallocating as data - this would cause a second revoke,
915 * which hits an assertion error.
917 JBUFFER_TRACE(jh
, "cancelling revoke");
918 jbd2_journal_cancel_revoke(handle
, jh
);
919 jbd2_journal_put_journal_head(jh
);
925 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
926 * non-rewindable consequences
927 * @handle: transaction
928 * @bh: buffer to undo
929 * @credits: store the number of taken credits here (if not NULL)
931 * Sometimes there is a need to distinguish between metadata which has
932 * been committed to disk and that which has not. The ext3fs code uses
933 * this for freeing and allocating space, we have to make sure that we
934 * do not reuse freed space until the deallocation has been committed,
935 * since if we overwrote that space we would make the delete
936 * un-rewindable in case of a crash.
938 * To deal with that, jbd2_journal_get_undo_access requests write access to a
939 * buffer for parts of non-rewindable operations such as delete
940 * operations on the bitmaps. The journaling code must keep a copy of
941 * the buffer's contents prior to the undo_access call until such time
942 * as we know that the buffer has definitely been committed to disk.
944 * We never need to know which transaction the committed data is part
945 * of, buffers touched here are guaranteed to be dirtied later and so
946 * will be committed to a new transaction in due course, at which point
947 * we can discard the old committed data pointer.
949 * Returns error number or 0 on success.
951 int jbd2_journal_get_undo_access(handle_t
*handle
, struct buffer_head
*bh
)
954 struct journal_head
*jh
= jbd2_journal_add_journal_head(bh
);
955 char *committed_data
= NULL
;
957 JBUFFER_TRACE(jh
, "entry");
960 * Do this first --- it can drop the journal lock, so we want to
961 * make sure that obtaining the committed_data is done
962 * atomically wrt. completion of any outstanding commits.
964 err
= do_get_write_access(handle
, jh
, 1);
969 if (!jh
->b_committed_data
) {
970 committed_data
= jbd2_alloc(jh2bh(jh
)->b_size
, GFP_NOFS
);
971 if (!committed_data
) {
972 printk(KERN_EMERG
"%s: No memory for committed data\n",
979 jbd_lock_bh_state(bh
);
980 if (!jh
->b_committed_data
) {
981 /* Copy out the current buffer contents into the
982 * preserved, committed copy. */
983 JBUFFER_TRACE(jh
, "generate b_committed data");
984 if (!committed_data
) {
985 jbd_unlock_bh_state(bh
);
989 jh
->b_committed_data
= committed_data
;
990 committed_data
= NULL
;
991 memcpy(jh
->b_committed_data
, bh
->b_data
, bh
->b_size
);
993 jbd_unlock_bh_state(bh
);
995 jbd2_journal_put_journal_head(jh
);
996 if (unlikely(committed_data
))
997 jbd2_free(committed_data
, bh
->b_size
);
1002 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1003 * @bh: buffer to trigger on
1004 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1006 * Set any triggers on this journal_head. This is always safe, because
1007 * triggers for a committing buffer will be saved off, and triggers for
1008 * a running transaction will match the buffer in that transaction.
1010 * Call with NULL to clear the triggers.
1012 void jbd2_journal_set_triggers(struct buffer_head
*bh
,
1013 struct jbd2_buffer_trigger_type
*type
)
1015 struct journal_head
*jh
= bh2jh(bh
);
1017 jh
->b_triggers
= type
;
1020 void jbd2_buffer_frozen_trigger(struct journal_head
*jh
, void *mapped_data
,
1021 struct jbd2_buffer_trigger_type
*triggers
)
1023 struct buffer_head
*bh
= jh2bh(jh
);
1025 if (!triggers
|| !triggers
->t_frozen
)
1028 triggers
->t_frozen(triggers
, bh
, mapped_data
, bh
->b_size
);
1031 void jbd2_buffer_abort_trigger(struct journal_head
*jh
,
1032 struct jbd2_buffer_trigger_type
*triggers
)
1034 if (!triggers
|| !triggers
->t_abort
)
1037 triggers
->t_abort(triggers
, jh2bh(jh
));
1043 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1044 * @handle: transaction to add buffer to.
1045 * @bh: buffer to mark
1047 * mark dirty metadata which needs to be journaled as part of the current
1050 * The buffer is placed on the transaction's metadata list and is marked
1051 * as belonging to the transaction.
1053 * Returns error number or 0 on success.
1055 * Special care needs to be taken if the buffer already belongs to the
1056 * current committing transaction (in which case we should have frozen
1057 * data present for that commit). In that case, we don't relink the
1058 * buffer: that only gets done when the old transaction finally
1059 * completes its commit.
1061 int jbd2_journal_dirty_metadata(handle_t
*handle
, struct buffer_head
*bh
)
1063 transaction_t
*transaction
= handle
->h_transaction
;
1064 journal_t
*journal
= transaction
->t_journal
;
1065 struct journal_head
*jh
= bh2jh(bh
);
1067 jbd_debug(5, "journal_head %p\n", jh
);
1068 JBUFFER_TRACE(jh
, "entry");
1069 if (is_handle_aborted(handle
))
1072 jbd_lock_bh_state(bh
);
1074 if (jh
->b_modified
== 0) {
1076 * This buffer's got modified and becoming part
1077 * of the transaction. This needs to be done
1078 * once a transaction -bzzz
1081 J_ASSERT_JH(jh
, handle
->h_buffer_credits
> 0);
1082 handle
->h_buffer_credits
--;
1086 * fastpath, to avoid expensive locking. If this buffer is already
1087 * on the running transaction's metadata list there is nothing to do.
1088 * Nobody can take it off again because there is a handle open.
1089 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1090 * result in this test being false, so we go in and take the locks.
1092 if (jh
->b_transaction
== transaction
&& jh
->b_jlist
== BJ_Metadata
) {
1093 JBUFFER_TRACE(jh
, "fastpath");
1094 J_ASSERT_JH(jh
, jh
->b_transaction
==
1095 journal
->j_running_transaction
);
1099 set_buffer_jbddirty(bh
);
1102 * Metadata already on the current transaction list doesn't
1103 * need to be filed. Metadata on another transaction's list must
1104 * be committing, and will be refiled once the commit completes:
1105 * leave it alone for now.
1107 if (jh
->b_transaction
!= transaction
) {
1108 JBUFFER_TRACE(jh
, "already on other transaction");
1109 J_ASSERT_JH(jh
, jh
->b_transaction
==
1110 journal
->j_committing_transaction
);
1111 J_ASSERT_JH(jh
, jh
->b_next_transaction
== transaction
);
1112 /* And this case is illegal: we can't reuse another
1113 * transaction's data buffer, ever. */
1117 /* That test should have eliminated the following case: */
1118 J_ASSERT_JH(jh
, jh
->b_frozen_data
== NULL
);
1120 JBUFFER_TRACE(jh
, "file as BJ_Metadata");
1121 spin_lock(&journal
->j_list_lock
);
1122 __jbd2_journal_file_buffer(jh
, handle
->h_transaction
, BJ_Metadata
);
1123 spin_unlock(&journal
->j_list_lock
);
1125 jbd_unlock_bh_state(bh
);
1127 JBUFFER_TRACE(jh
, "exit");
1132 * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1133 * updates, if the update decided in the end that it didn't need access.
1137 jbd2_journal_release_buffer(handle_t
*handle
, struct buffer_head
*bh
)
1139 BUFFER_TRACE(bh
, "entry");
1143 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1144 * @handle: transaction handle
1145 * @bh: bh to 'forget'
1147 * We can only do the bforget if there are no commits pending against the
1148 * buffer. If the buffer is dirty in the current running transaction we
1149 * can safely unlink it.
1151 * bh may not be a journalled buffer at all - it may be a non-JBD
1152 * buffer which came off the hashtable. Check for this.
1154 * Decrements bh->b_count by one.
1156 * Allow this call even if the handle has aborted --- it may be part of
1157 * the caller's cleanup after an abort.
1159 int jbd2_journal_forget (handle_t
*handle
, struct buffer_head
*bh
)
1161 transaction_t
*transaction
= handle
->h_transaction
;
1162 journal_t
*journal
= transaction
->t_journal
;
1163 struct journal_head
*jh
;
1164 int drop_reserve
= 0;
1166 int was_modified
= 0;
1168 BUFFER_TRACE(bh
, "entry");
1170 jbd_lock_bh_state(bh
);
1171 spin_lock(&journal
->j_list_lock
);
1173 if (!buffer_jbd(bh
))
1177 /* Critical error: attempting to delete a bitmap buffer, maybe?
1178 * Don't do any jbd operations, and return an error. */
1179 if (!J_EXPECT_JH(jh
, !jh
->b_committed_data
,
1180 "inconsistent data on disk")) {
1185 /* keep track of wether or not this transaction modified us */
1186 was_modified
= jh
->b_modified
;
1189 * The buffer's going from the transaction, we must drop
1190 * all references -bzzz
1194 if (jh
->b_transaction
== handle
->h_transaction
) {
1195 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
1197 /* If we are forgetting a buffer which is already part
1198 * of this transaction, then we can just drop it from
1199 * the transaction immediately. */
1200 clear_buffer_dirty(bh
);
1201 clear_buffer_jbddirty(bh
);
1203 JBUFFER_TRACE(jh
, "belongs to current transaction: unfile");
1206 * we only want to drop a reference if this transaction
1207 * modified the buffer
1213 * We are no longer going to journal this buffer.
1214 * However, the commit of this transaction is still
1215 * important to the buffer: the delete that we are now
1216 * processing might obsolete an old log entry, so by
1217 * committing, we can satisfy the buffer's checkpoint.
1219 * So, if we have a checkpoint on the buffer, we should
1220 * now refile the buffer on our BJ_Forget list so that
1221 * we know to remove the checkpoint after we commit.
1224 if (jh
->b_cp_transaction
) {
1225 __jbd2_journal_temp_unlink_buffer(jh
);
1226 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Forget
);
1228 __jbd2_journal_unfile_buffer(jh
);
1229 jbd2_journal_remove_journal_head(bh
);
1231 if (!buffer_jbd(bh
)) {
1232 spin_unlock(&journal
->j_list_lock
);
1233 jbd_unlock_bh_state(bh
);
1238 } else if (jh
->b_transaction
) {
1239 J_ASSERT_JH(jh
, (jh
->b_transaction
==
1240 journal
->j_committing_transaction
));
1241 /* However, if the buffer is still owned by a prior
1242 * (committing) transaction, we can't drop it yet... */
1243 JBUFFER_TRACE(jh
, "belongs to older transaction");
1244 /* ... but we CAN drop it from the new transaction if we
1245 * have also modified it since the original commit. */
1247 if (jh
->b_next_transaction
) {
1248 J_ASSERT(jh
->b_next_transaction
== transaction
);
1249 jh
->b_next_transaction
= NULL
;
1252 * only drop a reference if this transaction modified
1261 spin_unlock(&journal
->j_list_lock
);
1262 jbd_unlock_bh_state(bh
);
1266 /* no need to reserve log space for this block -bzzz */
1267 handle
->h_buffer_credits
++;
1273 * int jbd2_journal_stop() - complete a transaction
1274 * @handle: tranaction to complete.
1276 * All done for a particular handle.
1278 * There is not much action needed here. We just return any remaining
1279 * buffer credits to the transaction and remove the handle. The only
1280 * complication is that we need to start a commit operation if the
1281 * filesystem is marked for synchronous update.
1283 * jbd2_journal_stop itself will not usually return an error, but it may
1284 * do so in unusual circumstances. In particular, expect it to
1285 * return -EIO if a jbd2_journal_abort has been executed since the
1286 * transaction began.
1288 int jbd2_journal_stop(handle_t
*handle
)
1290 transaction_t
*transaction
= handle
->h_transaction
;
1291 journal_t
*journal
= transaction
->t_journal
;
1292 int err
, wait_for_commit
= 0;
1296 J_ASSERT(journal_current_handle() == handle
);
1298 if (is_handle_aborted(handle
))
1301 J_ASSERT(atomic_read(&transaction
->t_updates
) > 0);
1305 if (--handle
->h_ref
> 0) {
1306 jbd_debug(4, "h_ref %d -> %d\n", handle
->h_ref
+ 1,
1311 jbd_debug(4, "Handle %p going down\n", handle
);
1314 * Implement synchronous transaction batching. If the handle
1315 * was synchronous, don't force a commit immediately. Let's
1316 * yield and let another thread piggyback onto this
1317 * transaction. Keep doing that while new threads continue to
1318 * arrive. It doesn't cost much - we're about to run a commit
1319 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1320 * operations by 30x or more...
1322 * We try and optimize the sleep time against what the
1323 * underlying disk can do, instead of having a static sleep
1324 * time. This is useful for the case where our storage is so
1325 * fast that it is more optimal to go ahead and force a flush
1326 * and wait for the transaction to be committed than it is to
1327 * wait for an arbitrary amount of time for new writers to
1328 * join the transaction. We achieve this by measuring how
1329 * long it takes to commit a transaction, and compare it with
1330 * how long this transaction has been running, and if run time
1331 * < commit time then we sleep for the delta and commit. This
1332 * greatly helps super fast disks that would see slowdowns as
1333 * more threads started doing fsyncs.
1335 * But don't do this if this process was the most recent one
1336 * to perform a synchronous write. We do this to detect the
1337 * case where a single process is doing a stream of sync
1338 * writes. No point in waiting for joiners in that case.
1341 if (handle
->h_sync
&& journal
->j_last_sync_writer
!= pid
) {
1342 u64 commit_time
, trans_time
;
1344 journal
->j_last_sync_writer
= pid
;
1346 read_lock(&journal
->j_state_lock
);
1347 commit_time
= journal
->j_average_commit_time
;
1348 read_unlock(&journal
->j_state_lock
);
1350 trans_time
= ktime_to_ns(ktime_sub(ktime_get(),
1351 transaction
->t_start_time
));
1353 commit_time
= max_t(u64
, commit_time
,
1354 1000*journal
->j_min_batch_time
);
1355 commit_time
= min_t(u64
, commit_time
,
1356 1000*journal
->j_max_batch_time
);
1358 if (trans_time
< commit_time
) {
1359 ktime_t expires
= ktime_add_ns(ktime_get(),
1361 set_current_state(TASK_UNINTERRUPTIBLE
);
1362 schedule_hrtimeout(&expires
, HRTIMER_MODE_ABS
);
1367 transaction
->t_synchronous_commit
= 1;
1368 current
->journal_info
= NULL
;
1369 atomic_sub(handle
->h_buffer_credits
,
1370 &transaction
->t_outstanding_credits
);
1373 * If the handle is marked SYNC, we need to set another commit
1374 * going! We also want to force a commit if the current
1375 * transaction is occupying too much of the log, or if the
1376 * transaction is too old now.
1378 if (handle
->h_sync
||
1379 (atomic_read(&transaction
->t_outstanding_credits
) >
1380 journal
->j_max_transaction_buffers
) ||
1381 time_after_eq(jiffies
, transaction
->t_expires
)) {
1382 /* Do this even for aborted journals: an abort still
1383 * completes the commit thread, it just doesn't write
1384 * anything to disk. */
1386 jbd_debug(2, "transaction too old, requesting commit for "
1387 "handle %p\n", handle
);
1388 /* This is non-blocking */
1389 jbd2_log_start_commit(journal
, transaction
->t_tid
);
1392 * Special case: JBD2_SYNC synchronous updates require us
1393 * to wait for the commit to complete.
1395 if (handle
->h_sync
&& !(current
->flags
& PF_MEMALLOC
))
1396 wait_for_commit
= 1;
1400 * Once we drop t_updates, if it goes to zero the transaction
1401 * could start commiting on us and eventually disappear. So
1402 * once we do this, we must not dereference transaction
1405 tid
= transaction
->t_tid
;
1406 if (atomic_dec_and_test(&transaction
->t_updates
)) {
1407 wake_up(&journal
->j_wait_updates
);
1408 if (journal
->j_barrier_count
)
1409 wake_up(&journal
->j_wait_transaction_locked
);
1412 if (wait_for_commit
)
1413 err
= jbd2_log_wait_commit(journal
, tid
);
1415 lock_map_release(&handle
->h_lockdep_map
);
1417 jbd2_free_handle(handle
);
1422 * int jbd2_journal_force_commit() - force any uncommitted transactions
1423 * @journal: journal to force
1425 * For synchronous operations: force any uncommitted transactions
1426 * to disk. May seem kludgy, but it reuses all the handle batching
1427 * code in a very simple manner.
1429 int jbd2_journal_force_commit(journal_t
*journal
)
1434 handle
= jbd2_journal_start(journal
, 1);
1435 if (IS_ERR(handle
)) {
1436 ret
= PTR_ERR(handle
);
1439 ret
= jbd2_journal_stop(handle
);
1446 * List management code snippets: various functions for manipulating the
1447 * transaction buffer lists.
1452 * Append a buffer to a transaction list, given the transaction's list head
1455 * j_list_lock is held.
1457 * jbd_lock_bh_state(jh2bh(jh)) is held.
1461 __blist_add_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1464 jh
->b_tnext
= jh
->b_tprev
= jh
;
1467 /* Insert at the tail of the list to preserve order */
1468 struct journal_head
*first
= *list
, *last
= first
->b_tprev
;
1470 jh
->b_tnext
= first
;
1471 last
->b_tnext
= first
->b_tprev
= jh
;
1476 * Remove a buffer from a transaction list, given the transaction's list
1479 * Called with j_list_lock held, and the journal may not be locked.
1481 * jbd_lock_bh_state(jh2bh(jh)) is held.
1485 __blist_del_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1488 *list
= jh
->b_tnext
;
1492 jh
->b_tprev
->b_tnext
= jh
->b_tnext
;
1493 jh
->b_tnext
->b_tprev
= jh
->b_tprev
;
1497 * Remove a buffer from the appropriate transaction list.
1499 * Note that this function can *change* the value of
1500 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1501 * t_log_list or t_reserved_list. If the caller is holding onto a copy of one
1502 * of these pointers, it could go bad. Generally the caller needs to re-read
1503 * the pointer from the transaction_t.
1505 * Called under j_list_lock. The journal may not be locked.
1507 void __jbd2_journal_temp_unlink_buffer(struct journal_head
*jh
)
1509 struct journal_head
**list
= NULL
;
1510 transaction_t
*transaction
;
1511 struct buffer_head
*bh
= jh2bh(jh
);
1513 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
1514 transaction
= jh
->b_transaction
;
1516 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
1518 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
1519 if (jh
->b_jlist
!= BJ_None
)
1520 J_ASSERT_JH(jh
, transaction
!= NULL
);
1522 switch (jh
->b_jlist
) {
1526 transaction
->t_nr_buffers
--;
1527 J_ASSERT_JH(jh
, transaction
->t_nr_buffers
>= 0);
1528 list
= &transaction
->t_buffers
;
1531 list
= &transaction
->t_forget
;
1534 list
= &transaction
->t_iobuf_list
;
1537 list
= &transaction
->t_shadow_list
;
1540 list
= &transaction
->t_log_list
;
1543 list
= &transaction
->t_reserved_list
;
1547 __blist_del_buffer(list
, jh
);
1548 jh
->b_jlist
= BJ_None
;
1549 if (test_clear_buffer_jbddirty(bh
))
1550 mark_buffer_dirty(bh
); /* Expose it to the VM */
1553 void __jbd2_journal_unfile_buffer(struct journal_head
*jh
)
1555 __jbd2_journal_temp_unlink_buffer(jh
);
1556 jh
->b_transaction
= NULL
;
1559 void jbd2_journal_unfile_buffer(journal_t
*journal
, struct journal_head
*jh
)
1561 jbd_lock_bh_state(jh2bh(jh
));
1562 spin_lock(&journal
->j_list_lock
);
1563 __jbd2_journal_unfile_buffer(jh
);
1564 spin_unlock(&journal
->j_list_lock
);
1565 jbd_unlock_bh_state(jh2bh(jh
));
1569 * Called from jbd2_journal_try_to_free_buffers().
1571 * Called under jbd_lock_bh_state(bh)
1574 __journal_try_to_free_buffer(journal_t
*journal
, struct buffer_head
*bh
)
1576 struct journal_head
*jh
;
1580 if (buffer_locked(bh
) || buffer_dirty(bh
))
1583 if (jh
->b_next_transaction
!= NULL
)
1586 spin_lock(&journal
->j_list_lock
);
1587 if (jh
->b_cp_transaction
!= NULL
&& jh
->b_transaction
== NULL
) {
1588 /* written-back checkpointed metadata buffer */
1589 if (jh
->b_jlist
== BJ_None
) {
1590 JBUFFER_TRACE(jh
, "remove from checkpoint list");
1591 __jbd2_journal_remove_checkpoint(jh
);
1592 jbd2_journal_remove_journal_head(bh
);
1596 spin_unlock(&journal
->j_list_lock
);
1602 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1603 * @journal: journal for operation
1604 * @page: to try and free
1605 * @gfp_mask: we use the mask to detect how hard should we try to release
1606 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1607 * release the buffers.
1610 * For all the buffers on this page,
1611 * if they are fully written out ordered data, move them onto BUF_CLEAN
1612 * so try_to_free_buffers() can reap them.
1614 * This function returns non-zero if we wish try_to_free_buffers()
1615 * to be called. We do this if the page is releasable by try_to_free_buffers().
1616 * We also do it if the page has locked or dirty buffers and the caller wants
1617 * us to perform sync or async writeout.
1619 * This complicates JBD locking somewhat. We aren't protected by the
1620 * BKL here. We wish to remove the buffer from its committing or
1621 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1623 * This may *change* the value of transaction_t->t_datalist, so anyone
1624 * who looks at t_datalist needs to lock against this function.
1626 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1627 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
1628 * will come out of the lock with the buffer dirty, which makes it
1629 * ineligible for release here.
1631 * Who else is affected by this? hmm... Really the only contender
1632 * is do_get_write_access() - it could be looking at the buffer while
1633 * journal_try_to_free_buffer() is changing its state. But that
1634 * cannot happen because we never reallocate freed data as metadata
1635 * while the data is part of a transaction. Yes?
1637 * Return 0 on failure, 1 on success
1639 int jbd2_journal_try_to_free_buffers(journal_t
*journal
,
1640 struct page
*page
, gfp_t gfp_mask
)
1642 struct buffer_head
*head
;
1643 struct buffer_head
*bh
;
1646 J_ASSERT(PageLocked(page
));
1648 head
= page_buffers(page
);
1651 struct journal_head
*jh
;
1654 * We take our own ref against the journal_head here to avoid
1655 * having to add tons of locking around each instance of
1656 * jbd2_journal_remove_journal_head() and
1657 * jbd2_journal_put_journal_head().
1659 jh
= jbd2_journal_grab_journal_head(bh
);
1663 jbd_lock_bh_state(bh
);
1664 __journal_try_to_free_buffer(journal
, bh
);
1665 jbd2_journal_put_journal_head(jh
);
1666 jbd_unlock_bh_state(bh
);
1669 } while ((bh
= bh
->b_this_page
) != head
);
1671 ret
= try_to_free_buffers(page
);
1678 * This buffer is no longer needed. If it is on an older transaction's
1679 * checkpoint list we need to record it on this transaction's forget list
1680 * to pin this buffer (and hence its checkpointing transaction) down until
1681 * this transaction commits. If the buffer isn't on a checkpoint list, we
1683 * Returns non-zero if JBD no longer has an interest in the buffer.
1685 * Called under j_list_lock.
1687 * Called under jbd_lock_bh_state(bh).
1689 static int __dispose_buffer(struct journal_head
*jh
, transaction_t
*transaction
)
1692 struct buffer_head
*bh
= jh2bh(jh
);
1694 __jbd2_journal_unfile_buffer(jh
);
1696 if (jh
->b_cp_transaction
) {
1697 JBUFFER_TRACE(jh
, "on running+cp transaction");
1699 * We don't want to write the buffer anymore, clear the
1700 * bit so that we don't confuse checks in
1701 * __journal_file_buffer
1703 clear_buffer_dirty(bh
);
1704 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Forget
);
1707 JBUFFER_TRACE(jh
, "on running transaction");
1708 jbd2_journal_remove_journal_head(bh
);
1715 * jbd2_journal_invalidatepage
1717 * This code is tricky. It has a number of cases to deal with.
1719 * There are two invariants which this code relies on:
1721 * i_size must be updated on disk before we start calling invalidatepage on the
1724 * This is done in ext3 by defining an ext3_setattr method which
1725 * updates i_size before truncate gets going. By maintaining this
1726 * invariant, we can be sure that it is safe to throw away any buffers
1727 * attached to the current transaction: once the transaction commits,
1728 * we know that the data will not be needed.
1730 * Note however that we can *not* throw away data belonging to the
1731 * previous, committing transaction!
1733 * Any disk blocks which *are* part of the previous, committing
1734 * transaction (and which therefore cannot be discarded immediately) are
1735 * not going to be reused in the new running transaction
1737 * The bitmap committed_data images guarantee this: any block which is
1738 * allocated in one transaction and removed in the next will be marked
1739 * as in-use in the committed_data bitmap, so cannot be reused until
1740 * the next transaction to delete the block commits. This means that
1741 * leaving committing buffers dirty is quite safe: the disk blocks
1742 * cannot be reallocated to a different file and so buffer aliasing is
1746 * The above applies mainly to ordered data mode. In writeback mode we
1747 * don't make guarantees about the order in which data hits disk --- in
1748 * particular we don't guarantee that new dirty data is flushed before
1749 * transaction commit --- so it is always safe just to discard data
1750 * immediately in that mode. --sct
1754 * The journal_unmap_buffer helper function returns zero if the buffer
1755 * concerned remains pinned as an anonymous buffer belonging to an older
1758 * We're outside-transaction here. Either or both of j_running_transaction
1759 * and j_committing_transaction may be NULL.
1761 static int journal_unmap_buffer(journal_t
*journal
, struct buffer_head
*bh
)
1763 transaction_t
*transaction
;
1764 struct journal_head
*jh
;
1768 BUFFER_TRACE(bh
, "entry");
1771 * It is safe to proceed here without the j_list_lock because the
1772 * buffers cannot be stolen by try_to_free_buffers as long as we are
1773 * holding the page lock. --sct
1776 if (!buffer_jbd(bh
))
1777 goto zap_buffer_unlocked
;
1779 /* OK, we have data buffer in journaled mode */
1780 write_lock(&journal
->j_state_lock
);
1781 jbd_lock_bh_state(bh
);
1782 spin_lock(&journal
->j_list_lock
);
1784 jh
= jbd2_journal_grab_journal_head(bh
);
1786 goto zap_buffer_no_jh
;
1789 * We cannot remove the buffer from checkpoint lists until the
1790 * transaction adding inode to orphan list (let's call it T)
1791 * is committed. Otherwise if the transaction changing the
1792 * buffer would be cleaned from the journal before T is
1793 * committed, a crash will cause that the correct contents of
1794 * the buffer will be lost. On the other hand we have to
1795 * clear the buffer dirty bit at latest at the moment when the
1796 * transaction marking the buffer as freed in the filesystem
1797 * structures is committed because from that moment on the
1798 * buffer can be reallocated and used by a different page.
1799 * Since the block hasn't been freed yet but the inode has
1800 * already been added to orphan list, it is safe for us to add
1801 * the buffer to BJ_Forget list of the newest transaction.
1803 transaction
= jh
->b_transaction
;
1804 if (transaction
== NULL
) {
1805 /* First case: not on any transaction. If it
1806 * has no checkpoint link, then we can zap it:
1807 * it's a writeback-mode buffer so we don't care
1808 * if it hits disk safely. */
1809 if (!jh
->b_cp_transaction
) {
1810 JBUFFER_TRACE(jh
, "not on any transaction: zap");
1814 if (!buffer_dirty(bh
)) {
1815 /* bdflush has written it. We can drop it now */
1819 /* OK, it must be in the journal but still not
1820 * written fully to disk: it's metadata or
1821 * journaled data... */
1823 if (journal
->j_running_transaction
) {
1824 /* ... and once the current transaction has
1825 * committed, the buffer won't be needed any
1827 JBUFFER_TRACE(jh
, "checkpointed: add to BJ_Forget");
1828 ret
= __dispose_buffer(jh
,
1829 journal
->j_running_transaction
);
1830 jbd2_journal_put_journal_head(jh
);
1831 spin_unlock(&journal
->j_list_lock
);
1832 jbd_unlock_bh_state(bh
);
1833 write_unlock(&journal
->j_state_lock
);
1836 /* There is no currently-running transaction. So the
1837 * orphan record which we wrote for this file must have
1838 * passed into commit. We must attach this buffer to
1839 * the committing transaction, if it exists. */
1840 if (journal
->j_committing_transaction
) {
1841 JBUFFER_TRACE(jh
, "give to committing trans");
1842 ret
= __dispose_buffer(jh
,
1843 journal
->j_committing_transaction
);
1844 jbd2_journal_put_journal_head(jh
);
1845 spin_unlock(&journal
->j_list_lock
);
1846 jbd_unlock_bh_state(bh
);
1847 write_unlock(&journal
->j_state_lock
);
1850 /* The orphan record's transaction has
1851 * committed. We can cleanse this buffer */
1852 clear_buffer_jbddirty(bh
);
1856 } else if (transaction
== journal
->j_committing_transaction
) {
1857 JBUFFER_TRACE(jh
, "on committing transaction");
1859 * The buffer is committing, we simply cannot touch
1860 * it. So we just set j_next_transaction to the
1861 * running transaction (if there is one) and mark
1862 * buffer as freed so that commit code knows it should
1863 * clear dirty bits when it is done with the buffer.
1865 set_buffer_freed(bh
);
1866 if (journal
->j_running_transaction
&& buffer_jbddirty(bh
))
1867 jh
->b_next_transaction
= journal
->j_running_transaction
;
1868 jbd2_journal_put_journal_head(jh
);
1869 spin_unlock(&journal
->j_list_lock
);
1870 jbd_unlock_bh_state(bh
);
1871 write_unlock(&journal
->j_state_lock
);
1874 /* Good, the buffer belongs to the running transaction.
1875 * We are writing our own transaction's data, not any
1876 * previous one's, so it is safe to throw it away
1877 * (remember that we expect the filesystem to have set
1878 * i_size already for this truncate so recovery will not
1879 * expose the disk blocks we are discarding here.) */
1880 J_ASSERT_JH(jh
, transaction
== journal
->j_running_transaction
);
1881 JBUFFER_TRACE(jh
, "on running transaction");
1882 may_free
= __dispose_buffer(jh
, transaction
);
1886 jbd2_journal_put_journal_head(jh
);
1888 spin_unlock(&journal
->j_list_lock
);
1889 jbd_unlock_bh_state(bh
);
1890 write_unlock(&journal
->j_state_lock
);
1891 zap_buffer_unlocked
:
1892 clear_buffer_dirty(bh
);
1893 J_ASSERT_BH(bh
, !buffer_jbddirty(bh
));
1894 clear_buffer_mapped(bh
);
1895 clear_buffer_req(bh
);
1896 clear_buffer_new(bh
);
1902 * void jbd2_journal_invalidatepage()
1903 * @journal: journal to use for flush...
1904 * @page: page to flush
1905 * @offset: length of page to invalidate.
1907 * Reap page buffers containing data after offset in page.
1910 void jbd2_journal_invalidatepage(journal_t
*journal
,
1912 unsigned long offset
)
1914 struct buffer_head
*head
, *bh
, *next
;
1915 unsigned int curr_off
= 0;
1918 if (!PageLocked(page
))
1920 if (!page_has_buffers(page
))
1923 /* We will potentially be playing with lists other than just the
1924 * data lists (especially for journaled data mode), so be
1925 * cautious in our locking. */
1927 head
= bh
= page_buffers(page
);
1929 unsigned int next_off
= curr_off
+ bh
->b_size
;
1930 next
= bh
->b_this_page
;
1932 if (offset
<= curr_off
) {
1933 /* This block is wholly outside the truncation point */
1935 may_free
&= journal_unmap_buffer(journal
, bh
);
1938 curr_off
= next_off
;
1941 } while (bh
!= head
);
1944 if (may_free
&& try_to_free_buffers(page
))
1945 J_ASSERT(!page_has_buffers(page
));
1950 * File a buffer on the given transaction list.
1952 void __jbd2_journal_file_buffer(struct journal_head
*jh
,
1953 transaction_t
*transaction
, int jlist
)
1955 struct journal_head
**list
= NULL
;
1957 struct buffer_head
*bh
= jh2bh(jh
);
1959 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
1960 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
1962 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
1963 J_ASSERT_JH(jh
, jh
->b_transaction
== transaction
||
1964 jh
->b_transaction
== NULL
);
1966 if (jh
->b_transaction
&& jh
->b_jlist
== jlist
)
1969 if (jlist
== BJ_Metadata
|| jlist
== BJ_Reserved
||
1970 jlist
== BJ_Shadow
|| jlist
== BJ_Forget
) {
1972 * For metadata buffers, we track dirty bit in buffer_jbddirty
1973 * instead of buffer_dirty. We should not see a dirty bit set
1974 * here because we clear it in do_get_write_access but e.g.
1975 * tune2fs can modify the sb and set the dirty bit at any time
1976 * so we try to gracefully handle that.
1978 if (buffer_dirty(bh
))
1979 warn_dirty_buffer(bh
);
1980 if (test_clear_buffer_dirty(bh
) ||
1981 test_clear_buffer_jbddirty(bh
))
1985 if (jh
->b_transaction
)
1986 __jbd2_journal_temp_unlink_buffer(jh
);
1987 jh
->b_transaction
= transaction
;
1991 J_ASSERT_JH(jh
, !jh
->b_committed_data
);
1992 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
1995 transaction
->t_nr_buffers
++;
1996 list
= &transaction
->t_buffers
;
1999 list
= &transaction
->t_forget
;
2002 list
= &transaction
->t_iobuf_list
;
2005 list
= &transaction
->t_shadow_list
;
2008 list
= &transaction
->t_log_list
;
2011 list
= &transaction
->t_reserved_list
;
2015 __blist_add_buffer(list
, jh
);
2016 jh
->b_jlist
= jlist
;
2019 set_buffer_jbddirty(bh
);
2022 void jbd2_journal_file_buffer(struct journal_head
*jh
,
2023 transaction_t
*transaction
, int jlist
)
2025 jbd_lock_bh_state(jh2bh(jh
));
2026 spin_lock(&transaction
->t_journal
->j_list_lock
);
2027 __jbd2_journal_file_buffer(jh
, transaction
, jlist
);
2028 spin_unlock(&transaction
->t_journal
->j_list_lock
);
2029 jbd_unlock_bh_state(jh2bh(jh
));
2033 * Remove a buffer from its current buffer list in preparation for
2034 * dropping it from its current transaction entirely. If the buffer has
2035 * already started to be used by a subsequent transaction, refile the
2036 * buffer on that transaction's metadata list.
2038 * Called under journal->j_list_lock
2040 * Called under jbd_lock_bh_state(jh2bh(jh))
2042 void __jbd2_journal_refile_buffer(struct journal_head
*jh
)
2044 int was_dirty
, jlist
;
2045 struct buffer_head
*bh
= jh2bh(jh
);
2047 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
2048 if (jh
->b_transaction
)
2049 assert_spin_locked(&jh
->b_transaction
->t_journal
->j_list_lock
);
2051 /* If the buffer is now unused, just drop it. */
2052 if (jh
->b_next_transaction
== NULL
) {
2053 __jbd2_journal_unfile_buffer(jh
);
2058 * It has been modified by a later transaction: add it to the new
2059 * transaction's metadata list.
2062 was_dirty
= test_clear_buffer_jbddirty(bh
);
2063 __jbd2_journal_temp_unlink_buffer(jh
);
2064 jh
->b_transaction
= jh
->b_next_transaction
;
2065 jh
->b_next_transaction
= NULL
;
2066 if (buffer_freed(bh
))
2068 else if (jh
->b_modified
)
2069 jlist
= BJ_Metadata
;
2071 jlist
= BJ_Reserved
;
2072 __jbd2_journal_file_buffer(jh
, jh
->b_transaction
, jlist
);
2073 J_ASSERT_JH(jh
, jh
->b_transaction
->t_state
== T_RUNNING
);
2076 set_buffer_jbddirty(bh
);
2080 * For the unlocked version of this call, also make sure that any
2081 * hanging journal_head is cleaned up if necessary.
2083 * __jbd2_journal_refile_buffer is usually called as part of a single locked
2084 * operation on a buffer_head, in which the caller is probably going to
2085 * be hooking the journal_head onto other lists. In that case it is up
2086 * to the caller to remove the journal_head if necessary. For the
2087 * unlocked jbd2_journal_refile_buffer call, the caller isn't going to be
2088 * doing anything else to the buffer so we need to do the cleanup
2089 * ourselves to avoid a jh leak.
2091 * *** The journal_head may be freed by this call! ***
2093 void jbd2_journal_refile_buffer(journal_t
*journal
, struct journal_head
*jh
)
2095 struct buffer_head
*bh
= jh2bh(jh
);
2097 jbd_lock_bh_state(bh
);
2098 spin_lock(&journal
->j_list_lock
);
2100 __jbd2_journal_refile_buffer(jh
);
2101 jbd_unlock_bh_state(bh
);
2102 jbd2_journal_remove_journal_head(bh
);
2104 spin_unlock(&journal
->j_list_lock
);
2109 * File inode in the inode list of the handle's transaction
2111 int jbd2_journal_file_inode(handle_t
*handle
, struct jbd2_inode
*jinode
)
2113 transaction_t
*transaction
= handle
->h_transaction
;
2114 journal_t
*journal
= transaction
->t_journal
;
2116 if (is_handle_aborted(handle
))
2119 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode
->i_vfs_inode
->i_ino
,
2120 transaction
->t_tid
);
2123 * First check whether inode isn't already on the transaction's
2124 * lists without taking the lock. Note that this check is safe
2125 * without the lock as we cannot race with somebody removing inode
2126 * from the transaction. The reason is that we remove inode from the
2127 * transaction only in journal_release_jbd_inode() and when we commit
2128 * the transaction. We are guarded from the first case by holding
2129 * a reference to the inode. We are safe against the second case
2130 * because if jinode->i_transaction == transaction, commit code
2131 * cannot touch the transaction because we hold reference to it,
2132 * and if jinode->i_next_transaction == transaction, commit code
2133 * will only file the inode where we want it.
2135 if (jinode
->i_transaction
== transaction
||
2136 jinode
->i_next_transaction
== transaction
)
2139 spin_lock(&journal
->j_list_lock
);
2141 if (jinode
->i_transaction
== transaction
||
2142 jinode
->i_next_transaction
== transaction
)
2145 /* On some different transaction's list - should be
2146 * the committing one */
2147 if (jinode
->i_transaction
) {
2148 J_ASSERT(jinode
->i_next_transaction
== NULL
);
2149 J_ASSERT(jinode
->i_transaction
==
2150 journal
->j_committing_transaction
);
2151 jinode
->i_next_transaction
= transaction
;
2154 /* Not on any transaction list... */
2155 J_ASSERT(!jinode
->i_next_transaction
);
2156 jinode
->i_transaction
= transaction
;
2157 list_add(&jinode
->i_list
, &transaction
->t_inode_list
);
2159 spin_unlock(&journal
->j_list_lock
);
2165 * File truncate and transaction commit interact with each other in a
2166 * non-trivial way. If a transaction writing data block A is
2167 * committing, we cannot discard the data by truncate until we have
2168 * written them. Otherwise if we crashed after the transaction with
2169 * write has committed but before the transaction with truncate has
2170 * committed, we could see stale data in block A. This function is a
2171 * helper to solve this problem. It starts writeout of the truncated
2172 * part in case it is in the committing transaction.
2174 * Filesystem code must call this function when inode is journaled in
2175 * ordered mode before truncation happens and after the inode has been
2176 * placed on orphan list with the new inode size. The second condition
2177 * avoids the race that someone writes new data and we start
2178 * committing the transaction after this function has been called but
2179 * before a transaction for truncate is started (and furthermore it
2180 * allows us to optimize the case where the addition to orphan list
2181 * happens in the same transaction as write --- we don't have to write
2182 * any data in such case).
2184 int jbd2_journal_begin_ordered_truncate(journal_t
*journal
,
2185 struct jbd2_inode
*jinode
,
2188 transaction_t
*inode_trans
, *commit_trans
;
2191 /* This is a quick check to avoid locking if not necessary */
2192 if (!jinode
->i_transaction
)
2194 /* Locks are here just to force reading of recent values, it is
2195 * enough that the transaction was not committing before we started
2196 * a transaction adding the inode to orphan list */
2197 read_lock(&journal
->j_state_lock
);
2198 commit_trans
= journal
->j_committing_transaction
;
2199 read_unlock(&journal
->j_state_lock
);
2200 spin_lock(&journal
->j_list_lock
);
2201 inode_trans
= jinode
->i_transaction
;
2202 spin_unlock(&journal
->j_list_lock
);
2203 if (inode_trans
== commit_trans
) {
2204 ret
= filemap_fdatawrite_range(jinode
->i_vfs_inode
->i_mapping
,
2205 new_size
, LLONG_MAX
);
2207 jbd2_journal_abort(journal
, ret
);