2 * linux/fs/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/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/smp_lock.h>
28 #include <linux/highmem.h>
31 * get_transaction: obtain a new transaction_t object.
33 * Simply allocate and initialise a new transaction. Create it in
34 * RUNNING state and add it to the current journal (which should not
35 * have an existing running transaction: we only make a new transaction
36 * once we have started to commit the old one).
39 * The journal MUST be locked. We don't perform atomic mallocs on the
40 * new transaction and we can't block without protecting against other
41 * processes trying to touch the journal while it is in transition.
43 * Called under j_state_lock
46 static transaction_t
*
47 get_transaction(journal_t
*journal
, transaction_t
*transaction
)
49 transaction
->t_journal
= journal
;
50 transaction
->t_state
= T_RUNNING
;
51 transaction
->t_tid
= journal
->j_transaction_sequence
++;
52 transaction
->t_expires
= jiffies
+ journal
->j_commit_interval
;
53 INIT_LIST_HEAD(&transaction
->t_jcb
);
54 spin_lock_init(&transaction
->t_handle_lock
);
55 spin_lock_init(&transaction
->t_jcb_lock
);
57 /* Set up the commit timer for the new transaction. */
58 journal
->j_commit_timer
->expires
= transaction
->t_expires
;
59 add_timer(journal
->j_commit_timer
);
61 J_ASSERT(journal
->j_running_transaction
== NULL
);
62 journal
->j_running_transaction
= transaction
;
70 * A handle_t is an object which represents a single atomic update to a
71 * filesystem, and which tracks all of the modifications which form part
76 * start_this_handle: Given a handle, deal with any locking or stalling
77 * needed to make sure that there is enough journal space for the handle
78 * to begin. Attach the handle to a transaction and set up the
79 * transaction's buffer credits.
82 static int start_this_handle(journal_t
*journal
, handle_t
*handle
)
84 transaction_t
*transaction
;
86 int nblocks
= handle
->h_buffer_credits
;
87 transaction_t
*new_transaction
= NULL
;
90 if (nblocks
> journal
->j_max_transaction_buffers
) {
91 printk(KERN_ERR
"JBD: %s wants too many credits (%d > %d)\n",
92 current
->comm
, nblocks
,
93 journal
->j_max_transaction_buffers
);
99 if (!journal
->j_running_transaction
) {
100 new_transaction
= jbd_kmalloc(sizeof(*new_transaction
),
102 if (!new_transaction
) {
106 memset(new_transaction
, 0, sizeof(*new_transaction
));
109 jbd_debug(3, "New handle %p going live.\n", handle
);
114 * We need to hold j_state_lock until t_updates has been incremented,
115 * for proper journal barrier handling
117 spin_lock(&journal
->j_state_lock
);
119 if (is_journal_aborted(journal
) ||
120 (journal
->j_errno
!= 0 && !(journal
->j_flags
& JFS_ACK_ERR
))) {
121 spin_unlock(&journal
->j_state_lock
);
126 /* Wait on the journal's transaction barrier if necessary */
127 if (journal
->j_barrier_count
) {
128 spin_unlock(&journal
->j_state_lock
);
129 wait_event(journal
->j_wait_transaction_locked
,
130 journal
->j_barrier_count
== 0);
134 if (!journal
->j_running_transaction
) {
135 if (!new_transaction
) {
136 spin_unlock(&journal
->j_state_lock
);
137 goto alloc_transaction
;
139 get_transaction(journal
, new_transaction
);
140 new_transaction
= NULL
;
143 transaction
= journal
->j_running_transaction
;
146 * If the current transaction is locked down for commit, wait for the
147 * lock to be released.
149 if (transaction
->t_state
== T_LOCKED
) {
152 prepare_to_wait(&journal
->j_wait_transaction_locked
,
153 &wait
, TASK_UNINTERRUPTIBLE
);
154 spin_unlock(&journal
->j_state_lock
);
156 finish_wait(&journal
->j_wait_transaction_locked
, &wait
);
161 * If there is not enough space left in the log to write all potential
162 * buffers requested by this operation, we need to stall pending a log
163 * checkpoint to free some more log space.
165 spin_lock(&transaction
->t_handle_lock
);
166 needed
= transaction
->t_outstanding_credits
+ nblocks
;
168 if (needed
> journal
->j_max_transaction_buffers
) {
170 * If the current transaction is already too large, then start
171 * to commit it: we can then go back and attach this handle to
176 jbd_debug(2, "Handle %p starting new commit...\n", handle
);
177 spin_unlock(&transaction
->t_handle_lock
);
178 prepare_to_wait(&journal
->j_wait_transaction_locked
, &wait
,
179 TASK_UNINTERRUPTIBLE
);
180 __log_start_commit(journal
, transaction
->t_tid
);
181 spin_unlock(&journal
->j_state_lock
);
183 finish_wait(&journal
->j_wait_transaction_locked
, &wait
);
188 * The commit code assumes that it can get enough log space
189 * without forcing a checkpoint. This is *critical* for
190 * correctness: a checkpoint of a buffer which is also
191 * associated with a committing transaction creates a deadlock,
192 * so commit simply cannot force through checkpoints.
194 * We must therefore ensure the necessary space in the journal
195 * *before* starting to dirty potentially checkpointed buffers
196 * in the new transaction.
198 * The worst part is, any transaction currently committing can
199 * reduce the free space arbitrarily. Be careful to account for
200 * those buffers when checkpointing.
204 * @@@ AKPM: This seems rather over-defensive. We're giving commit
205 * a _lot_ of headroom: 1/4 of the journal plus the size of
206 * the committing transaction. Really, we only need to give it
207 * committing_transaction->t_outstanding_credits plus "enough" for
208 * the log control blocks.
209 * Also, this test is inconsitent with the matching one in
212 if (__log_space_left(journal
) < jbd_space_needed(journal
)) {
213 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle
);
214 spin_unlock(&transaction
->t_handle_lock
);
215 __log_wait_for_space(journal
);
219 /* OK, account for the buffers that this operation expects to
220 * use and add the handle to the running transaction. */
222 handle
->h_transaction
= transaction
;
223 transaction
->t_outstanding_credits
+= nblocks
;
224 transaction
->t_updates
++;
225 transaction
->t_handle_count
++;
226 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
227 handle
, nblocks
, transaction
->t_outstanding_credits
,
228 __log_space_left(journal
));
229 spin_unlock(&transaction
->t_handle_lock
);
230 spin_unlock(&journal
->j_state_lock
);
233 kfree(new_transaction
);
237 /* Allocate a new handle. This should probably be in a slab... */
238 static handle_t
*new_handle(int nblocks
)
240 handle_t
*handle
= jbd_alloc_handle(GFP_NOFS
);
243 memset(handle
, 0, sizeof(*handle
));
244 handle
->h_buffer_credits
= nblocks
;
246 INIT_LIST_HEAD(&handle
->h_jcb
);
252 * handle_t *journal_start() - Obtain a new handle.
253 * @journal: Journal to start transaction on.
254 * @nblocks: number of block buffer we might modify
256 * We make sure that the transaction can guarantee at least nblocks of
257 * modified buffers in the log. We block until the log can guarantee
260 * This function is visible to journal users (like ext3fs), so is not
261 * called with the journal already locked.
263 * Return a pointer to a newly allocated handle, or NULL on failure
265 handle_t
*journal_start(journal_t
*journal
, int nblocks
)
267 handle_t
*handle
= journal_current_handle();
271 return ERR_PTR(-EROFS
);
274 J_ASSERT(handle
->h_transaction
->t_journal
== journal
);
279 handle
= new_handle(nblocks
);
281 return ERR_PTR(-ENOMEM
);
283 current
->journal_info
= handle
;
285 err
= start_this_handle(journal
, handle
);
287 jbd_free_handle(handle
);
288 current
->journal_info
= NULL
;
289 handle
= ERR_PTR(err
);
295 * int journal_extend() - extend buffer credits.
296 * @handle: handle to 'extend'
297 * @nblocks: nr blocks to try to extend by.
299 * Some transactions, such as large extends and truncates, can be done
300 * atomically all at once or in several stages. The operation requests
301 * a credit for a number of buffer modications in advance, but can
302 * extend its credit if it needs more.
304 * journal_extend tries to give the running handle more buffer credits.
305 * It does not guarantee that allocation - this is a best-effort only.
306 * The calling process MUST be able to deal cleanly with a failure to
309 * Return 0 on success, non-zero on failure.
311 * return code < 0 implies an error
312 * return code > 0 implies normal transaction-full status.
314 int journal_extend(handle_t
*handle
, int nblocks
)
316 transaction_t
*transaction
= handle
->h_transaction
;
317 journal_t
*journal
= transaction
->t_journal
;
322 if (is_handle_aborted(handle
))
327 spin_lock(&journal
->j_state_lock
);
329 /* Don't extend a locked-down transaction! */
330 if (handle
->h_transaction
->t_state
!= T_RUNNING
) {
331 jbd_debug(3, "denied handle %p %d blocks: "
332 "transaction not running\n", handle
, nblocks
);
336 spin_lock(&transaction
->t_handle_lock
);
337 wanted
= transaction
->t_outstanding_credits
+ nblocks
;
339 if (wanted
> journal
->j_max_transaction_buffers
) {
340 jbd_debug(3, "denied handle %p %d blocks: "
341 "transaction too large\n", handle
, nblocks
);
345 if (wanted
> __log_space_left(journal
)) {
346 jbd_debug(3, "denied handle %p %d blocks: "
347 "insufficient log space\n", handle
, nblocks
);
351 handle
->h_buffer_credits
+= nblocks
;
352 transaction
->t_outstanding_credits
+= nblocks
;
355 jbd_debug(3, "extended handle %p by %d\n", handle
, nblocks
);
357 spin_unlock(&transaction
->t_handle_lock
);
359 spin_unlock(&journal
->j_state_lock
);
366 * int journal_restart() - restart a handle .
367 * @handle: handle to restart
368 * @nblocks: nr credits requested
370 * Restart a handle for a multi-transaction filesystem
373 * If the journal_extend() call above fails to grant new buffer credits
374 * to a running handle, a call to journal_restart will commit the
375 * handle's transaction so far and reattach the handle to a new
376 * transaction capabable of guaranteeing the requested number of
380 int journal_restart(handle_t
*handle
, int nblocks
)
382 transaction_t
*transaction
= handle
->h_transaction
;
383 journal_t
*journal
= transaction
->t_journal
;
386 /* If we've had an abort of any type, don't even think about
387 * actually doing the restart! */
388 if (is_handle_aborted(handle
))
392 * First unlink the handle from its current transaction, and start the
395 J_ASSERT(transaction
->t_updates
> 0);
396 J_ASSERT(journal_current_handle() == handle
);
398 spin_lock(&journal
->j_state_lock
);
399 spin_lock(&transaction
->t_handle_lock
);
400 transaction
->t_outstanding_credits
-= handle
->h_buffer_credits
;
401 transaction
->t_updates
--;
403 if (!transaction
->t_updates
)
404 wake_up(&journal
->j_wait_updates
);
405 spin_unlock(&transaction
->t_handle_lock
);
407 jbd_debug(2, "restarting handle %p\n", handle
);
408 __log_start_commit(journal
, transaction
->t_tid
);
409 spin_unlock(&journal
->j_state_lock
);
411 handle
->h_buffer_credits
= nblocks
;
412 ret
= start_this_handle(journal
, handle
);
418 * void journal_lock_updates () - establish a transaction barrier.
419 * @journal: Journal to establish a barrier on.
421 * This locks out any further updates from being started, and blocks
422 * until all existing updates have completed, returning only once the
423 * journal is in a quiescent state with no updates running.
425 * The journal lock should not be held on entry.
427 void journal_lock_updates(journal_t
*journal
)
431 spin_lock(&journal
->j_state_lock
);
432 ++journal
->j_barrier_count
;
434 /* Wait until there are no running updates */
436 transaction_t
*transaction
= journal
->j_running_transaction
;
441 spin_lock(&transaction
->t_handle_lock
);
442 if (!transaction
->t_updates
) {
443 spin_unlock(&transaction
->t_handle_lock
);
446 prepare_to_wait(&journal
->j_wait_updates
, &wait
,
447 TASK_UNINTERRUPTIBLE
);
448 spin_unlock(&transaction
->t_handle_lock
);
449 spin_unlock(&journal
->j_state_lock
);
451 finish_wait(&journal
->j_wait_updates
, &wait
);
452 spin_lock(&journal
->j_state_lock
);
454 spin_unlock(&journal
->j_state_lock
);
457 * We have now established a barrier against other normal updates, but
458 * we also need to barrier against other journal_lock_updates() calls
459 * to make sure that we serialise special journal-locked operations
462 down(&journal
->j_barrier
);
466 * void journal_unlock_updates (journal_t* journal) - release barrier
467 * @journal: Journal to release the barrier on.
469 * Release a transaction barrier obtained with journal_lock_updates().
471 * Should be called without the journal lock held.
473 void journal_unlock_updates (journal_t
*journal
)
475 J_ASSERT(journal
->j_barrier_count
!= 0);
477 up(&journal
->j_barrier
);
478 spin_lock(&journal
->j_state_lock
);
479 --journal
->j_barrier_count
;
480 spin_unlock(&journal
->j_state_lock
);
481 wake_up(&journal
->j_wait_transaction_locked
);
485 * Report any unexpected dirty buffers which turn up. Normally those
486 * indicate an error, but they can occur if the user is running (say)
487 * tune2fs to modify the live filesystem, so we need the option of
488 * continuing as gracefully as possible. #
490 * The caller should already hold the journal lock and
491 * j_list_lock spinlock: most callers will need those anyway
492 * in order to probe the buffer's journaling state safely.
494 static void jbd_unexpected_dirty_buffer(struct journal_head
*jh
)
496 struct buffer_head
*bh
= jh2bh(jh
);
499 if (buffer_dirty(bh
)) {
500 /* If this buffer is one which might reasonably be dirty
501 * --- ie. data, or not part of this journal --- then
502 * we're OK to leave it alone, but otherwise we need to
503 * move the dirty bit to the journal's own internal
507 if (jlist
== BJ_Metadata
|| jlist
== BJ_Reserved
||
508 jlist
== BJ_Shadow
|| jlist
== BJ_Forget
) {
509 if (test_clear_buffer_dirty(jh2bh(jh
))) {
510 set_bit(BH_JBDDirty
, &jh2bh(jh
)->b_state
);
517 * If the buffer is already part of the current transaction, then there
518 * is nothing we need to do. If it is already part of a prior
519 * transaction which we are still committing to disk, then we need to
520 * make sure that we do not overwrite the old copy: we do copy-out to
521 * preserve the copy going to disk. We also account the buffer against
522 * the handle's metadata buffer credits (unless the buffer is already
523 * part of the transaction, that is).
527 do_get_write_access(handle_t
*handle
, struct journal_head
*jh
,
528 int force_copy
, int *credits
)
530 struct buffer_head
*bh
;
531 transaction_t
*transaction
;
534 char *frozen_buffer
= NULL
;
537 if (is_handle_aborted(handle
))
540 transaction
= handle
->h_transaction
;
541 journal
= transaction
->t_journal
;
543 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh
, force_copy
);
545 JBUFFER_TRACE(jh
, "entry");
549 /* @@@ Need to check for errors here at some point. */
552 jbd_lock_bh_state(bh
);
554 /* We now hold the buffer lock so it is safe to query the buffer
555 * state. Is the buffer dirty?
557 * If so, there are two possibilities. The buffer may be
558 * non-journaled, and undergoing a quite legitimate writeback.
559 * Otherwise, it is journaled, and we don't expect dirty buffers
560 * in that state (the buffers should be marked JBD_Dirty
561 * instead.) So either the IO is being done under our own
562 * control and this is a bug, or it's a third party IO such as
563 * dump(8) (which may leave the buffer scheduled for read ---
564 * ie. locked but not dirty) or tune2fs (which may actually have
565 * the buffer dirtied, ugh.) */
567 if (buffer_dirty(bh
)) {
569 * First question: is this buffer already part of the current
570 * transaction or the existing committing transaction?
572 if (jh
->b_transaction
) {
574 jh
->b_transaction
== transaction
||
576 journal
->j_committing_transaction
);
577 if (jh
->b_next_transaction
)
578 J_ASSERT_JH(jh
, jh
->b_next_transaction
==
580 JBUFFER_TRACE(jh
, "Unexpected dirty buffer");
581 jbd_unexpected_dirty_buffer(jh
);
588 if (is_handle_aborted(handle
)) {
589 jbd_unlock_bh_state(bh
);
595 * The buffer is already part of this transaction if b_transaction or
596 * b_next_transaction points to it
598 if (jh
->b_transaction
== transaction
||
599 jh
->b_next_transaction
== transaction
)
603 * If there is already a copy-out version of this buffer, then we don't
604 * need to make another one
606 if (jh
->b_frozen_data
) {
607 JBUFFER_TRACE(jh
, "has frozen data");
608 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
609 jh
->b_next_transaction
= transaction
;
611 J_ASSERT_JH(jh
, handle
->h_buffer_credits
> 0);
612 handle
->h_buffer_credits
--;
618 /* Is there data here we need to preserve? */
620 if (jh
->b_transaction
&& jh
->b_transaction
!= transaction
) {
621 JBUFFER_TRACE(jh
, "owned by older transaction");
622 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
623 J_ASSERT_JH(jh
, jh
->b_transaction
==
624 journal
->j_committing_transaction
);
626 /* There is one case we have to be very careful about.
627 * If the committing transaction is currently writing
628 * this buffer out to disk and has NOT made a copy-out,
629 * then we cannot modify the buffer contents at all
630 * right now. The essence of copy-out is that it is the
631 * extra copy, not the primary copy, which gets
632 * journaled. If the primary copy is already going to
633 * disk then we cannot do copy-out here. */
635 if (jh
->b_jlist
== BJ_Shadow
) {
636 wait_queue_head_t
*wqh
;
639 JBUFFER_TRACE(jh
, "on shadow: sleep");
640 jbd_unlock_bh_state(bh
);
641 /* commit wakes up all shadow buffers after IO */
642 wqh
= bh_waitq_head(bh
);
644 prepare_to_wait(wqh
, &wait
,
645 TASK_UNINTERRUPTIBLE
);
646 if (jh
->b_jlist
!= BJ_Shadow
)
650 finish_wait(wqh
, &wait
);
654 /* Only do the copy if the currently-owning transaction
655 * still needs it. If it is on the Forget list, the
656 * committing transaction is past that stage. The
657 * buffer had better remain locked during the kmalloc,
658 * but that should be true --- we hold the journal lock
659 * still and the buffer is already on the BUF_JOURNAL
660 * list so won't be flushed.
662 * Subtle point, though: if this is a get_undo_access,
663 * then we will be relying on the frozen_data to contain
664 * the new value of the committed_data record after the
665 * transaction, so we HAVE to force the frozen_data copy
668 if (jh
->b_jlist
!= BJ_Forget
|| force_copy
) {
669 JBUFFER_TRACE(jh
, "generate frozen data");
670 if (!frozen_buffer
) {
671 JBUFFER_TRACE(jh
, "allocate memory for buffer");
672 jbd_unlock_bh_state(bh
);
673 frozen_buffer
= jbd_kmalloc(jh2bh(jh
)->b_size
,
675 if (!frozen_buffer
) {
677 "%s: OOM for frozen_buffer\n",
679 JBUFFER_TRACE(jh
, "oom!");
681 jbd_lock_bh_state(bh
);
686 jh
->b_frozen_data
= frozen_buffer
;
687 frozen_buffer
= NULL
;
690 jh
->b_next_transaction
= transaction
;
693 J_ASSERT(handle
->h_buffer_credits
> 0);
694 handle
->h_buffer_credits
--;
699 * Finally, if the buffer is not journaled right now, we need to make
700 * sure it doesn't get written to disk before the caller actually
701 * commits the new data
703 if (!jh
->b_transaction
) {
704 JBUFFER_TRACE(jh
, "no transaction");
705 J_ASSERT_JH(jh
, !jh
->b_next_transaction
);
706 jh
->b_transaction
= transaction
;
707 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
708 spin_lock(&journal
->j_list_lock
);
709 __journal_file_buffer(jh
, transaction
, BJ_Reserved
);
710 spin_unlock(&journal
->j_list_lock
);
719 J_EXPECT_JH(jh
, buffer_uptodate(jh2bh(jh
)),
720 "Possible IO failure.\n");
721 page
= jh2bh(jh
)->b_page
;
722 offset
= ((unsigned long) jh2bh(jh
)->b_data
) & ~PAGE_MASK
;
723 source
= kmap_atomic(page
, KM_USER0
);
724 memcpy(jh
->b_frozen_data
, source
+offset
, jh2bh(jh
)->b_size
);
725 kunmap_atomic(source
, KM_USER0
);
727 jbd_unlock_bh_state(bh
);
730 * If we are about to journal a buffer, then any revoke pending on it is
733 journal_cancel_revoke(handle
, jh
);
737 kfree(frozen_buffer
);
739 JBUFFER_TRACE(jh
, "exit");
744 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
745 * @handle: transaction to add buffer modifications to
746 * @bh: bh to be used for metadata writes
748 * Returns an error code or 0 on success.
750 * In full data journalling mode the buffer may be of type BJ_AsyncData,
751 * because we're write()ing a buffer which is also part of a shared mapping.
754 int journal_get_write_access(handle_t
*handle
,
755 struct buffer_head
*bh
, int *credits
)
757 struct journal_head
*jh
= journal_add_journal_head(bh
);
760 /* We do not want to get caught playing with fields which the
761 * log thread also manipulates. Make sure that the buffer
762 * completes any outstanding IO before proceeding. */
763 rc
= do_get_write_access(handle
, jh
, 0, credits
);
764 journal_put_journal_head(jh
);
770 * When the user wants to journal a newly created buffer_head
771 * (ie. getblk() returned a new buffer and we are going to populate it
772 * manually rather than reading off disk), then we need to keep the
773 * buffer_head locked until it has been completely filled with new
774 * data. In this case, we should be able to make the assertion that
775 * the bh is not already part of an existing transaction.
777 * The buffer should already be locked by the caller by this point.
778 * There is no lock ranking violation: it was a newly created,
779 * unlocked buffer beforehand. */
782 * int journal_get_create_access () - notify intent to use newly created bh
783 * @handle: transaction to new buffer to
786 * Call this if you create a new bh.
788 int journal_get_create_access(handle_t
*handle
, struct buffer_head
*bh
)
790 transaction_t
*transaction
= handle
->h_transaction
;
791 journal_t
*journal
= transaction
->t_journal
;
792 struct journal_head
*jh
= journal_add_journal_head(bh
);
795 jbd_debug(5, "journal_head %p\n", jh
);
797 if (is_handle_aborted(handle
))
801 JBUFFER_TRACE(jh
, "entry");
803 * The buffer may already belong to this transaction due to pre-zeroing
804 * in the filesystem's new_block code. It may also be on the previous,
805 * committing transaction's lists, but it HAS to be in Forget state in
806 * that case: the transaction must have deleted the buffer for it to be
809 jbd_lock_bh_state(bh
);
810 spin_lock(&journal
->j_list_lock
);
811 J_ASSERT_JH(jh
, (jh
->b_transaction
== transaction
||
812 jh
->b_transaction
== NULL
||
813 (jh
->b_transaction
== journal
->j_committing_transaction
&&
814 jh
->b_jlist
== BJ_Forget
)));
816 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
817 J_ASSERT_JH(jh
, buffer_locked(jh2bh(jh
)));
819 J_ASSERT_JH(jh
, handle
->h_buffer_credits
> 0);
820 handle
->h_buffer_credits
--;
822 if (jh
->b_transaction
== NULL
) {
823 jh
->b_transaction
= transaction
;
824 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
825 __journal_file_buffer(jh
, transaction
, BJ_Reserved
);
826 } else if (jh
->b_transaction
== journal
->j_committing_transaction
) {
827 JBUFFER_TRACE(jh
, "set next transaction");
828 jh
->b_next_transaction
= transaction
;
830 spin_unlock(&journal
->j_list_lock
);
831 jbd_unlock_bh_state(bh
);
834 * akpm: I added this. ext3_alloc_branch can pick up new indirect
835 * blocks which contain freed but then revoked metadata. We need
836 * to cancel the revoke in case we end up freeing it yet again
837 * and the reallocating as data - this would cause a second revoke,
838 * which hits an assertion error.
840 JBUFFER_TRACE(jh
, "cancelling revoke");
841 journal_cancel_revoke(handle
, jh
);
842 journal_put_journal_head(jh
);
848 * int journal_get_undo_access() - Notify intent to modify metadata with
849 * non-rewindable consequences
850 * @handle: transaction
851 * @bh: buffer to undo
852 * @credits: store the number of taken credits here (if not NULL)
854 * Sometimes there is a need to distinguish between metadata which has
855 * been committed to disk and that which has not. The ext3fs code uses
856 * this for freeing and allocating space, we have to make sure that we
857 * do not reuse freed space until the deallocation has been committed,
858 * since if we overwrote that space we would make the delete
859 * un-rewindable in case of a crash.
861 * To deal with that, journal_get_undo_access requests write access to a
862 * buffer for parts of non-rewindable operations such as delete
863 * operations on the bitmaps. The journaling code must keep a copy of
864 * the buffer's contents prior to the undo_access call until such time
865 * as we know that the buffer has definitely been committed to disk.
867 * We never need to know which transaction the committed data is part
868 * of, buffers touched here are guaranteed to be dirtied later and so
869 * will be committed to a new transaction in due course, at which point
870 * we can discard the old committed data pointer.
872 * Returns error number or 0 on success.
874 int journal_get_undo_access(handle_t
*handle
, struct buffer_head
*bh
,
878 struct journal_head
*jh
= journal_add_journal_head(bh
);
879 char *committed_data
= NULL
;
881 JBUFFER_TRACE(jh
, "entry");
884 * Do this first --- it can drop the journal lock, so we want to
885 * make sure that obtaining the committed_data is done
886 * atomically wrt. completion of any outstanding commits.
888 err
= do_get_write_access(handle
, jh
, 1, credits
);
893 if (!jh
->b_committed_data
) {
894 committed_data
= jbd_kmalloc(jh2bh(jh
)->b_size
, GFP_NOFS
);
895 if (!committed_data
) {
896 printk(KERN_EMERG
"%s: No memory for committed data\n",
903 jbd_lock_bh_state(bh
);
904 if (!jh
->b_committed_data
) {
905 /* Copy out the current buffer contents into the
906 * preserved, committed copy. */
907 JBUFFER_TRACE(jh
, "generate b_committed data");
908 if (!committed_data
) {
909 jbd_unlock_bh_state(bh
);
913 jh
->b_committed_data
= committed_data
;
914 committed_data
= NULL
;
915 memcpy(jh
->b_committed_data
, bh
->b_data
, bh
->b_size
);
917 jbd_unlock_bh_state(bh
);
919 journal_put_journal_head(jh
);
921 kfree(committed_data
);
926 * int journal_dirty_data() - mark a buffer as containing dirty data which
927 * needs to be flushed before we can commit the
928 * current transaction.
929 * @handle: transaction
930 * @bh: bufferhead to mark
932 * The buffer is placed on the transaction's data list and is marked as
933 * belonging to the transaction.
935 * Returns error number or 0 on success.
937 * journal_dirty_data() can be called via page_launder->ext3_writepage
940 int journal_dirty_data(handle_t
*handle
, struct buffer_head
*bh
)
942 journal_t
*journal
= handle
->h_transaction
->t_journal
;
944 struct journal_head
*jh
;
946 if (is_handle_aborted(handle
))
949 jh
= journal_add_journal_head(bh
);
950 JBUFFER_TRACE(jh
, "entry");
953 * The buffer could *already* be dirty. Writeout can start
956 jbd_debug(4, "jh: %p, tid:%d\n", jh
, handle
->h_transaction
->t_tid
);
959 * What if the buffer is already part of a running transaction?
961 * There are two cases:
962 * 1) It is part of the current running transaction. Refile it,
963 * just in case we have allocated it as metadata, deallocated
964 * it, then reallocated it as data.
965 * 2) It is part of the previous, still-committing transaction.
966 * If all we want to do is to guarantee that the buffer will be
967 * written to disk before this new transaction commits, then
968 * being sure that the *previous* transaction has this same
969 * property is sufficient for us! Just leave it on its old
972 * In case (2), the buffer must not already exist as metadata
973 * --- that would violate write ordering (a transaction is free
974 * to write its data at any point, even before the previous
975 * committing transaction has committed). The caller must
976 * never, ever allow this to happen: there's nothing we can do
977 * about it in this layer.
979 jbd_lock_bh_state(bh
);
980 spin_lock(&journal
->j_list_lock
);
981 if (jh
->b_transaction
) {
982 JBUFFER_TRACE(jh
, "has transaction");
983 if (jh
->b_transaction
!= handle
->h_transaction
) {
984 JBUFFER_TRACE(jh
, "belongs to older transaction");
985 J_ASSERT_JH(jh
, jh
->b_transaction
==
986 journal
->j_committing_transaction
);
988 /* @@@ IS THIS TRUE ? */
990 * Not any more. Scenario: someone does a write()
991 * in data=journal mode. The buffer's transaction has
992 * moved into commit. Then someone does another
993 * write() to the file. We do the frozen data copyout
994 * and set b_next_transaction to point to j_running_t.
995 * And while we're in that state, someone does a
996 * writepage() in an attempt to pageout the same area
997 * of the file via a shared mapping. At present that
998 * calls journal_dirty_data(), and we get right here.
999 * It may be too late to journal the data. Simply
1000 * falling through to the next test will suffice: the
1001 * data will be dirty and wil be checkpointed. The
1002 * ordering comments in the next comment block still
1005 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1008 * If we're journalling data, and this buffer was
1009 * subject to a write(), it could be metadata, forget
1010 * or shadow against the committing transaction. Now,
1011 * someone has dirtied the same darn page via a mapping
1012 * and it is being writepage()'d.
1013 * We *could* just steal the page from commit, with some
1014 * fancy locking there. Instead, we just skip it -
1015 * don't tie the page's buffers to the new transaction
1017 * Implication: if we crash before the writepage() data
1018 * is written into the filesystem, recovery will replay
1021 if (jh
->b_jlist
!= BJ_None
&&
1022 jh
->b_jlist
!= BJ_SyncData
&&
1023 jh
->b_jlist
!= BJ_Locked
) {
1024 JBUFFER_TRACE(jh
, "Not stealing");
1029 * This buffer may be undergoing writeout in commit. We
1030 * can't return from here and let the caller dirty it
1031 * again because that can cause the write-out loop in
1032 * commit to never terminate.
1034 if (buffer_dirty(bh
)) {
1036 spin_unlock(&journal
->j_list_lock
);
1037 jbd_unlock_bh_state(bh
);
1039 sync_dirty_buffer(bh
);
1040 jbd_lock_bh_state(bh
);
1041 spin_lock(&journal
->j_list_lock
);
1042 /* The buffer may become locked again at any
1043 time if it is redirtied */
1046 /* journal_clean_data_list() may have got there first */
1047 if (jh
->b_transaction
!= NULL
) {
1048 JBUFFER_TRACE(jh
, "unfile from commit");
1049 __journal_unfile_buffer(jh
);
1051 /* The buffer will be refiled below */
1055 * Special case --- the buffer might actually have been
1056 * allocated and then immediately deallocated in the previous,
1057 * committing transaction, so might still be left on that
1058 * transaction's metadata lists.
1060 if (jh
->b_jlist
!= BJ_SyncData
&& jh
->b_jlist
!= BJ_Locked
) {
1061 JBUFFER_TRACE(jh
, "not on correct data list: unfile");
1062 J_ASSERT_JH(jh
, jh
->b_jlist
!= BJ_Shadow
);
1063 __journal_unfile_buffer(jh
);
1064 JBUFFER_TRACE(jh
, "file as data");
1065 __journal_file_buffer(jh
, handle
->h_transaction
,
1069 JBUFFER_TRACE(jh
, "not on a transaction");
1070 __journal_file_buffer(jh
, handle
->h_transaction
, BJ_SyncData
);
1073 spin_unlock(&journal
->j_list_lock
);
1074 jbd_unlock_bh_state(bh
);
1076 BUFFER_TRACE(bh
, "brelse");
1079 JBUFFER_TRACE(jh
, "exit");
1080 journal_put_journal_head(jh
);
1085 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1086 * @handle: transaction to add buffer to.
1087 * @bh: buffer to mark
1089 * mark dirty metadata which needs to be journaled as part of the current
1092 * The buffer is placed on the transaction's metadata list and is marked
1093 * as belonging to the transaction.
1095 * Returns error number or 0 on success.
1097 * Special care needs to be taken if the buffer already belongs to the
1098 * current committing transaction (in which case we should have frozen
1099 * data present for that commit). In that case, we don't relink the
1100 * buffer: that only gets done when the old transaction finally
1101 * completes its commit.
1103 int journal_dirty_metadata(handle_t
*handle
, struct buffer_head
*bh
)
1105 transaction_t
*transaction
= handle
->h_transaction
;
1106 journal_t
*journal
= transaction
->t_journal
;
1107 struct journal_head
*jh
= bh2jh(bh
);
1109 jbd_debug(5, "journal_head %p\n", jh
);
1110 JBUFFER_TRACE(jh
, "entry");
1111 if (is_handle_aborted(handle
))
1114 jbd_lock_bh_state(bh
);
1117 * fastpath, to avoid expensive locking. If this buffer is already
1118 * on the running transaction's metadata list there is nothing to do.
1119 * Nobody can take it off again because there is a handle open.
1120 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1121 * result in this test being false, so we go in and take the locks.
1123 if (jh
->b_transaction
== transaction
&& jh
->b_jlist
== BJ_Metadata
) {
1124 JBUFFER_TRACE(jh
, "fastpath");
1125 J_ASSERT_JH(jh
, jh
->b_transaction
==
1126 journal
->j_running_transaction
);
1130 set_buffer_jbddirty(bh
);
1133 * Metadata already on the current transaction list doesn't
1134 * need to be filed. Metadata on another transaction's list must
1135 * be committing, and will be refiled once the commit completes:
1136 * leave it alone for now.
1138 if (jh
->b_transaction
!= transaction
) {
1139 JBUFFER_TRACE(jh
, "already on other transaction");
1140 J_ASSERT_JH(jh
, jh
->b_transaction
==
1141 journal
->j_committing_transaction
);
1142 J_ASSERT_JH(jh
, jh
->b_next_transaction
== transaction
);
1143 /* And this case is illegal: we can't reuse another
1144 * transaction's data buffer, ever. */
1148 /* That test should have eliminated the following case: */
1149 J_ASSERT_JH(jh
, jh
->b_frozen_data
== 0);
1151 JBUFFER_TRACE(jh
, "file as BJ_Metadata");
1152 spin_lock(&journal
->j_list_lock
);
1153 __journal_file_buffer(jh
, handle
->h_transaction
, BJ_Metadata
);
1154 spin_unlock(&journal
->j_list_lock
);
1156 jbd_unlock_bh_state(bh
);
1158 JBUFFER_TRACE(jh
, "exit");
1163 * journal_release_buffer: undo a get_write_access without any buffer
1164 * updates, if the update decided in the end that it didn't need access.
1166 * The caller passes in the number of credits which should be put back for
1167 * this buffer (zero or one).
1169 * We leave the buffer attached to t_reserved_list because even though this
1170 * handle doesn't want it, some other concurrent handle may want to journal
1171 * this buffer. If that handle is curently in between get_write_access() and
1172 * journal_dirty_metadata() then it expects the buffer to be reserved. If
1173 * we were to rip it off t_reserved_list here, the other handle will explode
1174 * when journal_dirty_metadata is presented with a non-reserved buffer.
1176 * If nobody really wants to journal this buffer then it will be thrown
1177 * away at the start of commit.
1180 journal_release_buffer(handle_t
*handle
, struct buffer_head
*bh
, int credits
)
1182 BUFFER_TRACE(bh
, "entry");
1183 handle
->h_buffer_credits
+= credits
;
1187 * void journal_forget() - bforget() for potentially-journaled buffers.
1188 * @handle: transaction handle
1189 * @bh: bh to 'forget'
1191 * We can only do the bforget if there are no commits pending against the
1192 * buffer. If the buffer is dirty in the current running transaction we
1193 * can safely unlink it.
1195 * bh may not be a journalled buffer at all - it may be a non-JBD
1196 * buffer which came off the hashtable. Check for this.
1198 * Decrements bh->b_count by one.
1200 * Allow this call even if the handle has aborted --- it may be part of
1201 * the caller's cleanup after an abort.
1203 void journal_forget(handle_t
*handle
, struct buffer_head
*bh
)
1205 transaction_t
*transaction
= handle
->h_transaction
;
1206 journal_t
*journal
= transaction
->t_journal
;
1207 struct journal_head
*jh
;
1209 BUFFER_TRACE(bh
, "entry");
1211 jbd_lock_bh_state(bh
);
1212 spin_lock(&journal
->j_list_lock
);
1214 if (!buffer_jbd(bh
))
1218 if (jh
->b_transaction
== handle
->h_transaction
) {
1219 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
1221 /* If we are forgetting a buffer which is already part
1222 * of this transaction, then we can just drop it from
1223 * the transaction immediately. */
1224 clear_buffer_dirty(bh
);
1225 clear_buffer_jbddirty(bh
);
1227 JBUFFER_TRACE(jh
, "belongs to current transaction: unfile");
1228 J_ASSERT_JH(jh
, !jh
->b_committed_data
);
1230 __journal_unfile_buffer(jh
);
1233 * We are no longer going to journal this buffer.
1234 * However, the commit of this transaction is still
1235 * important to the buffer: the delete that we are now
1236 * processing might obsolete an old log entry, so by
1237 * committing, we can satisfy the buffer's checkpoint.
1239 * So, if we have a checkpoint on the buffer, we should
1240 * now refile the buffer on our BJ_Forget list so that
1241 * we know to remove the checkpoint after we commit.
1244 if (jh
->b_cp_transaction
) {
1245 __journal_file_buffer(jh
, transaction
, BJ_Forget
);
1247 journal_remove_journal_head(bh
);
1249 if (!buffer_jbd(bh
)) {
1250 spin_unlock(&journal
->j_list_lock
);
1251 jbd_unlock_bh_state(bh
);
1256 } else if (jh
->b_transaction
) {
1257 J_ASSERT_JH(jh
, (jh
->b_transaction
==
1258 journal
->j_committing_transaction
));
1259 /* However, if the buffer is still owned by a prior
1260 * (committing) transaction, we can't drop it yet... */
1261 JBUFFER_TRACE(jh
, "belongs to older transaction");
1262 /* ... but we CAN drop it from the new transaction if we
1263 * have also modified it since the original commit. */
1265 if (jh
->b_next_transaction
) {
1266 J_ASSERT(jh
->b_next_transaction
== transaction
);
1267 jh
->b_next_transaction
= NULL
;
1272 spin_unlock(&journal
->j_list_lock
);
1273 jbd_unlock_bh_state(bh
);
1279 * void journal_callback_set() - Register a callback function for this handle.
1280 * @handle: handle to attach the callback to.
1281 * @func: function to callback.
1282 * @jcb: structure with additional information required by func() , and
1283 * some space for jbd internal information.
1285 * The function will be
1286 * called when the transaction that this handle is part of has been
1287 * committed to disk with the original callback data struct and the
1288 * error status of the journal as parameters. There is no guarantee of
1289 * ordering between handles within a single transaction, nor between
1290 * callbacks registered on the same handle.
1292 * The caller is responsible for allocating the journal_callback struct.
1293 * This is to allow the caller to add as much extra data to the callback
1294 * as needed, but reduce the overhead of multiple allocations. The caller
1295 * allocated struct must start with a struct journal_callback at offset 0,
1296 * and has the caller-specific data afterwards.
1298 void journal_callback_set(handle_t
*handle
,
1299 void (*func
)(struct journal_callback
*jcb
, int error
),
1300 struct journal_callback
*jcb
)
1302 spin_lock(&handle
->h_transaction
->t_jcb_lock
);
1303 list_add_tail(&jcb
->jcb_list
, &handle
->h_jcb
);
1304 spin_unlock(&handle
->h_transaction
->t_jcb_lock
);
1305 jcb
->jcb_func
= func
;
1309 * int journal_stop() - complete a transaction
1310 * @handle: tranaction to complete.
1312 * All done for a particular handle.
1314 * There is not much action needed here. We just return any remaining
1315 * buffer credits to the transaction and remove the handle. The only
1316 * complication is that we need to start a commit operation if the
1317 * filesystem is marked for synchronous update.
1319 * journal_stop itself will not usually return an error, but it may
1320 * do so in unusual circumstances. In particular, expect it to
1321 * return -EIO if a journal_abort has been executed since the
1322 * transaction began.
1324 int journal_stop(handle_t
*handle
)
1326 transaction_t
*transaction
= handle
->h_transaction
;
1327 journal_t
*journal
= transaction
->t_journal
;
1328 int old_handle_count
, err
;
1330 J_ASSERT(transaction
->t_updates
> 0);
1331 J_ASSERT(journal_current_handle() == handle
);
1333 if (is_handle_aborted(handle
))
1338 if (--handle
->h_ref
> 0) {
1339 jbd_debug(4, "h_ref %d -> %d\n", handle
->h_ref
+ 1,
1344 jbd_debug(4, "Handle %p going down\n", handle
);
1347 * Implement synchronous transaction batching. If the handle
1348 * was synchronous, don't force a commit immediately. Let's
1349 * yield and let another thread piggyback onto this transaction.
1350 * Keep doing that while new threads continue to arrive.
1351 * It doesn't cost much - we're about to run a commit and sleep
1352 * on IO anyway. Speeds up many-threaded, many-dir operations
1355 if (handle
->h_sync
) {
1357 old_handle_count
= transaction
->t_handle_count
;
1358 set_current_state(TASK_UNINTERRUPTIBLE
);
1359 schedule_timeout(1);
1360 } while (old_handle_count
!= transaction
->t_handle_count
);
1363 current
->journal_info
= NULL
;
1364 spin_lock(&journal
->j_state_lock
);
1365 spin_lock(&transaction
->t_handle_lock
);
1366 transaction
->t_outstanding_credits
-= handle
->h_buffer_credits
;
1367 transaction
->t_updates
--;
1368 if (!transaction
->t_updates
) {
1369 wake_up(&journal
->j_wait_updates
);
1370 if (journal
->j_barrier_count
)
1371 wake_up(&journal
->j_wait_transaction_locked
);
1374 /* Move callbacks from the handle to the transaction. */
1375 spin_lock(&transaction
->t_jcb_lock
);
1376 list_splice(&handle
->h_jcb
, &transaction
->t_jcb
);
1377 spin_unlock(&transaction
->t_jcb_lock
);
1380 * If the handle is marked SYNC, we need to set another commit
1381 * going! We also want to force a commit if the current
1382 * transaction is occupying too much of the log, or if the
1383 * transaction is too old now.
1385 if (handle
->h_sync
||
1386 transaction
->t_outstanding_credits
>
1387 journal
->j_max_transaction_buffers
||
1388 time_after_eq(jiffies
, transaction
->t_expires
)) {
1389 /* Do this even for aborted journals: an abort still
1390 * completes the commit thread, it just doesn't write
1391 * anything to disk. */
1392 tid_t tid
= transaction
->t_tid
;
1394 spin_unlock(&transaction
->t_handle_lock
);
1395 jbd_debug(2, "transaction too old, requesting commit for "
1396 "handle %p\n", handle
);
1397 /* This is non-blocking */
1398 __log_start_commit(journal
, transaction
->t_tid
);
1399 spin_unlock(&journal
->j_state_lock
);
1402 * Special case: JFS_SYNC synchronous updates require us
1403 * to wait for the commit to complete.
1405 if (handle
->h_sync
&& !(current
->flags
& PF_MEMALLOC
))
1406 err
= log_wait_commit(journal
, tid
);
1408 spin_unlock(&transaction
->t_handle_lock
);
1409 spin_unlock(&journal
->j_state_lock
);
1412 jbd_free_handle(handle
);
1416 /**int journal_force_commit() - force any uncommitted transactions
1417 * @journal: journal to force
1419 * For synchronous operations: force any uncommitted transactions
1420 * to disk. May seem kludgy, but it reuses all the handle batching
1421 * code in a very simple manner.
1423 int journal_force_commit(journal_t
*journal
)
1428 handle
= journal_start(journal
, 1);
1429 if (IS_ERR(handle
)) {
1430 ret
= PTR_ERR(handle
);
1433 ret
= journal_stop(handle
);
1440 * List management code snippets: various functions for manipulating the
1441 * transaction buffer lists.
1446 * Append a buffer to a transaction list, given the transaction's list head
1449 * j_list_lock is held.
1451 * jbd_lock_bh_state(jh2bh(jh)) is held.
1455 __blist_add_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1458 jh
->b_tnext
= jh
->b_tprev
= jh
;
1461 /* Insert at the tail of the list to preserve order */
1462 struct journal_head
*first
= *list
, *last
= first
->b_tprev
;
1464 jh
->b_tnext
= first
;
1465 last
->b_tnext
= first
->b_tprev
= jh
;
1470 * Remove a buffer from a transaction list, given the transaction's list
1473 * Called with j_list_lock held, and the journal may not be locked.
1475 * jbd_lock_bh_state(jh2bh(jh)) is held.
1479 __blist_del_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1482 *list
= jh
->b_tnext
;
1486 jh
->b_tprev
->b_tnext
= jh
->b_tnext
;
1487 jh
->b_tnext
->b_tprev
= jh
->b_tprev
;
1491 * Remove a buffer from the appropriate transaction list.
1493 * Note that this function can *change* the value of
1494 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1495 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1496 * is holding onto a copy of one of thee pointers, it could go bad.
1497 * Generally the caller needs to re-read the pointer from the transaction_t.
1499 * Called under j_list_lock. The journal may not be locked.
1501 void __journal_unfile_buffer(struct journal_head
*jh
)
1503 struct journal_head
**list
= NULL
;
1504 transaction_t
*transaction
;
1505 struct buffer_head
*bh
= jh2bh(jh
);
1507 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
1508 transaction
= jh
->b_transaction
;
1510 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
1512 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
1513 if (jh
->b_jlist
!= BJ_None
)
1514 J_ASSERT_JH(jh
, transaction
!= 0);
1516 switch (jh
->b_jlist
) {
1520 list
= &transaction
->t_sync_datalist
;
1523 transaction
->t_nr_buffers
--;
1524 J_ASSERT_JH(jh
, transaction
->t_nr_buffers
>= 0);
1525 list
= &transaction
->t_buffers
;
1528 list
= &transaction
->t_forget
;
1531 list
= &transaction
->t_iobuf_list
;
1534 list
= &transaction
->t_shadow_list
;
1537 list
= &transaction
->t_log_list
;
1540 list
= &transaction
->t_reserved_list
;
1543 list
= &transaction
->t_locked_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 */
1552 jh
->b_transaction
= NULL
;
1555 void journal_unfile_buffer(journal_t
*journal
, struct journal_head
*jh
)
1557 jbd_lock_bh_state(jh2bh(jh
));
1558 spin_lock(&journal
->j_list_lock
);
1559 __journal_unfile_buffer(jh
);
1560 spin_unlock(&journal
->j_list_lock
);
1561 jbd_unlock_bh_state(jh2bh(jh
));
1565 * Called from journal_try_to_free_buffers().
1567 * Called under jbd_lock_bh_state(bh)
1570 __journal_try_to_free_buffer(journal_t
*journal
, struct buffer_head
*bh
)
1572 struct journal_head
*jh
;
1576 if (buffer_locked(bh
) || buffer_dirty(bh
))
1579 if (jh
->b_next_transaction
!= 0)
1582 spin_lock(&journal
->j_list_lock
);
1583 if (jh
->b_transaction
!= 0 && jh
->b_cp_transaction
== 0) {
1584 if (jh
->b_jlist
== BJ_SyncData
|| jh
->b_jlist
== BJ_Locked
) {
1585 /* A written-back ordered data buffer */
1586 JBUFFER_TRACE(jh
, "release data");
1587 __journal_unfile_buffer(jh
);
1588 journal_remove_journal_head(bh
);
1591 } else if (jh
->b_cp_transaction
!= 0 && jh
->b_transaction
== 0) {
1592 /* written-back checkpointed metadata buffer */
1593 if (jh
->b_jlist
== BJ_None
) {
1594 JBUFFER_TRACE(jh
, "remove from checkpoint list");
1595 __journal_remove_checkpoint(jh
);
1596 journal_remove_journal_head(bh
);
1600 spin_unlock(&journal
->j_list_lock
);
1607 * int journal_try_to_free_buffers() - try to free page buffers.
1608 * @journal: journal for operation
1609 * @page: to try and free
1610 * @gfp_mask: 'IO' mode for try_to_free_buffers()
1613 * For all the buffers on this page,
1614 * if they are fully written out ordered data, move them onto BUF_CLEAN
1615 * so try_to_free_buffers() can reap them.
1617 * This function returns non-zero if we wish try_to_free_buffers()
1618 * to be called. We do this if the page is releasable by try_to_free_buffers().
1619 * We also do it if the page has locked or dirty buffers and the caller wants
1620 * us to perform sync or async writeout.
1622 * This complicates JBD locking somewhat. We aren't protected by the
1623 * BKL here. We wish to remove the buffer from its committing or
1624 * running transaction's ->t_datalist via __journal_unfile_buffer.
1626 * This may *change* the value of transaction_t->t_datalist, so anyone
1627 * who looks at t_datalist needs to lock against this function.
1629 * Even worse, someone may be doing a journal_dirty_data on this
1630 * buffer. So we need to lock against that. journal_dirty_data()
1631 * will come out of the lock with the buffer dirty, which makes it
1632 * ineligible for release here.
1634 * Who else is affected by this? hmm... Really the only contender
1635 * is do_get_write_access() - it could be looking at the buffer while
1636 * journal_try_to_free_buffer() is changing its state. But that
1637 * cannot happen because we never reallocate freed data as metadata
1638 * while the data is part of a transaction. Yes?
1640 int journal_try_to_free_buffers(journal_t
*journal
,
1641 struct page
*page
, int unused_gfp_mask
)
1643 struct buffer_head
*head
;
1644 struct buffer_head
*bh
;
1647 J_ASSERT(PageLocked(page
));
1649 head
= page_buffers(page
);
1652 struct journal_head
*jh
;
1655 * We take our own ref against the journal_head here to avoid
1656 * having to add tons of locking around each instance of
1657 * journal_remove_journal_head() and journal_put_journal_head().
1659 jh
= journal_grab_journal_head(bh
);
1663 jbd_lock_bh_state(bh
);
1664 __journal_try_to_free_buffer(journal
, bh
);
1665 journal_put_journal_head(jh
);
1666 jbd_unlock_bh_state(bh
);
1669 } while ((bh
= bh
->b_this_page
) != head
);
1670 ret
= try_to_free_buffers(page
);
1676 * This buffer is no longer needed. If it is on an older transaction's
1677 * checkpoint list we need to record it on this transaction's forget list
1678 * to pin this buffer (and hence its checkpointing transaction) down until
1679 * this transaction commits. If the buffer isn't on a checkpoint list, we
1681 * Returns non-zero if JBD no longer has an interest in the buffer.
1683 * Called under j_list_lock.
1685 * Called under jbd_lock_bh_state(bh).
1687 static int __dispose_buffer(struct journal_head
*jh
, transaction_t
*transaction
)
1690 struct buffer_head
*bh
= jh2bh(jh
);
1692 __journal_unfile_buffer(jh
);
1694 if (jh
->b_cp_transaction
) {
1695 JBUFFER_TRACE(jh
, "on running+cp transaction");
1696 __journal_file_buffer(jh
, transaction
, BJ_Forget
);
1697 clear_buffer_jbddirty(bh
);
1700 JBUFFER_TRACE(jh
, "on running transaction");
1701 journal_remove_journal_head(bh
);
1708 * journal_invalidatepage
1710 * This code is tricky. It has a number of cases to deal with.
1712 * There are two invariants which this code relies on:
1714 * i_size must be updated on disk before we start calling invalidatepage on the
1717 * This is done in ext3 by defining an ext3_setattr method which
1718 * updates i_size before truncate gets going. By maintaining this
1719 * invariant, we can be sure that it is safe to throw away any buffers
1720 * attached to the current transaction: once the transaction commits,
1721 * we know that the data will not be needed.
1723 * Note however that we can *not* throw away data belonging to the
1724 * previous, committing transaction!
1726 * Any disk blocks which *are* part of the previous, committing
1727 * transaction (and which therefore cannot be discarded immediately) are
1728 * not going to be reused in the new running transaction
1730 * The bitmap committed_data images guarantee this: any block which is
1731 * allocated in one transaction and removed in the next will be marked
1732 * as in-use in the committed_data bitmap, so cannot be reused until
1733 * the next transaction to delete the block commits. This means that
1734 * leaving committing buffers dirty is quite safe: the disk blocks
1735 * cannot be reallocated to a different file and so buffer aliasing is
1739 * The above applies mainly to ordered data mode. In writeback mode we
1740 * don't make guarantees about the order in which data hits disk --- in
1741 * particular we don't guarantee that new dirty data is flushed before
1742 * transaction commit --- so it is always safe just to discard data
1743 * immediately in that mode. --sct
1747 * The journal_unmap_buffer helper function returns zero if the buffer
1748 * concerned remains pinned as an anonymous buffer belonging to an older
1751 * We're outside-transaction here. Either or both of j_running_transaction
1752 * and j_committing_transaction may be NULL.
1754 static int journal_unmap_buffer(journal_t
*journal
, struct buffer_head
*bh
)
1756 transaction_t
*transaction
;
1757 struct journal_head
*jh
;
1761 BUFFER_TRACE(bh
, "entry");
1764 * It is safe to proceed here without the j_list_lock because the
1765 * buffers cannot be stolen by try_to_free_buffers as long as we are
1766 * holding the page lock. --sct
1769 if (!buffer_jbd(bh
))
1770 goto zap_buffer_unlocked
;
1772 spin_lock(&journal
->j_state_lock
);
1773 jbd_lock_bh_state(bh
);
1774 spin_lock(&journal
->j_list_lock
);
1776 jh
= journal_grab_journal_head(bh
);
1778 goto zap_buffer_no_jh
;
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 spin_unlock(&journal
->j_list_lock
);
1808 jbd_unlock_bh_state(bh
);
1809 spin_unlock(&journal
->j_state_lock
);
1810 journal_put_journal_head(jh
);
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 spin_unlock(&journal
->j_list_lock
);
1822 jbd_unlock_bh_state(bh
);
1823 spin_unlock(&journal
->j_state_lock
);
1824 journal_put_journal_head(jh
);
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 /* If it is committing, we simply cannot touch it. We
1835 * can remove it's next_transaction pointer from the
1836 * running transaction if that is set, but nothing
1838 JBUFFER_TRACE(jh
, "on committing transaction");
1839 set_buffer_freed(bh
);
1840 if (jh
->b_next_transaction
) {
1841 J_ASSERT(jh
->b_next_transaction
==
1842 journal
->j_running_transaction
);
1843 jh
->b_next_transaction
= NULL
;
1845 spin_unlock(&journal
->j_list_lock
);
1846 jbd_unlock_bh_state(bh
);
1847 spin_unlock(&journal
->j_state_lock
);
1848 journal_put_journal_head(jh
);
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 may_free
= __dispose_buffer(jh
, transaction
);
1862 journal_put_journal_head(jh
);
1864 spin_unlock(&journal
->j_list_lock
);
1865 jbd_unlock_bh_state(bh
);
1866 spin_unlock(&journal
->j_state_lock
);
1867 zap_buffer_unlocked
:
1868 clear_buffer_dirty(bh
);
1869 J_ASSERT_BH(bh
, !buffer_jbddirty(bh
));
1870 clear_buffer_mapped(bh
);
1871 clear_buffer_req(bh
);
1872 clear_buffer_new(bh
);
1878 * int journal_invalidatepage()
1879 * @journal: journal to use for flush...
1880 * @page: page to flush
1881 * @offset: length of page to invalidate.
1883 * Reap page buffers containing data after offset in page.
1885 * Return non-zero if the page's buffers were successfully reaped.
1887 int 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 /* AKPM: doing lock_buffer here may be overly paranoid */
1910 if (offset
<= curr_off
) {
1911 /* This block is wholly outside the truncation point */
1913 may_free
&= journal_unmap_buffer(journal
, bh
);
1916 curr_off
= next_off
;
1919 } while (bh
!= head
);
1922 if (!may_free
|| !try_to_free_buffers(page
))
1924 J_ASSERT(!page_has_buffers(page
));
1930 * File a buffer on the given transaction list.
1932 void __journal_file_buffer(struct journal_head
*jh
,
1933 transaction_t
*transaction
, int jlist
)
1935 struct journal_head
**list
= NULL
;
1937 struct buffer_head
*bh
= jh2bh(jh
);
1939 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
1940 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
1942 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
1943 J_ASSERT_JH(jh
, jh
->b_transaction
== transaction
||
1944 jh
->b_transaction
== 0);
1946 if (jh
->b_transaction
&& jh
->b_jlist
== jlist
)
1949 /* The following list of buffer states needs to be consistent
1950 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1953 if (jlist
== BJ_Metadata
|| jlist
== BJ_Reserved
||
1954 jlist
== BJ_Shadow
|| jlist
== BJ_Forget
) {
1955 if (test_clear_buffer_dirty(bh
) ||
1956 test_clear_buffer_jbddirty(bh
))
1960 if (jh
->b_transaction
)
1961 __journal_unfile_buffer(jh
);
1962 jh
->b_transaction
= transaction
;
1966 J_ASSERT_JH(jh
, !jh
->b_committed_data
);
1967 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
1970 list
= &transaction
->t_sync_datalist
;
1973 transaction
->t_nr_buffers
++;
1974 list
= &transaction
->t_buffers
;
1977 list
= &transaction
->t_forget
;
1980 list
= &transaction
->t_iobuf_list
;
1983 list
= &transaction
->t_shadow_list
;
1986 list
= &transaction
->t_log_list
;
1989 list
= &transaction
->t_reserved_list
;
1992 list
= &transaction
->t_locked_list
;
1996 __blist_add_buffer(list
, jh
);
1997 jh
->b_jlist
= jlist
;
2000 set_buffer_jbddirty(bh
);
2003 void journal_file_buffer(struct journal_head
*jh
,
2004 transaction_t
*transaction
, int jlist
)
2006 jbd_lock_bh_state(jh2bh(jh
));
2007 spin_lock(&transaction
->t_journal
->j_list_lock
);
2008 __journal_file_buffer(jh
, transaction
, jlist
);
2009 spin_unlock(&transaction
->t_journal
->j_list_lock
);
2010 jbd_unlock_bh_state(jh2bh(jh
));
2014 * Remove a buffer from its current buffer list in preparation for
2015 * dropping it from its current transaction entirely. If the buffer has
2016 * already started to be used by a subsequent transaction, refile the
2017 * buffer on that transaction's metadata list.
2019 * Called under journal->j_list_lock
2021 * Called under jbd_lock_bh_state(jh2bh(jh))
2023 void __journal_refile_buffer(struct journal_head
*jh
)
2026 struct buffer_head
*bh
= jh2bh(jh
);
2028 J_ASSERT_JH(jh
, jbd_is_locked_bh_state(bh
));
2029 if (jh
->b_transaction
)
2030 assert_spin_locked(&jh
->b_transaction
->t_journal
->j_list_lock
);
2032 /* If the buffer is now unused, just drop it. */
2033 if (jh
->b_next_transaction
== NULL
) {
2034 __journal_unfile_buffer(jh
);
2039 * It has been modified by a later transaction: add it to the new
2040 * transaction's metadata list.
2043 was_dirty
= test_clear_buffer_jbddirty(bh
);
2044 __journal_unfile_buffer(jh
);
2045 jh
->b_transaction
= jh
->b_next_transaction
;
2046 jh
->b_next_transaction
= NULL
;
2047 __journal_file_buffer(jh
, jh
->b_transaction
, BJ_Metadata
);
2048 J_ASSERT_JH(jh
, jh
->b_transaction
->t_state
== T_RUNNING
);
2051 set_buffer_jbddirty(bh
);
2055 * For the unlocked version of this call, also make sure that any
2056 * hanging journal_head is cleaned up if necessary.
2058 * __journal_refile_buffer is usually called as part of a single locked
2059 * operation on a buffer_head, in which the caller is probably going to
2060 * be hooking the journal_head onto other lists. In that case it is up
2061 * to the caller to remove the journal_head if necessary. For the
2062 * unlocked journal_refile_buffer call, the caller isn't going to be
2063 * doing anything else to the buffer so we need to do the cleanup
2064 * ourselves to avoid a jh leak.
2066 * *** The journal_head may be freed by this call! ***
2068 void journal_refile_buffer(journal_t
*journal
, struct journal_head
*jh
)
2070 struct buffer_head
*bh
= jh2bh(jh
);
2072 jbd_lock_bh_state(bh
);
2073 spin_lock(&journal
->j_list_lock
);
2075 __journal_refile_buffer(jh
);
2076 jbd_unlock_bh_state(bh
);
2077 journal_remove_journal_head(bh
);
2079 spin_unlock(&journal
->j_list_lock
);