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);
365 * int journal_restart() - restart a handle .
366 * @handle: handle to restart
367 * @nblocks: nr credits requested
369 * Restart a handle for a multi-transaction filesystem
372 * If the journal_extend() call above fails to grant new buffer credits
373 * to a running handle, a call to journal_restart will commit the
374 * handle's transaction so far and reattach the handle to a new
375 * transaction capabable of guaranteeing the requested number of
379 int journal_restart(handle_t *handle, int nblocks)
381 transaction_t *transaction = handle->h_transaction;
382 journal_t *journal = transaction->t_journal;
385 /* If we've had an abort of any type, don't even think about
386 * actually doing the restart! */
387 if (is_handle_aborted(handle))
391 * First unlink the handle from its current transaction, and start the
394 J_ASSERT(transaction->t_updates > 0);
395 J_ASSERT(journal_current_handle() == handle);
397 spin_lock(&journal->j_state_lock);
398 spin_lock(&transaction->t_handle_lock);
399 transaction->t_outstanding_credits -= handle->h_buffer_credits;
400 transaction->t_updates--;
402 if (!transaction->t_updates)
403 wake_up(&journal->j_wait_updates);
404 spin_unlock(&transaction->t_handle_lock);
406 jbd_debug(2, "restarting handle %p\n", handle);
407 __log_start_commit(journal, transaction->t_tid);
408 spin_unlock(&journal->j_state_lock);
410 handle->h_buffer_credits = nblocks;
411 ret = start_this_handle(journal, handle);
417 * void journal_lock_updates () - establish a transaction barrier.
418 * @journal: Journal to establish a barrier on.
420 * This locks out any further updates from being started, and blocks
421 * until all existing updates have completed, returning only once the
422 * journal is in a quiescent state with no updates running.
424 * The journal lock should not be held on entry.
426 void journal_lock_updates(journal_t *journal)
430 spin_lock(&journal->j_state_lock);
431 ++journal->j_barrier_count;
433 /* Wait until there are no running updates */
435 transaction_t *transaction = journal->j_running_transaction;
440 spin_lock(&transaction->t_handle_lock);
441 if (!transaction->t_updates) {
442 spin_unlock(&transaction->t_handle_lock);
445 prepare_to_wait(&journal->j_wait_updates, &wait,
446 TASK_UNINTERRUPTIBLE);
447 spin_unlock(&transaction->t_handle_lock);
448 spin_unlock(&journal->j_state_lock);
450 finish_wait(&journal->j_wait_updates, &wait);
451 spin_lock(&journal->j_state_lock);
453 spin_unlock(&journal->j_state_lock);
456 * We have now established a barrier against other normal updates, but
457 * we also need to barrier against other journal_lock_updates() calls
458 * to make sure that we serialise special journal-locked operations
461 down(&journal->j_barrier);
465 * void journal_unlock_updates (journal_t* journal) - release barrier
466 * @journal: Journal to release the barrier on.
468 * Release a transaction barrier obtained with journal_lock_updates().
470 * Should be called without the journal lock held.
472 void journal_unlock_updates (journal_t *journal)
474 J_ASSERT(journal->j_barrier_count != 0);
476 up(&journal->j_barrier);
477 spin_lock(&journal->j_state_lock);
478 --journal->j_barrier_count;
479 spin_unlock(&journal->j_state_lock);
480 wake_up(&journal->j_wait_transaction_locked);
484 * Report any unexpected dirty buffers which turn up. Normally those
485 * indicate an error, but they can occur if the user is running (say)
486 * tune2fs to modify the live filesystem, so we need the option of
487 * continuing as gracefully as possible. #
489 * The caller should already hold the journal lock and
490 * j_list_lock spinlock: most callers will need those anyway
491 * in order to probe the buffer's journaling state safely.
493 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
495 struct buffer_head *bh = jh2bh(jh);
498 if (buffer_dirty(bh)) {
499 /* If this buffer is one which might reasonably be dirty
500 * --- ie. data, or not part of this journal --- then
501 * we're OK to leave it alone, but otherwise we need to
502 * move the dirty bit to the journal's own internal
506 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
507 jlist == BJ_Shadow || jlist == BJ_Forget) {
508 if (test_clear_buffer_dirty(jh2bh(jh))) {
509 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
516 * If the buffer is already part of the current transaction, then there
517 * is nothing we need to do. If it is already part of a prior
518 * transaction which we are still committing to disk, then we need to
519 * make sure that we do not overwrite the old copy: we do copy-out to
520 * preserve the copy going to disk. We also account the buffer against
521 * the handle's metadata buffer credits (unless the buffer is already
522 * part of the transaction, that is).
526 do_get_write_access(handle_t *handle, struct journal_head *jh,
527 int force_copy, int *credits)
529 struct buffer_head *bh;
530 transaction_t *transaction;
533 char *frozen_buffer = NULL;
536 if (is_handle_aborted(handle))
539 transaction = handle->h_transaction;
540 journal = transaction->t_journal;
542 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
544 JBUFFER_TRACE(jh, "entry");
548 /* @@@ Need to check for errors here at some point. */
551 jbd_lock_bh_state(bh);
553 /* We now hold the buffer lock so it is safe to query the buffer
554 * state. Is the buffer dirty?
556 * If so, there are two possibilities. The buffer may be
557 * non-journaled, and undergoing a quite legitimate writeback.
558 * Otherwise, it is journaled, and we don't expect dirty buffers
559 * in that state (the buffers should be marked JBD_Dirty
560 * instead.) So either the IO is being done under our own
561 * control and this is a bug, or it's a third party IO such as
562 * dump(8) (which may leave the buffer scheduled for read ---
563 * ie. locked but not dirty) or tune2fs (which may actually have
564 * the buffer dirtied, ugh.) */
566 if (buffer_dirty(bh)) {
568 * First question: is this buffer already part of the current
569 * transaction or the existing committing transaction?
571 if (jh->b_transaction) {
573 jh->b_transaction == transaction ||
575 journal->j_committing_transaction);
576 if (jh->b_next_transaction)
577 J_ASSERT_JH(jh, jh->b_next_transaction ==
579 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
580 jbd_unexpected_dirty_buffer(jh);
587 if (is_handle_aborted(handle)) {
588 jbd_unlock_bh_state(bh);
594 * The buffer is already part of this transaction if b_transaction or
595 * b_next_transaction points to it
597 if (jh->b_transaction == transaction ||
598 jh->b_next_transaction == transaction)
602 * If there is already a copy-out version of this buffer, then we don't
603 * need to make another one
605 if (jh->b_frozen_data) {
606 JBUFFER_TRACE(jh, "has frozen data");
607 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
608 jh->b_next_transaction = transaction;
610 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
611 handle->h_buffer_credits--;
617 /* Is there data here we need to preserve? */
619 if (jh->b_transaction && jh->b_transaction != transaction) {
620 JBUFFER_TRACE(jh, "owned by older transaction");
621 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
622 J_ASSERT_JH(jh, jh->b_transaction ==
623 journal->j_committing_transaction);
625 /* There is one case we have to be very careful about.
626 * If the committing transaction is currently writing
627 * this buffer out to disk and has NOT made a copy-out,
628 * then we cannot modify the buffer contents at all
629 * right now. The essence of copy-out is that it is the
630 * extra copy, not the primary copy, which gets
631 * journaled. If the primary copy is already going to
632 * disk then we cannot do copy-out here. */
634 if (jh->b_jlist == BJ_Shadow) {
635 wait_queue_head_t *wqh;
638 JBUFFER_TRACE(jh, "on shadow: sleep");
639 jbd_unlock_bh_state(bh);
640 /* commit wakes up all shadow buffers after IO */
641 wqh = bh_waitq_head(bh);
643 prepare_to_wait(wqh, &wait,
644 TASK_UNINTERRUPTIBLE);
645 if (jh->b_jlist != BJ_Shadow)
649 finish_wait(wqh, &wait);
653 /* Only do the copy if the currently-owning transaction
654 * still needs it. If it is on the Forget list, the
655 * committing transaction is past that stage. The
656 * buffer had better remain locked during the kmalloc,
657 * but that should be true --- we hold the journal lock
658 * still and the buffer is already on the BUF_JOURNAL
659 * list so won't be flushed.
661 * Subtle point, though: if this is a get_undo_access,
662 * then we will be relying on the frozen_data to contain
663 * the new value of the committed_data record after the
664 * transaction, so we HAVE to force the frozen_data copy
667 if (jh->b_jlist != BJ_Forget || force_copy) {
668 JBUFFER_TRACE(jh, "generate frozen data");
669 if (!frozen_buffer) {
670 JBUFFER_TRACE(jh, "allocate memory for buffer");
671 jbd_unlock_bh_state(bh);
672 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
674 if (!frozen_buffer) {
676 "%s: OOM for frozen_buffer\n",
678 JBUFFER_TRACE(jh, "oom!");
680 jbd_lock_bh_state(bh);
685 jh->b_frozen_data = frozen_buffer;
686 frozen_buffer = NULL;
689 jh->b_next_transaction = transaction;
692 J_ASSERT(handle->h_buffer_credits > 0);
693 handle->h_buffer_credits--;
698 * Finally, if the buffer is not journaled right now, we need to make
699 * sure it doesn't get written to disk before the caller actually
700 * commits the new data
702 if (!jh->b_transaction) {
703 JBUFFER_TRACE(jh, "no transaction");
704 J_ASSERT_JH(jh, !jh->b_next_transaction);
705 jh->b_transaction = transaction;
706 JBUFFER_TRACE(jh, "file as BJ_Reserved");
707 spin_lock(&journal->j_list_lock);
708 __journal_file_buffer(jh, transaction, BJ_Reserved);
709 spin_unlock(&journal->j_list_lock);
718 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
719 "Possible IO failure.\n");
720 page = jh2bh(jh)->b_page;
721 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
722 source = kmap_atomic(page, KM_USER0);
723 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
724 kunmap_atomic(source, KM_USER0);
726 jbd_unlock_bh_state(bh);
729 * If we are about to journal a buffer, then any revoke pending on it is
732 journal_cancel_revoke(handle, jh);
736 kfree(frozen_buffer);
738 JBUFFER_TRACE(jh, "exit");
743 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
744 * @handle: transaction to add buffer modifications to
745 * @bh: bh to be used for metadata writes
747 * Returns an error code or 0 on success.
749 * In full data journalling mode the buffer may be of type BJ_AsyncData,
750 * because we're write()ing a buffer which is also part of a shared mapping.
753 int journal_get_write_access(handle_t *handle,
754 struct buffer_head *bh, int *credits)
756 struct journal_head *jh = journal_add_journal_head(bh);
759 /* We do not want to get caught playing with fields which the
760 * log thread also manipulates. Make sure that the buffer
761 * completes any outstanding IO before proceeding. */
762 rc = do_get_write_access(handle, jh, 0, credits);
763 journal_put_journal_head(jh);
769 * When the user wants to journal a newly created buffer_head
770 * (ie. getblk() returned a new buffer and we are going to populate it
771 * manually rather than reading off disk), then we need to keep the
772 * buffer_head locked until it has been completely filled with new
773 * data. In this case, we should be able to make the assertion that
774 * the bh is not already part of an existing transaction.
776 * The buffer should already be locked by the caller by this point.
777 * There is no lock ranking violation: it was a newly created,
778 * unlocked buffer beforehand. */
781 * int journal_get_create_access () - notify intent to use newly created bh
782 * @handle: transaction to new buffer to
785 * Call this if you create a new bh.
787 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
789 transaction_t *transaction = handle->h_transaction;
790 journal_t *journal = transaction->t_journal;
791 struct journal_head *jh = journal_add_journal_head(bh);
794 jbd_debug(5, "journal_head %p\n", jh);
796 if (is_handle_aborted(handle))
800 JBUFFER_TRACE(jh, "entry");
802 * The buffer may already belong to this transaction due to pre-zeroing
803 * in the filesystem's new_block code. It may also be on the previous,
804 * committing transaction's lists, but it HAS to be in Forget state in
805 * that case: the transaction must have deleted the buffer for it to be
808 jbd_lock_bh_state(bh);
809 spin_lock(&journal->j_list_lock);
810 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
811 jh->b_transaction == NULL ||
812 (jh->b_transaction == journal->j_committing_transaction &&
813 jh->b_jlist == BJ_Forget)));
815 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
816 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
818 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
819 handle->h_buffer_credits--;
821 if (jh->b_transaction == NULL) {
822 jh->b_transaction = transaction;
823 JBUFFER_TRACE(jh, "file as BJ_Reserved");
824 __journal_file_buffer(jh, transaction, BJ_Reserved);
825 } else if (jh->b_transaction == journal->j_committing_transaction) {
826 JBUFFER_TRACE(jh, "set next transaction");
827 jh->b_next_transaction = transaction;
829 spin_unlock(&journal->j_list_lock);
830 jbd_unlock_bh_state(bh);
833 * akpm: I added this. ext3_alloc_branch can pick up new indirect
834 * blocks which contain freed but then revoked metadata. We need
835 * to cancel the revoke in case we end up freeing it yet again
836 * and the reallocating as data - this would cause a second revoke,
837 * which hits an assertion error.
839 JBUFFER_TRACE(jh, "cancelling revoke");
840 journal_cancel_revoke(handle, jh);
841 journal_put_journal_head(jh);
847 * int journal_get_undo_access() - Notify intent to modify metadata with
848 * non-rewindable consequences
849 * @handle: transaction
850 * @bh: buffer to undo
851 * @credits: store the number of taken credits here (if not NULL)
853 * Sometimes there is a need to distinguish between metadata which has
854 * been committed to disk and that which has not. The ext3fs code uses
855 * this for freeing and allocating space, we have to make sure that we
856 * do not reuse freed space until the deallocation has been committed,
857 * since if we overwrote that space we would make the delete
858 * un-rewindable in case of a crash.
860 * To deal with that, journal_get_undo_access requests write access to a
861 * buffer for parts of non-rewindable operations such as delete
862 * operations on the bitmaps. The journaling code must keep a copy of
863 * the buffer's contents prior to the undo_access call until such time
864 * as we know that the buffer has definitely been committed to disk.
866 * We never need to know which transaction the committed data is part
867 * of, buffers touched here are guaranteed to be dirtied later and so
868 * will be committed to a new transaction in due course, at which point
869 * we can discard the old committed data pointer.
871 * Returns error number or 0 on success.
873 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh,
877 struct journal_head *jh = journal_add_journal_head(bh);
878 char *committed_data = NULL;
880 JBUFFER_TRACE(jh, "entry");
883 * Do this first --- it can drop the journal lock, so we want to
884 * make sure that obtaining the committed_data is done
885 * atomically wrt. completion of any outstanding commits.
887 err = do_get_write_access(handle, jh, 1, credits);
892 if (!jh->b_committed_data) {
893 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
894 if (!committed_data) {
895 printk(KERN_EMERG "%s: No memory for committed data\n",
902 jbd_lock_bh_state(bh);
903 if (!jh->b_committed_data) {
904 /* Copy out the current buffer contents into the
905 * preserved, committed copy. */
906 JBUFFER_TRACE(jh, "generate b_committed data");
907 if (!committed_data) {
908 jbd_unlock_bh_state(bh);
912 jh->b_committed_data = committed_data;
913 committed_data = NULL;
914 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
916 jbd_unlock_bh_state(bh);
918 journal_put_journal_head(jh);
920 kfree(committed_data);
925 * int journal_dirty_data() - mark a buffer as containing dirty data which
926 * needs to be flushed before we can commit the
927 * current transaction.
928 * @handle: transaction
929 * @bh: bufferhead to mark
931 * The buffer is placed on the transaction's data list and is marked as
932 * belonging to the transaction.
934 * Returns error number or 0 on success.
936 * journal_dirty_data() can be called via page_launder->ext3_writepage
939 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
941 journal_t *journal = handle->h_transaction->t_journal;
943 struct journal_head *jh;
945 if (is_handle_aborted(handle))
948 jh = journal_add_journal_head(bh);
949 JBUFFER_TRACE(jh, "entry");
952 * The buffer could *already* be dirty. Writeout can start
955 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
958 * What if the buffer is already part of a running transaction?
960 * There are two cases:
961 * 1) It is part of the current running transaction. Refile it,
962 * just in case we have allocated it as metadata, deallocated
963 * it, then reallocated it as data.
964 * 2) It is part of the previous, still-committing transaction.
965 * If all we want to do is to guarantee that the buffer will be
966 * written to disk before this new transaction commits, then
967 * being sure that the *previous* transaction has this same
968 * property is sufficient for us! Just leave it on its old
971 * In case (2), the buffer must not already exist as metadata
972 * --- that would violate write ordering (a transaction is free
973 * to write its data at any point, even before the previous
974 * committing transaction has committed). The caller must
975 * never, ever allow this to happen: there's nothing we can do
976 * about it in this layer.
978 jbd_lock_bh_state(bh);
979 spin_lock(&journal->j_list_lock);
980 if (jh->b_transaction) {
981 JBUFFER_TRACE(jh, "has transaction");
982 if (jh->b_transaction != handle->h_transaction) {
983 JBUFFER_TRACE(jh, "belongs to older transaction");
984 J_ASSERT_JH(jh, jh->b_transaction ==
985 journal->j_committing_transaction);
987 /* @@@ IS THIS TRUE ? */
989 * Not any more. Scenario: someone does a write()
990 * in data=journal mode. The buffer's transaction has
991 * moved into commit. Then someone does another
992 * write() to the file. We do the frozen data copyout
993 * and set b_next_transaction to point to j_running_t.
994 * And while we're in that state, someone does a
995 * writepage() in an attempt to pageout the same area
996 * of the file via a shared mapping. At present that
997 * calls journal_dirty_data(), and we get right here.
998 * It may be too late to journal the data. Simply
999 * falling through to the next test will suffice: the
1000 * data will be dirty and wil be checkpointed. The
1001 * ordering comments in the next comment block still
1004 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1007 * If we're journalling data, and this buffer was
1008 * subject to a write(), it could be metadata, forget
1009 * or shadow against the committing transaction. Now,
1010 * someone has dirtied the same darn page via a mapping
1011 * and it is being writepage()'d.
1012 * We *could* just steal the page from commit, with some
1013 * fancy locking there. Instead, we just skip it -
1014 * don't tie the page's buffers to the new transaction
1016 * Implication: if we crash before the writepage() data
1017 * is written into the filesystem, recovery will replay
1020 if (jh->b_jlist != BJ_None &&
1021 jh->b_jlist != BJ_SyncData &&
1022 jh->b_jlist != BJ_Locked) {
1023 JBUFFER_TRACE(jh, "Not stealing");
1028 * This buffer may be undergoing writeout in commit. We
1029 * can't return from here and let the caller dirty it
1030 * again because that can cause the write-out loop in
1031 * commit to never terminate.
1033 if (buffer_dirty(bh)) {
1035 spin_unlock(&journal->j_list_lock);
1036 jbd_unlock_bh_state(bh);
1038 sync_dirty_buffer(bh);
1039 jbd_lock_bh_state(bh);
1040 spin_lock(&journal->j_list_lock);
1041 /* The buffer may become locked again at any
1042 time if it is redirtied */
1045 /* journal_clean_data_list() may have got there first */
1046 if (jh->b_transaction != NULL) {
1047 JBUFFER_TRACE(jh, "unfile from commit");
1048 __journal_unfile_buffer(jh);
1050 /* The buffer will be refiled below */
1054 * Special case --- the buffer might actually have been
1055 * allocated and then immediately deallocated in the previous,
1056 * committing transaction, so might still be left on that
1057 * transaction's metadata lists.
1059 if (jh->b_jlist != BJ_SyncData && jh->b_jlist != BJ_Locked) {
1060 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1061 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1062 __journal_unfile_buffer(jh);
1063 JBUFFER_TRACE(jh, "file as data");
1064 __journal_file_buffer(jh, handle->h_transaction,
1068 JBUFFER_TRACE(jh, "not on a transaction");
1069 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1072 spin_unlock(&journal->j_list_lock);
1073 jbd_unlock_bh_state(bh);
1075 BUFFER_TRACE(bh, "brelse");
1078 JBUFFER_TRACE(jh, "exit");
1079 journal_put_journal_head(jh);
1084 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1085 * @handle: transaction to add buffer to.
1086 * @bh: buffer to mark
1088 * mark dirty metadata which needs to be journaled as part of the current
1091 * The buffer is placed on the transaction's metadata list and is marked
1092 * as belonging to the transaction.
1094 * Returns error number or 0 on success.
1096 * Special care needs to be taken if the buffer already belongs to the
1097 * current committing transaction (in which case we should have frozen
1098 * data present for that commit). In that case, we don't relink the
1099 * buffer: that only gets done when the old transaction finally
1100 * completes its commit.
1102 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1104 transaction_t *transaction = handle->h_transaction;
1105 journal_t *journal = transaction->t_journal;
1106 struct journal_head *jh = bh2jh(bh);
1108 jbd_debug(5, "journal_head %p\n", jh);
1109 JBUFFER_TRACE(jh, "entry");
1110 if (is_handle_aborted(handle))
1113 jbd_lock_bh_state(bh);
1116 * fastpath, to avoid expensive locking. If this buffer is already
1117 * on the running transaction's metadata list there is nothing to do.
1118 * Nobody can take it off again because there is a handle open.
1119 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1120 * result in this test being false, so we go in and take the locks.
1122 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1123 JBUFFER_TRACE(jh, "fastpath");
1124 J_ASSERT_JH(jh, jh->b_transaction ==
1125 journal->j_running_transaction);
1129 set_buffer_jbddirty(bh);
1132 * Metadata already on the current transaction list doesn't
1133 * need to be filed. Metadata on another transaction's list must
1134 * be committing, and will be refiled once the commit completes:
1135 * leave it alone for now.
1137 if (jh->b_transaction != transaction) {
1138 JBUFFER_TRACE(jh, "already on other transaction");
1139 J_ASSERT_JH(jh, jh->b_transaction ==
1140 journal->j_committing_transaction);
1141 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1142 /* And this case is illegal: we can't reuse another
1143 * transaction's data buffer, ever. */
1147 /* That test should have eliminated the following case: */
1148 J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1150 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1151 spin_lock(&journal->j_list_lock);
1152 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1153 spin_unlock(&journal->j_list_lock);
1155 jbd_unlock_bh_state(bh);
1157 JBUFFER_TRACE(jh, "exit");
1162 * journal_release_buffer: undo a get_write_access without any buffer
1163 * updates, if the update decided in the end that it didn't need access.
1165 * The caller passes in the number of credits which should be put back for
1166 * this buffer (zero or one).
1168 * We leave the buffer attached to t_reserved_list because even though this
1169 * handle doesn't want it, some other concurrent handle may want to journal
1170 * this buffer. If that handle is curently in between get_write_access() and
1171 * journal_dirty_metadata() then it expects the buffer to be reserved. If
1172 * we were to rip it off t_reserved_list here, the other handle will explode
1173 * when journal_dirty_metadata is presented with a non-reserved buffer.
1175 * If nobody really wants to journal this buffer then it will be thrown
1176 * away at the start of commit.
1179 journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
1181 BUFFER_TRACE(bh, "entry");
1182 handle->h_buffer_credits += credits;
1186 * void journal_forget() - bforget() for potentially-journaled buffers.
1187 * @handle: transaction handle
1188 * @bh: bh to 'forget'
1190 * We can only do the bforget if there are no commits pending against the
1191 * buffer. If the buffer is dirty in the current running transaction we
1192 * can safely unlink it.
1194 * bh may not be a journalled buffer at all - it may be a non-JBD
1195 * buffer which came off the hashtable. Check for this.
1197 * Decrements bh->b_count by one.
1199 * Allow this call even if the handle has aborted --- it may be part of
1200 * the caller's cleanup after an abort.
1202 void journal_forget(handle_t *handle, struct buffer_head *bh)
1204 transaction_t *transaction = handle->h_transaction;
1205 journal_t *journal = transaction->t_journal;
1206 struct journal_head *jh;
1208 BUFFER_TRACE(bh, "entry");
1210 jbd_lock_bh_state(bh);
1211 spin_lock(&journal->j_list_lock);
1213 if (!buffer_jbd(bh))
1217 if (jh->b_transaction == handle->h_transaction) {
1218 J_ASSERT_JH(jh, !jh->b_frozen_data);
1220 /* If we are forgetting a buffer which is already part
1221 * of this transaction, then we can just drop it from
1222 * the transaction immediately. */
1223 clear_buffer_dirty(bh);
1224 clear_buffer_jbddirty(bh);
1226 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1227 J_ASSERT_JH(jh, !jh->b_committed_data);
1229 __journal_unfile_buffer(jh);
1232 * We are no longer going to journal this buffer.
1233 * However, the commit of this transaction is still
1234 * important to the buffer: the delete that we are now
1235 * processing might obsolete an old log entry, so by
1236 * committing, we can satisfy the buffer's checkpoint.
1238 * So, if we have a checkpoint on the buffer, we should
1239 * now refile the buffer on our BJ_Forget list so that
1240 * we know to remove the checkpoint after we commit.
1243 if (jh->b_cp_transaction) {
1244 __journal_file_buffer(jh, transaction, BJ_Forget);
1246 journal_remove_journal_head(bh);
1248 if (!buffer_jbd(bh)) {
1249 spin_unlock(&journal->j_list_lock);
1250 jbd_unlock_bh_state(bh);
1255 } else if (jh->b_transaction) {
1256 J_ASSERT_JH(jh, (jh->b_transaction ==
1257 journal->j_committing_transaction));
1258 /* However, if the buffer is still owned by a prior
1259 * (committing) transaction, we can't drop it yet... */
1260 JBUFFER_TRACE(jh, "belongs to older transaction");
1261 /* ... but we CAN drop it from the new transaction if we
1262 * have also modified it since the original commit. */
1264 if (jh->b_next_transaction) {
1265 J_ASSERT(jh->b_next_transaction == transaction);
1266 jh->b_next_transaction = NULL;
1271 spin_unlock(&journal->j_list_lock);
1272 jbd_unlock_bh_state(bh);
1278 * void journal_callback_set() - Register a callback function for this handle.
1279 * @handle: handle to attach the callback to.
1280 * @func: function to callback.
1281 * @jcb: structure with additional information required by func() , and
1282 * some space for jbd internal information.
1284 * The function will be
1285 * called when the transaction that this handle is part of has been
1286 * committed to disk with the original callback data struct and the
1287 * error status of the journal as parameters. There is no guarantee of
1288 * ordering between handles within a single transaction, nor between
1289 * callbacks registered on the same handle.
1291 * The caller is responsible for allocating the journal_callback struct.
1292 * This is to allow the caller to add as much extra data to the callback
1293 * as needed, but reduce the overhead of multiple allocations. The caller
1294 * allocated struct must start with a struct journal_callback at offset 0,
1295 * and has the caller-specific data afterwards.
1297 void journal_callback_set(handle_t *handle,
1298 void (*func)(struct journal_callback *jcb, int error),
1299 struct journal_callback *jcb)
1301 spin_lock(&handle->h_transaction->t_jcb_lock);
1302 list_add_tail(&jcb->jcb_list, &handle->h_jcb);
1303 spin_unlock(&handle->h_transaction->t_jcb_lock);
1304 jcb->jcb_func = func;
1308 * int journal_stop() - complete a transaction
1309 * @handle: tranaction to complete.
1311 * All done for a particular handle.
1313 * There is not much action needed here. We just return any remaining
1314 * buffer credits to the transaction and remove the handle. The only
1315 * complication is that we need to start a commit operation if the
1316 * filesystem is marked for synchronous update.
1318 * journal_stop itself will not usually return an error, but it may
1319 * do so in unusual circumstances. In particular, expect it to
1320 * return -EIO if a journal_abort has been executed since the
1321 * transaction began.
1323 int journal_stop(handle_t *handle)
1325 transaction_t *transaction = handle->h_transaction;
1326 journal_t *journal = transaction->t_journal;
1327 int old_handle_count, err;
1329 J_ASSERT(transaction->t_updates > 0);
1330 J_ASSERT(journal_current_handle() == handle);
1332 if (is_handle_aborted(handle))
1337 if (--handle->h_ref > 0) {
1338 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1343 jbd_debug(4, "Handle %p going down\n", handle);
1346 * Implement synchronous transaction batching. If the handle
1347 * was synchronous, don't force a commit immediately. Let's
1348 * yield and let another thread piggyback onto this transaction.
1349 * Keep doing that while new threads continue to arrive.
1350 * It doesn't cost much - we're about to run a commit and sleep
1351 * on IO anyway. Speeds up many-threaded, many-dir operations
1354 if (handle->h_sync) {
1356 old_handle_count = transaction->t_handle_count;
1357 set_current_state(TASK_UNINTERRUPTIBLE);
1358 schedule_timeout(1);
1359 } while (old_handle_count != transaction->t_handle_count);
1362 current->journal_info = NULL;
1363 spin_lock(&journal->j_state_lock);
1364 spin_lock(&transaction->t_handle_lock);
1365 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1366 transaction->t_updates--;
1367 if (!transaction->t_updates) {
1368 wake_up(&journal->j_wait_updates);
1369 if (journal->j_barrier_count)
1370 wake_up(&journal->j_wait_transaction_locked);
1373 /* Move callbacks from the handle to the transaction. */
1374 spin_lock(&transaction->t_jcb_lock);
1375 list_splice(&handle->h_jcb, &transaction->t_jcb);
1376 spin_unlock(&transaction->t_jcb_lock);
1379 * If the handle is marked SYNC, we need to set another commit
1380 * going! We also want to force a commit if the current
1381 * transaction is occupying too much of the log, or if the
1382 * transaction is too old now.
1384 if (handle->h_sync ||
1385 transaction->t_outstanding_credits >
1386 journal->j_max_transaction_buffers ||
1387 time_after_eq(jiffies, transaction->t_expires)) {
1388 /* Do this even for aborted journals: an abort still
1389 * completes the commit thread, it just doesn't write
1390 * anything to disk. */
1391 tid_t tid = transaction->t_tid;
1393 spin_unlock(&transaction->t_handle_lock);
1394 jbd_debug(2, "transaction too old, requesting commit for "
1395 "handle %p\n", handle);
1396 /* This is non-blocking */
1397 __log_start_commit(journal, transaction->t_tid);
1398 spin_unlock(&journal->j_state_lock);
1401 * Special case: JFS_SYNC synchronous updates require us
1402 * to wait for the commit to complete.
1404 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1405 err = log_wait_commit(journal, tid);
1407 spin_unlock(&transaction->t_handle_lock);
1408 spin_unlock(&journal->j_state_lock);
1411 jbd_free_handle(handle);
1415 /**int journal_force_commit() - force any uncommitted transactions
1416 * @journal: journal to force
1418 * For synchronous operations: force any uncommitted transactions
1419 * to disk. May seem kludgy, but it reuses all the handle batching
1420 * code in a very simple manner.
1422 int journal_force_commit(journal_t *journal)
1427 handle = journal_start(journal, 1);
1428 if (IS_ERR(handle)) {
1429 ret = PTR_ERR(handle);
1432 ret = journal_stop(handle);
1439 * List management code snippets: various functions for manipulating the
1440 * transaction buffer lists.
1445 * Append a buffer to a transaction list, given the transaction's list head
1448 * j_list_lock is held.
1450 * jbd_lock_bh_state(jh2bh(jh)) is held.
1454 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1457 jh->b_tnext = jh->b_tprev = jh;
1460 /* Insert at the tail of the list to preserve order */
1461 struct journal_head *first = *list, *last = first->b_tprev;
1463 jh->b_tnext = first;
1464 last->b_tnext = first->b_tprev = jh;
1469 * Remove a buffer from a transaction list, given the transaction's list
1472 * Called with j_list_lock held, and the journal may not be locked.
1474 * jbd_lock_bh_state(jh2bh(jh)) is held.
1478 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1481 *list = jh->b_tnext;
1485 jh->b_tprev->b_tnext = jh->b_tnext;
1486 jh->b_tnext->b_tprev = jh->b_tprev;
1490 * Remove a buffer from the appropriate transaction list.
1492 * Note that this function can *change* the value of
1493 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1494 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1495 * is holding onto a copy of one of thee pointers, it could go bad.
1496 * Generally the caller needs to re-read the pointer from the transaction_t.
1498 * Called under j_list_lock. The journal may not be locked.
1500 void __journal_unfile_buffer(struct journal_head *jh)
1502 struct journal_head **list = 0;
1503 transaction_t *transaction;
1504 struct buffer_head *bh = jh2bh(jh);
1506 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1507 transaction = jh->b_transaction;
1509 assert_spin_locked(&transaction->t_journal->j_list_lock);
1511 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1512 if (jh->b_jlist != BJ_None)
1513 J_ASSERT_JH(jh, transaction != 0);
1515 switch (jh->b_jlist) {
1519 list = &transaction->t_sync_datalist;
1522 transaction->t_nr_buffers--;
1523 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1524 list = &transaction->t_buffers;
1527 list = &transaction->t_forget;
1530 list = &transaction->t_iobuf_list;
1533 list = &transaction->t_shadow_list;
1536 list = &transaction->t_log_list;
1539 list = &transaction->t_reserved_list;
1542 list = &transaction->t_locked_list;
1546 __blist_del_buffer(list, jh);
1547 jh->b_jlist = BJ_None;
1548 if (test_clear_buffer_jbddirty(bh))
1549 mark_buffer_dirty(bh); /* Expose it to the VM */
1551 jh->b_transaction = NULL;
1554 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1556 jbd_lock_bh_state(jh2bh(jh));
1557 spin_lock(&journal->j_list_lock);
1558 __journal_unfile_buffer(jh);
1559 spin_unlock(&journal->j_list_lock);
1560 jbd_unlock_bh_state(jh2bh(jh));
1564 * Called from journal_try_to_free_buffers().
1566 * Called under jbd_lock_bh_state(bh)
1569 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1571 struct journal_head *jh;
1575 if (buffer_locked(bh) || buffer_dirty(bh))
1578 if (jh->b_next_transaction != 0)
1581 spin_lock(&journal->j_list_lock);
1582 if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1583 if (jh->b_jlist == BJ_SyncData || jh->b_jlist == BJ_Locked) {
1584 /* A written-back ordered data buffer */
1585 JBUFFER_TRACE(jh, "release data");
1586 __journal_unfile_buffer(jh);
1587 journal_remove_journal_head(bh);
1590 } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1591 /* written-back checkpointed metadata buffer */
1592 if (jh->b_jlist == BJ_None) {
1593 JBUFFER_TRACE(jh, "remove from checkpoint list");
1594 __journal_remove_checkpoint(jh);
1595 journal_remove_journal_head(bh);
1599 spin_unlock(&journal->j_list_lock);
1606 * int journal_try_to_free_buffers() - try to free page buffers.
1607 * @journal: journal for operation
1608 * @page: to try and free
1609 * @gfp_mask: 'IO' mode for try_to_free_buffers()
1612 * For all the buffers on this page,
1613 * if they are fully written out ordered data, move them onto BUF_CLEAN
1614 * so try_to_free_buffers() can reap them.
1616 * This function returns non-zero if we wish try_to_free_buffers()
1617 * to be called. We do this if the page is releasable by try_to_free_buffers().
1618 * We also do it if the page has locked or dirty buffers and the caller wants
1619 * us to perform sync or async writeout.
1621 * This complicates JBD locking somewhat. We aren't protected by the
1622 * BKL here. We wish to remove the buffer from its committing or
1623 * running transaction's ->t_datalist via __journal_unfile_buffer.
1625 * This may *change* the value of transaction_t->t_datalist, so anyone
1626 * who looks at t_datalist needs to lock against this function.
1628 * Even worse, someone may be doing a journal_dirty_data on this
1629 * buffer. So we need to lock against that. journal_dirty_data()
1630 * will come out of the lock with the buffer dirty, which makes it
1631 * ineligible for release here.
1633 * Who else is affected by this? hmm... Really the only contender
1634 * is do_get_write_access() - it could be looking at the buffer while
1635 * journal_try_to_free_buffer() is changing its state. But that
1636 * cannot happen because we never reallocate freed data as metadata
1637 * while the data is part of a transaction. Yes?
1639 int journal_try_to_free_buffers(journal_t *journal,
1640 struct page *page, int unused_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 * journal_remove_journal_head() and journal_put_journal_head().
1658 jh = journal_grab_journal_head(bh);
1662 jbd_lock_bh_state(bh);
1663 __journal_try_to_free_buffer(journal, bh);
1664 journal_put_journal_head(jh);
1665 jbd_unlock_bh_state(bh);
1668 } while ((bh = bh->b_this_page) != head);
1669 ret = try_to_free_buffers(page);
1675 * This buffer is no longer needed. If it is on an older transaction's
1676 * checkpoint list we need to record it on this transaction's forget list
1677 * to pin this buffer (and hence its checkpointing transaction) down until
1678 * this transaction commits. If the buffer isn't on a checkpoint list, we
1680 * Returns non-zero if JBD no longer has an interest in the buffer.
1682 * Called under j_list_lock.
1684 * Called under jbd_lock_bh_state(bh).
1686 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1689 struct buffer_head *bh = jh2bh(jh);
1691 __journal_unfile_buffer(jh);
1693 if (jh->b_cp_transaction) {
1694 JBUFFER_TRACE(jh, "on running+cp transaction");
1695 __journal_file_buffer(jh, transaction, BJ_Forget);
1696 clear_buffer_jbddirty(bh);
1699 JBUFFER_TRACE(jh, "on running transaction");
1700 journal_remove_journal_head(bh);
1707 * journal_invalidatepage
1709 * This code is tricky. It has a number of cases to deal with.
1711 * There are two invariants which this code relies on:
1713 * i_size must be updated on disk before we start calling invalidatepage on the
1716 * This is done in ext3 by defining an ext3_setattr method which
1717 * updates i_size before truncate gets going. By maintaining this
1718 * invariant, we can be sure that it is safe to throw away any buffers
1719 * attached to the current transaction: once the transaction commits,
1720 * we know that the data will not be needed.
1722 * Note however that we can *not* throw away data belonging to the
1723 * previous, committing transaction!
1725 * Any disk blocks which *are* part of the previous, committing
1726 * transaction (and which therefore cannot be discarded immediately) are
1727 * not going to be reused in the new running transaction
1729 * The bitmap committed_data images guarantee this: any block which is
1730 * allocated in one transaction and removed in the next will be marked
1731 * as in-use in the committed_data bitmap, so cannot be reused until
1732 * the next transaction to delete the block commits. This means that
1733 * leaving committing buffers dirty is quite safe: the disk blocks
1734 * cannot be reallocated to a different file and so buffer aliasing is
1738 * The above applies mainly to ordered data mode. In writeback mode we
1739 * don't make guarantees about the order in which data hits disk --- in
1740 * particular we don't guarantee that new dirty data is flushed before
1741 * transaction commit --- so it is always safe just to discard data
1742 * immediately in that mode. --sct
1746 * The journal_unmap_buffer helper function returns zero if the buffer
1747 * concerned remains pinned as an anonymous buffer belonging to an older
1750 * We're outside-transaction here. Either or both of j_running_transaction
1751 * and j_committing_transaction may be NULL.
1753 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1755 transaction_t *transaction;
1756 struct journal_head *jh;
1760 BUFFER_TRACE(bh, "entry");
1763 * It is safe to proceed here without the j_list_lock because the
1764 * buffers cannot be stolen by try_to_free_buffers as long as we are
1765 * holding the page lock. --sct
1768 if (!buffer_jbd(bh))
1769 goto zap_buffer_unlocked;
1771 spin_lock(&journal->j_state_lock);
1772 jbd_lock_bh_state(bh);
1773 spin_lock(&journal->j_list_lock);
1776 * Now we have the locks, check again to see whether kjournald has
1777 * taken the buffer off the transaction.
1779 if (!buffer_jbd(bh))
1783 transaction = jh->b_transaction;
1784 if (transaction == NULL) {
1785 /* First case: not on any transaction. If it
1786 * has no checkpoint link, then we can zap it:
1787 * it's a writeback-mode buffer so we don't care
1788 * if it hits disk safely. */
1789 if (!jh->b_cp_transaction) {
1790 JBUFFER_TRACE(jh, "not on any transaction: zap");
1794 if (!buffer_dirty(bh)) {
1795 /* bdflush has written it. We can drop it now */
1799 /* OK, it must be in the journal but still not
1800 * written fully to disk: it's metadata or
1801 * journaled data... */
1803 if (journal->j_running_transaction) {
1804 /* ... and once the current transaction has
1805 * committed, the buffer won't be needed any
1807 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1808 ret = __dispose_buffer(jh,
1809 journal->j_running_transaction);
1810 spin_unlock(&journal->j_list_lock);
1811 jbd_unlock_bh_state(bh);
1812 spin_unlock(&journal->j_state_lock);
1815 /* There is no currently-running transaction. So the
1816 * orphan record which we wrote for this file must have
1817 * passed into commit. We must attach this buffer to
1818 * the committing transaction, if it exists. */
1819 if (journal->j_committing_transaction) {
1820 JBUFFER_TRACE(jh, "give to committing trans");
1821 ret = __dispose_buffer(jh,
1822 journal->j_committing_transaction);
1823 spin_unlock(&journal->j_list_lock);
1824 jbd_unlock_bh_state(bh);
1825 spin_unlock(&journal->j_state_lock);
1828 /* The orphan record's transaction has
1829 * committed. We can cleanse this buffer */
1830 clear_buffer_jbddirty(bh);
1834 } else if (transaction == journal->j_committing_transaction) {
1835 /* If it is committing, we simply cannot touch it. We
1836 * can remove it's next_transaction pointer from the
1837 * running transaction if that is set, but nothing
1839 JBUFFER_TRACE(jh, "on committing transaction");
1840 set_buffer_freed(bh);
1841 if (jh->b_next_transaction) {
1842 J_ASSERT(jh->b_next_transaction ==
1843 journal->j_running_transaction);
1844 jh->b_next_transaction = NULL;
1846 spin_unlock(&journal->j_list_lock);
1847 jbd_unlock_bh_state(bh);
1848 spin_unlock(&journal->j_state_lock);
1851 /* Good, the buffer belongs to the running transaction.
1852 * We are writing our own transaction's data, not any
1853 * previous one's, so it is safe to throw it away
1854 * (remember that we expect the filesystem to have set
1855 * i_size already for this truncate so recovery will not
1856 * expose the disk blocks we are discarding here.) */
1857 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1858 may_free = __dispose_buffer(jh, transaction);
1862 spin_unlock(&journal->j_list_lock);
1863 jbd_unlock_bh_state(bh);
1864 spin_unlock(&journal->j_state_lock);
1865 zap_buffer_unlocked:
1866 clear_buffer_dirty(bh);
1867 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1868 clear_buffer_mapped(bh);
1869 clear_buffer_req(bh);
1870 clear_buffer_new(bh);
1876 * int journal_invalidatepage()
1877 * @journal: journal to use for flush...
1878 * @page: page to flush
1879 * @offset: length of page to invalidate.
1881 * Reap page buffers containing data after offset in page.
1883 * Return non-zero if the page's buffers were successfully reaped.
1885 int journal_invalidatepage(journal_t *journal,
1887 unsigned long offset)
1889 struct buffer_head *head, *bh, *next;
1890 unsigned int curr_off = 0;
1893 if (!PageLocked(page))
1895 if (!page_has_buffers(page))
1898 /* We will potentially be playing with lists other than just the
1899 * data lists (especially for journaled data mode), so be
1900 * cautious in our locking. */
1902 head = bh = page_buffers(page);
1904 unsigned int next_off = curr_off + bh->b_size;
1905 next = bh->b_this_page;
1907 /* AKPM: doing lock_buffer here may be overly paranoid */
1908 if (offset <= curr_off) {
1909 /* This block is wholly outside the truncation point */
1911 may_free &= journal_unmap_buffer(journal, bh);
1914 curr_off = next_off;
1917 } while (bh != head);
1920 if (!may_free || !try_to_free_buffers(page))
1922 J_ASSERT(!page_has_buffers(page));
1928 * File a buffer on the given transaction list.
1930 void __journal_file_buffer(struct journal_head *jh,
1931 transaction_t *transaction, int jlist)
1933 struct journal_head **list = 0;
1935 struct buffer_head *bh = jh2bh(jh);
1937 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1938 assert_spin_locked(&transaction->t_journal->j_list_lock);
1940 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1941 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1942 jh->b_transaction == 0);
1944 if (jh->b_transaction && jh->b_jlist == jlist)
1947 /* The following list of buffer states needs to be consistent
1948 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1951 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1952 jlist == BJ_Shadow || jlist == BJ_Forget) {
1953 if (test_clear_buffer_dirty(bh) ||
1954 test_clear_buffer_jbddirty(bh))
1958 if (jh->b_transaction)
1959 __journal_unfile_buffer(jh);
1960 jh->b_transaction = transaction;
1964 J_ASSERT_JH(jh, !jh->b_committed_data);
1965 J_ASSERT_JH(jh, !jh->b_frozen_data);
1968 list = &transaction->t_sync_datalist;
1971 transaction->t_nr_buffers++;
1972 list = &transaction->t_buffers;
1975 list = &transaction->t_forget;
1978 list = &transaction->t_iobuf_list;
1981 list = &transaction->t_shadow_list;
1984 list = &transaction->t_log_list;
1987 list = &transaction->t_reserved_list;
1990 list = &transaction->t_locked_list;
1994 __blist_add_buffer(list, jh);
1995 jh->b_jlist = jlist;
1998 set_buffer_jbddirty(bh);
2001 void journal_file_buffer(struct journal_head *jh,
2002 transaction_t *transaction, int jlist)
2004 jbd_lock_bh_state(jh2bh(jh));
2005 spin_lock(&transaction->t_journal->j_list_lock);
2006 __journal_file_buffer(jh, transaction, jlist);
2007 spin_unlock(&transaction->t_journal->j_list_lock);
2008 jbd_unlock_bh_state(jh2bh(jh));
2012 * Remove a buffer from its current buffer list in preparation for
2013 * dropping it from its current transaction entirely. If the buffer has
2014 * already started to be used by a subsequent transaction, refile the
2015 * buffer on that transaction's metadata list.
2017 * Called under journal->j_list_lock
2019 * Called under jbd_lock_bh_state(jh2bh(jh))
2021 void __journal_refile_buffer(struct journal_head *jh)
2024 struct buffer_head *bh = jh2bh(jh);
2026 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2027 if (jh->b_transaction)
2028 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2030 /* If the buffer is now unused, just drop it. */
2031 if (jh->b_next_transaction == NULL) {
2032 __journal_unfile_buffer(jh);
2037 * It has been modified by a later transaction: add it to the new
2038 * transaction's metadata list.
2041 was_dirty = test_clear_buffer_jbddirty(bh);
2042 __journal_unfile_buffer(jh);
2043 jh->b_transaction = jh->b_next_transaction;
2044 jh->b_next_transaction = NULL;
2045 __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2046 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2049 set_buffer_jbddirty(bh);
2053 * For the unlocked version of this call, also make sure that any
2054 * hanging journal_head is cleaned up if necessary.
2056 * __journal_refile_buffer is usually called as part of a single locked
2057 * operation on a buffer_head, in which the caller is probably going to
2058 * be hooking the journal_head onto other lists. In that case it is up
2059 * to the caller to remove the journal_head if necessary. For the
2060 * unlocked journal_refile_buffer call, the caller isn't going to be
2061 * doing anything else to the buffer so we need to do the cleanup
2062 * ourselves to avoid a jh leak.
2064 * *** The journal_head may be freed by this call! ***
2066 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2068 struct buffer_head *bh = jh2bh(jh);
2070 jbd_lock_bh_state(bh);
2071 spin_lock(&journal->j_list_lock);
2073 __journal_refile_buffer(jh);
2074 jbd_unlock_bh_state(bh);
2075 journal_remove_journal_head(bh);
2077 spin_unlock(&journal->j_list_lock);