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) {
150 spin_unlock(&journal->j_state_lock);
151 jbd_debug(3, "Handle %p stalling...\n", handle);
152 wait_event(journal->j_wait_transaction_locked,
153 transaction->t_state != T_LOCKED);
158 * If there is not enough space left in the log to write all potential
159 * buffers requested by this operation, we need to stall pending a log
160 * checkpoint to free some more log space.
162 spin_lock(&transaction->t_handle_lock);
163 needed = transaction->t_outstanding_credits + nblocks;
165 if (needed > journal->j_max_transaction_buffers) {
167 * If the current transaction is already too large, then start
168 * to commit it: we can then go back and attach this handle to
173 jbd_debug(2, "Handle %p starting new commit...\n", handle);
174 spin_unlock(&transaction->t_handle_lock);
175 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
176 TASK_UNINTERRUPTIBLE);
177 __log_start_commit(journal, transaction->t_tid);
178 spin_unlock(&journal->j_state_lock);
180 finish_wait(&journal->j_wait_transaction_locked, &wait);
185 * The commit code assumes that it can get enough log space
186 * without forcing a checkpoint. This is *critical* for
187 * correctness: a checkpoint of a buffer which is also
188 * associated with a committing transaction creates a deadlock,
189 * so commit simply cannot force through checkpoints.
191 * We must therefore ensure the necessary space in the journal
192 * *before* starting to dirty potentially checkpointed buffers
193 * in the new transaction.
195 * The worst part is, any transaction currently committing can
196 * reduce the free space arbitrarily. Be careful to account for
197 * those buffers when checkpointing.
201 * @@@ AKPM: This seems rather over-defensive. We're giving commit
202 * a _lot_ of headroom: 1/4 of the journal plus the size of
203 * the committing transaction. Really, we only need to give it
204 * committing_transaction->t_outstanding_credits plus "enough" for
205 * the log control blocks.
206 * Also, this test is inconsitent with the matching one in
209 if (__log_space_left(journal) < jbd_space_needed(journal)) {
210 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
211 spin_unlock(&transaction->t_handle_lock);
212 __log_wait_for_space(journal);
216 /* OK, account for the buffers that this operation expects to
217 * use and add the handle to the running transaction. */
219 handle->h_transaction = transaction;
220 transaction->t_outstanding_credits += nblocks;
221 transaction->t_updates++;
222 transaction->t_handle_count++;
223 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
224 handle, nblocks, transaction->t_outstanding_credits,
225 __log_space_left(journal));
226 spin_unlock(&transaction->t_handle_lock);
227 spin_unlock(&journal->j_state_lock);
230 kfree(new_transaction);
234 /* Allocate a new handle. This should probably be in a slab... */
235 static handle_t *new_handle(int nblocks)
237 handle_t *handle = jbd_alloc_handle(GFP_NOFS);
240 memset(handle, 0, sizeof(*handle));
241 handle->h_buffer_credits = nblocks;
243 INIT_LIST_HEAD(&handle->h_jcb);
249 * handle_t *journal_start() - Obtain a new handle.
250 * @journal: Journal to start transaction on.
251 * @nblocks: number of block buffer we might modify
253 * We make sure that the transaction can guarantee at least nblocks of
254 * modified buffers in the log. We block until the log can guarantee
257 * This function is visible to journal users (like ext3fs), so is not
258 * called with the journal already locked.
260 * Return a pointer to a newly allocated handle, or NULL on failure
262 handle_t *journal_start(journal_t *journal, int nblocks)
264 handle_t *handle = journal_current_handle();
268 return ERR_PTR(-EROFS);
271 J_ASSERT(handle->h_transaction->t_journal == journal);
276 handle = new_handle(nblocks);
278 return ERR_PTR(-ENOMEM);
280 current->journal_info = handle;
282 err = start_this_handle(journal, handle);
284 jbd_free_handle(handle);
285 current->journal_info = NULL;
286 handle = ERR_PTR(err);
292 * int journal_extend() - extend buffer credits.
293 * @handle: handle to 'extend'
294 * @nblocks: nr blocks to try to extend by.
296 * Some transactions, such as large extends and truncates, can be done
297 * atomically all at once or in several stages. The operation requests
298 * a credit for a number of buffer modications in advance, but can
299 * extend its credit if it needs more.
301 * journal_extend tries to give the running handle more buffer credits.
302 * It does not guarantee that allocation - this is a best-effort only.
303 * The calling process MUST be able to deal cleanly with a failure to
306 * Return 0 on success, non-zero on failure.
308 * return code < 0 implies an error
309 * return code > 0 implies normal transaction-full status.
311 int journal_extend(handle_t *handle, int nblocks)
313 transaction_t *transaction = handle->h_transaction;
314 journal_t *journal = transaction->t_journal;
319 if (is_handle_aborted(handle))
324 spin_lock(&journal->j_state_lock);
326 /* Don't extend a locked-down transaction! */
327 if (handle->h_transaction->t_state != T_RUNNING) {
328 jbd_debug(3, "denied handle %p %d blocks: "
329 "transaction not running\n", handle, nblocks);
333 spin_lock(&transaction->t_handle_lock);
334 wanted = transaction->t_outstanding_credits + nblocks;
336 if (wanted > journal->j_max_transaction_buffers) {
337 jbd_debug(3, "denied handle %p %d blocks: "
338 "transaction too large\n", handle, nblocks);
342 if (wanted > __log_space_left(journal)) {
343 jbd_debug(3, "denied handle %p %d blocks: "
344 "insufficient log space\n", handle, nblocks);
348 handle->h_buffer_credits += nblocks;
349 transaction->t_outstanding_credits += nblocks;
352 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
354 spin_unlock(&transaction->t_handle_lock);
356 spin_unlock(&journal->j_state_lock);
362 * int journal_restart() - restart a handle .
363 * @handle: handle to restart
364 * @nblocks: nr credits requested
366 * Restart a handle for a multi-transaction filesystem
369 * If the journal_extend() call above fails to grant new buffer credits
370 * to a running handle, a call to journal_restart will commit the
371 * handle's transaction so far and reattach the handle to a new
372 * transaction capabable of guaranteeing the requested number of
376 int journal_restart(handle_t *handle, int nblocks)
378 transaction_t *transaction = handle->h_transaction;
379 journal_t *journal = transaction->t_journal;
382 /* If we've had an abort of any type, don't even think about
383 * actually doing the restart! */
384 if (is_handle_aborted(handle))
388 * First unlink the handle from its current transaction, and start the
391 J_ASSERT(transaction->t_updates > 0);
392 J_ASSERT(journal_current_handle() == handle);
394 spin_lock(&journal->j_state_lock);
395 spin_lock(&transaction->t_handle_lock);
396 transaction->t_outstanding_credits -= handle->h_buffer_credits;
397 transaction->t_updates--;
399 if (!transaction->t_updates)
400 wake_up(&journal->j_wait_updates);
401 spin_unlock(&transaction->t_handle_lock);
403 jbd_debug(2, "restarting handle %p\n", handle);
404 __log_start_commit(journal, transaction->t_tid);
405 spin_unlock(&journal->j_state_lock);
407 handle->h_buffer_credits = nblocks;
408 ret = start_this_handle(journal, handle);
414 * void journal_lock_updates () - establish a transaction barrier.
415 * @journal: Journal to establish a barrier on.
417 * This locks out any further updates from being started, and blocks
418 * until all existing updates have completed, returning only once the
419 * journal is in a quiescent state with no updates running.
421 * The journal lock should not be held on entry.
423 void journal_lock_updates(journal_t *journal)
427 spin_lock(&journal->j_state_lock);
428 ++journal->j_barrier_count;
430 /* Wait until there are no running updates */
432 transaction_t *transaction = journal->j_running_transaction;
437 spin_lock(&transaction->t_handle_lock);
438 if (!transaction->t_updates) {
439 spin_unlock(&transaction->t_handle_lock);
442 prepare_to_wait(&journal->j_wait_updates, &wait,
443 TASK_UNINTERRUPTIBLE);
444 spin_unlock(&transaction->t_handle_lock);
445 spin_unlock(&journal->j_state_lock);
447 finish_wait(&journal->j_wait_updates, &wait);
448 spin_lock(&journal->j_state_lock);
450 spin_unlock(&journal->j_state_lock);
453 * We have now established a barrier against other normal updates, but
454 * we also need to barrier against other journal_lock_updates() calls
455 * to make sure that we serialise special journal-locked operations
458 down(&journal->j_barrier);
462 * void journal_unlock_updates (journal_t* journal) - release barrier
463 * @journal: Journal to release the barrier on.
465 * Release a transaction barrier obtained with journal_lock_updates().
467 * Should be called without the journal lock held.
469 void journal_unlock_updates (journal_t *journal)
471 J_ASSERT(journal->j_barrier_count != 0);
473 up(&journal->j_barrier);
474 spin_lock(&journal->j_state_lock);
475 --journal->j_barrier_count;
476 spin_unlock(&journal->j_state_lock);
477 wake_up(&journal->j_wait_transaction_locked);
481 * Report any unexpected dirty buffers which turn up. Normally those
482 * indicate an error, but they can occur if the user is running (say)
483 * tune2fs to modify the live filesystem, so we need the option of
484 * continuing as gracefully as possible. #
486 * The caller should already hold the journal lock and
487 * j_list_lock spinlock: most callers will need those anyway
488 * in order to probe the buffer's journaling state safely.
490 static void jbd_unexpected_dirty_buffer(struct journal_head *jh)
492 struct buffer_head *bh = jh2bh(jh);
495 if (buffer_dirty(bh)) {
496 /* If this buffer is one which might reasonably be dirty
497 * --- ie. data, or not part of this journal --- then
498 * we're OK to leave it alone, but otherwise we need to
499 * move the dirty bit to the journal's own internal
503 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
504 jlist == BJ_Shadow || jlist == BJ_Forget) {
505 if (test_clear_buffer_dirty(jh2bh(jh))) {
506 set_bit(BH_JBDDirty, &jh2bh(jh)->b_state);
513 * If the buffer is already part of the current transaction, then there
514 * is nothing we need to do. If it is already part of a prior
515 * transaction which we are still committing to disk, then we need to
516 * make sure that we do not overwrite the old copy: we do copy-out to
517 * preserve the copy going to disk. We also account the buffer against
518 * the handle's metadata buffer credits (unless the buffer is already
519 * part of the transaction, that is).
524 do_get_write_access(handle_t *handle, struct journal_head *jh,
525 int force_copy, int *credits)
527 struct buffer_head *bh;
528 transaction_t *transaction = handle->h_transaction;
529 journal_t *journal = transaction->t_journal;
531 char *frozen_buffer = NULL;
534 jbd_debug(5, "buffer_head %p, force_copy %d\n", jh, force_copy);
536 JBUFFER_TRACE(jh, "entry");
540 /* @@@ Need to check for errors here at some point. */
543 jbd_lock_bh_state(bh);
544 spin_lock(&journal->j_list_lock);
546 /* We now hold the buffer lock so it is safe to query the buffer
547 * state. Is the buffer dirty?
549 * If so, there are two possibilities. The buffer may be
550 * non-journaled, and undergoing a quite legitimate writeback.
551 * Otherwise, it is journaled, and we don't expect dirty buffers
552 * in that state (the buffers should be marked JBD_Dirty
553 * instead.) So either the IO is being done under our own
554 * control and this is a bug, or it's a third party IO such as
555 * dump(8) (which may leave the buffer scheduled for read ---
556 * ie. locked but not dirty) or tune2fs (which may actually have
557 * the buffer dirtied, ugh.) */
559 if (buffer_dirty(bh)) {
560 /* First question: is this buffer already part of the
561 * current transaction or the existing committing
563 if (jh->b_transaction) {
565 jh->b_transaction == transaction ||
567 journal->j_committing_transaction);
568 if (jh->b_next_transaction)
569 J_ASSERT_JH(jh, jh->b_next_transaction ==
571 JBUFFER_TRACE(jh, "Unexpected dirty buffer");
572 jbd_unexpected_dirty_buffer(jh);
579 if (is_handle_aborted(handle)) {
580 spin_unlock(&journal->j_list_lock);
581 jbd_unlock_bh_state(bh);
586 /* The buffer is already part of this transaction if
587 * b_transaction or b_next_transaction points to it. */
589 if (jh->b_transaction == transaction ||
590 jh->b_next_transaction == transaction)
593 /* If there is already a copy-out version of this buffer, then
594 * we don't need to make another one. */
596 if (jh->b_frozen_data) {
597 JBUFFER_TRACE(jh, "has frozen data");
598 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
599 jh->b_next_transaction = transaction;
601 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
602 handle->h_buffer_credits--;
608 /* Is there data here we need to preserve? */
610 if (jh->b_transaction && jh->b_transaction != transaction) {
611 JBUFFER_TRACE(jh, "owned by older transaction");
612 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
613 J_ASSERT_JH(jh, jh->b_transaction ==
614 journal->j_committing_transaction);
616 /* There is one case we have to be very careful about.
617 * If the committing transaction is currently writing
618 * this buffer out to disk and has NOT made a copy-out,
619 * then we cannot modify the buffer contents at all
620 * right now. The essence of copy-out is that it is the
621 * extra copy, not the primary copy, which gets
622 * journaled. If the primary copy is already going to
623 * disk then we cannot do copy-out here. */
625 if (jh->b_jlist == BJ_Shadow) {
626 wait_queue_head_t *wqh;
628 JBUFFER_TRACE(jh, "on shadow: sleep");
629 spin_unlock(&journal->j_list_lock);
630 jbd_unlock_bh_state(bh);
631 /* commit wakes up all shadow buffers after IO */
632 wqh = bh_waitq_head(jh2bh(jh));
633 wait_event(*wqh, (jh->b_jlist != BJ_Shadow));
637 /* Only do the copy if the currently-owning transaction
638 * still needs it. If it is on the Forget list, the
639 * committing transaction is past that stage. The
640 * buffer had better remain locked during the kmalloc,
641 * but that should be true --- we hold the journal lock
642 * still and the buffer is already on the BUF_JOURNAL
643 * list so won't be flushed.
645 * Subtle point, though: if this is a get_undo_access,
646 * then we will be relying on the frozen_data to contain
647 * the new value of the committed_data record after the
648 * transaction, so we HAVE to force the frozen_data copy
651 if (jh->b_jlist != BJ_Forget || force_copy) {
652 JBUFFER_TRACE(jh, "generate frozen data");
653 if (!frozen_buffer) {
654 JBUFFER_TRACE(jh, "allocate memory for buffer");
655 spin_unlock(&journal->j_list_lock);
656 jbd_unlock_bh_state(bh);
657 frozen_buffer = jbd_kmalloc(jh2bh(jh)->b_size,
659 if (!frozen_buffer) {
661 "%s: OOM for frozen_buffer\n",
663 JBUFFER_TRACE(jh, "oom!");
665 jbd_lock_bh_state(bh);
666 spin_lock(&journal->j_list_lock);
672 jh->b_frozen_data = frozen_buffer;
673 frozen_buffer = NULL;
676 jh->b_next_transaction = transaction;
679 J_ASSERT(handle->h_buffer_credits > 0);
680 handle->h_buffer_credits--;
684 /* Finally, if the buffer is not journaled right now, we need to
685 * make sure it doesn't get written to disk before the caller
686 * actually commits the new data. */
688 if (!jh->b_transaction) {
689 JBUFFER_TRACE(jh, "no transaction");
690 J_ASSERT_JH(jh, !jh->b_next_transaction);
691 jh->b_transaction = transaction;
692 JBUFFER_TRACE(jh, "file as BJ_Reserved");
693 __journal_file_buffer(jh, transaction, BJ_Reserved);
697 spin_unlock(&journal->j_list_lock);
703 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
704 "Possible IO failure.\n");
705 page = jh2bh(jh)->b_page;
706 offset = ((unsigned long) jh2bh(jh)->b_data) & ~PAGE_MASK;
707 source = kmap_atomic(page, KM_USER0);
708 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
709 kunmap_atomic(source, KM_USER0);
711 jbd_unlock_bh_state(bh);
713 /* If we are about to journal a buffer, then any revoke pending
714 on it is no longer valid. */
715 journal_cancel_revoke(handle, jh);
719 kfree(frozen_buffer);
721 JBUFFER_TRACE(jh, "exit");
726 * int journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
727 * @handle: transaction to add buffer modifications to
728 * @bh: bh to be used for metadata writes
730 * Returns an error code or 0 on success.
732 * In full data journalling mode the buffer may be of type BJ_AsyncData,
733 * because we're write()ing a buffer which is also part of a shared mapping.
736 int journal_get_write_access(handle_t *handle,
737 struct buffer_head *bh, int *credits)
739 struct journal_head *jh = journal_add_journal_head(bh);
742 /* We do not want to get caught playing with fields which the
743 * log thread also manipulates. Make sure that the buffer
744 * completes any outstanding IO before proceeding. */
745 rc = do_get_write_access(handle, jh, 0, credits);
746 journal_put_journal_head(jh);
752 * When the user wants to journal a newly created buffer_head
753 * (ie. getblk() returned a new buffer and we are going to populate it
754 * manually rather than reading off disk), then we need to keep the
755 * buffer_head locked until it has been completely filled with new
756 * data. In this case, we should be able to make the assertion that
757 * the bh is not already part of an existing transaction.
759 * The buffer should already be locked by the caller by this point.
760 * There is no lock ranking violation: it was a newly created,
761 * unlocked buffer beforehand. */
764 * int journal_get_create_access () - notify intent to use newly created bh
765 * @handle: transaction to new buffer to
768 * Call this if you create a new bh.
770 int journal_get_create_access(handle_t *handle, struct buffer_head *bh)
772 transaction_t *transaction = handle->h_transaction;
773 journal_t *journal = transaction->t_journal;
774 struct journal_head *jh = journal_add_journal_head(bh);
777 jbd_debug(5, "journal_head %p\n", jh);
779 if (is_handle_aborted(handle))
783 JBUFFER_TRACE(jh, "entry");
785 * The buffer may already belong to this transaction due to pre-zeroing
786 * in the filesystem's new_block code. It may also be on the previous,
787 * committing transaction's lists, but it HAS to be in Forget state in
788 * that case: the transaction must have deleted the buffer for it to be
791 jbd_lock_bh_state(bh);
792 spin_lock(&journal->j_list_lock);
793 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
794 jh->b_transaction == NULL ||
795 (jh->b_transaction == journal->j_committing_transaction &&
796 jh->b_jlist == BJ_Forget)));
798 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
799 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
801 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
802 handle->h_buffer_credits--;
804 if (jh->b_transaction == NULL) {
805 jh->b_transaction = transaction;
806 JBUFFER_TRACE(jh, "file as BJ_Reserved");
807 __journal_file_buffer(jh, transaction, BJ_Reserved);
808 } else if (jh->b_transaction == journal->j_committing_transaction) {
809 JBUFFER_TRACE(jh, "set next transaction");
810 jh->b_next_transaction = transaction;
812 spin_unlock(&journal->j_list_lock);
813 jbd_unlock_bh_state(bh);
816 * akpm: I added this. ext3_alloc_branch can pick up new indirect
817 * blocks which contain freed but then revoked metadata. We need
818 * to cancel the revoke in case we end up freeing it yet again
819 * and the reallocating as data - this would cause a second revoke,
820 * which hits an assertion error.
822 JBUFFER_TRACE(jh, "cancelling revoke");
823 journal_cancel_revoke(handle, jh);
824 journal_put_journal_head(jh);
830 * int journal_get_undo_access() - Notify intent to modify metadata with
831 * non-rewindable consequences
832 * @handle: transaction
833 * @bh: buffer to undo
834 * @credits: store the number of taken credits here (if not NULL)
836 * Sometimes there is a need to distinguish between metadata which has
837 * been committed to disk and that which has not. The ext3fs code uses
838 * this for freeing and allocating space, we have to make sure that we
839 * do not reuse freed space until the deallocation has been committed,
840 * since if we overwrote that space we would make the delete
841 * un-rewindable in case of a crash.
843 * To deal with that, journal_get_undo_access requests write access to a
844 * buffer for parts of non-rewindable operations such as delete
845 * operations on the bitmaps. The journaling code must keep a copy of
846 * the buffer's contents prior to the undo_access call until such time
847 * as we know that the buffer has definitely been committed to disk.
849 * We never need to know which transaction the committed data is part
850 * of, buffers touched here are guaranteed to be dirtied later and so
851 * will be committed to a new transaction in due course, at which point
852 * we can discard the old committed data pointer.
854 * Returns error number or 0 on success.
856 int journal_get_undo_access(handle_t *handle, struct buffer_head *bh,
860 struct journal_head *jh = journal_add_journal_head(bh);
861 char *committed_data = NULL;
863 JBUFFER_TRACE(jh, "entry");
866 * Do this first --- it can drop the journal lock, so we want to
867 * make sure that obtaining the committed_data is done
868 * atomically wrt. completion of any outstanding commits.
870 err = do_get_write_access(handle, jh, 1, credits);
875 if (!jh->b_committed_data) {
876 committed_data = jbd_kmalloc(jh2bh(jh)->b_size, GFP_NOFS);
877 if (!committed_data) {
878 printk(KERN_EMERG "%s: No memory for committed data\n",
885 jbd_lock_bh_state(bh);
886 if (!jh->b_committed_data) {
887 /* Copy out the current buffer contents into the
888 * preserved, committed copy. */
889 JBUFFER_TRACE(jh, "generate b_committed data");
890 if (!committed_data) {
891 jbd_unlock_bh_state(bh);
895 jh->b_committed_data = committed_data;
896 committed_data = NULL;
897 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
899 jbd_unlock_bh_state(bh);
901 journal_put_journal_head(jh);
903 kfree(committed_data);
908 * int journal_dirty_data() - mark a buffer as containing dirty data which
909 * needs to be flushed before we can commit the
910 * current transaction.
911 * @handle: transaction
912 * @bh: bufferhead to mark
914 * The buffer is placed on the transaction's data list and is marked as
915 * belonging to the transaction.
917 * Returns error number or 0 on success.
919 * journal_dirty_data() can be called via page_launder->ext3_writepage
922 int journal_dirty_data(handle_t *handle, struct buffer_head *bh)
924 journal_t *journal = handle->h_transaction->t_journal;
926 struct journal_head *jh;
928 if (is_handle_aborted(handle))
931 jh = journal_add_journal_head(bh);
932 JBUFFER_TRACE(jh, "entry");
935 * The buffer could *already* be dirty. Writeout can start
938 jbd_debug(4, "jh: %p, tid:%d\n", jh, handle->h_transaction->t_tid);
941 * What if the buffer is already part of a running transaction?
943 * There are two cases:
944 * 1) It is part of the current running transaction. Refile it,
945 * just in case we have allocated it as metadata, deallocated
946 * it, then reallocated it as data.
947 * 2) It is part of the previous, still-committing transaction.
948 * If all we want to do is to guarantee that the buffer will be
949 * written to disk before this new transaction commits, then
950 * being sure that the *previous* transaction has this same
951 * property is sufficient for us! Just leave it on its old
954 * In case (2), the buffer must not already exist as metadata
955 * --- that would violate write ordering (a transaction is free
956 * to write its data at any point, even before the previous
957 * committing transaction has committed). The caller must
958 * never, ever allow this to happen: there's nothing we can do
959 * about it in this layer.
961 jbd_lock_bh_state(bh);
962 spin_lock(&journal->j_list_lock);
963 if (jh->b_transaction) {
964 JBUFFER_TRACE(jh, "has transaction");
965 if (jh->b_transaction != handle->h_transaction) {
966 JBUFFER_TRACE(jh, "belongs to older transaction");
967 J_ASSERT_JH(jh, jh->b_transaction ==
968 journal->j_committing_transaction);
970 /* @@@ IS THIS TRUE ? */
972 * Not any more. Scenario: someone does a write()
973 * in data=journal mode. The buffer's transaction has
974 * moved into commit. Then someone does another
975 * write() to the file. We do the frozen data copyout
976 * and set b_next_transaction to point to j_running_t.
977 * And while we're in that state, someone does a
978 * writepage() in an attempt to pageout the same area
979 * of the file via a shared mapping. At present that
980 * calls journal_dirty_data(), and we get right here.
981 * It may be too late to journal the data. Simply
982 * falling through to the next test will suffice: the
983 * data will be dirty and wil be checkpointed. The
984 * ordering comments in the next comment block still
987 //J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
990 * If we're journalling data, and this buffer was
991 * subject to a write(), it could be metadata, forget
992 * or shadow against the committing transaction. Now,
993 * someone has dirtied the same darn page via a mapping
994 * and it is being writepage()'d.
995 * We *could* just steal the page from commit, with some
996 * fancy locking there. Instead, we just skip it -
997 * don't tie the page's buffers to the new transaction
999 * Implication: if we crash before the writepage() data
1000 * is written into the filesystem, recovery will replay
1003 if (jh->b_jlist != BJ_None &&
1004 jh->b_jlist != BJ_SyncData) {
1005 JBUFFER_TRACE(jh, "Not stealing");
1010 * This buffer may be undergoing writeout in commit. We
1011 * can't return from here and let the caller dirty it
1012 * again because that can cause the write-out loop in
1013 * commit to never terminate.
1015 if (buffer_dirty(bh)) {
1017 spin_unlock(&journal->j_list_lock);
1018 jbd_unlock_bh_state(bh);
1020 sync_dirty_buffer(bh);
1021 jbd_lock_bh_state(bh);
1022 spin_lock(&journal->j_list_lock);
1023 /* The buffer may become locked again at any
1024 time if it is redirtied */
1027 /* journal_clean_data_list() may have got there first */
1028 if (jh->b_transaction != NULL) {
1029 JBUFFER_TRACE(jh, "unfile from commit");
1030 __journal_unfile_buffer(jh);
1031 jh->b_transaction = NULL;
1033 /* The buffer will be refiled below */
1037 * Special case --- the buffer might actually have been
1038 * allocated and then immediately deallocated in the previous,
1039 * committing transaction, so might still be left on that
1040 * transaction's metadata lists.
1042 if (jh->b_jlist != BJ_SyncData) {
1043 JBUFFER_TRACE(jh, "not on correct data list: unfile");
1044 J_ASSERT_JH(jh, jh->b_jlist != BJ_Shadow);
1045 __journal_unfile_buffer(jh);
1046 jh->b_transaction = NULL;
1047 JBUFFER_TRACE(jh, "file as data");
1048 __journal_file_buffer(jh, handle->h_transaction,
1052 JBUFFER_TRACE(jh, "not on a transaction");
1053 __journal_file_buffer(jh, handle->h_transaction, BJ_SyncData);
1056 spin_unlock(&journal->j_list_lock);
1057 jbd_unlock_bh_state(bh);
1059 BUFFER_TRACE(bh, "brelse");
1062 JBUFFER_TRACE(jh, "exit");
1063 journal_put_journal_head(jh);
1068 * int journal_dirty_metadata() - mark a buffer as containing dirty metadata
1069 * @handle: transaction to add buffer to.
1070 * @bh: buffer to mark
1072 * mark dirty metadata which needs to be journaled as part of the current
1075 * The buffer is placed on the transaction's metadata list and is marked
1076 * as belonging to the transaction.
1078 * Returns error number or 0 on success.
1080 * Special care needs to be taken if the buffer already belongs to the
1081 * current committing transaction (in which case we should have frozen
1082 * data present for that commit). In that case, we don't relink the
1083 * buffer: that only gets done when the old transaction finally
1084 * completes its commit.
1086 int journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1088 transaction_t *transaction = handle->h_transaction;
1089 journal_t *journal = transaction->t_journal;
1090 struct journal_head *jh = bh2jh(bh);
1091 int console_loglevel_saved = console_loglevel;
1093 jbd_debug(5, "journal_head %p\n", jh);
1094 JBUFFER_TRACE(jh, "entry");
1095 if (is_handle_aborted(handle))
1098 jbd_lock_bh_state(bh);
1101 * fastpath, to avoid expensive locking. If this buffer is already
1102 * on the running transaction's metadata list there is nothing to do.
1103 * Nobody can take it off again because there is a handle open.
1104 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1105 * result in this test being false, so we go in and take the locks.
1107 if (jh->b_transaction == handle->h_transaction &&
1108 jh->b_jlist == BJ_Metadata) {
1109 JBUFFER_TRACE(jh, "fastpath");
1110 J_ASSERT_JH(jh, jh->b_transaction ==
1111 journal->j_running_transaction);
1115 spin_lock(&journal->j_list_lock);
1116 set_buffer_jbddirty(bh);
1118 J_ASSERT_JH(jh, jh->b_transaction != NULL);
1121 * Metadata already on the current transaction list doesn't
1122 * need to be filed. Metadata on another transaction's list must
1123 * be committing, and will be refiled once the commit completes:
1124 * leave it alone for now.
1127 if (jh->b_transaction != transaction) {
1128 JBUFFER_TRACE(jh, "already on other transaction");
1129 J_ASSERT_JH(jh, jh->b_transaction ==
1130 journal->j_committing_transaction);
1131 J_ASSERT_JH(jh, jh->b_next_transaction == transaction);
1132 /* And this case is illegal: we can't reuse another
1133 * transaction's data buffer, ever. */
1134 /* FIXME: writepage() should be journalled */
1135 goto out_unlock_list;
1138 /* That test should have eliminated the following case: */
1139 J_ASSERT_JH(jh, jh->b_frozen_data == 0);
1141 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1142 __journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1145 spin_unlock(&journal->j_list_lock);
1147 jbd_unlock_bh_state(bh);
1149 JBUFFER_TRACE(jh, "exit");
1150 console_loglevel = console_loglevel_saved;
1155 * journal_release_buffer: undo a get_write_access without any buffer
1156 * updates, if the update decided in the end that it didn't need access.
1158 * The caller passes in the number of credits which should be put back for
1159 * this buffer (zero or one).
1161 * We leave the buffer attached to t_reserved_list because even though this
1162 * handle doesn't want it, some other concurrent handle may want to journal
1163 * this buffer. If that handle is curently in between get_write_access() and
1164 * journal_dirty_metadata() then it expects the buffer to be reserved. If
1165 * we were to rip it off t_reserved_list here, the other handle will explode
1166 * when journal_dirty_metadata is presented with a non-reserved buffer.
1168 * If nobody really wants to journal this buffer then it will be thrown
1169 * away at the start of commit.
1172 journal_release_buffer(handle_t *handle, struct buffer_head *bh, int credits)
1174 BUFFER_TRACE(bh, "entry");
1175 handle->h_buffer_credits += credits;
1179 * void journal_forget() - bforget() for potentially-journaled buffers.
1180 * @handle: transaction handle
1181 * @bh: bh to 'forget'
1183 * We can only do the bforget if there are no commits pending against the
1184 * buffer. If the buffer is dirty in the current running transaction we
1185 * can safely unlink it.
1187 * bh may not be a journalled buffer at all - it may be a non-JBD
1188 * buffer which came off the hashtable. Check for this.
1190 * Decrements bh->b_count by one.
1192 * Allow this call even if the handle has aborted --- it may be part of
1193 * the caller's cleanup after an abort.
1195 void journal_forget(handle_t *handle, struct buffer_head *bh)
1197 transaction_t *transaction = handle->h_transaction;
1198 journal_t *journal = transaction->t_journal;
1199 struct journal_head *jh;
1201 BUFFER_TRACE(bh, "entry");
1203 jbd_lock_bh_state(bh);
1204 spin_lock(&journal->j_list_lock);
1206 if (!buffer_jbd(bh))
1210 if (jh->b_transaction == handle->h_transaction) {
1211 J_ASSERT_JH(jh, !jh->b_frozen_data);
1213 /* If we are forgetting a buffer which is already part
1214 * of this transaction, then we can just drop it from
1215 * the transaction immediately. */
1216 clear_buffer_dirty(bh);
1217 clear_buffer_jbddirty(bh);
1219 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1220 J_ASSERT_JH(jh, !jh->b_committed_data);
1222 __journal_unfile_buffer(jh);
1223 jh->b_transaction = 0;
1226 * We are no longer going to journal this buffer.
1227 * However, the commit of this transaction is still
1228 * important to the buffer: the delete that we are now
1229 * processing might obsolete an old log entry, so by
1230 * committing, we can satisfy the buffer's checkpoint.
1232 * So, if we have a checkpoint on the buffer, we should
1233 * now refile the buffer on our BJ_Forget list so that
1234 * we know to remove the checkpoint after we commit.
1237 if (jh->b_cp_transaction) {
1238 __journal_file_buffer(jh, transaction, BJ_Forget);
1240 journal_remove_journal_head(bh);
1242 if (!buffer_jbd(bh)) {
1243 spin_unlock(&journal->j_list_lock);
1244 jbd_unlock_bh_state(bh);
1249 } else if (jh->b_transaction) {
1250 J_ASSERT_JH(jh, (jh->b_transaction ==
1251 journal->j_committing_transaction));
1252 /* However, if the buffer is still owned by a prior
1253 * (committing) transaction, we can't drop it yet... */
1254 JBUFFER_TRACE(jh, "belongs to older transaction");
1255 /* ... but we CAN drop it from the new transaction if we
1256 * have also modified it since the original commit. */
1258 if (jh->b_next_transaction) {
1259 J_ASSERT(jh->b_next_transaction == transaction);
1260 jh->b_next_transaction = NULL;
1265 spin_unlock(&journal->j_list_lock);
1266 jbd_unlock_bh_state(bh);
1272 * void journal_callback_set() - Register a callback function for this handle.
1273 * @handle: handle to attach the callback to.
1274 * @func: function to callback.
1275 * @jcb: structure with additional information required by func() , and
1276 * some space for jbd internal information.
1278 * The function will be
1279 * called when the transaction that this handle is part of has been
1280 * committed to disk with the original callback data struct and the
1281 * error status of the journal as parameters. There is no guarantee of
1282 * ordering between handles within a single transaction, nor between
1283 * callbacks registered on the same handle.
1285 * The caller is responsible for allocating the journal_callback struct.
1286 * This is to allow the caller to add as much extra data to the callback
1287 * as needed, but reduce the overhead of multiple allocations. The caller
1288 * allocated struct must start with a struct journal_callback at offset 0,
1289 * and has the caller-specific data afterwards.
1291 void journal_callback_set(handle_t *handle,
1292 void (*func)(struct journal_callback *jcb, int error),
1293 struct journal_callback *jcb)
1295 spin_lock(&handle->h_transaction->t_jcb_lock);
1296 list_add_tail(&jcb->jcb_list, &handle->h_jcb);
1297 spin_unlock(&handle->h_transaction->t_jcb_lock);
1298 jcb->jcb_func = func;
1302 * int journal_stop() - complete a transaction
1303 * @handle: tranaction to complete.
1305 * All done for a particular handle.
1307 * There is not much action needed here. We just return any remaining
1308 * buffer credits to the transaction and remove the handle. The only
1309 * complication is that we need to start a commit operation if the
1310 * filesystem is marked for synchronous update.
1312 * journal_stop itself will not usually return an error, but it may
1313 * do so in unusual circumstances. In particular, expect it to
1314 * return -EIO if a journal_abort has been executed since the
1315 * transaction began.
1317 int journal_stop(handle_t *handle)
1319 transaction_t *transaction = handle->h_transaction;
1320 journal_t *journal = transaction->t_journal;
1321 int old_handle_count, err;
1323 J_ASSERT(transaction->t_updates > 0);
1324 J_ASSERT(journal_current_handle() == handle);
1326 if (is_handle_aborted(handle))
1331 if (--handle->h_ref > 0) {
1332 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1337 jbd_debug(4, "Handle %p going down\n", handle);
1340 * Implement synchronous transaction batching. If the handle
1341 * was synchronous, don't force a commit immediately. Let's
1342 * yield and let another thread piggyback onto this transaction.
1343 * Keep doing that while new threads continue to arrive.
1344 * It doesn't cost much - we're about to run a commit and sleep
1345 * on IO anyway. Speeds up many-threaded, many-dir operations
1348 if (handle->h_sync) {
1349 set_current_state(TASK_RUNNING);
1351 old_handle_count = transaction->t_handle_count;
1353 } while (old_handle_count != transaction->t_handle_count);
1356 current->journal_info = NULL;
1357 spin_lock(&journal->j_state_lock);
1358 spin_lock(&transaction->t_handle_lock);
1359 transaction->t_outstanding_credits -= handle->h_buffer_credits;
1360 transaction->t_updates--;
1361 if (!transaction->t_updates) {
1362 wake_up(&journal->j_wait_updates);
1363 if (journal->j_barrier_count)
1364 wake_up(&journal->j_wait_transaction_locked);
1367 /* Move callbacks from the handle to the transaction. */
1368 spin_lock(&transaction->t_jcb_lock);
1369 list_splice(&handle->h_jcb, &transaction->t_jcb);
1370 spin_unlock(&transaction->t_jcb_lock);
1373 * If the handle is marked SYNC, we need to set another commit
1374 * going! We also want to force a commit if the current
1375 * transaction is occupying too much of the log, or if the
1376 * transaction is too old now.
1378 if (handle->h_sync ||
1379 transaction->t_outstanding_credits >
1380 journal->j_max_transaction_buffers ||
1381 time_after_eq(jiffies, transaction->t_expires)) {
1382 /* Do this even for aborted journals: an abort still
1383 * completes the commit thread, it just doesn't write
1384 * anything to disk. */
1385 tid_t tid = transaction->t_tid;
1387 spin_unlock(&transaction->t_handle_lock);
1388 jbd_debug(2, "transaction too old, requesting commit for "
1389 "handle %p\n", handle);
1390 /* This is non-blocking */
1391 __log_start_commit(journal, transaction->t_tid);
1392 spin_unlock(&journal->j_state_lock);
1395 * Special case: JFS_SYNC synchronous updates require us
1396 * to wait for the commit to complete.
1398 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1399 err = log_wait_commit(journal, tid);
1401 spin_unlock(&transaction->t_handle_lock);
1402 spin_unlock(&journal->j_state_lock);
1405 jbd_free_handle(handle);
1409 /**int journal_force_commit() - force any uncommitted transactions
1410 * @journal: journal to force
1412 * For synchronous operations: force any uncommitted transactions
1413 * to disk. May seem kludgy, but it reuses all the handle batching
1414 * code in a very simple manner.
1416 int journal_force_commit(journal_t *journal)
1421 handle = journal_start(journal, 1);
1422 if (IS_ERR(handle)) {
1423 ret = PTR_ERR(handle);
1426 ret = journal_stop(handle);
1433 * List management code snippets: various functions for manipulating the
1434 * transaction buffer lists.
1439 * Append a buffer to a transaction list, given the transaction's list head
1442 * j_list_lock is held.
1444 * jbd_lock_bh_state(jh2bh(jh)) is held.
1448 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1451 jh->b_tnext = jh->b_tprev = jh;
1454 /* Insert at the tail of the list to preserve order */
1455 struct journal_head *first = *list, *last = first->b_tprev;
1457 jh->b_tnext = first;
1458 last->b_tnext = first->b_tprev = jh;
1463 * Remove a buffer from a transaction list, given the transaction's list
1466 * Called with j_list_lock held, and the journal may not be locked.
1468 * jbd_lock_bh_state(jh2bh(jh)) is held.
1472 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1475 *list = jh->b_tnext;
1479 jh->b_tprev->b_tnext = jh->b_tnext;
1480 jh->b_tnext->b_tprev = jh->b_tprev;
1484 * Remove a buffer from the appropriate transaction list.
1486 * Note that this function can *change* the value of
1487 * bh->b_transaction->t_sync_datalist, t_buffers, t_forget,
1488 * t_iobuf_list, t_shadow_list, t_log_list or t_reserved_list. If the caller
1489 * is holding onto a copy of one of thee pointers, it could go bad.
1490 * Generally the caller needs to re-read the pointer from the transaction_t.
1492 * Called under j_list_lock. The journal may not be locked.
1494 void __journal_unfile_buffer(struct journal_head *jh)
1496 struct journal_head **list = 0;
1497 transaction_t *transaction;
1498 struct buffer_head *bh = jh2bh(jh);
1500 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1501 transaction = jh->b_transaction;
1503 assert_spin_locked(&transaction->t_journal->j_list_lock);
1505 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1506 if (jh->b_jlist != BJ_None)
1507 J_ASSERT_JH(jh, transaction != 0);
1509 switch (jh->b_jlist) {
1513 list = &transaction->t_sync_datalist;
1516 transaction->t_nr_buffers--;
1517 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1518 list = &transaction->t_buffers;
1521 list = &transaction->t_forget;
1524 list = &transaction->t_iobuf_list;
1527 list = &transaction->t_shadow_list;
1530 list = &transaction->t_log_list;
1533 list = &transaction->t_reserved_list;
1537 __blist_del_buffer(list, jh);
1538 jh->b_jlist = BJ_None;
1539 if (test_clear_buffer_jbddirty(bh))
1540 mark_buffer_dirty(bh); /* Expose it to the VM */
1543 void journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1545 jbd_lock_bh_state(jh2bh(jh));
1546 spin_lock(&journal->j_list_lock);
1547 __journal_unfile_buffer(jh);
1548 spin_unlock(&journal->j_list_lock);
1549 jbd_unlock_bh_state(jh2bh(jh));
1553 * Called from journal_try_to_free_buffers().
1555 * Called under jbd_lock_bh_state(bh)
1558 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1560 struct journal_head *jh;
1564 if (buffer_locked(bh) || buffer_dirty(bh))
1567 if (jh->b_next_transaction != 0)
1570 spin_lock(&journal->j_list_lock);
1571 if (jh->b_transaction != 0 && jh->b_cp_transaction == 0) {
1572 if (jh->b_jlist == BJ_SyncData) {
1573 /* A written-back ordered data buffer */
1574 JBUFFER_TRACE(jh, "release data");
1575 __journal_unfile_buffer(jh);
1576 jh->b_transaction = 0;
1577 journal_remove_journal_head(bh);
1580 } else if (jh->b_cp_transaction != 0 && jh->b_transaction == 0) {
1581 /* written-back checkpointed metadata buffer */
1582 if (jh->b_jlist == BJ_None) {
1583 JBUFFER_TRACE(jh, "remove from checkpoint list");
1584 __journal_remove_checkpoint(jh);
1585 journal_remove_journal_head(bh);
1589 spin_unlock(&journal->j_list_lock);
1596 * int journal_try_to_free_buffers() - try to free page buffers.
1597 * @journal: journal for operation
1598 * @page: to try and free
1599 * @gfp_mask: 'IO' mode for try_to_free_buffers()
1602 * For all the buffers on this page,
1603 * if they are fully written out ordered data, move them onto BUF_CLEAN
1604 * so try_to_free_buffers() can reap them.
1606 * This function returns non-zero if we wish try_to_free_buffers()
1607 * to be called. We do this if the page is releasable by try_to_free_buffers().
1608 * We also do it if the page has locked or dirty buffers and the caller wants
1609 * us to perform sync or async writeout.
1611 * This complicates JBD locking somewhat. We aren't protected by the
1612 * BKL here. We wish to remove the buffer from its committing or
1613 * running transaction's ->t_datalist via __journal_unfile_buffer.
1615 * This may *change* the value of transaction_t->t_datalist, so anyone
1616 * who looks at t_datalist needs to lock against this function.
1618 * Even worse, someone may be doing a journal_dirty_data on this
1619 * buffer. So we need to lock against that. journal_dirty_data()
1620 * will come out of the lock with the buffer dirty, which makes it
1621 * ineligible for release here.
1623 * Who else is affected by this? hmm... Really the only contender
1624 * is do_get_write_access() - it could be looking at the buffer while
1625 * journal_try_to_free_buffer() is changing its state. But that
1626 * cannot happen because we never reallocate freed data as metadata
1627 * while the data is part of a transaction. Yes?
1629 int journal_try_to_free_buffers(journal_t *journal,
1630 struct page *page, int unused_gfp_mask)
1632 struct buffer_head *head;
1633 struct buffer_head *bh;
1636 J_ASSERT(PageLocked(page));
1638 head = page_buffers(page);
1641 struct journal_head *jh;
1644 * We take our own ref against the journal_head here to avoid
1645 * having to add tons of locking around each instance of
1646 * journal_remove_journal_head() and journal_put_journal_head().
1648 jh = journal_grab_journal_head(bh);
1652 jbd_lock_bh_state(bh);
1653 __journal_try_to_free_buffer(journal, bh);
1654 journal_put_journal_head(jh);
1655 jbd_unlock_bh_state(bh);
1658 } while ((bh = bh->b_this_page) != head);
1659 ret = try_to_free_buffers(page);
1665 * This buffer is no longer needed. If it is on an older transaction's
1666 * checkpoint list we need to record it on this transaction's forget list
1667 * to pin this buffer (and hence its checkpointing transaction) down until
1668 * this transaction commits. If the buffer isn't on a checkpoint list, we
1670 * Returns non-zero if JBD no longer has an interest in the buffer.
1672 * Called under j_list_lock.
1674 * Called under jbd_lock_bh_state(bh).
1676 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1679 struct buffer_head *bh = jh2bh(jh);
1681 __journal_unfile_buffer(jh);
1682 jh->b_transaction = 0;
1684 if (jh->b_cp_transaction) {
1685 JBUFFER_TRACE(jh, "on running+cp transaction");
1686 __journal_file_buffer(jh, transaction, BJ_Forget);
1687 clear_buffer_jbddirty(bh);
1690 JBUFFER_TRACE(jh, "on running transaction");
1691 journal_remove_journal_head(bh);
1698 * journal_invalidatepage
1700 * This code is tricky. It has a number of cases to deal with.
1702 * There are two invariants which this code relies on:
1704 * i_size must be updated on disk before we start calling invalidatepage on the
1707 * This is done in ext3 by defining an ext3_setattr method which
1708 * updates i_size before truncate gets going. By maintaining this
1709 * invariant, we can be sure that it is safe to throw away any buffers
1710 * attached to the current transaction: once the transaction commits,
1711 * we know that the data will not be needed.
1713 * Note however that we can *not* throw away data belonging to the
1714 * previous, committing transaction!
1716 * Any disk blocks which *are* part of the previous, committing
1717 * transaction (and which therefore cannot be discarded immediately) are
1718 * not going to be reused in the new running transaction
1720 * The bitmap committed_data images guarantee this: any block which is
1721 * allocated in one transaction and removed in the next will be marked
1722 * as in-use in the committed_data bitmap, so cannot be reused until
1723 * the next transaction to delete the block commits. This means that
1724 * leaving committing buffers dirty is quite safe: the disk blocks
1725 * cannot be reallocated to a different file and so buffer aliasing is
1729 * The above applies mainly to ordered data mode. In writeback mode we
1730 * don't make guarantees about the order in which data hits disk --- in
1731 * particular we don't guarantee that new dirty data is flushed before
1732 * transaction commit --- so it is always safe just to discard data
1733 * immediately in that mode. --sct
1737 * The journal_unmap_buffer helper function returns zero if the buffer
1738 * concerned remains pinned as an anonymous buffer belonging to an older
1741 * We're outside-transaction here. Either or both of j_running_transaction
1742 * and j_committing_transaction may be NULL.
1744 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1746 transaction_t *transaction;
1747 struct journal_head *jh;
1751 BUFFER_TRACE(bh, "entry");
1754 * It is safe to proceed here without the j_list_lock because the
1755 * buffers cannot be stolen by try_to_free_buffers as long as we are
1756 * holding the page lock. --sct
1759 if (!buffer_jbd(bh))
1760 goto zap_buffer_unlocked;
1762 spin_lock(&journal->j_state_lock);
1763 jbd_lock_bh_state(bh);
1764 spin_lock(&journal->j_list_lock);
1767 * Now we have the locks, check again to see whether kjournald has
1768 * taken the buffer off the transaction.
1770 if (!buffer_jbd(bh))
1774 transaction = jh->b_transaction;
1775 if (transaction == NULL) {
1776 /* First case: not on any transaction. If it
1777 * has no checkpoint link, then we can zap it:
1778 * it's a writeback-mode buffer so we don't care
1779 * if it hits disk safely. */
1780 if (!jh->b_cp_transaction) {
1781 JBUFFER_TRACE(jh, "not on any transaction: zap");
1785 if (!buffer_dirty(bh)) {
1786 /* bdflush has written it. We can drop it now */
1790 /* OK, it must be in the journal but still not
1791 * written fully to disk: it's metadata or
1792 * journaled data... */
1794 if (journal->j_running_transaction) {
1795 /* ... and once the current transaction has
1796 * committed, the buffer won't be needed any
1798 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1799 ret = __dispose_buffer(jh,
1800 journal->j_running_transaction);
1801 spin_unlock(&journal->j_list_lock);
1802 jbd_unlock_bh_state(bh);
1803 spin_unlock(&journal->j_state_lock);
1806 /* There is no currently-running transaction. So the
1807 * orphan record which we wrote for this file must have
1808 * passed into commit. We must attach this buffer to
1809 * the committing transaction, if it exists. */
1810 if (journal->j_committing_transaction) {
1811 JBUFFER_TRACE(jh, "give to committing trans");
1812 ret = __dispose_buffer(jh,
1813 journal->j_committing_transaction);
1814 spin_unlock(&journal->j_list_lock);
1815 jbd_unlock_bh_state(bh);
1816 spin_unlock(&journal->j_state_lock);
1819 /* The orphan record's transaction has
1820 * committed. We can cleanse this buffer */
1821 clear_buffer_jbddirty(bh);
1825 } else if (transaction == journal->j_committing_transaction) {
1826 /* If it is committing, we simply cannot touch it. We
1827 * can remove it's next_transaction pointer from the
1828 * running transaction if that is set, but nothing
1830 JBUFFER_TRACE(jh, "on committing transaction");
1831 set_buffer_freed(bh);
1832 if (jh->b_next_transaction) {
1833 J_ASSERT(jh->b_next_transaction ==
1834 journal->j_running_transaction);
1835 jh->b_next_transaction = NULL;
1837 spin_unlock(&journal->j_list_lock);
1838 jbd_unlock_bh_state(bh);
1839 spin_unlock(&journal->j_state_lock);
1842 /* Good, the buffer belongs to the running transaction.
1843 * We are writing our own transaction's data, not any
1844 * previous one's, so it is safe to throw it away
1845 * (remember that we expect the filesystem to have set
1846 * i_size already for this truncate so recovery will not
1847 * expose the disk blocks we are discarding here.) */
1848 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1849 may_free = __dispose_buffer(jh, transaction);
1853 spin_unlock(&journal->j_list_lock);
1854 jbd_unlock_bh_state(bh);
1855 spin_unlock(&journal->j_state_lock);
1856 zap_buffer_unlocked:
1857 clear_buffer_dirty(bh);
1858 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1859 clear_buffer_mapped(bh);
1860 clear_buffer_req(bh);
1861 clear_buffer_new(bh);
1867 * int journal_invalidatepage()
1868 * @journal: journal to use for flush...
1869 * @page: page to flush
1870 * @offset: length of page to invalidate.
1872 * Reap page buffers containing data after offset in page.
1874 * Return non-zero if the page's buffers were successfully reaped.
1876 int journal_invalidatepage(journal_t *journal,
1878 unsigned long offset)
1880 struct buffer_head *head, *bh, *next;
1881 unsigned int curr_off = 0;
1884 if (!PageLocked(page))
1886 if (!page_has_buffers(page))
1889 /* We will potentially be playing with lists other than just the
1890 * data lists (especially for journaled data mode), so be
1891 * cautious in our locking. */
1893 head = bh = page_buffers(page);
1895 unsigned int next_off = curr_off + bh->b_size;
1896 next = bh->b_this_page;
1898 /* AKPM: doing lock_buffer here may be overly paranoid */
1899 if (offset <= curr_off) {
1900 /* This block is wholly outside the truncation point */
1902 may_free &= journal_unmap_buffer(journal, bh);
1905 curr_off = next_off;
1908 } while (bh != head);
1911 if (!may_free || !try_to_free_buffers(page))
1913 J_ASSERT(!page_has_buffers(page));
1919 * File a buffer on the given transaction list.
1921 void __journal_file_buffer(struct journal_head *jh,
1922 transaction_t *transaction, int jlist)
1924 struct journal_head **list = 0;
1926 struct buffer_head *bh = jh2bh(jh);
1928 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1929 assert_spin_locked(&transaction->t_journal->j_list_lock);
1932 J_ASSERT (current->lock_depth >= 0);
1934 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1935 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1936 jh->b_transaction == 0);
1938 if (jh->b_transaction && jh->b_jlist == jlist)
1941 /* The following list of buffer states needs to be consistent
1942 * with __jbd_unexpected_dirty_buffer()'s handling of dirty
1945 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
1946 jlist == BJ_Shadow || jlist == BJ_Forget) {
1947 if (test_clear_buffer_dirty(bh) ||
1948 test_clear_buffer_jbddirty(bh))
1952 if (jh->b_transaction)
1953 __journal_unfile_buffer(jh);
1955 jh->b_transaction = transaction;
1959 J_ASSERT_JH(jh, !jh->b_committed_data);
1960 J_ASSERT_JH(jh, !jh->b_frozen_data);
1963 list = &transaction->t_sync_datalist;
1966 transaction->t_nr_buffers++;
1967 list = &transaction->t_buffers;
1970 list = &transaction->t_forget;
1973 list = &transaction->t_iobuf_list;
1976 list = &transaction->t_shadow_list;
1979 list = &transaction->t_log_list;
1982 list = &transaction->t_reserved_list;
1986 __blist_add_buffer(list, jh);
1987 jh->b_jlist = jlist;
1990 set_buffer_jbddirty(bh);
1993 void journal_file_buffer(struct journal_head *jh,
1994 transaction_t *transaction, int jlist)
1996 jbd_lock_bh_state(jh2bh(jh));
1997 spin_lock(&transaction->t_journal->j_list_lock);
1998 __journal_file_buffer(jh, transaction, jlist);
1999 spin_unlock(&transaction->t_journal->j_list_lock);
2000 jbd_unlock_bh_state(jh2bh(jh));
2004 * Remove a buffer from its current buffer list in preparation for
2005 * dropping it from its current transaction entirely. If the buffer has
2006 * already started to be used by a subsequent transaction, refile the
2007 * buffer on that transaction's metadata list.
2009 * Called under journal->j_list_lock
2011 * Called under jbd_lock_bh_state(jh2bh(jh))
2013 void __journal_refile_buffer(struct journal_head *jh)
2016 struct buffer_head *bh = jh2bh(jh);
2018 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2019 if (jh->b_transaction)
2020 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2022 /* If the buffer is now unused, just drop it. */
2023 if (jh->b_next_transaction == NULL) {
2024 __journal_unfile_buffer(jh);
2025 jh->b_transaction = NULL;
2030 * It has been modified by a later transaction: add it to the new
2031 * transaction's metadata list.
2034 was_dirty = test_clear_buffer_jbddirty(bh);
2035 __journal_unfile_buffer(jh);
2036 jh->b_transaction = jh->b_next_transaction;
2037 jh->b_next_transaction = NULL;
2038 __journal_file_buffer(jh, jh->b_transaction, BJ_Metadata);
2039 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2042 set_buffer_jbddirty(bh);
2046 * For the unlocked version of this call, also make sure that any
2047 * hanging journal_head is cleaned up if necessary.
2049 * __journal_refile_buffer is usually called as part of a single locked
2050 * operation on a buffer_head, in which the caller is probably going to
2051 * be hooking the journal_head onto other lists. In that case it is up
2052 * to the caller to remove the journal_head if necessary. For the
2053 * unlocked journal_refile_buffer call, the caller isn't going to be
2054 * doing anything else to the buffer so we need to do the cleanup
2055 * ourselves to avoid a jh leak.
2057 * *** The journal_head may be freed by this call! ***
2059 void journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2061 struct buffer_head *bh = jh2bh(jh);
2063 jbd_lock_bh_state(bh);
2064 spin_lock(&journal->j_list_lock);
2066 __journal_refile_buffer(jh);
2067 jbd_unlock_bh_state(bh);
2068 journal_remove_journal_head(bh);
2070 spin_unlock(&journal->j_list_lock);