2 * linux/fs/jbd2/transaction.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6 * Copyright 1998 Red Hat corp --- All Rights Reserved
8 * This file is part of the Linux kernel and is made available under
9 * the terms of the GNU General Public License, version 2, or at your
10 * option, any later version, incorporated herein by reference.
12 * Generic filesystem transaction handling code; part of the ext2fs
15 * This file manages transactions (compound commits managed by the
16 * journaling code) and handles (individual atomic operations by the
20 #include <linux/time.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29 #include <linux/backing-dev.h>
30 #include <linux/bug.h>
31 #include <linux/module.h>
33 #include <trace/events/jbd2.h>
35 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
36 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
38 static struct kmem_cache *transaction_cache;
39 int __init jbd2_journal_init_transaction_cache(void)
41 J_ASSERT(!transaction_cache);
42 transaction_cache = kmem_cache_create("jbd2_transaction_s",
43 sizeof(transaction_t),
45 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
47 if (transaction_cache)
52 void jbd2_journal_destroy_transaction_cache(void)
54 if (transaction_cache) {
55 kmem_cache_destroy(transaction_cache);
56 transaction_cache = NULL;
60 void jbd2_journal_free_transaction(transaction_t *transaction)
62 if (unlikely(ZERO_OR_NULL_PTR(transaction)))
64 kmem_cache_free(transaction_cache, transaction);
68 * jbd2_get_transaction: obtain a new transaction_t object.
70 * Simply allocate and initialise a new transaction. Create it in
71 * RUNNING state and add it to the current journal (which should not
72 * have an existing running transaction: we only make a new transaction
73 * once we have started to commit the old one).
76 * The journal MUST be locked. We don't perform atomic mallocs on the
77 * new transaction and we can't block without protecting against other
78 * processes trying to touch the journal while it is in transition.
82 static transaction_t *
83 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
85 transaction->t_journal = journal;
86 transaction->t_state = T_RUNNING;
87 transaction->t_start_time = ktime_get();
88 transaction->t_tid = journal->j_transaction_sequence++;
89 transaction->t_expires = jiffies + journal->j_commit_interval;
90 spin_lock_init(&transaction->t_handle_lock);
91 atomic_set(&transaction->t_updates, 0);
92 atomic_set(&transaction->t_outstanding_credits, 0);
93 atomic_set(&transaction->t_handle_count, 0);
94 INIT_LIST_HEAD(&transaction->t_inode_list);
95 INIT_LIST_HEAD(&transaction->t_private_list);
97 /* Set up the commit timer for the new transaction. */
98 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
99 add_timer(&journal->j_commit_timer);
101 J_ASSERT(journal->j_running_transaction == NULL);
102 journal->j_running_transaction = transaction;
103 transaction->t_max_wait = 0;
104 transaction->t_start = jiffies;
105 transaction->t_requested = 0;
113 * A handle_t is an object which represents a single atomic update to a
114 * filesystem, and which tracks all of the modifications which form part
115 * of that one update.
119 * Update transaction's maximum wait time, if debugging is enabled.
121 * In order for t_max_wait to be reliable, it must be protected by a
122 * lock. But doing so will mean that start_this_handle() can not be
123 * run in parallel on SMP systems, which limits our scalability. So
124 * unless debugging is enabled, we no longer update t_max_wait, which
125 * means that maximum wait time reported by the jbd2_run_stats
126 * tracepoint will always be zero.
128 static inline void update_t_max_wait(transaction_t *transaction,
131 #ifdef CONFIG_JBD2_DEBUG
132 if (jbd2_journal_enable_debug &&
133 time_after(transaction->t_start, ts)) {
134 ts = jbd2_time_diff(ts, transaction->t_start);
135 spin_lock(&transaction->t_handle_lock);
136 if (ts > transaction->t_max_wait)
137 transaction->t_max_wait = ts;
138 spin_unlock(&transaction->t_handle_lock);
144 * start_this_handle: Given a handle, deal with any locking or stalling
145 * needed to make sure that there is enough journal space for the handle
146 * to begin. Attach the handle to a transaction and set up the
147 * transaction's buffer credits.
150 static int start_this_handle(journal_t *journal, handle_t *handle,
153 transaction_t *transaction, *new_transaction = NULL;
155 int needed, need_to_start;
156 int nblocks = handle->h_buffer_credits;
157 unsigned long ts = jiffies;
159 if (nblocks > journal->j_max_transaction_buffers) {
160 printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
161 current->comm, nblocks,
162 journal->j_max_transaction_buffers);
167 if (!journal->j_running_transaction) {
168 new_transaction = kmem_cache_zalloc(transaction_cache,
170 if (!new_transaction) {
172 * If __GFP_FS is not present, then we may be
173 * being called from inside the fs writeback
174 * layer, so we MUST NOT fail. Since
175 * __GFP_NOFAIL is going away, we will arrange
176 * to retry the allocation ourselves.
178 if ((gfp_mask & __GFP_FS) == 0) {
179 congestion_wait(BLK_RW_ASYNC, HZ/50);
180 goto alloc_transaction;
186 jbd_debug(3, "New handle %p going live.\n", handle);
189 * We need to hold j_state_lock until t_updates has been incremented,
190 * for proper journal barrier handling
193 read_lock(&journal->j_state_lock);
194 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
195 if (is_journal_aborted(journal) ||
196 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
197 read_unlock(&journal->j_state_lock);
198 jbd2_journal_free_transaction(new_transaction);
202 /* Wait on the journal's transaction barrier if necessary */
203 if (journal->j_barrier_count) {
204 read_unlock(&journal->j_state_lock);
205 wait_event(journal->j_wait_transaction_locked,
206 journal->j_barrier_count == 0);
210 if (!journal->j_running_transaction) {
211 read_unlock(&journal->j_state_lock);
212 if (!new_transaction)
213 goto alloc_transaction;
214 write_lock(&journal->j_state_lock);
215 if (!journal->j_running_transaction &&
216 !journal->j_barrier_count) {
217 jbd2_get_transaction(journal, new_transaction);
218 new_transaction = NULL;
220 write_unlock(&journal->j_state_lock);
224 transaction = journal->j_running_transaction;
227 * If the current transaction is locked down for commit, wait for the
228 * lock to be released.
230 if (transaction->t_state == T_LOCKED) {
233 prepare_to_wait(&journal->j_wait_transaction_locked,
234 &wait, TASK_UNINTERRUPTIBLE);
235 read_unlock(&journal->j_state_lock);
237 finish_wait(&journal->j_wait_transaction_locked, &wait);
242 * If there is not enough space left in the log to write all potential
243 * buffers requested by this operation, we need to stall pending a log
244 * checkpoint to free some more log space.
246 needed = atomic_add_return(nblocks,
247 &transaction->t_outstanding_credits);
249 if (needed > journal->j_max_transaction_buffers) {
251 * If the current transaction is already too large, then start
252 * to commit it: we can then go back and attach this handle to
257 jbd_debug(2, "Handle %p starting new commit...\n", handle);
258 atomic_sub(nblocks, &transaction->t_outstanding_credits);
259 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
260 TASK_UNINTERRUPTIBLE);
261 tid = transaction->t_tid;
262 need_to_start = !tid_geq(journal->j_commit_request, tid);
263 read_unlock(&journal->j_state_lock);
265 jbd2_log_start_commit(journal, tid);
267 finish_wait(&journal->j_wait_transaction_locked, &wait);
272 * The commit code assumes that it can get enough log space
273 * without forcing a checkpoint. This is *critical* for
274 * correctness: a checkpoint of a buffer which is also
275 * associated with a committing transaction creates a deadlock,
276 * so commit simply cannot force through checkpoints.
278 * We must therefore ensure the necessary space in the journal
279 * *before* starting to dirty potentially checkpointed buffers
280 * in the new transaction.
282 * The worst part is, any transaction currently committing can
283 * reduce the free space arbitrarily. Be careful to account for
284 * those buffers when checkpointing.
288 * @@@ AKPM: This seems rather over-defensive. We're giving commit
289 * a _lot_ of headroom: 1/4 of the journal plus the size of
290 * the committing transaction. Really, we only need to give it
291 * committing_transaction->t_outstanding_credits plus "enough" for
292 * the log control blocks.
293 * Also, this test is inconsistent with the matching one in
294 * jbd2_journal_extend().
296 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
297 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
298 atomic_sub(nblocks, &transaction->t_outstanding_credits);
299 read_unlock(&journal->j_state_lock);
300 write_lock(&journal->j_state_lock);
301 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
302 __jbd2_log_wait_for_space(journal);
303 write_unlock(&journal->j_state_lock);
307 /* OK, account for the buffers that this operation expects to
308 * use and add the handle to the running transaction.
310 update_t_max_wait(transaction, ts);
311 handle->h_transaction = transaction;
312 handle->h_requested_credits = nblocks;
313 handle->h_start_jiffies = jiffies;
314 atomic_inc(&transaction->t_updates);
315 atomic_inc(&transaction->t_handle_count);
316 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
318 atomic_read(&transaction->t_outstanding_credits),
319 __jbd2_log_space_left(journal));
320 read_unlock(&journal->j_state_lock);
322 lock_map_acquire(&handle->h_lockdep_map);
323 jbd2_journal_free_transaction(new_transaction);
327 static struct lock_class_key jbd2_handle_key;
329 /* Allocate a new handle. This should probably be in a slab... */
330 static handle_t *new_handle(int nblocks)
332 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
335 handle->h_buffer_credits = nblocks;
338 lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
339 &jbd2_handle_key, 0);
345 * handle_t *jbd2_journal_start() - Obtain a new handle.
346 * @journal: Journal to start transaction on.
347 * @nblocks: number of block buffer we might modify
349 * We make sure that the transaction can guarantee at least nblocks of
350 * modified buffers in the log. We block until the log can guarantee
353 * This function is visible to journal users (like ext3fs), so is not
354 * called with the journal already locked.
356 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
359 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, gfp_t gfp_mask,
360 unsigned int type, unsigned int line_no)
362 handle_t *handle = journal_current_handle();
366 return ERR_PTR(-EROFS);
369 J_ASSERT(handle->h_transaction->t_journal == journal);
374 handle = new_handle(nblocks);
376 return ERR_PTR(-ENOMEM);
378 current->journal_info = handle;
380 err = start_this_handle(journal, handle, gfp_mask);
382 jbd2_free_handle(handle);
383 current->journal_info = NULL;
386 handle->h_type = type;
387 handle->h_line_no = line_no;
388 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
389 handle->h_transaction->t_tid, type,
393 EXPORT_SYMBOL(jbd2__journal_start);
396 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
398 return jbd2__journal_start(journal, nblocks, GFP_NOFS, 0, 0);
400 EXPORT_SYMBOL(jbd2_journal_start);
404 * int jbd2_journal_extend() - extend buffer credits.
405 * @handle: handle to 'extend'
406 * @nblocks: nr blocks to try to extend by.
408 * Some transactions, such as large extends and truncates, can be done
409 * atomically all at once or in several stages. The operation requests
410 * a credit for a number of buffer modications in advance, but can
411 * extend its credit if it needs more.
413 * jbd2_journal_extend tries to give the running handle more buffer credits.
414 * It does not guarantee that allocation - this is a best-effort only.
415 * The calling process MUST be able to deal cleanly with a failure to
418 * Return 0 on success, non-zero on failure.
420 * return code < 0 implies an error
421 * return code > 0 implies normal transaction-full status.
423 int jbd2_journal_extend(handle_t *handle, int nblocks)
425 transaction_t *transaction = handle->h_transaction;
426 journal_t *journal = transaction->t_journal;
431 if (is_handle_aborted(handle))
436 read_lock(&journal->j_state_lock);
438 /* Don't extend a locked-down transaction! */
439 if (handle->h_transaction->t_state != T_RUNNING) {
440 jbd_debug(3, "denied handle %p %d blocks: "
441 "transaction not running\n", handle, nblocks);
445 spin_lock(&transaction->t_handle_lock);
446 wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
448 if (wanted > journal->j_max_transaction_buffers) {
449 jbd_debug(3, "denied handle %p %d blocks: "
450 "transaction too large\n", handle, nblocks);
454 if (wanted > __jbd2_log_space_left(journal)) {
455 jbd_debug(3, "denied handle %p %d blocks: "
456 "insufficient log space\n", handle, nblocks);
460 trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
461 handle->h_transaction->t_tid,
462 handle->h_type, handle->h_line_no,
463 handle->h_buffer_credits,
466 handle->h_buffer_credits += nblocks;
467 handle->h_requested_credits += nblocks;
468 atomic_add(nblocks, &transaction->t_outstanding_credits);
471 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
473 spin_unlock(&transaction->t_handle_lock);
475 read_unlock(&journal->j_state_lock);
482 * int jbd2_journal_restart() - restart a handle .
483 * @handle: handle to restart
484 * @nblocks: nr credits requested
486 * Restart a handle for a multi-transaction filesystem
489 * If the jbd2_journal_extend() call above fails to grant new buffer credits
490 * to a running handle, a call to jbd2_journal_restart will commit the
491 * handle's transaction so far and reattach the handle to a new
492 * transaction capabable of guaranteeing the requested number of
495 int jbd2__journal_restart(handle_t *handle, int nblocks, gfp_t gfp_mask)
497 transaction_t *transaction = handle->h_transaction;
498 journal_t *journal = transaction->t_journal;
500 int need_to_start, ret;
502 /* If we've had an abort of any type, don't even think about
503 * actually doing the restart! */
504 if (is_handle_aborted(handle))
508 * First unlink the handle from its current transaction, and start the
511 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
512 J_ASSERT(journal_current_handle() == handle);
514 read_lock(&journal->j_state_lock);
515 spin_lock(&transaction->t_handle_lock);
516 atomic_sub(handle->h_buffer_credits,
517 &transaction->t_outstanding_credits);
518 if (atomic_dec_and_test(&transaction->t_updates))
519 wake_up(&journal->j_wait_updates);
520 spin_unlock(&transaction->t_handle_lock);
522 jbd_debug(2, "restarting handle %p\n", handle);
523 tid = transaction->t_tid;
524 need_to_start = !tid_geq(journal->j_commit_request, tid);
525 read_unlock(&journal->j_state_lock);
527 jbd2_log_start_commit(journal, tid);
529 lock_map_release(&handle->h_lockdep_map);
530 handle->h_buffer_credits = nblocks;
531 ret = start_this_handle(journal, handle, gfp_mask);
534 EXPORT_SYMBOL(jbd2__journal_restart);
537 int jbd2_journal_restart(handle_t *handle, int nblocks)
539 return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
541 EXPORT_SYMBOL(jbd2_journal_restart);
544 * void jbd2_journal_lock_updates () - establish a transaction barrier.
545 * @journal: Journal to establish a barrier on.
547 * This locks out any further updates from being started, and blocks
548 * until all existing updates have completed, returning only once the
549 * journal is in a quiescent state with no updates running.
551 * The journal lock should not be held on entry.
553 void jbd2_journal_lock_updates(journal_t *journal)
557 write_lock(&journal->j_state_lock);
558 ++journal->j_barrier_count;
560 /* Wait until there are no running updates */
562 transaction_t *transaction = journal->j_running_transaction;
567 spin_lock(&transaction->t_handle_lock);
568 prepare_to_wait(&journal->j_wait_updates, &wait,
569 TASK_UNINTERRUPTIBLE);
570 if (!atomic_read(&transaction->t_updates)) {
571 spin_unlock(&transaction->t_handle_lock);
572 finish_wait(&journal->j_wait_updates, &wait);
575 spin_unlock(&transaction->t_handle_lock);
576 write_unlock(&journal->j_state_lock);
578 finish_wait(&journal->j_wait_updates, &wait);
579 write_lock(&journal->j_state_lock);
581 write_unlock(&journal->j_state_lock);
584 * We have now established a barrier against other normal updates, but
585 * we also need to barrier against other jbd2_journal_lock_updates() calls
586 * to make sure that we serialise special journal-locked operations
589 mutex_lock(&journal->j_barrier);
593 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
594 * @journal: Journal to release the barrier on.
596 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
598 * Should be called without the journal lock held.
600 void jbd2_journal_unlock_updates (journal_t *journal)
602 J_ASSERT(journal->j_barrier_count != 0);
604 mutex_unlock(&journal->j_barrier);
605 write_lock(&journal->j_state_lock);
606 --journal->j_barrier_count;
607 write_unlock(&journal->j_state_lock);
608 wake_up(&journal->j_wait_transaction_locked);
611 static void warn_dirty_buffer(struct buffer_head *bh)
613 char b[BDEVNAME_SIZE];
616 "JBD2: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
617 "There's a risk of filesystem corruption in case of system "
619 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
623 * If the buffer is already part of the current transaction, then there
624 * is nothing we need to do. If it is already part of a prior
625 * transaction which we are still committing to disk, then we need to
626 * make sure that we do not overwrite the old copy: we do copy-out to
627 * preserve the copy going to disk. We also account the buffer against
628 * the handle's metadata buffer credits (unless the buffer is already
629 * part of the transaction, that is).
633 do_get_write_access(handle_t *handle, struct journal_head *jh,
636 struct buffer_head *bh;
637 transaction_t *transaction;
640 char *frozen_buffer = NULL;
642 unsigned long start_lock, time_lock;
644 if (is_handle_aborted(handle))
647 transaction = handle->h_transaction;
648 journal = transaction->t_journal;
650 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
652 JBUFFER_TRACE(jh, "entry");
656 /* @@@ Need to check for errors here at some point. */
658 start_lock = jiffies;
660 jbd_lock_bh_state(bh);
662 /* If it takes too long to lock the buffer, trace it */
663 time_lock = jbd2_time_diff(start_lock, jiffies);
664 if (time_lock > HZ/10)
665 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
666 jiffies_to_msecs(time_lock));
668 /* We now hold the buffer lock so it is safe to query the buffer
669 * state. Is the buffer dirty?
671 * If so, there are two possibilities. The buffer may be
672 * non-journaled, and undergoing a quite legitimate writeback.
673 * Otherwise, it is journaled, and we don't expect dirty buffers
674 * in that state (the buffers should be marked JBD_Dirty
675 * instead.) So either the IO is being done under our own
676 * control and this is a bug, or it's a third party IO such as
677 * dump(8) (which may leave the buffer scheduled for read ---
678 * ie. locked but not dirty) or tune2fs (which may actually have
679 * the buffer dirtied, ugh.) */
681 if (buffer_dirty(bh)) {
683 * First question: is this buffer already part of the current
684 * transaction or the existing committing transaction?
686 if (jh->b_transaction) {
688 jh->b_transaction == transaction ||
690 journal->j_committing_transaction);
691 if (jh->b_next_transaction)
692 J_ASSERT_JH(jh, jh->b_next_transaction ==
694 warn_dirty_buffer(bh);
697 * In any case we need to clean the dirty flag and we must
698 * do it under the buffer lock to be sure we don't race
699 * with running write-out.
701 JBUFFER_TRACE(jh, "Journalling dirty buffer");
702 clear_buffer_dirty(bh);
703 set_buffer_jbddirty(bh);
709 if (is_handle_aborted(handle)) {
710 jbd_unlock_bh_state(bh);
716 * The buffer is already part of this transaction if b_transaction or
717 * b_next_transaction points to it
719 if (jh->b_transaction == transaction ||
720 jh->b_next_transaction == transaction)
724 * this is the first time this transaction is touching this buffer,
725 * reset the modified flag
730 * If there is already a copy-out version of this buffer, then we don't
731 * need to make another one
733 if (jh->b_frozen_data) {
734 JBUFFER_TRACE(jh, "has frozen data");
735 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
736 jh->b_next_transaction = transaction;
740 /* Is there data here we need to preserve? */
742 if (jh->b_transaction && jh->b_transaction != transaction) {
743 JBUFFER_TRACE(jh, "owned by older transaction");
744 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
745 J_ASSERT_JH(jh, jh->b_transaction ==
746 journal->j_committing_transaction);
748 /* There is one case we have to be very careful about.
749 * If the committing transaction is currently writing
750 * this buffer out to disk and has NOT made a copy-out,
751 * then we cannot modify the buffer contents at all
752 * right now. The essence of copy-out is that it is the
753 * extra copy, not the primary copy, which gets
754 * journaled. If the primary copy is already going to
755 * disk then we cannot do copy-out here. */
757 if (jh->b_jlist == BJ_Shadow) {
758 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
759 wait_queue_head_t *wqh;
761 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
763 JBUFFER_TRACE(jh, "on shadow: sleep");
764 jbd_unlock_bh_state(bh);
765 /* commit wakes up all shadow buffers after IO */
767 prepare_to_wait(wqh, &wait.wait,
768 TASK_UNINTERRUPTIBLE);
769 if (jh->b_jlist != BJ_Shadow)
773 finish_wait(wqh, &wait.wait);
777 /* Only do the copy if the currently-owning transaction
778 * still needs it. If it is on the Forget list, the
779 * committing transaction is past that stage. The
780 * buffer had better remain locked during the kmalloc,
781 * but that should be true --- we hold the journal lock
782 * still and the buffer is already on the BUF_JOURNAL
783 * list so won't be flushed.
785 * Subtle point, though: if this is a get_undo_access,
786 * then we will be relying on the frozen_data to contain
787 * the new value of the committed_data record after the
788 * transaction, so we HAVE to force the frozen_data copy
791 if (jh->b_jlist != BJ_Forget || force_copy) {
792 JBUFFER_TRACE(jh, "generate frozen data");
793 if (!frozen_buffer) {
794 JBUFFER_TRACE(jh, "allocate memory for buffer");
795 jbd_unlock_bh_state(bh);
797 jbd2_alloc(jh2bh(jh)->b_size,
799 if (!frozen_buffer) {
801 "%s: OOM for frozen_buffer\n",
803 JBUFFER_TRACE(jh, "oom!");
805 jbd_lock_bh_state(bh);
810 jh->b_frozen_data = frozen_buffer;
811 frozen_buffer = NULL;
814 jh->b_next_transaction = transaction;
819 * Finally, if the buffer is not journaled right now, we need to make
820 * sure it doesn't get written to disk before the caller actually
821 * commits the new data
823 if (!jh->b_transaction) {
824 JBUFFER_TRACE(jh, "no transaction");
825 J_ASSERT_JH(jh, !jh->b_next_transaction);
826 JBUFFER_TRACE(jh, "file as BJ_Reserved");
827 spin_lock(&journal->j_list_lock);
828 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
829 spin_unlock(&journal->j_list_lock);
838 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
839 "Possible IO failure.\n");
840 page = jh2bh(jh)->b_page;
841 offset = offset_in_page(jh2bh(jh)->b_data);
842 source = kmap_atomic(page);
843 /* Fire data frozen trigger just before we copy the data */
844 jbd2_buffer_frozen_trigger(jh, source + offset,
846 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
847 kunmap_atomic(source);
850 * Now that the frozen data is saved off, we need to store
851 * any matching triggers.
853 jh->b_frozen_triggers = jh->b_triggers;
855 jbd_unlock_bh_state(bh);
858 * If we are about to journal a buffer, then any revoke pending on it is
861 jbd2_journal_cancel_revoke(handle, jh);
864 if (unlikely(frozen_buffer)) /* It's usually NULL */
865 jbd2_free(frozen_buffer, bh->b_size);
867 JBUFFER_TRACE(jh, "exit");
872 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
873 * @handle: transaction to add buffer modifications to
874 * @bh: bh to be used for metadata writes
876 * Returns an error code or 0 on success.
878 * In full data journalling mode the buffer may be of type BJ_AsyncData,
879 * because we're write()ing a buffer which is also part of a shared mapping.
882 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
884 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
887 /* We do not want to get caught playing with fields which the
888 * log thread also manipulates. Make sure that the buffer
889 * completes any outstanding IO before proceeding. */
890 rc = do_get_write_access(handle, jh, 0);
891 jbd2_journal_put_journal_head(jh);
897 * When the user wants to journal a newly created buffer_head
898 * (ie. getblk() returned a new buffer and we are going to populate it
899 * manually rather than reading off disk), then we need to keep the
900 * buffer_head locked until it has been completely filled with new
901 * data. In this case, we should be able to make the assertion that
902 * the bh is not already part of an existing transaction.
904 * The buffer should already be locked by the caller by this point.
905 * There is no lock ranking violation: it was a newly created,
906 * unlocked buffer beforehand. */
909 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
910 * @handle: transaction to new buffer to
913 * Call this if you create a new bh.
915 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
917 transaction_t *transaction = handle->h_transaction;
918 journal_t *journal = transaction->t_journal;
919 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
922 jbd_debug(5, "journal_head %p\n", jh);
924 if (is_handle_aborted(handle))
928 JBUFFER_TRACE(jh, "entry");
930 * The buffer may already belong to this transaction due to pre-zeroing
931 * in the filesystem's new_block code. It may also be on the previous,
932 * committing transaction's lists, but it HAS to be in Forget state in
933 * that case: the transaction must have deleted the buffer for it to be
936 jbd_lock_bh_state(bh);
937 spin_lock(&journal->j_list_lock);
938 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
939 jh->b_transaction == NULL ||
940 (jh->b_transaction == journal->j_committing_transaction &&
941 jh->b_jlist == BJ_Forget)));
943 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
944 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
946 if (jh->b_transaction == NULL) {
948 * Previous jbd2_journal_forget() could have left the buffer
949 * with jbddirty bit set because it was being committed. When
950 * the commit finished, we've filed the buffer for
951 * checkpointing and marked it dirty. Now we are reallocating
952 * the buffer so the transaction freeing it must have
953 * committed and so it's safe to clear the dirty bit.
955 clear_buffer_dirty(jh2bh(jh));
956 /* first access by this transaction */
959 JBUFFER_TRACE(jh, "file as BJ_Reserved");
960 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
961 } else if (jh->b_transaction == journal->j_committing_transaction) {
962 /* first access by this transaction */
965 JBUFFER_TRACE(jh, "set next transaction");
966 jh->b_next_transaction = transaction;
968 spin_unlock(&journal->j_list_lock);
969 jbd_unlock_bh_state(bh);
972 * akpm: I added this. ext3_alloc_branch can pick up new indirect
973 * blocks which contain freed but then revoked metadata. We need
974 * to cancel the revoke in case we end up freeing it yet again
975 * and the reallocating as data - this would cause a second revoke,
976 * which hits an assertion error.
978 JBUFFER_TRACE(jh, "cancelling revoke");
979 jbd2_journal_cancel_revoke(handle, jh);
981 jbd2_journal_put_journal_head(jh);
986 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
987 * non-rewindable consequences
988 * @handle: transaction
989 * @bh: buffer to undo
991 * Sometimes there is a need to distinguish between metadata which has
992 * been committed to disk and that which has not. The ext3fs code uses
993 * this for freeing and allocating space, we have to make sure that we
994 * do not reuse freed space until the deallocation has been committed,
995 * since if we overwrote that space we would make the delete
996 * un-rewindable in case of a crash.
998 * To deal with that, jbd2_journal_get_undo_access requests write access to a
999 * buffer for parts of non-rewindable operations such as delete
1000 * operations on the bitmaps. The journaling code must keep a copy of
1001 * the buffer's contents prior to the undo_access call until such time
1002 * as we know that the buffer has definitely been committed to disk.
1004 * We never need to know which transaction the committed data is part
1005 * of, buffers touched here are guaranteed to be dirtied later and so
1006 * will be committed to a new transaction in due course, at which point
1007 * we can discard the old committed data pointer.
1009 * Returns error number or 0 on success.
1011 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1014 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1015 char *committed_data = NULL;
1017 JBUFFER_TRACE(jh, "entry");
1020 * Do this first --- it can drop the journal lock, so we want to
1021 * make sure that obtaining the committed_data is done
1022 * atomically wrt. completion of any outstanding commits.
1024 err = do_get_write_access(handle, jh, 1);
1029 if (!jh->b_committed_data) {
1030 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
1031 if (!committed_data) {
1032 printk(KERN_EMERG "%s: No memory for committed data\n",
1039 jbd_lock_bh_state(bh);
1040 if (!jh->b_committed_data) {
1041 /* Copy out the current buffer contents into the
1042 * preserved, committed copy. */
1043 JBUFFER_TRACE(jh, "generate b_committed data");
1044 if (!committed_data) {
1045 jbd_unlock_bh_state(bh);
1049 jh->b_committed_data = committed_data;
1050 committed_data = NULL;
1051 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1053 jbd_unlock_bh_state(bh);
1055 jbd2_journal_put_journal_head(jh);
1056 if (unlikely(committed_data))
1057 jbd2_free(committed_data, bh->b_size);
1062 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1063 * @bh: buffer to trigger on
1064 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1066 * Set any triggers on this journal_head. This is always safe, because
1067 * triggers for a committing buffer will be saved off, and triggers for
1068 * a running transaction will match the buffer in that transaction.
1070 * Call with NULL to clear the triggers.
1072 void jbd2_journal_set_triggers(struct buffer_head *bh,
1073 struct jbd2_buffer_trigger_type *type)
1075 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1079 jh->b_triggers = type;
1080 jbd2_journal_put_journal_head(jh);
1083 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1084 struct jbd2_buffer_trigger_type *triggers)
1086 struct buffer_head *bh = jh2bh(jh);
1088 if (!triggers || !triggers->t_frozen)
1091 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1094 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1095 struct jbd2_buffer_trigger_type *triggers)
1097 if (!triggers || !triggers->t_abort)
1100 triggers->t_abort(triggers, jh2bh(jh));
1106 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1107 * @handle: transaction to add buffer to.
1108 * @bh: buffer to mark
1110 * mark dirty metadata which needs to be journaled as part of the current
1113 * The buffer must have previously had jbd2_journal_get_write_access()
1114 * called so that it has a valid journal_head attached to the buffer
1117 * The buffer is placed on the transaction's metadata list and is marked
1118 * as belonging to the transaction.
1120 * Returns error number or 0 on success.
1122 * Special care needs to be taken if the buffer already belongs to the
1123 * current committing transaction (in which case we should have frozen
1124 * data present for that commit). In that case, we don't relink the
1125 * buffer: that only gets done when the old transaction finally
1126 * completes its commit.
1128 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1130 transaction_t *transaction = handle->h_transaction;
1131 journal_t *journal = transaction->t_journal;
1132 struct journal_head *jh;
1135 if (is_handle_aborted(handle))
1137 jh = jbd2_journal_grab_journal_head(bh);
1142 jbd_debug(5, "journal_head %p\n", jh);
1143 JBUFFER_TRACE(jh, "entry");
1145 jbd_lock_bh_state(bh);
1147 if (jh->b_modified == 0) {
1149 * This buffer's got modified and becoming part
1150 * of the transaction. This needs to be done
1151 * once a transaction -bzzz
1154 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1155 handle->h_buffer_credits--;
1159 * fastpath, to avoid expensive locking. If this buffer is already
1160 * on the running transaction's metadata list there is nothing to do.
1161 * Nobody can take it off again because there is a handle open.
1162 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1163 * result in this test being false, so we go in and take the locks.
1165 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1166 JBUFFER_TRACE(jh, "fastpath");
1167 if (unlikely(jh->b_transaction !=
1168 journal->j_running_transaction)) {
1169 printk(KERN_EMERG "JBD: %s: "
1170 "jh->b_transaction (%llu, %p, %u) != "
1171 "journal->j_running_transaction (%p, %u)",
1173 (unsigned long long) bh->b_blocknr,
1175 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1176 journal->j_running_transaction,
1177 journal->j_running_transaction ?
1178 journal->j_running_transaction->t_tid : 0);
1184 set_buffer_jbddirty(bh);
1187 * Metadata already on the current transaction list doesn't
1188 * need to be filed. Metadata on another transaction's list must
1189 * be committing, and will be refiled once the commit completes:
1190 * leave it alone for now.
1192 if (jh->b_transaction != transaction) {
1193 JBUFFER_TRACE(jh, "already on other transaction");
1194 if (unlikely(jh->b_transaction !=
1195 journal->j_committing_transaction)) {
1196 printk(KERN_EMERG "JBD: %s: "
1197 "jh->b_transaction (%llu, %p, %u) != "
1198 "journal->j_committing_transaction (%p, %u)",
1200 (unsigned long long) bh->b_blocknr,
1202 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1203 journal->j_committing_transaction,
1204 journal->j_committing_transaction ?
1205 journal->j_committing_transaction->t_tid : 0);
1208 if (unlikely(jh->b_next_transaction != transaction)) {
1209 printk(KERN_EMERG "JBD: %s: "
1210 "jh->b_next_transaction (%llu, %p, %u) != "
1211 "transaction (%p, %u)",
1213 (unsigned long long) bh->b_blocknr,
1214 jh->b_next_transaction,
1215 jh->b_next_transaction ?
1216 jh->b_next_transaction->t_tid : 0,
1217 transaction, transaction->t_tid);
1220 /* And this case is illegal: we can't reuse another
1221 * transaction's data buffer, ever. */
1225 /* That test should have eliminated the following case: */
1226 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1228 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1229 spin_lock(&journal->j_list_lock);
1230 __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1231 spin_unlock(&journal->j_list_lock);
1233 jbd_unlock_bh_state(bh);
1234 jbd2_journal_put_journal_head(jh);
1236 JBUFFER_TRACE(jh, "exit");
1237 WARN_ON(ret); /* All errors are bugs, so dump the stack */
1242 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1243 * @handle: transaction handle
1244 * @bh: bh to 'forget'
1246 * We can only do the bforget if there are no commits pending against the
1247 * buffer. If the buffer is dirty in the current running transaction we
1248 * can safely unlink it.
1250 * bh may not be a journalled buffer at all - it may be a non-JBD
1251 * buffer which came off the hashtable. Check for this.
1253 * Decrements bh->b_count by one.
1255 * Allow this call even if the handle has aborted --- it may be part of
1256 * the caller's cleanup after an abort.
1258 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1260 transaction_t *transaction = handle->h_transaction;
1261 journal_t *journal = transaction->t_journal;
1262 struct journal_head *jh;
1263 int drop_reserve = 0;
1265 int was_modified = 0;
1267 BUFFER_TRACE(bh, "entry");
1269 jbd_lock_bh_state(bh);
1270 spin_lock(&journal->j_list_lock);
1272 if (!buffer_jbd(bh))
1276 /* Critical error: attempting to delete a bitmap buffer, maybe?
1277 * Don't do any jbd operations, and return an error. */
1278 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1279 "inconsistent data on disk")) {
1284 /* keep track of whether or not this transaction modified us */
1285 was_modified = jh->b_modified;
1288 * The buffer's going from the transaction, we must drop
1289 * all references -bzzz
1293 if (jh->b_transaction == handle->h_transaction) {
1294 J_ASSERT_JH(jh, !jh->b_frozen_data);
1296 /* If we are forgetting a buffer which is already part
1297 * of this transaction, then we can just drop it from
1298 * the transaction immediately. */
1299 clear_buffer_dirty(bh);
1300 clear_buffer_jbddirty(bh);
1302 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1305 * we only want to drop a reference if this transaction
1306 * modified the buffer
1312 * We are no longer going to journal this buffer.
1313 * However, the commit of this transaction is still
1314 * important to the buffer: the delete that we are now
1315 * processing might obsolete an old log entry, so by
1316 * committing, we can satisfy the buffer's checkpoint.
1318 * So, if we have a checkpoint on the buffer, we should
1319 * now refile the buffer on our BJ_Forget list so that
1320 * we know to remove the checkpoint after we commit.
1323 if (jh->b_cp_transaction) {
1324 __jbd2_journal_temp_unlink_buffer(jh);
1325 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1327 __jbd2_journal_unfile_buffer(jh);
1328 if (!buffer_jbd(bh)) {
1329 spin_unlock(&journal->j_list_lock);
1330 jbd_unlock_bh_state(bh);
1335 } else if (jh->b_transaction) {
1336 J_ASSERT_JH(jh, (jh->b_transaction ==
1337 journal->j_committing_transaction));
1338 /* However, if the buffer is still owned by a prior
1339 * (committing) transaction, we can't drop it yet... */
1340 JBUFFER_TRACE(jh, "belongs to older transaction");
1341 /* ... but we CAN drop it from the new transaction if we
1342 * have also modified it since the original commit. */
1344 if (jh->b_next_transaction) {
1345 J_ASSERT(jh->b_next_transaction == transaction);
1346 jh->b_next_transaction = NULL;
1349 * only drop a reference if this transaction modified
1358 spin_unlock(&journal->j_list_lock);
1359 jbd_unlock_bh_state(bh);
1363 /* no need to reserve log space for this block -bzzz */
1364 handle->h_buffer_credits++;
1370 * int jbd2_journal_stop() - complete a transaction
1371 * @handle: tranaction to complete.
1373 * All done for a particular handle.
1375 * There is not much action needed here. We just return any remaining
1376 * buffer credits to the transaction and remove the handle. The only
1377 * complication is that we need to start a commit operation if the
1378 * filesystem is marked for synchronous update.
1380 * jbd2_journal_stop itself will not usually return an error, but it may
1381 * do so in unusual circumstances. In particular, expect it to
1382 * return -EIO if a jbd2_journal_abort has been executed since the
1383 * transaction began.
1385 int jbd2_journal_stop(handle_t *handle)
1387 transaction_t *transaction = handle->h_transaction;
1388 journal_t *journal = transaction->t_journal;
1389 int err, wait_for_commit = 0;
1393 J_ASSERT(journal_current_handle() == handle);
1395 if (is_handle_aborted(handle))
1398 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1402 if (--handle->h_ref > 0) {
1403 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1408 jbd_debug(4, "Handle %p going down\n", handle);
1409 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1410 handle->h_transaction->t_tid,
1411 handle->h_type, handle->h_line_no,
1412 jiffies - handle->h_start_jiffies,
1413 handle->h_sync, handle->h_requested_credits,
1414 (handle->h_requested_credits -
1415 handle->h_buffer_credits));
1418 * Implement synchronous transaction batching. If the handle
1419 * was synchronous, don't force a commit immediately. Let's
1420 * yield and let another thread piggyback onto this
1421 * transaction. Keep doing that while new threads continue to
1422 * arrive. It doesn't cost much - we're about to run a commit
1423 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1424 * operations by 30x or more...
1426 * We try and optimize the sleep time against what the
1427 * underlying disk can do, instead of having a static sleep
1428 * time. This is useful for the case where our storage is so
1429 * fast that it is more optimal to go ahead and force a flush
1430 * and wait for the transaction to be committed than it is to
1431 * wait for an arbitrary amount of time for new writers to
1432 * join the transaction. We achieve this by measuring how
1433 * long it takes to commit a transaction, and compare it with
1434 * how long this transaction has been running, and if run time
1435 * < commit time then we sleep for the delta and commit. This
1436 * greatly helps super fast disks that would see slowdowns as
1437 * more threads started doing fsyncs.
1439 * But don't do this if this process was the most recent one
1440 * to perform a synchronous write. We do this to detect the
1441 * case where a single process is doing a stream of sync
1442 * writes. No point in waiting for joiners in that case.
1445 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1446 u64 commit_time, trans_time;
1448 journal->j_last_sync_writer = pid;
1450 read_lock(&journal->j_state_lock);
1451 commit_time = journal->j_average_commit_time;
1452 read_unlock(&journal->j_state_lock);
1454 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1455 transaction->t_start_time));
1457 commit_time = max_t(u64, commit_time,
1458 1000*journal->j_min_batch_time);
1459 commit_time = min_t(u64, commit_time,
1460 1000*journal->j_max_batch_time);
1462 if (trans_time < commit_time) {
1463 ktime_t expires = ktime_add_ns(ktime_get(),
1465 set_current_state(TASK_UNINTERRUPTIBLE);
1466 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1471 transaction->t_synchronous_commit = 1;
1472 current->journal_info = NULL;
1473 atomic_sub(handle->h_buffer_credits,
1474 &transaction->t_outstanding_credits);
1477 * If the handle is marked SYNC, we need to set another commit
1478 * going! We also want to force a commit if the current
1479 * transaction is occupying too much of the log, or if the
1480 * transaction is too old now.
1482 if (handle->h_sync ||
1483 (atomic_read(&transaction->t_outstanding_credits) >
1484 journal->j_max_transaction_buffers) ||
1485 time_after_eq(jiffies, transaction->t_expires)) {
1486 /* Do this even for aborted journals: an abort still
1487 * completes the commit thread, it just doesn't write
1488 * anything to disk. */
1490 jbd_debug(2, "transaction too old, requesting commit for "
1491 "handle %p\n", handle);
1492 /* This is non-blocking */
1493 jbd2_log_start_commit(journal, transaction->t_tid);
1496 * Special case: JBD2_SYNC synchronous updates require us
1497 * to wait for the commit to complete.
1499 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1500 wait_for_commit = 1;
1504 * Once we drop t_updates, if it goes to zero the transaction
1505 * could start committing on us and eventually disappear. So
1506 * once we do this, we must not dereference transaction
1509 tid = transaction->t_tid;
1510 if (atomic_dec_and_test(&transaction->t_updates)) {
1511 wake_up(&journal->j_wait_updates);
1512 if (journal->j_barrier_count)
1513 wake_up(&journal->j_wait_transaction_locked);
1516 if (wait_for_commit)
1517 err = jbd2_log_wait_commit(journal, tid);
1519 lock_map_release(&handle->h_lockdep_map);
1521 jbd2_free_handle(handle);
1526 * int jbd2_journal_force_commit() - force any uncommitted transactions
1527 * @journal: journal to force
1529 * For synchronous operations: force any uncommitted transactions
1530 * to disk. May seem kludgy, but it reuses all the handle batching
1531 * code in a very simple manner.
1533 int jbd2_journal_force_commit(journal_t *journal)
1538 handle = jbd2_journal_start(journal, 1);
1539 if (IS_ERR(handle)) {
1540 ret = PTR_ERR(handle);
1543 ret = jbd2_journal_stop(handle);
1550 * List management code snippets: various functions for manipulating the
1551 * transaction buffer lists.
1556 * Append a buffer to a transaction list, given the transaction's list head
1559 * j_list_lock is held.
1561 * jbd_lock_bh_state(jh2bh(jh)) is held.
1565 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1568 jh->b_tnext = jh->b_tprev = jh;
1571 /* Insert at the tail of the list to preserve order */
1572 struct journal_head *first = *list, *last = first->b_tprev;
1574 jh->b_tnext = first;
1575 last->b_tnext = first->b_tprev = jh;
1580 * Remove a buffer from a transaction list, given the transaction's list
1583 * Called with j_list_lock held, and the journal may not be locked.
1585 * jbd_lock_bh_state(jh2bh(jh)) is held.
1589 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1592 *list = jh->b_tnext;
1596 jh->b_tprev->b_tnext = jh->b_tnext;
1597 jh->b_tnext->b_tprev = jh->b_tprev;
1601 * Remove a buffer from the appropriate transaction list.
1603 * Note that this function can *change* the value of
1604 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1605 * t_log_list or t_reserved_list. If the caller is holding onto a copy of one
1606 * of these pointers, it could go bad. Generally the caller needs to re-read
1607 * the pointer from the transaction_t.
1609 * Called under j_list_lock.
1611 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1613 struct journal_head **list = NULL;
1614 transaction_t *transaction;
1615 struct buffer_head *bh = jh2bh(jh);
1617 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1618 transaction = jh->b_transaction;
1620 assert_spin_locked(&transaction->t_journal->j_list_lock);
1622 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1623 if (jh->b_jlist != BJ_None)
1624 J_ASSERT_JH(jh, transaction != NULL);
1626 switch (jh->b_jlist) {
1630 transaction->t_nr_buffers--;
1631 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1632 list = &transaction->t_buffers;
1635 list = &transaction->t_forget;
1638 list = &transaction->t_iobuf_list;
1641 list = &transaction->t_shadow_list;
1644 list = &transaction->t_log_list;
1647 list = &transaction->t_reserved_list;
1651 __blist_del_buffer(list, jh);
1652 jh->b_jlist = BJ_None;
1653 if (test_clear_buffer_jbddirty(bh))
1654 mark_buffer_dirty(bh); /* Expose it to the VM */
1658 * Remove buffer from all transactions.
1660 * Called with bh_state lock and j_list_lock
1662 * jh and bh may be already freed when this function returns.
1664 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1666 __jbd2_journal_temp_unlink_buffer(jh);
1667 jh->b_transaction = NULL;
1668 jbd2_journal_put_journal_head(jh);
1671 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1673 struct buffer_head *bh = jh2bh(jh);
1675 /* Get reference so that buffer cannot be freed before we unlock it */
1677 jbd_lock_bh_state(bh);
1678 spin_lock(&journal->j_list_lock);
1679 __jbd2_journal_unfile_buffer(jh);
1680 spin_unlock(&journal->j_list_lock);
1681 jbd_unlock_bh_state(bh);
1686 * Called from jbd2_journal_try_to_free_buffers().
1688 * Called under jbd_lock_bh_state(bh)
1691 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1693 struct journal_head *jh;
1697 if (buffer_locked(bh) || buffer_dirty(bh))
1700 if (jh->b_next_transaction != NULL)
1703 spin_lock(&journal->j_list_lock);
1704 if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1705 /* written-back checkpointed metadata buffer */
1706 JBUFFER_TRACE(jh, "remove from checkpoint list");
1707 __jbd2_journal_remove_checkpoint(jh);
1709 spin_unlock(&journal->j_list_lock);
1715 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1716 * @journal: journal for operation
1717 * @page: to try and free
1718 * @gfp_mask: we use the mask to detect how hard should we try to release
1719 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1720 * release the buffers.
1723 * For all the buffers on this page,
1724 * if they are fully written out ordered data, move them onto BUF_CLEAN
1725 * so try_to_free_buffers() can reap them.
1727 * This function returns non-zero if we wish try_to_free_buffers()
1728 * to be called. We do this if the page is releasable by try_to_free_buffers().
1729 * We also do it if the page has locked or dirty buffers and the caller wants
1730 * us to perform sync or async writeout.
1732 * This complicates JBD locking somewhat. We aren't protected by the
1733 * BKL here. We wish to remove the buffer from its committing or
1734 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1736 * This may *change* the value of transaction_t->t_datalist, so anyone
1737 * who looks at t_datalist needs to lock against this function.
1739 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1740 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
1741 * will come out of the lock with the buffer dirty, which makes it
1742 * ineligible for release here.
1744 * Who else is affected by this? hmm... Really the only contender
1745 * is do_get_write_access() - it could be looking at the buffer while
1746 * journal_try_to_free_buffer() is changing its state. But that
1747 * cannot happen because we never reallocate freed data as metadata
1748 * while the data is part of a transaction. Yes?
1750 * Return 0 on failure, 1 on success
1752 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1753 struct page *page, gfp_t gfp_mask)
1755 struct buffer_head *head;
1756 struct buffer_head *bh;
1759 J_ASSERT(PageLocked(page));
1761 head = page_buffers(page);
1764 struct journal_head *jh;
1767 * We take our own ref against the journal_head here to avoid
1768 * having to add tons of locking around each instance of
1769 * jbd2_journal_put_journal_head().
1771 jh = jbd2_journal_grab_journal_head(bh);
1775 jbd_lock_bh_state(bh);
1776 __journal_try_to_free_buffer(journal, bh);
1777 jbd2_journal_put_journal_head(jh);
1778 jbd_unlock_bh_state(bh);
1781 } while ((bh = bh->b_this_page) != head);
1783 ret = try_to_free_buffers(page);
1790 * This buffer is no longer needed. If it is on an older transaction's
1791 * checkpoint list we need to record it on this transaction's forget list
1792 * to pin this buffer (and hence its checkpointing transaction) down until
1793 * this transaction commits. If the buffer isn't on a checkpoint list, we
1795 * Returns non-zero if JBD no longer has an interest in the buffer.
1797 * Called under j_list_lock.
1799 * Called under jbd_lock_bh_state(bh).
1801 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1804 struct buffer_head *bh = jh2bh(jh);
1806 if (jh->b_cp_transaction) {
1807 JBUFFER_TRACE(jh, "on running+cp transaction");
1808 __jbd2_journal_temp_unlink_buffer(jh);
1810 * We don't want to write the buffer anymore, clear the
1811 * bit so that we don't confuse checks in
1812 * __journal_file_buffer
1814 clear_buffer_dirty(bh);
1815 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1818 JBUFFER_TRACE(jh, "on running transaction");
1819 __jbd2_journal_unfile_buffer(jh);
1825 * jbd2_journal_invalidatepage
1827 * This code is tricky. It has a number of cases to deal with.
1829 * There are two invariants which this code relies on:
1831 * i_size must be updated on disk before we start calling invalidatepage on the
1834 * This is done in ext3 by defining an ext3_setattr method which
1835 * updates i_size before truncate gets going. By maintaining this
1836 * invariant, we can be sure that it is safe to throw away any buffers
1837 * attached to the current transaction: once the transaction commits,
1838 * we know that the data will not be needed.
1840 * Note however that we can *not* throw away data belonging to the
1841 * previous, committing transaction!
1843 * Any disk blocks which *are* part of the previous, committing
1844 * transaction (and which therefore cannot be discarded immediately) are
1845 * not going to be reused in the new running transaction
1847 * The bitmap committed_data images guarantee this: any block which is
1848 * allocated in one transaction and removed in the next will be marked
1849 * as in-use in the committed_data bitmap, so cannot be reused until
1850 * the next transaction to delete the block commits. This means that
1851 * leaving committing buffers dirty is quite safe: the disk blocks
1852 * cannot be reallocated to a different file and so buffer aliasing is
1856 * The above applies mainly to ordered data mode. In writeback mode we
1857 * don't make guarantees about the order in which data hits disk --- in
1858 * particular we don't guarantee that new dirty data is flushed before
1859 * transaction commit --- so it is always safe just to discard data
1860 * immediately in that mode. --sct
1864 * The journal_unmap_buffer helper function returns zero if the buffer
1865 * concerned remains pinned as an anonymous buffer belonging to an older
1868 * We're outside-transaction here. Either or both of j_running_transaction
1869 * and j_committing_transaction may be NULL.
1871 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
1874 transaction_t *transaction;
1875 struct journal_head *jh;
1878 BUFFER_TRACE(bh, "entry");
1881 * It is safe to proceed here without the j_list_lock because the
1882 * buffers cannot be stolen by try_to_free_buffers as long as we are
1883 * holding the page lock. --sct
1886 if (!buffer_jbd(bh))
1887 goto zap_buffer_unlocked;
1889 /* OK, we have data buffer in journaled mode */
1890 write_lock(&journal->j_state_lock);
1891 jbd_lock_bh_state(bh);
1892 spin_lock(&journal->j_list_lock);
1894 jh = jbd2_journal_grab_journal_head(bh);
1896 goto zap_buffer_no_jh;
1899 * We cannot remove the buffer from checkpoint lists until the
1900 * transaction adding inode to orphan list (let's call it T)
1901 * is committed. Otherwise if the transaction changing the
1902 * buffer would be cleaned from the journal before T is
1903 * committed, a crash will cause that the correct contents of
1904 * the buffer will be lost. On the other hand we have to
1905 * clear the buffer dirty bit at latest at the moment when the
1906 * transaction marking the buffer as freed in the filesystem
1907 * structures is committed because from that moment on the
1908 * block can be reallocated and used by a different page.
1909 * Since the block hasn't been freed yet but the inode has
1910 * already been added to orphan list, it is safe for us to add
1911 * the buffer to BJ_Forget list of the newest transaction.
1913 * Also we have to clear buffer_mapped flag of a truncated buffer
1914 * because the buffer_head may be attached to the page straddling
1915 * i_size (can happen only when blocksize < pagesize) and thus the
1916 * buffer_head can be reused when the file is extended again. So we end
1917 * up keeping around invalidated buffers attached to transactions'
1918 * BJ_Forget list just to stop checkpointing code from cleaning up
1919 * the transaction this buffer was modified in.
1921 transaction = jh->b_transaction;
1922 if (transaction == NULL) {
1923 /* First case: not on any transaction. If it
1924 * has no checkpoint link, then we can zap it:
1925 * it's a writeback-mode buffer so we don't care
1926 * if it hits disk safely. */
1927 if (!jh->b_cp_transaction) {
1928 JBUFFER_TRACE(jh, "not on any transaction: zap");
1932 if (!buffer_dirty(bh)) {
1933 /* bdflush has written it. We can drop it now */
1937 /* OK, it must be in the journal but still not
1938 * written fully to disk: it's metadata or
1939 * journaled data... */
1941 if (journal->j_running_transaction) {
1942 /* ... and once the current transaction has
1943 * committed, the buffer won't be needed any
1945 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1946 may_free = __dispose_buffer(jh,
1947 journal->j_running_transaction);
1950 /* There is no currently-running transaction. So the
1951 * orphan record which we wrote for this file must have
1952 * passed into commit. We must attach this buffer to
1953 * the committing transaction, if it exists. */
1954 if (journal->j_committing_transaction) {
1955 JBUFFER_TRACE(jh, "give to committing trans");
1956 may_free = __dispose_buffer(jh,
1957 journal->j_committing_transaction);
1960 /* The orphan record's transaction has
1961 * committed. We can cleanse this buffer */
1962 clear_buffer_jbddirty(bh);
1966 } else if (transaction == journal->j_committing_transaction) {
1967 JBUFFER_TRACE(jh, "on committing transaction");
1969 * The buffer is committing, we simply cannot touch
1970 * it. If the page is straddling i_size we have to wait
1971 * for commit and try again.
1974 jbd2_journal_put_journal_head(jh);
1975 spin_unlock(&journal->j_list_lock);
1976 jbd_unlock_bh_state(bh);
1977 write_unlock(&journal->j_state_lock);
1981 * OK, buffer won't be reachable after truncate. We just set
1982 * j_next_transaction to the running transaction (if there is
1983 * one) and mark buffer as freed so that commit code knows it
1984 * should clear dirty bits when it is done with the buffer.
1986 set_buffer_freed(bh);
1987 if (journal->j_running_transaction && buffer_jbddirty(bh))
1988 jh->b_next_transaction = journal->j_running_transaction;
1989 jbd2_journal_put_journal_head(jh);
1990 spin_unlock(&journal->j_list_lock);
1991 jbd_unlock_bh_state(bh);
1992 write_unlock(&journal->j_state_lock);
1995 /* Good, the buffer belongs to the running transaction.
1996 * We are writing our own transaction's data, not any
1997 * previous one's, so it is safe to throw it away
1998 * (remember that we expect the filesystem to have set
1999 * i_size already for this truncate so recovery will not
2000 * expose the disk blocks we are discarding here.) */
2001 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2002 JBUFFER_TRACE(jh, "on running transaction");
2003 may_free = __dispose_buffer(jh, transaction);
2008 * This is tricky. Although the buffer is truncated, it may be reused
2009 * if blocksize < pagesize and it is attached to the page straddling
2010 * EOF. Since the buffer might have been added to BJ_Forget list of the
2011 * running transaction, journal_get_write_access() won't clear
2012 * b_modified and credit accounting gets confused. So clear b_modified
2016 jbd2_journal_put_journal_head(jh);
2018 spin_unlock(&journal->j_list_lock);
2019 jbd_unlock_bh_state(bh);
2020 write_unlock(&journal->j_state_lock);
2021 zap_buffer_unlocked:
2022 clear_buffer_dirty(bh);
2023 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2024 clear_buffer_mapped(bh);
2025 clear_buffer_req(bh);
2026 clear_buffer_new(bh);
2027 clear_buffer_delay(bh);
2028 clear_buffer_unwritten(bh);
2034 * void jbd2_journal_invalidatepage()
2035 * @journal: journal to use for flush...
2036 * @page: page to flush
2037 * @offset: length of page to invalidate.
2039 * Reap page buffers containing data after offset in page. Can return -EBUSY
2040 * if buffers are part of the committing transaction and the page is straddling
2041 * i_size. Caller then has to wait for current commit and try again.
2043 int jbd2_journal_invalidatepage(journal_t *journal,
2045 unsigned long offset)
2047 struct buffer_head *head, *bh, *next;
2048 unsigned int curr_off = 0;
2052 if (!PageLocked(page))
2054 if (!page_has_buffers(page))
2057 /* We will potentially be playing with lists other than just the
2058 * data lists (especially for journaled data mode), so be
2059 * cautious in our locking. */
2061 head = bh = page_buffers(page);
2063 unsigned int next_off = curr_off + bh->b_size;
2064 next = bh->b_this_page;
2066 if (offset <= curr_off) {
2067 /* This block is wholly outside the truncation point */
2069 ret = journal_unmap_buffer(journal, bh, offset > 0);
2075 curr_off = next_off;
2078 } while (bh != head);
2081 if (may_free && try_to_free_buffers(page))
2082 J_ASSERT(!page_has_buffers(page));
2088 * File a buffer on the given transaction list.
2090 void __jbd2_journal_file_buffer(struct journal_head *jh,
2091 transaction_t *transaction, int jlist)
2093 struct journal_head **list = NULL;
2095 struct buffer_head *bh = jh2bh(jh);
2097 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2098 assert_spin_locked(&transaction->t_journal->j_list_lock);
2100 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2101 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2102 jh->b_transaction == NULL);
2104 if (jh->b_transaction && jh->b_jlist == jlist)
2107 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2108 jlist == BJ_Shadow || jlist == BJ_Forget) {
2110 * For metadata buffers, we track dirty bit in buffer_jbddirty
2111 * instead of buffer_dirty. We should not see a dirty bit set
2112 * here because we clear it in do_get_write_access but e.g.
2113 * tune2fs can modify the sb and set the dirty bit at any time
2114 * so we try to gracefully handle that.
2116 if (buffer_dirty(bh))
2117 warn_dirty_buffer(bh);
2118 if (test_clear_buffer_dirty(bh) ||
2119 test_clear_buffer_jbddirty(bh))
2123 if (jh->b_transaction)
2124 __jbd2_journal_temp_unlink_buffer(jh);
2126 jbd2_journal_grab_journal_head(bh);
2127 jh->b_transaction = transaction;
2131 J_ASSERT_JH(jh, !jh->b_committed_data);
2132 J_ASSERT_JH(jh, !jh->b_frozen_data);
2135 transaction->t_nr_buffers++;
2136 list = &transaction->t_buffers;
2139 list = &transaction->t_forget;
2142 list = &transaction->t_iobuf_list;
2145 list = &transaction->t_shadow_list;
2148 list = &transaction->t_log_list;
2151 list = &transaction->t_reserved_list;
2155 __blist_add_buffer(list, jh);
2156 jh->b_jlist = jlist;
2159 set_buffer_jbddirty(bh);
2162 void jbd2_journal_file_buffer(struct journal_head *jh,
2163 transaction_t *transaction, int jlist)
2165 jbd_lock_bh_state(jh2bh(jh));
2166 spin_lock(&transaction->t_journal->j_list_lock);
2167 __jbd2_journal_file_buffer(jh, transaction, jlist);
2168 spin_unlock(&transaction->t_journal->j_list_lock);
2169 jbd_unlock_bh_state(jh2bh(jh));
2173 * Remove a buffer from its current buffer list in preparation for
2174 * dropping it from its current transaction entirely. If the buffer has
2175 * already started to be used by a subsequent transaction, refile the
2176 * buffer on that transaction's metadata list.
2178 * Called under j_list_lock
2179 * Called under jbd_lock_bh_state(jh2bh(jh))
2181 * jh and bh may be already free when this function returns
2183 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2185 int was_dirty, jlist;
2186 struct buffer_head *bh = jh2bh(jh);
2188 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2189 if (jh->b_transaction)
2190 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2192 /* If the buffer is now unused, just drop it. */
2193 if (jh->b_next_transaction == NULL) {
2194 __jbd2_journal_unfile_buffer(jh);
2199 * It has been modified by a later transaction: add it to the new
2200 * transaction's metadata list.
2203 was_dirty = test_clear_buffer_jbddirty(bh);
2204 __jbd2_journal_temp_unlink_buffer(jh);
2206 * We set b_transaction here because b_next_transaction will inherit
2207 * our jh reference and thus __jbd2_journal_file_buffer() must not
2210 jh->b_transaction = jh->b_next_transaction;
2211 jh->b_next_transaction = NULL;
2212 if (buffer_freed(bh))
2214 else if (jh->b_modified)
2215 jlist = BJ_Metadata;
2217 jlist = BJ_Reserved;
2218 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2219 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2222 set_buffer_jbddirty(bh);
2226 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2227 * bh reference so that we can safely unlock bh.
2229 * The jh and bh may be freed by this call.
2231 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2233 struct buffer_head *bh = jh2bh(jh);
2235 /* Get reference so that buffer cannot be freed before we unlock it */
2237 jbd_lock_bh_state(bh);
2238 spin_lock(&journal->j_list_lock);
2239 __jbd2_journal_refile_buffer(jh);
2240 jbd_unlock_bh_state(bh);
2241 spin_unlock(&journal->j_list_lock);
2246 * File inode in the inode list of the handle's transaction
2248 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2250 transaction_t *transaction = handle->h_transaction;
2251 journal_t *journal = transaction->t_journal;
2253 if (is_handle_aborted(handle))
2256 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2257 transaction->t_tid);
2260 * First check whether inode isn't already on the transaction's
2261 * lists without taking the lock. Note that this check is safe
2262 * without the lock as we cannot race with somebody removing inode
2263 * from the transaction. The reason is that we remove inode from the
2264 * transaction only in journal_release_jbd_inode() and when we commit
2265 * the transaction. We are guarded from the first case by holding
2266 * a reference to the inode. We are safe against the second case
2267 * because if jinode->i_transaction == transaction, commit code
2268 * cannot touch the transaction because we hold reference to it,
2269 * and if jinode->i_next_transaction == transaction, commit code
2270 * will only file the inode where we want it.
2272 if (jinode->i_transaction == transaction ||
2273 jinode->i_next_transaction == transaction)
2276 spin_lock(&journal->j_list_lock);
2278 if (jinode->i_transaction == transaction ||
2279 jinode->i_next_transaction == transaction)
2283 * We only ever set this variable to 1 so the test is safe. Since
2284 * t_need_data_flush is likely to be set, we do the test to save some
2285 * cacheline bouncing
2287 if (!transaction->t_need_data_flush)
2288 transaction->t_need_data_flush = 1;
2289 /* On some different transaction's list - should be
2290 * the committing one */
2291 if (jinode->i_transaction) {
2292 J_ASSERT(jinode->i_next_transaction == NULL);
2293 J_ASSERT(jinode->i_transaction ==
2294 journal->j_committing_transaction);
2295 jinode->i_next_transaction = transaction;
2298 /* Not on any transaction list... */
2299 J_ASSERT(!jinode->i_next_transaction);
2300 jinode->i_transaction = transaction;
2301 list_add(&jinode->i_list, &transaction->t_inode_list);
2303 spin_unlock(&journal->j_list_lock);
2309 * File truncate and transaction commit interact with each other in a
2310 * non-trivial way. If a transaction writing data block A is
2311 * committing, we cannot discard the data by truncate until we have
2312 * written them. Otherwise if we crashed after the transaction with
2313 * write has committed but before the transaction with truncate has
2314 * committed, we could see stale data in block A. This function is a
2315 * helper to solve this problem. It starts writeout of the truncated
2316 * part in case it is in the committing transaction.
2318 * Filesystem code must call this function when inode is journaled in
2319 * ordered mode before truncation happens and after the inode has been
2320 * placed on orphan list with the new inode size. The second condition
2321 * avoids the race that someone writes new data and we start
2322 * committing the transaction after this function has been called but
2323 * before a transaction for truncate is started (and furthermore it
2324 * allows us to optimize the case where the addition to orphan list
2325 * happens in the same transaction as write --- we don't have to write
2326 * any data in such case).
2328 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2329 struct jbd2_inode *jinode,
2332 transaction_t *inode_trans, *commit_trans;
2335 /* This is a quick check to avoid locking if not necessary */
2336 if (!jinode->i_transaction)
2338 /* Locks are here just to force reading of recent values, it is
2339 * enough that the transaction was not committing before we started
2340 * a transaction adding the inode to orphan list */
2341 read_lock(&journal->j_state_lock);
2342 commit_trans = journal->j_committing_transaction;
2343 read_unlock(&journal->j_state_lock);
2344 spin_lock(&journal->j_list_lock);
2345 inode_trans = jinode->i_transaction;
2346 spin_unlock(&journal->j_list_lock);
2347 if (inode_trans == commit_trans) {
2348 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2349 new_size, LLONG_MAX);
2351 jbd2_journal_abort(journal, ret);