Merge branch 'akpm'
[linux-2.6/next.git] / fs / jbd2 / transaction.c
blob8c640f0c97bbfbbdcc3df2da94654972ab164f1e
1 /*
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
13 * journaling system.
15 * This file manages transactions (compound commits managed by the
16 * journaling code) and handles (individual atomic operations by the
17 * filesystem).
20 #include <linux/time.h>
21 #include <linux/fs.h>
22 #include <linux/jbd2.h>
23 #include <linux/errno.h>
24 #include <linux/slab.h>
25 #include <linux/timer.h>
26 #include <linux/mm.h>
27 #include <linux/highmem.h>
28 #include <linux/hrtimer.h>
29 #include <linux/backing-dev.h>
30 #include <linux/module.h>
32 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
33 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
36 * jbd2_get_transaction: obtain a new transaction_t object.
38 * Simply allocate and initialise a new transaction. Create it in
39 * RUNNING state and add it to the current journal (which should not
40 * have an existing running transaction: we only make a new transaction
41 * once we have started to commit the old one).
43 * Preconditions:
44 * The journal MUST be locked. We don't perform atomic mallocs on the
45 * new transaction and we can't block without protecting against other
46 * processes trying to touch the journal while it is in transition.
50 static transaction_t *
51 jbd2_get_transaction(journal_t *journal, transaction_t *transaction)
53 transaction->t_journal = journal;
54 transaction->t_state = T_RUNNING;
55 transaction->t_start_time = ktime_get();
56 transaction->t_tid = journal->j_transaction_sequence++;
57 transaction->t_expires = jiffies + journal->j_commit_interval;
58 spin_lock_init(&transaction->t_handle_lock);
59 atomic_set(&transaction->t_updates, 0);
60 atomic_set(&transaction->t_outstanding_credits, 0);
61 atomic_set(&transaction->t_handle_count, 0);
62 INIT_LIST_HEAD(&transaction->t_inode_list);
63 INIT_LIST_HEAD(&transaction->t_private_list);
65 /* Set up the commit timer for the new transaction. */
66 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
67 add_timer(&journal->j_commit_timer);
69 J_ASSERT(journal->j_running_transaction == NULL);
70 journal->j_running_transaction = transaction;
71 transaction->t_max_wait = 0;
72 transaction->t_start = jiffies;
74 return transaction;
78 * Handle management.
80 * A handle_t is an object which represents a single atomic update to a
81 * filesystem, and which tracks all of the modifications which form part
82 * of that one update.
86 * Update transaction's maximum wait time, if debugging is enabled.
88 * In order for t_max_wait to be reliable, it must be protected by a
89 * lock. But doing so will mean that start_this_handle() can not be
90 * run in parallel on SMP systems, which limits our scalability. So
91 * unless debugging is enabled, we no longer update t_max_wait, which
92 * means that maximum wait time reported by the jbd2_run_stats
93 * tracepoint will always be zero.
95 static inline void update_t_max_wait(transaction_t *transaction,
96 unsigned long ts)
98 #ifdef CONFIG_JBD2_DEBUG
99 if (jbd2_journal_enable_debug &&
100 time_after(transaction->t_start, ts)) {
101 ts = jbd2_time_diff(ts, transaction->t_start);
102 spin_lock(&transaction->t_handle_lock);
103 if (ts > transaction->t_max_wait)
104 transaction->t_max_wait = ts;
105 spin_unlock(&transaction->t_handle_lock);
107 #endif
111 * start_this_handle: Given a handle, deal with any locking or stalling
112 * needed to make sure that there is enough journal space for the handle
113 * to begin. Attach the handle to a transaction and set up the
114 * transaction's buffer credits.
117 static int start_this_handle(journal_t *journal, handle_t *handle,
118 int gfp_mask)
120 transaction_t *transaction, *new_transaction = NULL;
121 tid_t tid;
122 int needed, need_to_start;
123 int nblocks = handle->h_buffer_credits;
124 unsigned long ts = jiffies;
126 if (nblocks > journal->j_max_transaction_buffers) {
127 printk(KERN_ERR "JBD: %s wants too many credits (%d > %d)\n",
128 current->comm, nblocks,
129 journal->j_max_transaction_buffers);
130 return -ENOSPC;
133 alloc_transaction:
134 if (!journal->j_running_transaction) {
135 new_transaction = kzalloc(sizeof(*new_transaction), gfp_mask);
136 if (!new_transaction) {
138 * If __GFP_FS is not present, then we may be
139 * being called from inside the fs writeback
140 * layer, so we MUST NOT fail. Since
141 * __GFP_NOFAIL is going away, we will arrange
142 * to retry the allocation ourselves.
144 if ((gfp_mask & __GFP_FS) == 0) {
145 congestion_wait(BLK_RW_ASYNC, HZ/50);
146 goto alloc_transaction;
148 return -ENOMEM;
152 jbd_debug(3, "New handle %p going live.\n", handle);
155 * We need to hold j_state_lock until t_updates has been incremented,
156 * for proper journal barrier handling
158 repeat:
159 read_lock(&journal->j_state_lock);
160 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
161 if (is_journal_aborted(journal) ||
162 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
163 read_unlock(&journal->j_state_lock);
164 kfree(new_transaction);
165 return -EROFS;
168 /* Wait on the journal's transaction barrier if necessary */
169 if (journal->j_barrier_count) {
170 read_unlock(&journal->j_state_lock);
171 wait_event(journal->j_wait_transaction_locked,
172 journal->j_barrier_count == 0);
173 goto repeat;
176 if (!journal->j_running_transaction) {
177 read_unlock(&journal->j_state_lock);
178 if (!new_transaction)
179 goto alloc_transaction;
180 write_lock(&journal->j_state_lock);
181 if (!journal->j_running_transaction) {
182 jbd2_get_transaction(journal, new_transaction);
183 new_transaction = NULL;
185 write_unlock(&journal->j_state_lock);
186 goto repeat;
189 transaction = journal->j_running_transaction;
192 * If the current transaction is locked down for commit, wait for the
193 * lock to be released.
195 if (transaction->t_state == T_LOCKED) {
196 DEFINE_WAIT(wait);
198 prepare_to_wait(&journal->j_wait_transaction_locked,
199 &wait, TASK_UNINTERRUPTIBLE);
200 read_unlock(&journal->j_state_lock);
201 schedule();
202 finish_wait(&journal->j_wait_transaction_locked, &wait);
203 goto repeat;
207 * If there is not enough space left in the log to write all potential
208 * buffers requested by this operation, we need to stall pending a log
209 * checkpoint to free some more log space.
211 needed = atomic_add_return(nblocks,
212 &transaction->t_outstanding_credits);
214 if (needed > journal->j_max_transaction_buffers) {
216 * If the current transaction is already too large, then start
217 * to commit it: we can then go back and attach this handle to
218 * a new transaction.
220 DEFINE_WAIT(wait);
222 jbd_debug(2, "Handle %p starting new commit...\n", handle);
223 atomic_sub(nblocks, &transaction->t_outstanding_credits);
224 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
225 TASK_UNINTERRUPTIBLE);
226 tid = transaction->t_tid;
227 need_to_start = !tid_geq(journal->j_commit_request, tid);
228 read_unlock(&journal->j_state_lock);
229 if (need_to_start)
230 jbd2_log_start_commit(journal, tid);
231 schedule();
232 finish_wait(&journal->j_wait_transaction_locked, &wait);
233 goto repeat;
237 * The commit code assumes that it can get enough log space
238 * without forcing a checkpoint. This is *critical* for
239 * correctness: a checkpoint of a buffer which is also
240 * associated with a committing transaction creates a deadlock,
241 * so commit simply cannot force through checkpoints.
243 * We must therefore ensure the necessary space in the journal
244 * *before* starting to dirty potentially checkpointed buffers
245 * in the new transaction.
247 * The worst part is, any transaction currently committing can
248 * reduce the free space arbitrarily. Be careful to account for
249 * those buffers when checkpointing.
253 * @@@ AKPM: This seems rather over-defensive. We're giving commit
254 * a _lot_ of headroom: 1/4 of the journal plus the size of
255 * the committing transaction. Really, we only need to give it
256 * committing_transaction->t_outstanding_credits plus "enough" for
257 * the log control blocks.
258 * Also, this test is inconsistent with the matching one in
259 * jbd2_journal_extend().
261 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal)) {
262 jbd_debug(2, "Handle %p waiting for checkpoint...\n", handle);
263 atomic_sub(nblocks, &transaction->t_outstanding_credits);
264 read_unlock(&journal->j_state_lock);
265 write_lock(&journal->j_state_lock);
266 if (__jbd2_log_space_left(journal) < jbd_space_needed(journal))
267 __jbd2_log_wait_for_space(journal);
268 write_unlock(&journal->j_state_lock);
269 goto repeat;
272 /* OK, account for the buffers that this operation expects to
273 * use and add the handle to the running transaction.
275 update_t_max_wait(transaction, ts);
276 handle->h_transaction = transaction;
277 atomic_inc(&transaction->t_updates);
278 atomic_inc(&transaction->t_handle_count);
279 jbd_debug(4, "Handle %p given %d credits (total %d, free %d)\n",
280 handle, nblocks,
281 atomic_read(&transaction->t_outstanding_credits),
282 __jbd2_log_space_left(journal));
283 read_unlock(&journal->j_state_lock);
285 lock_map_acquire(&handle->h_lockdep_map);
286 kfree(new_transaction);
287 return 0;
290 static struct lock_class_key jbd2_handle_key;
292 /* Allocate a new handle. This should probably be in a slab... */
293 static handle_t *new_handle(int nblocks)
295 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
296 if (!handle)
297 return NULL;
298 memset(handle, 0, sizeof(*handle));
299 handle->h_buffer_credits = nblocks;
300 handle->h_ref = 1;
302 lockdep_init_map(&handle->h_lockdep_map, "jbd2_handle",
303 &jbd2_handle_key, 0);
305 return handle;
309 * handle_t *jbd2_journal_start() - Obtain a new handle.
310 * @journal: Journal to start transaction on.
311 * @nblocks: number of block buffer we might modify
313 * We make sure that the transaction can guarantee at least nblocks of
314 * modified buffers in the log. We block until the log can guarantee
315 * that much space.
317 * This function is visible to journal users (like ext3fs), so is not
318 * called with the journal already locked.
320 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
321 * on failure.
323 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int gfp_mask)
325 handle_t *handle = journal_current_handle();
326 int err;
328 if (!journal)
329 return ERR_PTR(-EROFS);
331 if (handle) {
332 J_ASSERT(handle->h_transaction->t_journal == journal);
333 handle->h_ref++;
334 return handle;
337 handle = new_handle(nblocks);
338 if (!handle)
339 return ERR_PTR(-ENOMEM);
341 current->journal_info = handle;
343 err = start_this_handle(journal, handle, gfp_mask);
344 if (err < 0) {
345 jbd2_free_handle(handle);
346 current->journal_info = NULL;
347 handle = ERR_PTR(err);
349 return handle;
351 EXPORT_SYMBOL(jbd2__journal_start);
354 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
356 return jbd2__journal_start(journal, nblocks, GFP_NOFS);
358 EXPORT_SYMBOL(jbd2_journal_start);
362 * int jbd2_journal_extend() - extend buffer credits.
363 * @handle: handle to 'extend'
364 * @nblocks: nr blocks to try to extend by.
366 * Some transactions, such as large extends and truncates, can be done
367 * atomically all at once or in several stages. The operation requests
368 * a credit for a number of buffer modications in advance, but can
369 * extend its credit if it needs more.
371 * jbd2_journal_extend tries to give the running handle more buffer credits.
372 * It does not guarantee that allocation - this is a best-effort only.
373 * The calling process MUST be able to deal cleanly with a failure to
374 * extend here.
376 * Return 0 on success, non-zero on failure.
378 * return code < 0 implies an error
379 * return code > 0 implies normal transaction-full status.
381 int jbd2_journal_extend(handle_t *handle, int nblocks)
383 transaction_t *transaction = handle->h_transaction;
384 journal_t *journal = transaction->t_journal;
385 int result;
386 int wanted;
388 result = -EIO;
389 if (is_handle_aborted(handle))
390 goto out;
392 result = 1;
394 read_lock(&journal->j_state_lock);
396 /* Don't extend a locked-down transaction! */
397 if (handle->h_transaction->t_state != T_RUNNING) {
398 jbd_debug(3, "denied handle %p %d blocks: "
399 "transaction not running\n", handle, nblocks);
400 goto error_out;
403 spin_lock(&transaction->t_handle_lock);
404 wanted = atomic_read(&transaction->t_outstanding_credits) + nblocks;
406 if (wanted > journal->j_max_transaction_buffers) {
407 jbd_debug(3, "denied handle %p %d blocks: "
408 "transaction too large\n", handle, nblocks);
409 goto unlock;
412 if (wanted > __jbd2_log_space_left(journal)) {
413 jbd_debug(3, "denied handle %p %d blocks: "
414 "insufficient log space\n", handle, nblocks);
415 goto unlock;
418 handle->h_buffer_credits += nblocks;
419 atomic_add(nblocks, &transaction->t_outstanding_credits);
420 result = 0;
422 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
423 unlock:
424 spin_unlock(&transaction->t_handle_lock);
425 error_out:
426 read_unlock(&journal->j_state_lock);
427 out:
428 return result;
433 * int jbd2_journal_restart() - restart a handle .
434 * @handle: handle to restart
435 * @nblocks: nr credits requested
437 * Restart a handle for a multi-transaction filesystem
438 * operation.
440 * If the jbd2_journal_extend() call above fails to grant new buffer credits
441 * to a running handle, a call to jbd2_journal_restart will commit the
442 * handle's transaction so far and reattach the handle to a new
443 * transaction capabable of guaranteeing the requested number of
444 * credits.
446 int jbd2__journal_restart(handle_t *handle, int nblocks, int gfp_mask)
448 transaction_t *transaction = handle->h_transaction;
449 journal_t *journal = transaction->t_journal;
450 tid_t tid;
451 int need_to_start, ret;
453 /* If we've had an abort of any type, don't even think about
454 * actually doing the restart! */
455 if (is_handle_aborted(handle))
456 return 0;
459 * First unlink the handle from its current transaction, and start the
460 * commit on that.
462 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
463 J_ASSERT(journal_current_handle() == handle);
465 read_lock(&journal->j_state_lock);
466 spin_lock(&transaction->t_handle_lock);
467 atomic_sub(handle->h_buffer_credits,
468 &transaction->t_outstanding_credits);
469 if (atomic_dec_and_test(&transaction->t_updates))
470 wake_up(&journal->j_wait_updates);
471 spin_unlock(&transaction->t_handle_lock);
473 jbd_debug(2, "restarting handle %p\n", handle);
474 tid = transaction->t_tid;
475 need_to_start = !tid_geq(journal->j_commit_request, tid);
476 read_unlock(&journal->j_state_lock);
477 if (need_to_start)
478 jbd2_log_start_commit(journal, tid);
480 lock_map_release(&handle->h_lockdep_map);
481 handle->h_buffer_credits = nblocks;
482 ret = start_this_handle(journal, handle, gfp_mask);
483 return ret;
485 EXPORT_SYMBOL(jbd2__journal_restart);
488 int jbd2_journal_restart(handle_t *handle, int nblocks)
490 return jbd2__journal_restart(handle, nblocks, GFP_NOFS);
492 EXPORT_SYMBOL(jbd2_journal_restart);
495 * void jbd2_journal_lock_updates () - establish a transaction barrier.
496 * @journal: Journal to establish a barrier on.
498 * This locks out any further updates from being started, and blocks
499 * until all existing updates have completed, returning only once the
500 * journal is in a quiescent state with no updates running.
502 * The journal lock should not be held on entry.
504 void jbd2_journal_lock_updates(journal_t *journal)
506 DEFINE_WAIT(wait);
508 write_lock(&journal->j_state_lock);
509 ++journal->j_barrier_count;
511 /* Wait until there are no running updates */
512 while (1) {
513 transaction_t *transaction = journal->j_running_transaction;
515 if (!transaction)
516 break;
518 spin_lock(&transaction->t_handle_lock);
519 if (!atomic_read(&transaction->t_updates)) {
520 spin_unlock(&transaction->t_handle_lock);
521 break;
523 prepare_to_wait(&journal->j_wait_updates, &wait,
524 TASK_UNINTERRUPTIBLE);
525 spin_unlock(&transaction->t_handle_lock);
526 write_unlock(&journal->j_state_lock);
527 schedule();
528 finish_wait(&journal->j_wait_updates, &wait);
529 write_lock(&journal->j_state_lock);
531 write_unlock(&journal->j_state_lock);
534 * We have now established a barrier against other normal updates, but
535 * we also need to barrier against other jbd2_journal_lock_updates() calls
536 * to make sure that we serialise special journal-locked operations
537 * too.
539 mutex_lock(&journal->j_barrier);
543 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
544 * @journal: Journal to release the barrier on.
546 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
548 * Should be called without the journal lock held.
550 void jbd2_journal_unlock_updates (journal_t *journal)
552 J_ASSERT(journal->j_barrier_count != 0);
554 mutex_unlock(&journal->j_barrier);
555 write_lock(&journal->j_state_lock);
556 --journal->j_barrier_count;
557 write_unlock(&journal->j_state_lock);
558 wake_up(&journal->j_wait_transaction_locked);
561 static void warn_dirty_buffer(struct buffer_head *bh)
563 char b[BDEVNAME_SIZE];
565 printk(KERN_WARNING
566 "JBD: Spotted dirty metadata buffer (dev = %s, blocknr = %llu). "
567 "There's a risk of filesystem corruption in case of system "
568 "crash.\n",
569 bdevname(bh->b_bdev, b), (unsigned long long)bh->b_blocknr);
573 * If the buffer is already part of the current transaction, then there
574 * is nothing we need to do. If it is already part of a prior
575 * transaction which we are still committing to disk, then we need to
576 * make sure that we do not overwrite the old copy: we do copy-out to
577 * preserve the copy going to disk. We also account the buffer against
578 * the handle's metadata buffer credits (unless the buffer is already
579 * part of the transaction, that is).
582 static int
583 do_get_write_access(handle_t *handle, struct journal_head *jh,
584 int force_copy)
586 struct buffer_head *bh;
587 transaction_t *transaction;
588 journal_t *journal;
589 int error;
590 char *frozen_buffer = NULL;
591 int need_copy = 0;
593 if (is_handle_aborted(handle))
594 return -EROFS;
596 transaction = handle->h_transaction;
597 journal = transaction->t_journal;
599 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
601 JBUFFER_TRACE(jh, "entry");
602 repeat:
603 bh = jh2bh(jh);
605 /* @@@ Need to check for errors here at some point. */
607 lock_buffer(bh);
608 jbd_lock_bh_state(bh);
610 /* We now hold the buffer lock so it is safe to query the buffer
611 * state. Is the buffer dirty?
613 * If so, there are two possibilities. The buffer may be
614 * non-journaled, and undergoing a quite legitimate writeback.
615 * Otherwise, it is journaled, and we don't expect dirty buffers
616 * in that state (the buffers should be marked JBD_Dirty
617 * instead.) So either the IO is being done under our own
618 * control and this is a bug, or it's a third party IO such as
619 * dump(8) (which may leave the buffer scheduled for read ---
620 * ie. locked but not dirty) or tune2fs (which may actually have
621 * the buffer dirtied, ugh.) */
623 if (buffer_dirty(bh)) {
625 * First question: is this buffer already part of the current
626 * transaction or the existing committing transaction?
628 if (jh->b_transaction) {
629 J_ASSERT_JH(jh,
630 jh->b_transaction == transaction ||
631 jh->b_transaction ==
632 journal->j_committing_transaction);
633 if (jh->b_next_transaction)
634 J_ASSERT_JH(jh, jh->b_next_transaction ==
635 transaction);
636 warn_dirty_buffer(bh);
639 * In any case we need to clean the dirty flag and we must
640 * do it under the buffer lock to be sure we don't race
641 * with running write-out.
643 JBUFFER_TRACE(jh, "Journalling dirty buffer");
644 clear_buffer_dirty(bh);
645 set_buffer_jbddirty(bh);
648 unlock_buffer(bh);
650 error = -EROFS;
651 if (is_handle_aborted(handle)) {
652 jbd_unlock_bh_state(bh);
653 goto out;
655 error = 0;
658 * The buffer is already part of this transaction if b_transaction or
659 * b_next_transaction points to it
661 if (jh->b_transaction == transaction ||
662 jh->b_next_transaction == transaction)
663 goto done;
666 * this is the first time this transaction is touching this buffer,
667 * reset the modified flag
669 jh->b_modified = 0;
672 * If there is already a copy-out version of this buffer, then we don't
673 * need to make another one
675 if (jh->b_frozen_data) {
676 JBUFFER_TRACE(jh, "has frozen data");
677 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
678 jh->b_next_transaction = transaction;
679 goto done;
682 /* Is there data here we need to preserve? */
684 if (jh->b_transaction && jh->b_transaction != transaction) {
685 JBUFFER_TRACE(jh, "owned by older transaction");
686 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
687 J_ASSERT_JH(jh, jh->b_transaction ==
688 journal->j_committing_transaction);
690 /* There is one case we have to be very careful about.
691 * If the committing transaction is currently writing
692 * this buffer out to disk and has NOT made a copy-out,
693 * then we cannot modify the buffer contents at all
694 * right now. The essence of copy-out is that it is the
695 * extra copy, not the primary copy, which gets
696 * journaled. If the primary copy is already going to
697 * disk then we cannot do copy-out here. */
699 if (jh->b_jlist == BJ_Shadow) {
700 DEFINE_WAIT_BIT(wait, &bh->b_state, BH_Unshadow);
701 wait_queue_head_t *wqh;
703 wqh = bit_waitqueue(&bh->b_state, BH_Unshadow);
705 JBUFFER_TRACE(jh, "on shadow: sleep");
706 jbd_unlock_bh_state(bh);
707 /* commit wakes up all shadow buffers after IO */
708 for ( ; ; ) {
709 prepare_to_wait(wqh, &wait.wait,
710 TASK_UNINTERRUPTIBLE);
711 if (jh->b_jlist != BJ_Shadow)
712 break;
713 schedule();
715 finish_wait(wqh, &wait.wait);
716 goto repeat;
719 /* Only do the copy if the currently-owning transaction
720 * still needs it. If it is on the Forget list, the
721 * committing transaction is past that stage. The
722 * buffer had better remain locked during the kmalloc,
723 * but that should be true --- we hold the journal lock
724 * still and the buffer is already on the BUF_JOURNAL
725 * list so won't be flushed.
727 * Subtle point, though: if this is a get_undo_access,
728 * then we will be relying on the frozen_data to contain
729 * the new value of the committed_data record after the
730 * transaction, so we HAVE to force the frozen_data copy
731 * in that case. */
733 if (jh->b_jlist != BJ_Forget || force_copy) {
734 JBUFFER_TRACE(jh, "generate frozen data");
735 if (!frozen_buffer) {
736 JBUFFER_TRACE(jh, "allocate memory for buffer");
737 jbd_unlock_bh_state(bh);
738 frozen_buffer =
739 jbd2_alloc(jh2bh(jh)->b_size,
740 GFP_NOFS);
741 if (!frozen_buffer) {
742 printk(KERN_EMERG
743 "%s: OOM for frozen_buffer\n",
744 __func__);
745 JBUFFER_TRACE(jh, "oom!");
746 error = -ENOMEM;
747 jbd_lock_bh_state(bh);
748 goto done;
750 goto repeat;
752 jh->b_frozen_data = frozen_buffer;
753 frozen_buffer = NULL;
754 need_copy = 1;
756 jh->b_next_transaction = transaction;
761 * Finally, if the buffer is not journaled right now, we need to make
762 * sure it doesn't get written to disk before the caller actually
763 * commits the new data
765 if (!jh->b_transaction) {
766 JBUFFER_TRACE(jh, "no transaction");
767 J_ASSERT_JH(jh, !jh->b_next_transaction);
768 JBUFFER_TRACE(jh, "file as BJ_Reserved");
769 spin_lock(&journal->j_list_lock);
770 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
771 spin_unlock(&journal->j_list_lock);
774 done:
775 if (need_copy) {
776 struct page *page;
777 int offset;
778 char *source;
780 J_EXPECT_JH(jh, buffer_uptodate(jh2bh(jh)),
781 "Possible IO failure.\n");
782 page = jh2bh(jh)->b_page;
783 offset = offset_in_page(jh2bh(jh)->b_data);
784 source = kmap_atomic(page, KM_USER0);
785 /* Fire data frozen trigger just before we copy the data */
786 jbd2_buffer_frozen_trigger(jh, source + offset,
787 jh->b_triggers);
788 memcpy(jh->b_frozen_data, source+offset, jh2bh(jh)->b_size);
789 kunmap_atomic(source, KM_USER0);
792 * Now that the frozen data is saved off, we need to store
793 * any matching triggers.
795 jh->b_frozen_triggers = jh->b_triggers;
797 jbd_unlock_bh_state(bh);
800 * If we are about to journal a buffer, then any revoke pending on it is
801 * no longer valid
803 jbd2_journal_cancel_revoke(handle, jh);
805 out:
806 if (unlikely(frozen_buffer)) /* It's usually NULL */
807 jbd2_free(frozen_buffer, bh->b_size);
809 JBUFFER_TRACE(jh, "exit");
810 return error;
814 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
815 * @handle: transaction to add buffer modifications to
816 * @bh: bh to be used for metadata writes
818 * Returns an error code or 0 on success.
820 * In full data journalling mode the buffer may be of type BJ_AsyncData,
821 * because we're write()ing a buffer which is also part of a shared mapping.
824 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
826 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
827 int rc;
829 /* We do not want to get caught playing with fields which the
830 * log thread also manipulates. Make sure that the buffer
831 * completes any outstanding IO before proceeding. */
832 rc = do_get_write_access(handle, jh, 0);
833 jbd2_journal_put_journal_head(jh);
834 return rc;
839 * When the user wants to journal a newly created buffer_head
840 * (ie. getblk() returned a new buffer and we are going to populate it
841 * manually rather than reading off disk), then we need to keep the
842 * buffer_head locked until it has been completely filled with new
843 * data. In this case, we should be able to make the assertion that
844 * the bh is not already part of an existing transaction.
846 * The buffer should already be locked by the caller by this point.
847 * There is no lock ranking violation: it was a newly created,
848 * unlocked buffer beforehand. */
851 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
852 * @handle: transaction to new buffer to
853 * @bh: new buffer.
855 * Call this if you create a new bh.
857 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
859 transaction_t *transaction = handle->h_transaction;
860 journal_t *journal = transaction->t_journal;
861 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
862 int err;
864 jbd_debug(5, "journal_head %p\n", jh);
865 err = -EROFS;
866 if (is_handle_aborted(handle))
867 goto out;
868 err = 0;
870 JBUFFER_TRACE(jh, "entry");
872 * The buffer may already belong to this transaction due to pre-zeroing
873 * in the filesystem's new_block code. It may also be on the previous,
874 * committing transaction's lists, but it HAS to be in Forget state in
875 * that case: the transaction must have deleted the buffer for it to be
876 * reused here.
878 jbd_lock_bh_state(bh);
879 spin_lock(&journal->j_list_lock);
880 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
881 jh->b_transaction == NULL ||
882 (jh->b_transaction == journal->j_committing_transaction &&
883 jh->b_jlist == BJ_Forget)));
885 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
886 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
888 if (jh->b_transaction == NULL) {
890 * Previous jbd2_journal_forget() could have left the buffer
891 * with jbddirty bit set because it was being committed. When
892 * the commit finished, we've filed the buffer for
893 * checkpointing and marked it dirty. Now we are reallocating
894 * the buffer so the transaction freeing it must have
895 * committed and so it's safe to clear the dirty bit.
897 clear_buffer_dirty(jh2bh(jh));
898 /* first access by this transaction */
899 jh->b_modified = 0;
901 JBUFFER_TRACE(jh, "file as BJ_Reserved");
902 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
903 } else if (jh->b_transaction == journal->j_committing_transaction) {
904 /* first access by this transaction */
905 jh->b_modified = 0;
907 JBUFFER_TRACE(jh, "set next transaction");
908 jh->b_next_transaction = transaction;
910 spin_unlock(&journal->j_list_lock);
911 jbd_unlock_bh_state(bh);
914 * akpm: I added this. ext3_alloc_branch can pick up new indirect
915 * blocks which contain freed but then revoked metadata. We need
916 * to cancel the revoke in case we end up freeing it yet again
917 * and the reallocating as data - this would cause a second revoke,
918 * which hits an assertion error.
920 JBUFFER_TRACE(jh, "cancelling revoke");
921 jbd2_journal_cancel_revoke(handle, jh);
922 out:
923 jbd2_journal_put_journal_head(jh);
924 return err;
928 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
929 * non-rewindable consequences
930 * @handle: transaction
931 * @bh: buffer to undo
933 * Sometimes there is a need to distinguish between metadata which has
934 * been committed to disk and that which has not. The ext3fs code uses
935 * this for freeing and allocating space, we have to make sure that we
936 * do not reuse freed space until the deallocation has been committed,
937 * since if we overwrote that space we would make the delete
938 * un-rewindable in case of a crash.
940 * To deal with that, jbd2_journal_get_undo_access requests write access to a
941 * buffer for parts of non-rewindable operations such as delete
942 * operations on the bitmaps. The journaling code must keep a copy of
943 * the buffer's contents prior to the undo_access call until such time
944 * as we know that the buffer has definitely been committed to disk.
946 * We never need to know which transaction the committed data is part
947 * of, buffers touched here are guaranteed to be dirtied later and so
948 * will be committed to a new transaction in due course, at which point
949 * we can discard the old committed data pointer.
951 * Returns error number or 0 on success.
953 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
955 int err;
956 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
957 char *committed_data = NULL;
959 JBUFFER_TRACE(jh, "entry");
962 * Do this first --- it can drop the journal lock, so we want to
963 * make sure that obtaining the committed_data is done
964 * atomically wrt. completion of any outstanding commits.
966 err = do_get_write_access(handle, jh, 1);
967 if (err)
968 goto out;
970 repeat:
971 if (!jh->b_committed_data) {
972 committed_data = jbd2_alloc(jh2bh(jh)->b_size, GFP_NOFS);
973 if (!committed_data) {
974 printk(KERN_EMERG "%s: No memory for committed data\n",
975 __func__);
976 err = -ENOMEM;
977 goto out;
981 jbd_lock_bh_state(bh);
982 if (!jh->b_committed_data) {
983 /* Copy out the current buffer contents into the
984 * preserved, committed copy. */
985 JBUFFER_TRACE(jh, "generate b_committed data");
986 if (!committed_data) {
987 jbd_unlock_bh_state(bh);
988 goto repeat;
991 jh->b_committed_data = committed_data;
992 committed_data = NULL;
993 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
995 jbd_unlock_bh_state(bh);
996 out:
997 jbd2_journal_put_journal_head(jh);
998 if (unlikely(committed_data))
999 jbd2_free(committed_data, bh->b_size);
1000 return err;
1004 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1005 * @bh: buffer to trigger on
1006 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1008 * Set any triggers on this journal_head. This is always safe, because
1009 * triggers for a committing buffer will be saved off, and triggers for
1010 * a running transaction will match the buffer in that transaction.
1012 * Call with NULL to clear the triggers.
1014 void jbd2_journal_set_triggers(struct buffer_head *bh,
1015 struct jbd2_buffer_trigger_type *type)
1017 struct journal_head *jh = bh2jh(bh);
1019 jh->b_triggers = type;
1022 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1023 struct jbd2_buffer_trigger_type *triggers)
1025 struct buffer_head *bh = jh2bh(jh);
1027 if (!triggers || !triggers->t_frozen)
1028 return;
1030 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1033 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1034 struct jbd2_buffer_trigger_type *triggers)
1036 if (!triggers || !triggers->t_abort)
1037 return;
1039 triggers->t_abort(triggers, jh2bh(jh));
1045 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1046 * @handle: transaction to add buffer to.
1047 * @bh: buffer to mark
1049 * mark dirty metadata which needs to be journaled as part of the current
1050 * transaction.
1052 * The buffer must have previously had jbd2_journal_get_write_access()
1053 * called so that it has a valid journal_head attached to the buffer
1054 * head.
1056 * The buffer is placed on the transaction's metadata list and is marked
1057 * as belonging to the transaction.
1059 * Returns error number or 0 on success.
1061 * Special care needs to be taken if the buffer already belongs to the
1062 * current committing transaction (in which case we should have frozen
1063 * data present for that commit). In that case, we don't relink the
1064 * buffer: that only gets done when the old transaction finally
1065 * completes its commit.
1067 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1069 transaction_t *transaction = handle->h_transaction;
1070 journal_t *journal = transaction->t_journal;
1071 struct journal_head *jh = bh2jh(bh);
1072 int ret = 0;
1074 jbd_debug(5, "journal_head %p\n", jh);
1075 JBUFFER_TRACE(jh, "entry");
1076 if (is_handle_aborted(handle))
1077 goto out;
1078 if (!buffer_jbd(bh)) {
1079 ret = -EUCLEAN;
1080 goto out;
1083 jbd_lock_bh_state(bh);
1085 if (jh->b_modified == 0) {
1087 * This buffer's got modified and becoming part
1088 * of the transaction. This needs to be done
1089 * once a transaction -bzzz
1091 jh->b_modified = 1;
1092 J_ASSERT_JH(jh, handle->h_buffer_credits > 0);
1093 handle->h_buffer_credits--;
1097 * fastpath, to avoid expensive locking. If this buffer is already
1098 * on the running transaction's metadata list there is nothing to do.
1099 * Nobody can take it off again because there is a handle open.
1100 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1101 * result in this test being false, so we go in and take the locks.
1103 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1104 JBUFFER_TRACE(jh, "fastpath");
1105 if (unlikely(jh->b_transaction !=
1106 journal->j_running_transaction)) {
1107 printk(KERN_EMERG "JBD: %s: "
1108 "jh->b_transaction (%llu, %p, %u) != "
1109 "journal->j_running_transaction (%p, %u)",
1110 journal->j_devname,
1111 (unsigned long long) bh->b_blocknr,
1112 jh->b_transaction,
1113 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1114 journal->j_running_transaction,
1115 journal->j_running_transaction ?
1116 journal->j_running_transaction->t_tid : 0);
1117 ret = -EINVAL;
1119 goto out_unlock_bh;
1122 set_buffer_jbddirty(bh);
1125 * Metadata already on the current transaction list doesn't
1126 * need to be filed. Metadata on another transaction's list must
1127 * be committing, and will be refiled once the commit completes:
1128 * leave it alone for now.
1130 if (jh->b_transaction != transaction) {
1131 JBUFFER_TRACE(jh, "already on other transaction");
1132 if (unlikely(jh->b_transaction !=
1133 journal->j_committing_transaction)) {
1134 printk(KERN_EMERG "JBD: %s: "
1135 "jh->b_transaction (%llu, %p, %u) != "
1136 "journal->j_committing_transaction (%p, %u)",
1137 journal->j_devname,
1138 (unsigned long long) bh->b_blocknr,
1139 jh->b_transaction,
1140 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1141 journal->j_committing_transaction,
1142 journal->j_committing_transaction ?
1143 journal->j_committing_transaction->t_tid : 0);
1144 ret = -EINVAL;
1146 if (unlikely(jh->b_next_transaction != transaction)) {
1147 printk(KERN_EMERG "JBD: %s: "
1148 "jh->b_next_transaction (%llu, %p, %u) != "
1149 "transaction (%p, %u)",
1150 journal->j_devname,
1151 (unsigned long long) bh->b_blocknr,
1152 jh->b_next_transaction,
1153 jh->b_next_transaction ?
1154 jh->b_next_transaction->t_tid : 0,
1155 transaction,
1156 transaction ? transaction->t_tid : 0);
1157 ret = -EINVAL;
1159 /* And this case is illegal: we can't reuse another
1160 * transaction's data buffer, ever. */
1161 goto out_unlock_bh;
1164 /* That test should have eliminated the following case: */
1165 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1167 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1168 spin_lock(&journal->j_list_lock);
1169 __jbd2_journal_file_buffer(jh, handle->h_transaction, BJ_Metadata);
1170 spin_unlock(&journal->j_list_lock);
1171 out_unlock_bh:
1172 jbd_unlock_bh_state(bh);
1173 out:
1174 JBUFFER_TRACE(jh, "exit");
1175 if (ret)
1176 __WARN(); /* All errors are bugs, so dump the stack */
1177 return ret;
1181 * jbd2_journal_release_buffer: undo a get_write_access without any buffer
1182 * updates, if the update decided in the end that it didn't need access.
1185 void
1186 jbd2_journal_release_buffer(handle_t *handle, struct buffer_head *bh)
1188 BUFFER_TRACE(bh, "entry");
1192 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1193 * @handle: transaction handle
1194 * @bh: bh to 'forget'
1196 * We can only do the bforget if there are no commits pending against the
1197 * buffer. If the buffer is dirty in the current running transaction we
1198 * can safely unlink it.
1200 * bh may not be a journalled buffer at all - it may be a non-JBD
1201 * buffer which came off the hashtable. Check for this.
1203 * Decrements bh->b_count by one.
1205 * Allow this call even if the handle has aborted --- it may be part of
1206 * the caller's cleanup after an abort.
1208 int jbd2_journal_forget (handle_t *handle, struct buffer_head *bh)
1210 transaction_t *transaction = handle->h_transaction;
1211 journal_t *journal = transaction->t_journal;
1212 struct journal_head *jh;
1213 int drop_reserve = 0;
1214 int err = 0;
1215 int was_modified = 0;
1217 BUFFER_TRACE(bh, "entry");
1219 jbd_lock_bh_state(bh);
1220 spin_lock(&journal->j_list_lock);
1222 if (!buffer_jbd(bh))
1223 goto not_jbd;
1224 jh = bh2jh(bh);
1226 /* Critical error: attempting to delete a bitmap buffer, maybe?
1227 * Don't do any jbd operations, and return an error. */
1228 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1229 "inconsistent data on disk")) {
1230 err = -EIO;
1231 goto not_jbd;
1234 /* keep track of wether or not this transaction modified us */
1235 was_modified = jh->b_modified;
1238 * The buffer's going from the transaction, we must drop
1239 * all references -bzzz
1241 jh->b_modified = 0;
1243 if (jh->b_transaction == handle->h_transaction) {
1244 J_ASSERT_JH(jh, !jh->b_frozen_data);
1246 /* If we are forgetting a buffer which is already part
1247 * of this transaction, then we can just drop it from
1248 * the transaction immediately. */
1249 clear_buffer_dirty(bh);
1250 clear_buffer_jbddirty(bh);
1252 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1255 * we only want to drop a reference if this transaction
1256 * modified the buffer
1258 if (was_modified)
1259 drop_reserve = 1;
1262 * We are no longer going to journal this buffer.
1263 * However, the commit of this transaction is still
1264 * important to the buffer: the delete that we are now
1265 * processing might obsolete an old log entry, so by
1266 * committing, we can satisfy the buffer's checkpoint.
1268 * So, if we have a checkpoint on the buffer, we should
1269 * now refile the buffer on our BJ_Forget list so that
1270 * we know to remove the checkpoint after we commit.
1273 if (jh->b_cp_transaction) {
1274 __jbd2_journal_temp_unlink_buffer(jh);
1275 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1276 } else {
1277 __jbd2_journal_unfile_buffer(jh);
1278 if (!buffer_jbd(bh)) {
1279 spin_unlock(&journal->j_list_lock);
1280 jbd_unlock_bh_state(bh);
1281 __bforget(bh);
1282 goto drop;
1285 } else if (jh->b_transaction) {
1286 J_ASSERT_JH(jh, (jh->b_transaction ==
1287 journal->j_committing_transaction));
1288 /* However, if the buffer is still owned by a prior
1289 * (committing) transaction, we can't drop it yet... */
1290 JBUFFER_TRACE(jh, "belongs to older transaction");
1291 /* ... but we CAN drop it from the new transaction if we
1292 * have also modified it since the original commit. */
1294 if (jh->b_next_transaction) {
1295 J_ASSERT(jh->b_next_transaction == transaction);
1296 jh->b_next_transaction = NULL;
1299 * only drop a reference if this transaction modified
1300 * the buffer
1302 if (was_modified)
1303 drop_reserve = 1;
1307 not_jbd:
1308 spin_unlock(&journal->j_list_lock);
1309 jbd_unlock_bh_state(bh);
1310 __brelse(bh);
1311 drop:
1312 if (drop_reserve) {
1313 /* no need to reserve log space for this block -bzzz */
1314 handle->h_buffer_credits++;
1316 return err;
1320 * int jbd2_journal_stop() - complete a transaction
1321 * @handle: tranaction to complete.
1323 * All done for a particular handle.
1325 * There is not much action needed here. We just return any remaining
1326 * buffer credits to the transaction and remove the handle. The only
1327 * complication is that we need to start a commit operation if the
1328 * filesystem is marked for synchronous update.
1330 * jbd2_journal_stop itself will not usually return an error, but it may
1331 * do so in unusual circumstances. In particular, expect it to
1332 * return -EIO if a jbd2_journal_abort has been executed since the
1333 * transaction began.
1335 int jbd2_journal_stop(handle_t *handle)
1337 transaction_t *transaction = handle->h_transaction;
1338 journal_t *journal = transaction->t_journal;
1339 int err, wait_for_commit = 0;
1340 tid_t tid;
1341 pid_t pid;
1343 J_ASSERT(journal_current_handle() == handle);
1345 if (is_handle_aborted(handle))
1346 err = -EIO;
1347 else {
1348 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
1349 err = 0;
1352 if (--handle->h_ref > 0) {
1353 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1354 handle->h_ref);
1355 return err;
1358 jbd_debug(4, "Handle %p going down\n", handle);
1361 * Implement synchronous transaction batching. If the handle
1362 * was synchronous, don't force a commit immediately. Let's
1363 * yield and let another thread piggyback onto this
1364 * transaction. Keep doing that while new threads continue to
1365 * arrive. It doesn't cost much - we're about to run a commit
1366 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1367 * operations by 30x or more...
1369 * We try and optimize the sleep time against what the
1370 * underlying disk can do, instead of having a static sleep
1371 * time. This is useful for the case where our storage is so
1372 * fast that it is more optimal to go ahead and force a flush
1373 * and wait for the transaction to be committed than it is to
1374 * wait for an arbitrary amount of time for new writers to
1375 * join the transaction. We achieve this by measuring how
1376 * long it takes to commit a transaction, and compare it with
1377 * how long this transaction has been running, and if run time
1378 * < commit time then we sleep for the delta and commit. This
1379 * greatly helps super fast disks that would see slowdowns as
1380 * more threads started doing fsyncs.
1382 * But don't do this if this process was the most recent one
1383 * to perform a synchronous write. We do this to detect the
1384 * case where a single process is doing a stream of sync
1385 * writes. No point in waiting for joiners in that case.
1387 pid = current->pid;
1388 if (handle->h_sync && journal->j_last_sync_writer != pid) {
1389 u64 commit_time, trans_time;
1391 journal->j_last_sync_writer = pid;
1393 read_lock(&journal->j_state_lock);
1394 commit_time = journal->j_average_commit_time;
1395 read_unlock(&journal->j_state_lock);
1397 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1398 transaction->t_start_time));
1400 commit_time = max_t(u64, commit_time,
1401 1000*journal->j_min_batch_time);
1402 commit_time = min_t(u64, commit_time,
1403 1000*journal->j_max_batch_time);
1405 if (trans_time < commit_time) {
1406 ktime_t expires = ktime_add_ns(ktime_get(),
1407 commit_time);
1408 set_current_state(TASK_UNINTERRUPTIBLE);
1409 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1413 if (handle->h_sync)
1414 transaction->t_synchronous_commit = 1;
1415 current->journal_info = NULL;
1416 atomic_sub(handle->h_buffer_credits,
1417 &transaction->t_outstanding_credits);
1420 * If the handle is marked SYNC, we need to set another commit
1421 * going! We also want to force a commit if the current
1422 * transaction is occupying too much of the log, or if the
1423 * transaction is too old now.
1425 if (handle->h_sync ||
1426 (atomic_read(&transaction->t_outstanding_credits) >
1427 journal->j_max_transaction_buffers) ||
1428 time_after_eq(jiffies, transaction->t_expires)) {
1429 /* Do this even for aborted journals: an abort still
1430 * completes the commit thread, it just doesn't write
1431 * anything to disk. */
1433 jbd_debug(2, "transaction too old, requesting commit for "
1434 "handle %p\n", handle);
1435 /* This is non-blocking */
1436 jbd2_log_start_commit(journal, transaction->t_tid);
1439 * Special case: JBD2_SYNC synchronous updates require us
1440 * to wait for the commit to complete.
1442 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1443 wait_for_commit = 1;
1447 * Once we drop t_updates, if it goes to zero the transaction
1448 * could start committing on us and eventually disappear. So
1449 * once we do this, we must not dereference transaction
1450 * pointer again.
1452 tid = transaction->t_tid;
1453 if (atomic_dec_and_test(&transaction->t_updates)) {
1454 wake_up(&journal->j_wait_updates);
1455 if (journal->j_barrier_count)
1456 wake_up(&journal->j_wait_transaction_locked);
1459 if (wait_for_commit)
1460 err = jbd2_log_wait_commit(journal, tid);
1462 lock_map_release(&handle->h_lockdep_map);
1464 jbd2_free_handle(handle);
1465 return err;
1469 * int jbd2_journal_force_commit() - force any uncommitted transactions
1470 * @journal: journal to force
1472 * For synchronous operations: force any uncommitted transactions
1473 * to disk. May seem kludgy, but it reuses all the handle batching
1474 * code in a very simple manner.
1476 int jbd2_journal_force_commit(journal_t *journal)
1478 handle_t *handle;
1479 int ret;
1481 handle = jbd2_journal_start(journal, 1);
1482 if (IS_ERR(handle)) {
1483 ret = PTR_ERR(handle);
1484 } else {
1485 handle->h_sync = 1;
1486 ret = jbd2_journal_stop(handle);
1488 return ret;
1493 * List management code snippets: various functions for manipulating the
1494 * transaction buffer lists.
1499 * Append a buffer to a transaction list, given the transaction's list head
1500 * pointer.
1502 * j_list_lock is held.
1504 * jbd_lock_bh_state(jh2bh(jh)) is held.
1507 static inline void
1508 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1510 if (!*list) {
1511 jh->b_tnext = jh->b_tprev = jh;
1512 *list = jh;
1513 } else {
1514 /* Insert at the tail of the list to preserve order */
1515 struct journal_head *first = *list, *last = first->b_tprev;
1516 jh->b_tprev = last;
1517 jh->b_tnext = first;
1518 last->b_tnext = first->b_tprev = jh;
1523 * Remove a buffer from a transaction list, given the transaction's list
1524 * head pointer.
1526 * Called with j_list_lock held, and the journal may not be locked.
1528 * jbd_lock_bh_state(jh2bh(jh)) is held.
1531 static inline void
1532 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1534 if (*list == jh) {
1535 *list = jh->b_tnext;
1536 if (*list == jh)
1537 *list = NULL;
1539 jh->b_tprev->b_tnext = jh->b_tnext;
1540 jh->b_tnext->b_tprev = jh->b_tprev;
1544 * Remove a buffer from the appropriate transaction list.
1546 * Note that this function can *change* the value of
1547 * bh->b_transaction->t_buffers, t_forget, t_iobuf_list, t_shadow_list,
1548 * t_log_list or t_reserved_list. If the caller is holding onto a copy of one
1549 * of these pointers, it could go bad. Generally the caller needs to re-read
1550 * the pointer from the transaction_t.
1552 * Called under j_list_lock. The journal may not be locked.
1554 void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1556 struct journal_head **list = NULL;
1557 transaction_t *transaction;
1558 struct buffer_head *bh = jh2bh(jh);
1560 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
1561 transaction = jh->b_transaction;
1562 if (transaction)
1563 assert_spin_locked(&transaction->t_journal->j_list_lock);
1565 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1566 if (jh->b_jlist != BJ_None)
1567 J_ASSERT_JH(jh, transaction != NULL);
1569 switch (jh->b_jlist) {
1570 case BJ_None:
1571 return;
1572 case BJ_Metadata:
1573 transaction->t_nr_buffers--;
1574 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
1575 list = &transaction->t_buffers;
1576 break;
1577 case BJ_Forget:
1578 list = &transaction->t_forget;
1579 break;
1580 case BJ_IO:
1581 list = &transaction->t_iobuf_list;
1582 break;
1583 case BJ_Shadow:
1584 list = &transaction->t_shadow_list;
1585 break;
1586 case BJ_LogCtl:
1587 list = &transaction->t_log_list;
1588 break;
1589 case BJ_Reserved:
1590 list = &transaction->t_reserved_list;
1591 break;
1594 __blist_del_buffer(list, jh);
1595 jh->b_jlist = BJ_None;
1596 if (test_clear_buffer_jbddirty(bh))
1597 mark_buffer_dirty(bh); /* Expose it to the VM */
1601 * Remove buffer from all transactions.
1603 * Called with bh_state lock and j_list_lock
1605 * jh and bh may be already freed when this function returns.
1607 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
1609 __jbd2_journal_temp_unlink_buffer(jh);
1610 jh->b_transaction = NULL;
1611 jbd2_journal_put_journal_head(jh);
1614 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
1616 struct buffer_head *bh = jh2bh(jh);
1618 /* Get reference so that buffer cannot be freed before we unlock it */
1619 get_bh(bh);
1620 jbd_lock_bh_state(bh);
1621 spin_lock(&journal->j_list_lock);
1622 __jbd2_journal_unfile_buffer(jh);
1623 spin_unlock(&journal->j_list_lock);
1624 jbd_unlock_bh_state(bh);
1625 __brelse(bh);
1629 * Called from jbd2_journal_try_to_free_buffers().
1631 * Called under jbd_lock_bh_state(bh)
1633 static void
1634 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
1636 struct journal_head *jh;
1638 jh = bh2jh(bh);
1640 if (buffer_locked(bh) || buffer_dirty(bh))
1641 goto out;
1643 if (jh->b_next_transaction != NULL)
1644 goto out;
1646 spin_lock(&journal->j_list_lock);
1647 if (jh->b_cp_transaction != NULL && jh->b_transaction == NULL) {
1648 /* written-back checkpointed metadata buffer */
1649 if (jh->b_jlist == BJ_None) {
1650 JBUFFER_TRACE(jh, "remove from checkpoint list");
1651 __jbd2_journal_remove_checkpoint(jh);
1654 spin_unlock(&journal->j_list_lock);
1655 out:
1656 return;
1660 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
1661 * @journal: journal for operation
1662 * @page: to try and free
1663 * @gfp_mask: we use the mask to detect how hard should we try to release
1664 * buffers. If __GFP_WAIT and __GFP_FS is set, we wait for commit code to
1665 * release the buffers.
1668 * For all the buffers on this page,
1669 * if they are fully written out ordered data, move them onto BUF_CLEAN
1670 * so try_to_free_buffers() can reap them.
1672 * This function returns non-zero if we wish try_to_free_buffers()
1673 * to be called. We do this if the page is releasable by try_to_free_buffers().
1674 * We also do it if the page has locked or dirty buffers and the caller wants
1675 * us to perform sync or async writeout.
1677 * This complicates JBD locking somewhat. We aren't protected by the
1678 * BKL here. We wish to remove the buffer from its committing or
1679 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
1681 * This may *change* the value of transaction_t->t_datalist, so anyone
1682 * who looks at t_datalist needs to lock against this function.
1684 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
1685 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
1686 * will come out of the lock with the buffer dirty, which makes it
1687 * ineligible for release here.
1689 * Who else is affected by this? hmm... Really the only contender
1690 * is do_get_write_access() - it could be looking at the buffer while
1691 * journal_try_to_free_buffer() is changing its state. But that
1692 * cannot happen because we never reallocate freed data as metadata
1693 * while the data is part of a transaction. Yes?
1695 * Return 0 on failure, 1 on success
1697 int jbd2_journal_try_to_free_buffers(journal_t *journal,
1698 struct page *page, gfp_t gfp_mask)
1700 struct buffer_head *head;
1701 struct buffer_head *bh;
1702 int ret = 0;
1704 J_ASSERT(PageLocked(page));
1706 head = page_buffers(page);
1707 bh = head;
1708 do {
1709 struct journal_head *jh;
1712 * We take our own ref against the journal_head here to avoid
1713 * having to add tons of locking around each instance of
1714 * jbd2_journal_put_journal_head().
1716 jh = jbd2_journal_grab_journal_head(bh);
1717 if (!jh)
1718 continue;
1720 jbd_lock_bh_state(bh);
1721 __journal_try_to_free_buffer(journal, bh);
1722 jbd2_journal_put_journal_head(jh);
1723 jbd_unlock_bh_state(bh);
1724 if (buffer_jbd(bh))
1725 goto busy;
1726 } while ((bh = bh->b_this_page) != head);
1728 ret = try_to_free_buffers(page);
1730 busy:
1731 return ret;
1735 * This buffer is no longer needed. If it is on an older transaction's
1736 * checkpoint list we need to record it on this transaction's forget list
1737 * to pin this buffer (and hence its checkpointing transaction) down until
1738 * this transaction commits. If the buffer isn't on a checkpoint list, we
1739 * release it.
1740 * Returns non-zero if JBD no longer has an interest in the buffer.
1742 * Called under j_list_lock.
1744 * Called under jbd_lock_bh_state(bh).
1746 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
1748 int may_free = 1;
1749 struct buffer_head *bh = jh2bh(jh);
1751 if (jh->b_cp_transaction) {
1752 JBUFFER_TRACE(jh, "on running+cp transaction");
1753 __jbd2_journal_temp_unlink_buffer(jh);
1755 * We don't want to write the buffer anymore, clear the
1756 * bit so that we don't confuse checks in
1757 * __journal_file_buffer
1759 clear_buffer_dirty(bh);
1760 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1761 may_free = 0;
1762 } else {
1763 JBUFFER_TRACE(jh, "on running transaction");
1764 __jbd2_journal_unfile_buffer(jh);
1766 return may_free;
1770 * jbd2_journal_invalidatepage
1772 * This code is tricky. It has a number of cases to deal with.
1774 * There are two invariants which this code relies on:
1776 * i_size must be updated on disk before we start calling invalidatepage on the
1777 * data.
1779 * This is done in ext3 by defining an ext3_setattr method which
1780 * updates i_size before truncate gets going. By maintaining this
1781 * invariant, we can be sure that it is safe to throw away any buffers
1782 * attached to the current transaction: once the transaction commits,
1783 * we know that the data will not be needed.
1785 * Note however that we can *not* throw away data belonging to the
1786 * previous, committing transaction!
1788 * Any disk blocks which *are* part of the previous, committing
1789 * transaction (and which therefore cannot be discarded immediately) are
1790 * not going to be reused in the new running transaction
1792 * The bitmap committed_data images guarantee this: any block which is
1793 * allocated in one transaction and removed in the next will be marked
1794 * as in-use in the committed_data bitmap, so cannot be reused until
1795 * the next transaction to delete the block commits. This means that
1796 * leaving committing buffers dirty is quite safe: the disk blocks
1797 * cannot be reallocated to a different file and so buffer aliasing is
1798 * not possible.
1801 * The above applies mainly to ordered data mode. In writeback mode we
1802 * don't make guarantees about the order in which data hits disk --- in
1803 * particular we don't guarantee that new dirty data is flushed before
1804 * transaction commit --- so it is always safe just to discard data
1805 * immediately in that mode. --sct
1809 * The journal_unmap_buffer helper function returns zero if the buffer
1810 * concerned remains pinned as an anonymous buffer belonging to an older
1811 * transaction.
1813 * We're outside-transaction here. Either or both of j_running_transaction
1814 * and j_committing_transaction may be NULL.
1816 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh)
1818 transaction_t *transaction;
1819 struct journal_head *jh;
1820 int may_free = 1;
1821 int ret;
1823 BUFFER_TRACE(bh, "entry");
1826 * It is safe to proceed here without the j_list_lock because the
1827 * buffers cannot be stolen by try_to_free_buffers as long as we are
1828 * holding the page lock. --sct
1831 if (!buffer_jbd(bh))
1832 goto zap_buffer_unlocked;
1834 /* OK, we have data buffer in journaled mode */
1835 write_lock(&journal->j_state_lock);
1836 jbd_lock_bh_state(bh);
1837 spin_lock(&journal->j_list_lock);
1839 jh = jbd2_journal_grab_journal_head(bh);
1840 if (!jh)
1841 goto zap_buffer_no_jh;
1844 * We cannot remove the buffer from checkpoint lists until the
1845 * transaction adding inode to orphan list (let's call it T)
1846 * is committed. Otherwise if the transaction changing the
1847 * buffer would be cleaned from the journal before T is
1848 * committed, a crash will cause that the correct contents of
1849 * the buffer will be lost. On the other hand we have to
1850 * clear the buffer dirty bit at latest at the moment when the
1851 * transaction marking the buffer as freed in the filesystem
1852 * structures is committed because from that moment on the
1853 * buffer can be reallocated and used by a different page.
1854 * Since the block hasn't been freed yet but the inode has
1855 * already been added to orphan list, it is safe for us to add
1856 * the buffer to BJ_Forget list of the newest transaction.
1858 transaction = jh->b_transaction;
1859 if (transaction == NULL) {
1860 /* First case: not on any transaction. If it
1861 * has no checkpoint link, then we can zap it:
1862 * it's a writeback-mode buffer so we don't care
1863 * if it hits disk safely. */
1864 if (!jh->b_cp_transaction) {
1865 JBUFFER_TRACE(jh, "not on any transaction: zap");
1866 goto zap_buffer;
1869 if (!buffer_dirty(bh)) {
1870 /* bdflush has written it. We can drop it now */
1871 goto zap_buffer;
1874 /* OK, it must be in the journal but still not
1875 * written fully to disk: it's metadata or
1876 * journaled data... */
1878 if (journal->j_running_transaction) {
1879 /* ... and once the current transaction has
1880 * committed, the buffer won't be needed any
1881 * longer. */
1882 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
1883 ret = __dispose_buffer(jh,
1884 journal->j_running_transaction);
1885 jbd2_journal_put_journal_head(jh);
1886 spin_unlock(&journal->j_list_lock);
1887 jbd_unlock_bh_state(bh);
1888 write_unlock(&journal->j_state_lock);
1889 return ret;
1890 } else {
1891 /* There is no currently-running transaction. So the
1892 * orphan record which we wrote for this file must have
1893 * passed into commit. We must attach this buffer to
1894 * the committing transaction, if it exists. */
1895 if (journal->j_committing_transaction) {
1896 JBUFFER_TRACE(jh, "give to committing trans");
1897 ret = __dispose_buffer(jh,
1898 journal->j_committing_transaction);
1899 jbd2_journal_put_journal_head(jh);
1900 spin_unlock(&journal->j_list_lock);
1901 jbd_unlock_bh_state(bh);
1902 write_unlock(&journal->j_state_lock);
1903 return ret;
1904 } else {
1905 /* The orphan record's transaction has
1906 * committed. We can cleanse this buffer */
1907 clear_buffer_jbddirty(bh);
1908 goto zap_buffer;
1911 } else if (transaction == journal->j_committing_transaction) {
1912 JBUFFER_TRACE(jh, "on committing transaction");
1914 * The buffer is committing, we simply cannot touch
1915 * it. So we just set j_next_transaction to the
1916 * running transaction (if there is one) and mark
1917 * buffer as freed so that commit code knows it should
1918 * clear dirty bits when it is done with the buffer.
1920 set_buffer_freed(bh);
1921 if (journal->j_running_transaction && buffer_jbddirty(bh))
1922 jh->b_next_transaction = journal->j_running_transaction;
1923 jbd2_journal_put_journal_head(jh);
1924 spin_unlock(&journal->j_list_lock);
1925 jbd_unlock_bh_state(bh);
1926 write_unlock(&journal->j_state_lock);
1927 return 0;
1928 } else {
1929 /* Good, the buffer belongs to the running transaction.
1930 * We are writing our own transaction's data, not any
1931 * previous one's, so it is safe to throw it away
1932 * (remember that we expect the filesystem to have set
1933 * i_size already for this truncate so recovery will not
1934 * expose the disk blocks we are discarding here.) */
1935 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
1936 JBUFFER_TRACE(jh, "on running transaction");
1937 may_free = __dispose_buffer(jh, transaction);
1940 zap_buffer:
1941 jbd2_journal_put_journal_head(jh);
1942 zap_buffer_no_jh:
1943 spin_unlock(&journal->j_list_lock);
1944 jbd_unlock_bh_state(bh);
1945 write_unlock(&journal->j_state_lock);
1946 zap_buffer_unlocked:
1947 clear_buffer_dirty(bh);
1948 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
1949 clear_buffer_mapped(bh);
1950 clear_buffer_req(bh);
1951 clear_buffer_new(bh);
1952 bh->b_bdev = NULL;
1953 return may_free;
1957 * void jbd2_journal_invalidatepage()
1958 * @journal: journal to use for flush...
1959 * @page: page to flush
1960 * @offset: length of page to invalidate.
1962 * Reap page buffers containing data after offset in page.
1965 void jbd2_journal_invalidatepage(journal_t *journal,
1966 struct page *page,
1967 unsigned long offset)
1969 struct buffer_head *head, *bh, *next;
1970 unsigned int curr_off = 0;
1971 int may_free = 1;
1973 if (!PageLocked(page))
1974 BUG();
1975 if (!page_has_buffers(page))
1976 return;
1978 /* We will potentially be playing with lists other than just the
1979 * data lists (especially for journaled data mode), so be
1980 * cautious in our locking. */
1982 head = bh = page_buffers(page);
1983 do {
1984 unsigned int next_off = curr_off + bh->b_size;
1985 next = bh->b_this_page;
1987 if (offset <= curr_off) {
1988 /* This block is wholly outside the truncation point */
1989 lock_buffer(bh);
1990 may_free &= journal_unmap_buffer(journal, bh);
1991 unlock_buffer(bh);
1993 curr_off = next_off;
1994 bh = next;
1996 } while (bh != head);
1998 if (!offset) {
1999 if (may_free && try_to_free_buffers(page))
2000 J_ASSERT(!page_has_buffers(page));
2005 * File a buffer on the given transaction list.
2007 void __jbd2_journal_file_buffer(struct journal_head *jh,
2008 transaction_t *transaction, int jlist)
2010 struct journal_head **list = NULL;
2011 int was_dirty = 0;
2012 struct buffer_head *bh = jh2bh(jh);
2014 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2015 assert_spin_locked(&transaction->t_journal->j_list_lock);
2017 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2018 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2019 jh->b_transaction == NULL);
2021 if (jh->b_transaction && jh->b_jlist == jlist)
2022 return;
2024 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2025 jlist == BJ_Shadow || jlist == BJ_Forget) {
2027 * For metadata buffers, we track dirty bit in buffer_jbddirty
2028 * instead of buffer_dirty. We should not see a dirty bit set
2029 * here because we clear it in do_get_write_access but e.g.
2030 * tune2fs can modify the sb and set the dirty bit at any time
2031 * so we try to gracefully handle that.
2033 if (buffer_dirty(bh))
2034 warn_dirty_buffer(bh);
2035 if (test_clear_buffer_dirty(bh) ||
2036 test_clear_buffer_jbddirty(bh))
2037 was_dirty = 1;
2040 if (jh->b_transaction)
2041 __jbd2_journal_temp_unlink_buffer(jh);
2042 else
2043 jbd2_journal_grab_journal_head(bh);
2044 jh->b_transaction = transaction;
2046 switch (jlist) {
2047 case BJ_None:
2048 J_ASSERT_JH(jh, !jh->b_committed_data);
2049 J_ASSERT_JH(jh, !jh->b_frozen_data);
2050 return;
2051 case BJ_Metadata:
2052 transaction->t_nr_buffers++;
2053 list = &transaction->t_buffers;
2054 break;
2055 case BJ_Forget:
2056 list = &transaction->t_forget;
2057 break;
2058 case BJ_IO:
2059 list = &transaction->t_iobuf_list;
2060 break;
2061 case BJ_Shadow:
2062 list = &transaction->t_shadow_list;
2063 break;
2064 case BJ_LogCtl:
2065 list = &transaction->t_log_list;
2066 break;
2067 case BJ_Reserved:
2068 list = &transaction->t_reserved_list;
2069 break;
2072 __blist_add_buffer(list, jh);
2073 jh->b_jlist = jlist;
2075 if (was_dirty)
2076 set_buffer_jbddirty(bh);
2079 void jbd2_journal_file_buffer(struct journal_head *jh,
2080 transaction_t *transaction, int jlist)
2082 jbd_lock_bh_state(jh2bh(jh));
2083 spin_lock(&transaction->t_journal->j_list_lock);
2084 __jbd2_journal_file_buffer(jh, transaction, jlist);
2085 spin_unlock(&transaction->t_journal->j_list_lock);
2086 jbd_unlock_bh_state(jh2bh(jh));
2090 * Remove a buffer from its current buffer list in preparation for
2091 * dropping it from its current transaction entirely. If the buffer has
2092 * already started to be used by a subsequent transaction, refile the
2093 * buffer on that transaction's metadata list.
2095 * Called under j_list_lock
2096 * Called under jbd_lock_bh_state(jh2bh(jh))
2098 * jh and bh may be already free when this function returns
2100 void __jbd2_journal_refile_buffer(struct journal_head *jh)
2102 int was_dirty, jlist;
2103 struct buffer_head *bh = jh2bh(jh);
2105 J_ASSERT_JH(jh, jbd_is_locked_bh_state(bh));
2106 if (jh->b_transaction)
2107 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2109 /* If the buffer is now unused, just drop it. */
2110 if (jh->b_next_transaction == NULL) {
2111 __jbd2_journal_unfile_buffer(jh);
2112 return;
2116 * It has been modified by a later transaction: add it to the new
2117 * transaction's metadata list.
2120 was_dirty = test_clear_buffer_jbddirty(bh);
2121 __jbd2_journal_temp_unlink_buffer(jh);
2123 * We set b_transaction here because b_next_transaction will inherit
2124 * our jh reference and thus __jbd2_journal_file_buffer() must not
2125 * take a new one.
2127 jh->b_transaction = jh->b_next_transaction;
2128 jh->b_next_transaction = NULL;
2129 if (buffer_freed(bh))
2130 jlist = BJ_Forget;
2131 else if (jh->b_modified)
2132 jlist = BJ_Metadata;
2133 else
2134 jlist = BJ_Reserved;
2135 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2136 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2138 if (was_dirty)
2139 set_buffer_jbddirty(bh);
2143 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2144 * bh reference so that we can safely unlock bh.
2146 * The jh and bh may be freed by this call.
2148 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2150 struct buffer_head *bh = jh2bh(jh);
2152 /* Get reference so that buffer cannot be freed before we unlock it */
2153 get_bh(bh);
2154 jbd_lock_bh_state(bh);
2155 spin_lock(&journal->j_list_lock);
2156 __jbd2_journal_refile_buffer(jh);
2157 jbd_unlock_bh_state(bh);
2158 spin_unlock(&journal->j_list_lock);
2159 __brelse(bh);
2163 * File inode in the inode list of the handle's transaction
2165 int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode)
2167 transaction_t *transaction = handle->h_transaction;
2168 journal_t *journal = transaction->t_journal;
2170 if (is_handle_aborted(handle))
2171 return -EIO;
2173 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2174 transaction->t_tid);
2177 * First check whether inode isn't already on the transaction's
2178 * lists without taking the lock. Note that this check is safe
2179 * without the lock as we cannot race with somebody removing inode
2180 * from the transaction. The reason is that we remove inode from the
2181 * transaction only in journal_release_jbd_inode() and when we commit
2182 * the transaction. We are guarded from the first case by holding
2183 * a reference to the inode. We are safe against the second case
2184 * because if jinode->i_transaction == transaction, commit code
2185 * cannot touch the transaction because we hold reference to it,
2186 * and if jinode->i_next_transaction == transaction, commit code
2187 * will only file the inode where we want it.
2189 if (jinode->i_transaction == transaction ||
2190 jinode->i_next_transaction == transaction)
2191 return 0;
2193 spin_lock(&journal->j_list_lock);
2195 if (jinode->i_transaction == transaction ||
2196 jinode->i_next_transaction == transaction)
2197 goto done;
2200 * We only ever set this variable to 1 so the test is safe. Since
2201 * t_need_data_flush is likely to be set, we do the test to save some
2202 * cacheline bouncing
2204 if (!transaction->t_need_data_flush)
2205 transaction->t_need_data_flush = 1;
2206 /* On some different transaction's list - should be
2207 * the committing one */
2208 if (jinode->i_transaction) {
2209 J_ASSERT(jinode->i_next_transaction == NULL);
2210 J_ASSERT(jinode->i_transaction ==
2211 journal->j_committing_transaction);
2212 jinode->i_next_transaction = transaction;
2213 goto done;
2215 /* Not on any transaction list... */
2216 J_ASSERT(!jinode->i_next_transaction);
2217 jinode->i_transaction = transaction;
2218 list_add(&jinode->i_list, &transaction->t_inode_list);
2219 done:
2220 spin_unlock(&journal->j_list_lock);
2222 return 0;
2226 * File truncate and transaction commit interact with each other in a
2227 * non-trivial way. If a transaction writing data block A is
2228 * committing, we cannot discard the data by truncate until we have
2229 * written them. Otherwise if we crashed after the transaction with
2230 * write has committed but before the transaction with truncate has
2231 * committed, we could see stale data in block A. This function is a
2232 * helper to solve this problem. It starts writeout of the truncated
2233 * part in case it is in the committing transaction.
2235 * Filesystem code must call this function when inode is journaled in
2236 * ordered mode before truncation happens and after the inode has been
2237 * placed on orphan list with the new inode size. The second condition
2238 * avoids the race that someone writes new data and we start
2239 * committing the transaction after this function has been called but
2240 * before a transaction for truncate is started (and furthermore it
2241 * allows us to optimize the case where the addition to orphan list
2242 * happens in the same transaction as write --- we don't have to write
2243 * any data in such case).
2245 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2246 struct jbd2_inode *jinode,
2247 loff_t new_size)
2249 transaction_t *inode_trans, *commit_trans;
2250 int ret = 0;
2252 /* This is a quick check to avoid locking if not necessary */
2253 if (!jinode->i_transaction)
2254 goto out;
2255 /* Locks are here just to force reading of recent values, it is
2256 * enough that the transaction was not committing before we started
2257 * a transaction adding the inode to orphan list */
2258 read_lock(&journal->j_state_lock);
2259 commit_trans = journal->j_committing_transaction;
2260 read_unlock(&journal->j_state_lock);
2261 spin_lock(&journal->j_list_lock);
2262 inode_trans = jinode->i_transaction;
2263 spin_unlock(&journal->j_list_lock);
2264 if (inode_trans == commit_trans) {
2265 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2266 new_size, LLONG_MAX);
2267 if (ret)
2268 jbd2_journal_abort(journal, ret);
2270 out:
2271 return ret;