1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/transaction.c
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
7 * Copyright 1998 Red Hat corp --- All Rights Reserved
9 * Generic filesystem transaction handling code; part of the ext2fs
12 * This file manages transactions (compound commits managed by the
13 * journaling code) and handles (individual atomic operations by the
17 #include <linux/time.h>
19 #include <linux/jbd2.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/timer.h>
24 #include <linux/highmem.h>
25 #include <linux/hrtimer.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bug.h>
28 #include <linux/module.h>
29 #include <linux/sched/mm.h>
31 #include <trace/events/jbd2.h>
33 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh);
34 static void __jbd2_journal_unfile_buffer(struct journal_head *jh);
36 static struct kmem_cache *transaction_cache;
37 int __init jbd2_journal_init_transaction_cache(void)
39 J_ASSERT(!transaction_cache);
40 transaction_cache = kmem_cache_create("jbd2_transaction_s",
41 sizeof(transaction_t),
43 SLAB_HWCACHE_ALIGN|SLAB_TEMPORARY,
45 if (!transaction_cache) {
46 pr_emerg("JBD2: failed to create transaction cache\n");
52 void jbd2_journal_destroy_transaction_cache(void)
54 kmem_cache_destroy(transaction_cache);
55 transaction_cache = NULL;
58 void jbd2_journal_free_transaction(transaction_t *transaction)
60 if (unlikely(ZERO_OR_NULL_PTR(transaction)))
62 kmem_cache_free(transaction_cache, transaction);
66 * Base amount of descriptor blocks we reserve for each transaction.
68 static int jbd2_descriptor_blocks_per_trans(journal_t *journal)
70 int tag_space = journal->j_blocksize - sizeof(journal_header_t);
75 if (jbd2_journal_has_csum_v2or3(journal))
76 tag_space -= sizeof(struct jbd2_journal_block_tail);
77 /* Commit code leaves a slack space of 16 bytes at the end of block */
78 tags_per_block = (tag_space - 16) / journal_tag_bytes(journal);
80 * Revoke descriptors are accounted separately so we need to reserve
81 * space for commit block and normal transaction descriptor blocks.
83 return 1 + DIV_ROUND_UP(journal->j_max_transaction_buffers,
88 * jbd2_get_transaction: obtain a new transaction_t object.
90 * Simply initialise a new transaction. Initialize it in
91 * RUNNING state and add it to the current journal (which should not
92 * have an existing running transaction: we only make a new transaction
93 * once we have started to commit the old one).
96 * The journal MUST be locked. We don't perform atomic mallocs on the
97 * new transaction and we can't block without protecting against other
98 * processes trying to touch the journal while it is in transition.
102 static void jbd2_get_transaction(journal_t *journal,
103 transaction_t *transaction)
105 transaction->t_journal = journal;
106 transaction->t_state = T_RUNNING;
107 transaction->t_start_time = ktime_get();
108 transaction->t_tid = journal->j_transaction_sequence++;
109 transaction->t_expires = jiffies + journal->j_commit_interval;
110 spin_lock_init(&transaction->t_handle_lock);
111 atomic_set(&transaction->t_updates, 0);
112 atomic_set(&transaction->t_outstanding_credits,
113 jbd2_descriptor_blocks_per_trans(journal) +
114 atomic_read(&journal->j_reserved_credits));
115 atomic_set(&transaction->t_outstanding_revokes, 0);
116 atomic_set(&transaction->t_handle_count, 0);
117 INIT_LIST_HEAD(&transaction->t_inode_list);
118 INIT_LIST_HEAD(&transaction->t_private_list);
120 /* Set up the commit timer for the new transaction. */
121 journal->j_commit_timer.expires = round_jiffies_up(transaction->t_expires);
122 add_timer(&journal->j_commit_timer);
124 J_ASSERT(journal->j_running_transaction == NULL);
125 journal->j_running_transaction = transaction;
126 transaction->t_max_wait = 0;
127 transaction->t_start = jiffies;
128 transaction->t_requested = 0;
134 * A handle_t is an object which represents a single atomic update to a
135 * filesystem, and which tracks all of the modifications which form part
136 * of that one update.
140 * Update transaction's maximum wait time, if debugging is enabled.
142 * In order for t_max_wait to be reliable, it must be protected by a
143 * lock. But doing so will mean that start_this_handle() can not be
144 * run in parallel on SMP systems, which limits our scalability. So
145 * unless debugging is enabled, we no longer update t_max_wait, which
146 * means that maximum wait time reported by the jbd2_run_stats
147 * tracepoint will always be zero.
149 static inline void update_t_max_wait(transaction_t *transaction,
152 #ifdef CONFIG_JBD2_DEBUG
153 if (jbd2_journal_enable_debug &&
154 time_after(transaction->t_start, ts)) {
155 ts = jbd2_time_diff(ts, transaction->t_start);
156 spin_lock(&transaction->t_handle_lock);
157 if (ts > transaction->t_max_wait)
158 transaction->t_max_wait = ts;
159 spin_unlock(&transaction->t_handle_lock);
165 * Wait until running transaction passes to T_FLUSH state and new transaction
166 * can thus be started. Also starts the commit if needed. The function expects
167 * running transaction to exist and releases j_state_lock.
169 static void wait_transaction_locked(journal_t *journal)
170 __releases(journal->j_state_lock)
174 tid_t tid = journal->j_running_transaction->t_tid;
176 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
177 TASK_UNINTERRUPTIBLE);
178 need_to_start = !tid_geq(journal->j_commit_request, tid);
179 read_unlock(&journal->j_state_lock);
181 jbd2_log_start_commit(journal, tid);
182 jbd2_might_wait_for_commit(journal);
184 finish_wait(&journal->j_wait_transaction_locked, &wait);
188 * Wait until running transaction transitions from T_SWITCH to T_FLUSH
189 * state and new transaction can thus be started. The function releases
192 static void wait_transaction_switching(journal_t *journal)
193 __releases(journal->j_state_lock)
197 if (WARN_ON(!journal->j_running_transaction ||
198 journal->j_running_transaction->t_state != T_SWITCH)) {
199 read_unlock(&journal->j_state_lock);
202 prepare_to_wait(&journal->j_wait_transaction_locked, &wait,
203 TASK_UNINTERRUPTIBLE);
204 read_unlock(&journal->j_state_lock);
206 * We don't call jbd2_might_wait_for_commit() here as there's no
207 * waiting for outstanding handles happening anymore in T_SWITCH state
208 * and handling of reserved handles actually relies on that for
212 finish_wait(&journal->j_wait_transaction_locked, &wait);
215 static void sub_reserved_credits(journal_t *journal, int blocks)
217 atomic_sub(blocks, &journal->j_reserved_credits);
218 wake_up(&journal->j_wait_reserved);
222 * Wait until we can add credits for handle to the running transaction. Called
223 * with j_state_lock held for reading. Returns 0 if handle joined the running
224 * transaction. Returns 1 if we had to wait, j_state_lock is dropped, and
227 static int add_transaction_credits(journal_t *journal, int blocks,
230 transaction_t *t = journal->j_running_transaction;
232 int total = blocks + rsv_blocks;
235 * If the current transaction is locked down for commit, wait
236 * for the lock to be released.
238 if (t->t_state != T_RUNNING) {
239 WARN_ON_ONCE(t->t_state >= T_FLUSH);
240 wait_transaction_locked(journal);
245 * If there is not enough space left in the log to write all
246 * potential buffers requested by this operation, we need to
247 * stall pending a log checkpoint to free some more log space.
249 needed = atomic_add_return(total, &t->t_outstanding_credits);
250 if (needed > journal->j_max_transaction_buffers) {
252 * If the current transaction is already too large,
253 * then start to commit it: we can then go back and
254 * attach this handle to a new transaction.
256 atomic_sub(total, &t->t_outstanding_credits);
259 * Is the number of reserved credits in the current transaction too
260 * big to fit this handle? Wait until reserved credits are freed.
262 if (atomic_read(&journal->j_reserved_credits) + total >
263 journal->j_max_transaction_buffers) {
264 read_unlock(&journal->j_state_lock);
265 jbd2_might_wait_for_commit(journal);
266 wait_event(journal->j_wait_reserved,
267 atomic_read(&journal->j_reserved_credits) + total <=
268 journal->j_max_transaction_buffers);
272 wait_transaction_locked(journal);
277 * The commit code assumes that it can get enough log space
278 * without forcing a checkpoint. This is *critical* for
279 * correctness: a checkpoint of a buffer which is also
280 * associated with a committing transaction creates a deadlock,
281 * so commit simply cannot force through checkpoints.
283 * We must therefore ensure the necessary space in the journal
284 * *before* starting to dirty potentially checkpointed buffers
285 * in the new transaction.
287 if (jbd2_log_space_left(journal) < journal->j_max_transaction_buffers) {
288 atomic_sub(total, &t->t_outstanding_credits);
289 read_unlock(&journal->j_state_lock);
290 jbd2_might_wait_for_commit(journal);
291 write_lock(&journal->j_state_lock);
292 if (jbd2_log_space_left(journal) <
293 journal->j_max_transaction_buffers)
294 __jbd2_log_wait_for_space(journal);
295 write_unlock(&journal->j_state_lock);
299 /* No reservation? We are done... */
303 needed = atomic_add_return(rsv_blocks, &journal->j_reserved_credits);
304 /* We allow at most half of a transaction to be reserved */
305 if (needed > journal->j_max_transaction_buffers / 2) {
306 sub_reserved_credits(journal, rsv_blocks);
307 atomic_sub(total, &t->t_outstanding_credits);
308 read_unlock(&journal->j_state_lock);
309 jbd2_might_wait_for_commit(journal);
310 wait_event(journal->j_wait_reserved,
311 atomic_read(&journal->j_reserved_credits) + rsv_blocks
312 <= journal->j_max_transaction_buffers / 2);
319 * start_this_handle: Given a handle, deal with any locking or stalling
320 * needed to make sure that there is enough journal space for the handle
321 * to begin. Attach the handle to a transaction and set up the
322 * transaction's buffer credits.
325 static int start_this_handle(journal_t *journal, handle_t *handle,
328 transaction_t *transaction, *new_transaction = NULL;
329 int blocks = handle->h_total_credits;
331 unsigned long ts = jiffies;
333 if (handle->h_rsv_handle)
334 rsv_blocks = handle->h_rsv_handle->h_total_credits;
337 * Limit the number of reserved credits to 1/2 of maximum transaction
338 * size and limit the number of total credits to not exceed maximum
339 * transaction size per operation.
341 if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
342 (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
343 printk(KERN_ERR "JBD2: %s wants too many credits "
344 "credits:%d rsv_credits:%d max:%d\n",
345 current->comm, blocks, rsv_blocks,
346 journal->j_max_transaction_buffers);
352 if (!journal->j_running_transaction) {
354 * If __GFP_FS is not present, then we may be being called from
355 * inside the fs writeback layer, so we MUST NOT fail.
357 if ((gfp_mask & __GFP_FS) == 0)
358 gfp_mask |= __GFP_NOFAIL;
359 new_transaction = kmem_cache_zalloc(transaction_cache,
361 if (!new_transaction)
365 jbd_debug(3, "New handle %p going live.\n", handle);
368 * We need to hold j_state_lock until t_updates has been incremented,
369 * for proper journal barrier handling
372 read_lock(&journal->j_state_lock);
373 BUG_ON(journal->j_flags & JBD2_UNMOUNT);
374 if (is_journal_aborted(journal) ||
375 (journal->j_errno != 0 && !(journal->j_flags & JBD2_ACK_ERR))) {
376 read_unlock(&journal->j_state_lock);
377 jbd2_journal_free_transaction(new_transaction);
382 * Wait on the journal's transaction barrier if necessary. Specifically
383 * we allow reserved handles to proceed because otherwise commit could
384 * deadlock on page writeback not being able to complete.
386 if (!handle->h_reserved && journal->j_barrier_count) {
387 read_unlock(&journal->j_state_lock);
388 wait_event(journal->j_wait_transaction_locked,
389 journal->j_barrier_count == 0);
393 if (!journal->j_running_transaction) {
394 read_unlock(&journal->j_state_lock);
395 if (!new_transaction)
396 goto alloc_transaction;
397 write_lock(&journal->j_state_lock);
398 if (!journal->j_running_transaction &&
399 (handle->h_reserved || !journal->j_barrier_count)) {
400 jbd2_get_transaction(journal, new_transaction);
401 new_transaction = NULL;
403 write_unlock(&journal->j_state_lock);
407 transaction = journal->j_running_transaction;
409 if (!handle->h_reserved) {
410 /* We may have dropped j_state_lock - restart in that case */
411 if (add_transaction_credits(journal, blocks, rsv_blocks))
415 * We have handle reserved so we are allowed to join T_LOCKED
416 * transaction and we don't have to check for transaction size
417 * and journal space. But we still have to wait while running
418 * transaction is being switched to a committing one as it
419 * won't wait for any handles anymore.
421 if (transaction->t_state == T_SWITCH) {
422 wait_transaction_switching(journal);
425 sub_reserved_credits(journal, blocks);
426 handle->h_reserved = 0;
429 /* OK, account for the buffers that this operation expects to
430 * use and add the handle to the running transaction.
432 update_t_max_wait(transaction, ts);
433 handle->h_transaction = transaction;
434 handle->h_requested_credits = blocks;
435 handle->h_revoke_credits_requested = handle->h_revoke_credits;
436 handle->h_start_jiffies = jiffies;
437 atomic_inc(&transaction->t_updates);
438 atomic_inc(&transaction->t_handle_count);
439 jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
441 atomic_read(&transaction->t_outstanding_credits),
442 jbd2_log_space_left(journal));
443 read_unlock(&journal->j_state_lock);
444 current->journal_info = handle;
446 rwsem_acquire_read(&journal->j_trans_commit_map, 0, 0, _THIS_IP_);
447 jbd2_journal_free_transaction(new_transaction);
449 * Ensure that no allocations done while the transaction is open are
450 * going to recurse back to the fs layer.
452 handle->saved_alloc_context = memalloc_nofs_save();
456 /* Allocate a new handle. This should probably be in a slab... */
457 static handle_t *new_handle(int nblocks)
459 handle_t *handle = jbd2_alloc_handle(GFP_NOFS);
462 handle->h_total_credits = nblocks;
468 handle_t *jbd2__journal_start(journal_t *journal, int nblocks, int rsv_blocks,
469 int revoke_records, gfp_t gfp_mask,
470 unsigned int type, unsigned int line_no)
472 handle_t *handle = journal_current_handle();
476 return ERR_PTR(-EROFS);
479 J_ASSERT(handle->h_transaction->t_journal == journal);
484 nblocks += DIV_ROUND_UP(revoke_records,
485 journal->j_revoke_records_per_block);
486 handle = new_handle(nblocks);
488 return ERR_PTR(-ENOMEM);
490 handle_t *rsv_handle;
492 rsv_handle = new_handle(rsv_blocks);
494 jbd2_free_handle(handle);
495 return ERR_PTR(-ENOMEM);
497 rsv_handle->h_reserved = 1;
498 rsv_handle->h_journal = journal;
499 handle->h_rsv_handle = rsv_handle;
501 handle->h_revoke_credits = revoke_records;
503 err = start_this_handle(journal, handle, gfp_mask);
505 if (handle->h_rsv_handle)
506 jbd2_free_handle(handle->h_rsv_handle);
507 jbd2_free_handle(handle);
510 handle->h_type = type;
511 handle->h_line_no = line_no;
512 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
513 handle->h_transaction->t_tid, type,
518 EXPORT_SYMBOL(jbd2__journal_start);
522 * handle_t *jbd2_journal_start() - Obtain a new handle.
523 * @journal: Journal to start transaction on.
524 * @nblocks: number of block buffer we might modify
526 * We make sure that the transaction can guarantee at least nblocks of
527 * modified buffers in the log. We block until the log can guarantee
528 * that much space. Additionally, if rsv_blocks > 0, we also create another
529 * handle with rsv_blocks reserved blocks in the journal. This handle is
530 * stored in h_rsv_handle. It is not attached to any particular transaction
531 * and thus doesn't block transaction commit. If the caller uses this reserved
532 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
533 * on the parent handle will dispose the reserved one. Reserved handle has to
534 * be converted to a normal handle using jbd2_journal_start_reserved() before
537 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
540 handle_t *jbd2_journal_start(journal_t *journal, int nblocks)
542 return jbd2__journal_start(journal, nblocks, 0, 0, GFP_NOFS, 0, 0);
544 EXPORT_SYMBOL(jbd2_journal_start);
546 static void __jbd2_journal_unreserve_handle(handle_t *handle, transaction_t *t)
548 journal_t *journal = handle->h_journal;
550 WARN_ON(!handle->h_reserved);
551 sub_reserved_credits(journal, handle->h_total_credits);
553 atomic_sub(handle->h_total_credits, &t->t_outstanding_credits);
556 void jbd2_journal_free_reserved(handle_t *handle)
558 journal_t *journal = handle->h_journal;
560 /* Get j_state_lock to pin running transaction if it exists */
561 read_lock(&journal->j_state_lock);
562 __jbd2_journal_unreserve_handle(handle, journal->j_running_transaction);
563 read_unlock(&journal->j_state_lock);
564 jbd2_free_handle(handle);
566 EXPORT_SYMBOL(jbd2_journal_free_reserved);
569 * int jbd2_journal_start_reserved() - start reserved handle
570 * @handle: handle to start
571 * @type: for handle statistics
572 * @line_no: for handle statistics
574 * Start handle that has been previously reserved with jbd2_journal_reserve().
575 * This attaches @handle to the running transaction (or creates one if there's
576 * not transaction running). Unlike jbd2_journal_start() this function cannot
577 * block on journal commit, checkpointing, or similar stuff. It can block on
578 * memory allocation or frozen journal though.
580 * Return 0 on success, non-zero on error - handle is freed in that case.
582 int jbd2_journal_start_reserved(handle_t *handle, unsigned int type,
583 unsigned int line_no)
585 journal_t *journal = handle->h_journal;
588 if (WARN_ON(!handle->h_reserved)) {
589 /* Someone passed in normal handle? Just stop it. */
590 jbd2_journal_stop(handle);
594 * Usefulness of mixing of reserved and unreserved handles is
595 * questionable. So far nobody seems to need it so just error out.
597 if (WARN_ON(current->journal_info)) {
598 jbd2_journal_free_reserved(handle);
602 handle->h_journal = NULL;
604 * GFP_NOFS is here because callers are likely from writeback or
605 * similarly constrained call sites
607 ret = start_this_handle(journal, handle, GFP_NOFS);
609 handle->h_journal = journal;
610 jbd2_journal_free_reserved(handle);
613 handle->h_type = type;
614 handle->h_line_no = line_no;
615 trace_jbd2_handle_start(journal->j_fs_dev->bd_dev,
616 handle->h_transaction->t_tid, type,
617 line_no, handle->h_total_credits);
620 EXPORT_SYMBOL(jbd2_journal_start_reserved);
623 * int jbd2_journal_extend() - extend buffer credits.
624 * @handle: handle to 'extend'
625 * @nblocks: nr blocks to try to extend by.
626 * @revoke_records: number of revoke records to try to extend by.
628 * Some transactions, such as large extends and truncates, can be done
629 * atomically all at once or in several stages. The operation requests
630 * a credit for a number of buffer modifications in advance, but can
631 * extend its credit if it needs more.
633 * jbd2_journal_extend tries to give the running handle more buffer credits.
634 * It does not guarantee that allocation - this is a best-effort only.
635 * The calling process MUST be able to deal cleanly with a failure to
638 * Return 0 on success, non-zero on failure.
640 * return code < 0 implies an error
641 * return code > 0 implies normal transaction-full status.
643 int jbd2_journal_extend(handle_t *handle, int nblocks, int revoke_records)
645 transaction_t *transaction = handle->h_transaction;
650 if (is_handle_aborted(handle))
652 journal = transaction->t_journal;
656 read_lock(&journal->j_state_lock);
658 /* Don't extend a locked-down transaction! */
659 if (transaction->t_state != T_RUNNING) {
660 jbd_debug(3, "denied handle %p %d blocks: "
661 "transaction not running\n", handle, nblocks);
665 nblocks += DIV_ROUND_UP(
666 handle->h_revoke_credits_requested + revoke_records,
667 journal->j_revoke_records_per_block) -
669 handle->h_revoke_credits_requested,
670 journal->j_revoke_records_per_block);
671 spin_lock(&transaction->t_handle_lock);
672 wanted = atomic_add_return(nblocks,
673 &transaction->t_outstanding_credits);
675 if (wanted > journal->j_max_transaction_buffers) {
676 jbd_debug(3, "denied handle %p %d blocks: "
677 "transaction too large\n", handle, nblocks);
678 atomic_sub(nblocks, &transaction->t_outstanding_credits);
682 trace_jbd2_handle_extend(journal->j_fs_dev->bd_dev,
684 handle->h_type, handle->h_line_no,
685 handle->h_total_credits,
688 handle->h_total_credits += nblocks;
689 handle->h_requested_credits += nblocks;
690 handle->h_revoke_credits += revoke_records;
691 handle->h_revoke_credits_requested += revoke_records;
694 jbd_debug(3, "extended handle %p by %d\n", handle, nblocks);
696 spin_unlock(&transaction->t_handle_lock);
698 read_unlock(&journal->j_state_lock);
702 static void stop_this_handle(handle_t *handle)
704 transaction_t *transaction = handle->h_transaction;
705 journal_t *journal = transaction->t_journal;
708 J_ASSERT(journal_current_handle() == handle);
709 J_ASSERT(atomic_read(&transaction->t_updates) > 0);
710 current->journal_info = NULL;
712 * Subtract necessary revoke descriptor blocks from handle credits. We
713 * take care to account only for revoke descriptor blocks the
714 * transaction will really need as large sequences of transactions with
715 * small numbers of revokes are relatively common.
717 revokes = handle->h_revoke_credits_requested - handle->h_revoke_credits;
719 int t_revokes, revoke_descriptors;
720 int rr_per_blk = journal->j_revoke_records_per_block;
722 WARN_ON_ONCE(DIV_ROUND_UP(revokes, rr_per_blk)
723 > handle->h_total_credits);
724 t_revokes = atomic_add_return(revokes,
725 &transaction->t_outstanding_revokes);
727 DIV_ROUND_UP(t_revokes, rr_per_blk) -
728 DIV_ROUND_UP(t_revokes - revokes, rr_per_blk);
729 handle->h_total_credits -= revoke_descriptors;
731 atomic_sub(handle->h_total_credits,
732 &transaction->t_outstanding_credits);
733 if (handle->h_rsv_handle)
734 __jbd2_journal_unreserve_handle(handle->h_rsv_handle,
736 if (atomic_dec_and_test(&transaction->t_updates))
737 wake_up(&journal->j_wait_updates);
739 rwsem_release(&journal->j_trans_commit_map, _THIS_IP_);
741 * Scope of the GFP_NOFS context is over here and so we can restore the
742 * original alloc context.
744 memalloc_nofs_restore(handle->saved_alloc_context);
748 * int jbd2_journal_restart() - restart a handle .
749 * @handle: handle to restart
750 * @nblocks: nr credits requested
751 * @revoke_records: number of revoke record credits requested
752 * @gfp_mask: memory allocation flags (for start_this_handle)
754 * Restart a handle for a multi-transaction filesystem
757 * If the jbd2_journal_extend() call above fails to grant new buffer credits
758 * to a running handle, a call to jbd2_journal_restart will commit the
759 * handle's transaction so far and reattach the handle to a new
760 * transaction capable of guaranteeing the requested number of
761 * credits. We preserve reserved handle if there's any attached to the
764 int jbd2__journal_restart(handle_t *handle, int nblocks, int revoke_records,
767 transaction_t *transaction = handle->h_transaction;
773 /* If we've had an abort of any type, don't even think about
774 * actually doing the restart! */
775 if (is_handle_aborted(handle))
777 journal = transaction->t_journal;
778 tid = transaction->t_tid;
781 * First unlink the handle from its current transaction, and start the
784 jbd_debug(2, "restarting handle %p\n", handle);
785 stop_this_handle(handle);
786 handle->h_transaction = NULL;
789 * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can
790 * get rid of pointless j_state_lock traffic like this.
792 read_lock(&journal->j_state_lock);
793 need_to_start = !tid_geq(journal->j_commit_request, tid);
794 read_unlock(&journal->j_state_lock);
796 jbd2_log_start_commit(journal, tid);
797 handle->h_total_credits = nblocks +
798 DIV_ROUND_UP(revoke_records,
799 journal->j_revoke_records_per_block);
800 handle->h_revoke_credits = revoke_records;
801 ret = start_this_handle(journal, handle, gfp_mask);
802 trace_jbd2_handle_restart(journal->j_fs_dev->bd_dev,
803 ret ? 0 : handle->h_transaction->t_tid,
804 handle->h_type, handle->h_line_no,
805 handle->h_total_credits);
808 EXPORT_SYMBOL(jbd2__journal_restart);
811 int jbd2_journal_restart(handle_t *handle, int nblocks)
813 return jbd2__journal_restart(handle, nblocks, 0, GFP_NOFS);
815 EXPORT_SYMBOL(jbd2_journal_restart);
818 * void jbd2_journal_lock_updates () - establish a transaction barrier.
819 * @journal: Journal to establish a barrier on.
821 * This locks out any further updates from being started, and blocks
822 * until all existing updates have completed, returning only once the
823 * journal is in a quiescent state with no updates running.
825 * The journal lock should not be held on entry.
827 void jbd2_journal_lock_updates(journal_t *journal)
831 jbd2_might_wait_for_commit(journal);
833 write_lock(&journal->j_state_lock);
834 ++journal->j_barrier_count;
836 /* Wait until there are no reserved handles */
837 if (atomic_read(&journal->j_reserved_credits)) {
838 write_unlock(&journal->j_state_lock);
839 wait_event(journal->j_wait_reserved,
840 atomic_read(&journal->j_reserved_credits) == 0);
841 write_lock(&journal->j_state_lock);
844 /* Wait until there are no running updates */
846 transaction_t *transaction = journal->j_running_transaction;
851 spin_lock(&transaction->t_handle_lock);
852 prepare_to_wait(&journal->j_wait_updates, &wait,
853 TASK_UNINTERRUPTIBLE);
854 if (!atomic_read(&transaction->t_updates)) {
855 spin_unlock(&transaction->t_handle_lock);
856 finish_wait(&journal->j_wait_updates, &wait);
859 spin_unlock(&transaction->t_handle_lock);
860 write_unlock(&journal->j_state_lock);
862 finish_wait(&journal->j_wait_updates, &wait);
863 write_lock(&journal->j_state_lock);
865 write_unlock(&journal->j_state_lock);
868 * We have now established a barrier against other normal updates, but
869 * we also need to barrier against other jbd2_journal_lock_updates() calls
870 * to make sure that we serialise special journal-locked operations
873 mutex_lock(&journal->j_barrier);
877 * void jbd2_journal_unlock_updates (journal_t* journal) - release barrier
878 * @journal: Journal to release the barrier on.
880 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
882 * Should be called without the journal lock held.
884 void jbd2_journal_unlock_updates (journal_t *journal)
886 J_ASSERT(journal->j_barrier_count != 0);
888 mutex_unlock(&journal->j_barrier);
889 write_lock(&journal->j_state_lock);
890 --journal->j_barrier_count;
891 write_unlock(&journal->j_state_lock);
892 wake_up(&journal->j_wait_transaction_locked);
895 static void warn_dirty_buffer(struct buffer_head *bh)
898 "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
899 "There's a risk of filesystem corruption in case of system "
901 bh->b_bdev, (unsigned long long)bh->b_blocknr);
904 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
905 static void jbd2_freeze_jh_data(struct journal_head *jh)
910 struct buffer_head *bh = jh2bh(jh);
912 J_EXPECT_JH(jh, buffer_uptodate(bh), "Possible IO failure.\n");
914 offset = offset_in_page(bh->b_data);
915 source = kmap_atomic(page);
916 /* Fire data frozen trigger just before we copy the data */
917 jbd2_buffer_frozen_trigger(jh, source + offset, jh->b_triggers);
918 memcpy(jh->b_frozen_data, source + offset, bh->b_size);
919 kunmap_atomic(source);
922 * Now that the frozen data is saved off, we need to store any matching
925 jh->b_frozen_triggers = jh->b_triggers;
929 * If the buffer is already part of the current transaction, then there
930 * is nothing we need to do. If it is already part of a prior
931 * transaction which we are still committing to disk, then we need to
932 * make sure that we do not overwrite the old copy: we do copy-out to
933 * preserve the copy going to disk. We also account the buffer against
934 * the handle's metadata buffer credits (unless the buffer is already
935 * part of the transaction, that is).
939 do_get_write_access(handle_t *handle, struct journal_head *jh,
942 struct buffer_head *bh;
943 transaction_t *transaction = handle->h_transaction;
946 char *frozen_buffer = NULL;
947 unsigned long start_lock, time_lock;
949 journal = transaction->t_journal;
951 jbd_debug(5, "journal_head %p, force_copy %d\n", jh, force_copy);
953 JBUFFER_TRACE(jh, "entry");
957 /* @@@ Need to check for errors here at some point. */
959 start_lock = jiffies;
961 spin_lock(&jh->b_state_lock);
963 /* If it takes too long to lock the buffer, trace it */
964 time_lock = jbd2_time_diff(start_lock, jiffies);
965 if (time_lock > HZ/10)
966 trace_jbd2_lock_buffer_stall(bh->b_bdev->bd_dev,
967 jiffies_to_msecs(time_lock));
969 /* We now hold the buffer lock so it is safe to query the buffer
970 * state. Is the buffer dirty?
972 * If so, there are two possibilities. The buffer may be
973 * non-journaled, and undergoing a quite legitimate writeback.
974 * Otherwise, it is journaled, and we don't expect dirty buffers
975 * in that state (the buffers should be marked JBD_Dirty
976 * instead.) So either the IO is being done under our own
977 * control and this is a bug, or it's a third party IO such as
978 * dump(8) (which may leave the buffer scheduled for read ---
979 * ie. locked but not dirty) or tune2fs (which may actually have
980 * the buffer dirtied, ugh.) */
982 if (buffer_dirty(bh)) {
984 * First question: is this buffer already part of the current
985 * transaction or the existing committing transaction?
987 if (jh->b_transaction) {
989 jh->b_transaction == transaction ||
991 journal->j_committing_transaction);
992 if (jh->b_next_transaction)
993 J_ASSERT_JH(jh, jh->b_next_transaction ==
995 warn_dirty_buffer(bh);
998 * In any case we need to clean the dirty flag and we must
999 * do it under the buffer lock to be sure we don't race
1000 * with running write-out.
1002 JBUFFER_TRACE(jh, "Journalling dirty buffer");
1003 clear_buffer_dirty(bh);
1004 set_buffer_jbddirty(bh);
1010 if (is_handle_aborted(handle)) {
1011 spin_unlock(&jh->b_state_lock);
1017 * The buffer is already part of this transaction if b_transaction or
1018 * b_next_transaction points to it
1020 if (jh->b_transaction == transaction ||
1021 jh->b_next_transaction == transaction)
1025 * this is the first time this transaction is touching this buffer,
1026 * reset the modified flag
1031 * If the buffer is not journaled right now, we need to make sure it
1032 * doesn't get written to disk before the caller actually commits the
1035 if (!jh->b_transaction) {
1036 JBUFFER_TRACE(jh, "no transaction");
1037 J_ASSERT_JH(jh, !jh->b_next_transaction);
1038 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1040 * Make sure all stores to jh (b_modified, b_frozen_data) are
1041 * visible before attaching it to the running transaction.
1042 * Paired with barrier in jbd2_write_access_granted()
1045 spin_lock(&journal->j_list_lock);
1046 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1047 spin_unlock(&journal->j_list_lock);
1051 * If there is already a copy-out version of this buffer, then we don't
1052 * need to make another one
1054 if (jh->b_frozen_data) {
1055 JBUFFER_TRACE(jh, "has frozen data");
1056 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1060 JBUFFER_TRACE(jh, "owned by older transaction");
1061 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1062 J_ASSERT_JH(jh, jh->b_transaction == journal->j_committing_transaction);
1065 * There is one case we have to be very careful about. If the
1066 * committing transaction is currently writing this buffer out to disk
1067 * and has NOT made a copy-out, then we cannot modify the buffer
1068 * contents at all right now. The essence of copy-out is that it is
1069 * the extra copy, not the primary copy, which gets journaled. If the
1070 * primary copy is already going to disk then we cannot do copy-out
1073 if (buffer_shadow(bh)) {
1074 JBUFFER_TRACE(jh, "on shadow: sleep");
1075 spin_unlock(&jh->b_state_lock);
1076 wait_on_bit_io(&bh->b_state, BH_Shadow, TASK_UNINTERRUPTIBLE);
1081 * Only do the copy if the currently-owning transaction still needs it.
1082 * If buffer isn't on BJ_Metadata list, the committing transaction is
1083 * past that stage (here we use the fact that BH_Shadow is set under
1084 * bh_state lock together with refiling to BJ_Shadow list and at this
1085 * point we know the buffer doesn't have BH_Shadow set).
1087 * Subtle point, though: if this is a get_undo_access, then we will be
1088 * relying on the frozen_data to contain the new value of the
1089 * committed_data record after the transaction, so we HAVE to force the
1090 * frozen_data copy in that case.
1092 if (jh->b_jlist == BJ_Metadata || force_copy) {
1093 JBUFFER_TRACE(jh, "generate frozen data");
1094 if (!frozen_buffer) {
1095 JBUFFER_TRACE(jh, "allocate memory for buffer");
1096 spin_unlock(&jh->b_state_lock);
1097 frozen_buffer = jbd2_alloc(jh2bh(jh)->b_size,
1098 GFP_NOFS | __GFP_NOFAIL);
1101 jh->b_frozen_data = frozen_buffer;
1102 frozen_buffer = NULL;
1103 jbd2_freeze_jh_data(jh);
1107 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1108 * before attaching it to the running transaction. Paired with barrier
1109 * in jbd2_write_access_granted()
1112 jh->b_next_transaction = transaction;
1115 spin_unlock(&jh->b_state_lock);
1118 * If we are about to journal a buffer, then any revoke pending on it is
1121 jbd2_journal_cancel_revoke(handle, jh);
1124 if (unlikely(frozen_buffer)) /* It's usually NULL */
1125 jbd2_free(frozen_buffer, bh->b_size);
1127 JBUFFER_TRACE(jh, "exit");
1131 /* Fast check whether buffer is already attached to the required transaction */
1132 static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
1135 struct journal_head *jh;
1138 /* Dirty buffers require special handling... */
1139 if (buffer_dirty(bh))
1143 * RCU protects us from dereferencing freed pages. So the checks we do
1144 * are guaranteed not to oops. However the jh slab object can get freed
1145 * & reallocated while we work with it. So we have to be careful. When
1146 * we see jh attached to the running transaction, we know it must stay
1147 * so until the transaction is committed. Thus jh won't be freed and
1148 * will be attached to the same bh while we run. However it can
1149 * happen jh gets freed, reallocated, and attached to the transaction
1150 * just after we get pointer to it from bh. So we have to be careful
1151 * and recheck jh still belongs to our bh before we return success.
1154 if (!buffer_jbd(bh))
1156 /* This should be bh2jh() but that doesn't work with inline functions */
1157 jh = READ_ONCE(bh->b_private);
1160 /* For undo access buffer must have data copied */
1161 if (undo && !jh->b_committed_data)
1163 if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
1164 READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
1167 * There are two reasons for the barrier here:
1168 * 1) Make sure to fetch b_bh after we did previous checks so that we
1169 * detect when jh went through free, realloc, attach to transaction
1170 * while we were checking. Paired with implicit barrier in that path.
1171 * 2) So that access to bh done after jbd2_write_access_granted()
1172 * doesn't get reordered and see inconsistent state of concurrent
1173 * do_get_write_access().
1176 if (unlikely(jh->b_bh != bh))
1185 * int jbd2_journal_get_write_access() - notify intent to modify a buffer for metadata (not data) update.
1186 * @handle: transaction to add buffer modifications to
1187 * @bh: bh to be used for metadata writes
1189 * Returns: error code or 0 on success.
1191 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1192 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1195 int jbd2_journal_get_write_access(handle_t *handle, struct buffer_head *bh)
1197 struct journal_head *jh;
1200 if (is_handle_aborted(handle))
1203 if (jbd2_write_access_granted(handle, bh, false))
1206 jh = jbd2_journal_add_journal_head(bh);
1207 /* We do not want to get caught playing with fields which the
1208 * log thread also manipulates. Make sure that the buffer
1209 * completes any outstanding IO before proceeding. */
1210 rc = do_get_write_access(handle, jh, 0);
1211 jbd2_journal_put_journal_head(jh);
1217 * When the user wants to journal a newly created buffer_head
1218 * (ie. getblk() returned a new buffer and we are going to populate it
1219 * manually rather than reading off disk), then we need to keep the
1220 * buffer_head locked until it has been completely filled with new
1221 * data. In this case, we should be able to make the assertion that
1222 * the bh is not already part of an existing transaction.
1224 * The buffer should already be locked by the caller by this point.
1225 * There is no lock ranking violation: it was a newly created,
1226 * unlocked buffer beforehand. */
1229 * int jbd2_journal_get_create_access () - notify intent to use newly created bh
1230 * @handle: transaction to new buffer to
1233 * Call this if you create a new bh.
1235 int jbd2_journal_get_create_access(handle_t *handle, struct buffer_head *bh)
1237 transaction_t *transaction = handle->h_transaction;
1239 struct journal_head *jh = jbd2_journal_add_journal_head(bh);
1242 jbd_debug(5, "journal_head %p\n", jh);
1244 if (is_handle_aborted(handle))
1246 journal = transaction->t_journal;
1249 JBUFFER_TRACE(jh, "entry");
1251 * The buffer may already belong to this transaction due to pre-zeroing
1252 * in the filesystem's new_block code. It may also be on the previous,
1253 * committing transaction's lists, but it HAS to be in Forget state in
1254 * that case: the transaction must have deleted the buffer for it to be
1257 spin_lock(&jh->b_state_lock);
1258 J_ASSERT_JH(jh, (jh->b_transaction == transaction ||
1259 jh->b_transaction == NULL ||
1260 (jh->b_transaction == journal->j_committing_transaction &&
1261 jh->b_jlist == BJ_Forget)));
1263 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
1264 J_ASSERT_JH(jh, buffer_locked(jh2bh(jh)));
1266 if (jh->b_transaction == NULL) {
1268 * Previous jbd2_journal_forget() could have left the buffer
1269 * with jbddirty bit set because it was being committed. When
1270 * the commit finished, we've filed the buffer for
1271 * checkpointing and marked it dirty. Now we are reallocating
1272 * the buffer so the transaction freeing it must have
1273 * committed and so it's safe to clear the dirty bit.
1275 clear_buffer_dirty(jh2bh(jh));
1276 /* first access by this transaction */
1279 JBUFFER_TRACE(jh, "file as BJ_Reserved");
1280 spin_lock(&journal->j_list_lock);
1281 __jbd2_journal_file_buffer(jh, transaction, BJ_Reserved);
1282 spin_unlock(&journal->j_list_lock);
1283 } else if (jh->b_transaction == journal->j_committing_transaction) {
1284 /* first access by this transaction */
1287 JBUFFER_TRACE(jh, "set next transaction");
1288 spin_lock(&journal->j_list_lock);
1289 jh->b_next_transaction = transaction;
1290 spin_unlock(&journal->j_list_lock);
1292 spin_unlock(&jh->b_state_lock);
1295 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1296 * blocks which contain freed but then revoked metadata. We need
1297 * to cancel the revoke in case we end up freeing it yet again
1298 * and the reallocating as data - this would cause a second revoke,
1299 * which hits an assertion error.
1301 JBUFFER_TRACE(jh, "cancelling revoke");
1302 jbd2_journal_cancel_revoke(handle, jh);
1304 jbd2_journal_put_journal_head(jh);
1309 * int jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1310 * non-rewindable consequences
1311 * @handle: transaction
1312 * @bh: buffer to undo
1314 * Sometimes there is a need to distinguish between metadata which has
1315 * been committed to disk and that which has not. The ext3fs code uses
1316 * this for freeing and allocating space, we have to make sure that we
1317 * do not reuse freed space until the deallocation has been committed,
1318 * since if we overwrote that space we would make the delete
1319 * un-rewindable in case of a crash.
1321 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1322 * buffer for parts of non-rewindable operations such as delete
1323 * operations on the bitmaps. The journaling code must keep a copy of
1324 * the buffer's contents prior to the undo_access call until such time
1325 * as we know that the buffer has definitely been committed to disk.
1327 * We never need to know which transaction the committed data is part
1328 * of, buffers touched here are guaranteed to be dirtied later and so
1329 * will be committed to a new transaction in due course, at which point
1330 * we can discard the old committed data pointer.
1332 * Returns error number or 0 on success.
1334 int jbd2_journal_get_undo_access(handle_t *handle, struct buffer_head *bh)
1337 struct journal_head *jh;
1338 char *committed_data = NULL;
1340 if (is_handle_aborted(handle))
1343 if (jbd2_write_access_granted(handle, bh, true))
1346 jh = jbd2_journal_add_journal_head(bh);
1347 JBUFFER_TRACE(jh, "entry");
1350 * Do this first --- it can drop the journal lock, so we want to
1351 * make sure that obtaining the committed_data is done
1352 * atomically wrt. completion of any outstanding commits.
1354 err = do_get_write_access(handle, jh, 1);
1359 if (!jh->b_committed_data)
1360 committed_data = jbd2_alloc(jh2bh(jh)->b_size,
1361 GFP_NOFS|__GFP_NOFAIL);
1363 spin_lock(&jh->b_state_lock);
1364 if (!jh->b_committed_data) {
1365 /* Copy out the current buffer contents into the
1366 * preserved, committed copy. */
1367 JBUFFER_TRACE(jh, "generate b_committed data");
1368 if (!committed_data) {
1369 spin_unlock(&jh->b_state_lock);
1373 jh->b_committed_data = committed_data;
1374 committed_data = NULL;
1375 memcpy(jh->b_committed_data, bh->b_data, bh->b_size);
1377 spin_unlock(&jh->b_state_lock);
1379 jbd2_journal_put_journal_head(jh);
1380 if (unlikely(committed_data))
1381 jbd2_free(committed_data, bh->b_size);
1386 * void jbd2_journal_set_triggers() - Add triggers for commit writeout
1387 * @bh: buffer to trigger on
1388 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1390 * Set any triggers on this journal_head. This is always safe, because
1391 * triggers for a committing buffer will be saved off, and triggers for
1392 * a running transaction will match the buffer in that transaction.
1394 * Call with NULL to clear the triggers.
1396 void jbd2_journal_set_triggers(struct buffer_head *bh,
1397 struct jbd2_buffer_trigger_type *type)
1399 struct journal_head *jh = jbd2_journal_grab_journal_head(bh);
1403 jh->b_triggers = type;
1404 jbd2_journal_put_journal_head(jh);
1407 void jbd2_buffer_frozen_trigger(struct journal_head *jh, void *mapped_data,
1408 struct jbd2_buffer_trigger_type *triggers)
1410 struct buffer_head *bh = jh2bh(jh);
1412 if (!triggers || !triggers->t_frozen)
1415 triggers->t_frozen(triggers, bh, mapped_data, bh->b_size);
1418 void jbd2_buffer_abort_trigger(struct journal_head *jh,
1419 struct jbd2_buffer_trigger_type *triggers)
1421 if (!triggers || !triggers->t_abort)
1424 triggers->t_abort(triggers, jh2bh(jh));
1428 * int jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1429 * @handle: transaction to add buffer to.
1430 * @bh: buffer to mark
1432 * mark dirty metadata which needs to be journaled as part of the current
1435 * The buffer must have previously had jbd2_journal_get_write_access()
1436 * called so that it has a valid journal_head attached to the buffer
1439 * The buffer is placed on the transaction's metadata list and is marked
1440 * as belonging to the transaction.
1442 * Returns error number or 0 on success.
1444 * Special care needs to be taken if the buffer already belongs to the
1445 * current committing transaction (in which case we should have frozen
1446 * data present for that commit). In that case, we don't relink the
1447 * buffer: that only gets done when the old transaction finally
1448 * completes its commit.
1450 int jbd2_journal_dirty_metadata(handle_t *handle, struct buffer_head *bh)
1452 transaction_t *transaction = handle->h_transaction;
1454 struct journal_head *jh;
1457 if (is_handle_aborted(handle))
1459 if (!buffer_jbd(bh))
1463 * We don't grab jh reference here since the buffer must be part
1464 * of the running transaction.
1467 jbd_debug(5, "journal_head %p\n", jh);
1468 JBUFFER_TRACE(jh, "entry");
1471 * This and the following assertions are unreliable since we may see jh
1472 * in inconsistent state unless we grab bh_state lock. But this is
1473 * crucial to catch bugs so let's do a reliable check until the
1474 * lockless handling is fully proven.
1476 if (jh->b_transaction != transaction &&
1477 jh->b_next_transaction != transaction) {
1478 spin_lock(&jh->b_state_lock);
1479 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
1480 jh->b_next_transaction == transaction);
1481 spin_unlock(&jh->b_state_lock);
1483 if (jh->b_modified == 1) {
1484 /* If it's in our transaction it must be in BJ_Metadata list. */
1485 if (jh->b_transaction == transaction &&
1486 jh->b_jlist != BJ_Metadata) {
1487 spin_lock(&jh->b_state_lock);
1488 if (jh->b_transaction == transaction &&
1489 jh->b_jlist != BJ_Metadata)
1490 pr_err("JBD2: assertion failure: h_type=%u "
1491 "h_line_no=%u block_no=%llu jlist=%u\n",
1492 handle->h_type, handle->h_line_no,
1493 (unsigned long long) bh->b_blocknr,
1495 J_ASSERT_JH(jh, jh->b_transaction != transaction ||
1496 jh->b_jlist == BJ_Metadata);
1497 spin_unlock(&jh->b_state_lock);
1502 journal = transaction->t_journal;
1503 spin_lock(&jh->b_state_lock);
1505 if (jh->b_modified == 0) {
1507 * This buffer's got modified and becoming part
1508 * of the transaction. This needs to be done
1509 * once a transaction -bzzz
1511 if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle) <= 0)) {
1516 handle->h_total_credits--;
1520 * fastpath, to avoid expensive locking. If this buffer is already
1521 * on the running transaction's metadata list there is nothing to do.
1522 * Nobody can take it off again because there is a handle open.
1523 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1524 * result in this test being false, so we go in and take the locks.
1526 if (jh->b_transaction == transaction && jh->b_jlist == BJ_Metadata) {
1527 JBUFFER_TRACE(jh, "fastpath");
1528 if (unlikely(jh->b_transaction !=
1529 journal->j_running_transaction)) {
1530 printk(KERN_ERR "JBD2: %s: "
1531 "jh->b_transaction (%llu, %p, %u) != "
1532 "journal->j_running_transaction (%p, %u)\n",
1534 (unsigned long long) bh->b_blocknr,
1536 jh->b_transaction ? jh->b_transaction->t_tid : 0,
1537 journal->j_running_transaction,
1538 journal->j_running_transaction ?
1539 journal->j_running_transaction->t_tid : 0);
1545 set_buffer_jbddirty(bh);
1548 * Metadata already on the current transaction list doesn't
1549 * need to be filed. Metadata on another transaction's list must
1550 * be committing, and will be refiled once the commit completes:
1551 * leave it alone for now.
1553 if (jh->b_transaction != transaction) {
1554 JBUFFER_TRACE(jh, "already on other transaction");
1555 if (unlikely(((jh->b_transaction !=
1556 journal->j_committing_transaction)) ||
1557 (jh->b_next_transaction != transaction))) {
1558 printk(KERN_ERR "jbd2_journal_dirty_metadata: %s: "
1559 "bad jh for block %llu: "
1560 "transaction (%p, %u), "
1561 "jh->b_transaction (%p, %u), "
1562 "jh->b_next_transaction (%p, %u), jlist %u\n",
1564 (unsigned long long) bh->b_blocknr,
1565 transaction, transaction->t_tid,
1568 jh->b_transaction->t_tid : 0,
1569 jh->b_next_transaction,
1570 jh->b_next_transaction ?
1571 jh->b_next_transaction->t_tid : 0,
1576 /* And this case is illegal: we can't reuse another
1577 * transaction's data buffer, ever. */
1581 /* That test should have eliminated the following case: */
1582 J_ASSERT_JH(jh, jh->b_frozen_data == NULL);
1584 JBUFFER_TRACE(jh, "file as BJ_Metadata");
1585 spin_lock(&journal->j_list_lock);
1586 __jbd2_journal_file_buffer(jh, transaction, BJ_Metadata);
1587 spin_unlock(&journal->j_list_lock);
1589 spin_unlock(&jh->b_state_lock);
1591 JBUFFER_TRACE(jh, "exit");
1596 * void jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1597 * @handle: transaction handle
1598 * @bh: bh to 'forget'
1600 * We can only do the bforget if there are no commits pending against the
1601 * buffer. If the buffer is dirty in the current running transaction we
1602 * can safely unlink it.
1604 * bh may not be a journalled buffer at all - it may be a non-JBD
1605 * buffer which came off the hashtable. Check for this.
1607 * Decrements bh->b_count by one.
1609 * Allow this call even if the handle has aborted --- it may be part of
1610 * the caller's cleanup after an abort.
1612 int jbd2_journal_forget(handle_t *handle, struct buffer_head *bh)
1614 transaction_t *transaction = handle->h_transaction;
1616 struct journal_head *jh;
1617 int drop_reserve = 0;
1619 int was_modified = 0;
1621 if (is_handle_aborted(handle))
1623 journal = transaction->t_journal;
1625 BUFFER_TRACE(bh, "entry");
1627 jh = jbd2_journal_grab_journal_head(bh);
1633 spin_lock(&jh->b_state_lock);
1635 /* Critical error: attempting to delete a bitmap buffer, maybe?
1636 * Don't do any jbd operations, and return an error. */
1637 if (!J_EXPECT_JH(jh, !jh->b_committed_data,
1638 "inconsistent data on disk")) {
1643 /* keep track of whether or not this transaction modified us */
1644 was_modified = jh->b_modified;
1647 * The buffer's going from the transaction, we must drop
1648 * all references -bzzz
1652 if (jh->b_transaction == transaction) {
1653 J_ASSERT_JH(jh, !jh->b_frozen_data);
1655 /* If we are forgetting a buffer which is already part
1656 * of this transaction, then we can just drop it from
1657 * the transaction immediately. */
1658 clear_buffer_dirty(bh);
1659 clear_buffer_jbddirty(bh);
1661 JBUFFER_TRACE(jh, "belongs to current transaction: unfile");
1664 * we only want to drop a reference if this transaction
1665 * modified the buffer
1671 * We are no longer going to journal this buffer.
1672 * However, the commit of this transaction is still
1673 * important to the buffer: the delete that we are now
1674 * processing might obsolete an old log entry, so by
1675 * committing, we can satisfy the buffer's checkpoint.
1677 * So, if we have a checkpoint on the buffer, we should
1678 * now refile the buffer on our BJ_Forget list so that
1679 * we know to remove the checkpoint after we commit.
1682 spin_lock(&journal->j_list_lock);
1683 if (jh->b_cp_transaction) {
1684 __jbd2_journal_temp_unlink_buffer(jh);
1685 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1687 __jbd2_journal_unfile_buffer(jh);
1688 jbd2_journal_put_journal_head(jh);
1690 spin_unlock(&journal->j_list_lock);
1691 } else if (jh->b_transaction) {
1692 J_ASSERT_JH(jh, (jh->b_transaction ==
1693 journal->j_committing_transaction));
1694 /* However, if the buffer is still owned by a prior
1695 * (committing) transaction, we can't drop it yet... */
1696 JBUFFER_TRACE(jh, "belongs to older transaction");
1697 /* ... but we CAN drop it from the new transaction through
1698 * marking the buffer as freed and set j_next_transaction to
1699 * the new transaction, so that not only the commit code
1700 * knows it should clear dirty bits when it is done with the
1701 * buffer, but also the buffer can be checkpointed only
1702 * after the new transaction commits. */
1704 set_buffer_freed(bh);
1706 if (!jh->b_next_transaction) {
1707 spin_lock(&journal->j_list_lock);
1708 jh->b_next_transaction = transaction;
1709 spin_unlock(&journal->j_list_lock);
1711 J_ASSERT(jh->b_next_transaction == transaction);
1714 * only drop a reference if this transaction modified
1722 * Finally, if the buffer is not belongs to any
1723 * transaction, we can just drop it now if it has no
1726 spin_lock(&journal->j_list_lock);
1727 if (!jh->b_cp_transaction) {
1728 JBUFFER_TRACE(jh, "belongs to none transaction");
1729 spin_unlock(&journal->j_list_lock);
1734 * Otherwise, if the buffer has been written to disk,
1735 * it is safe to remove the checkpoint and drop it.
1737 if (!buffer_dirty(bh)) {
1738 __jbd2_journal_remove_checkpoint(jh);
1739 spin_unlock(&journal->j_list_lock);
1744 * The buffer is still not written to disk, we should
1745 * attach this buffer to current transaction so that the
1746 * buffer can be checkpointed only after the current
1747 * transaction commits.
1749 clear_buffer_dirty(bh);
1750 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
1751 spin_unlock(&journal->j_list_lock);
1755 spin_unlock(&jh->b_state_lock);
1756 jbd2_journal_put_journal_head(jh);
1758 /* no need to reserve log space for this block -bzzz */
1759 handle->h_total_credits++;
1765 * int jbd2_journal_stop() - complete a transaction
1766 * @handle: transaction to complete.
1768 * All done for a particular handle.
1770 * There is not much action needed here. We just return any remaining
1771 * buffer credits to the transaction and remove the handle. The only
1772 * complication is that we need to start a commit operation if the
1773 * filesystem is marked for synchronous update.
1775 * jbd2_journal_stop itself will not usually return an error, but it may
1776 * do so in unusual circumstances. In particular, expect it to
1777 * return -EIO if a jbd2_journal_abort has been executed since the
1778 * transaction began.
1780 int jbd2_journal_stop(handle_t *handle)
1782 transaction_t *transaction = handle->h_transaction;
1784 int err = 0, wait_for_commit = 0;
1788 if (--handle->h_ref > 0) {
1789 jbd_debug(4, "h_ref %d -> %d\n", handle->h_ref + 1,
1791 if (is_handle_aborted(handle))
1797 * Handle is already detached from the transaction so there is
1798 * nothing to do other than free the handle.
1800 memalloc_nofs_restore(handle->saved_alloc_context);
1803 journal = transaction->t_journal;
1804 tid = transaction->t_tid;
1806 if (is_handle_aborted(handle))
1809 jbd_debug(4, "Handle %p going down\n", handle);
1810 trace_jbd2_handle_stats(journal->j_fs_dev->bd_dev,
1811 tid, handle->h_type, handle->h_line_no,
1812 jiffies - handle->h_start_jiffies,
1813 handle->h_sync, handle->h_requested_credits,
1814 (handle->h_requested_credits -
1815 handle->h_total_credits));
1818 * Implement synchronous transaction batching. If the handle
1819 * was synchronous, don't force a commit immediately. Let's
1820 * yield and let another thread piggyback onto this
1821 * transaction. Keep doing that while new threads continue to
1822 * arrive. It doesn't cost much - we're about to run a commit
1823 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1824 * operations by 30x or more...
1826 * We try and optimize the sleep time against what the
1827 * underlying disk can do, instead of having a static sleep
1828 * time. This is useful for the case where our storage is so
1829 * fast that it is more optimal to go ahead and force a flush
1830 * and wait for the transaction to be committed than it is to
1831 * wait for an arbitrary amount of time for new writers to
1832 * join the transaction. We achieve this by measuring how
1833 * long it takes to commit a transaction, and compare it with
1834 * how long this transaction has been running, and if run time
1835 * < commit time then we sleep for the delta and commit. This
1836 * greatly helps super fast disks that would see slowdowns as
1837 * more threads started doing fsyncs.
1839 * But don't do this if this process was the most recent one
1840 * to perform a synchronous write. We do this to detect the
1841 * case where a single process is doing a stream of sync
1842 * writes. No point in waiting for joiners in that case.
1844 * Setting max_batch_time to 0 disables this completely.
1847 if (handle->h_sync && journal->j_last_sync_writer != pid &&
1848 journal->j_max_batch_time) {
1849 u64 commit_time, trans_time;
1851 journal->j_last_sync_writer = pid;
1853 read_lock(&journal->j_state_lock);
1854 commit_time = journal->j_average_commit_time;
1855 read_unlock(&journal->j_state_lock);
1857 trans_time = ktime_to_ns(ktime_sub(ktime_get(),
1858 transaction->t_start_time));
1860 commit_time = max_t(u64, commit_time,
1861 1000*journal->j_min_batch_time);
1862 commit_time = min_t(u64, commit_time,
1863 1000*journal->j_max_batch_time);
1865 if (trans_time < commit_time) {
1866 ktime_t expires = ktime_add_ns(ktime_get(),
1868 set_current_state(TASK_UNINTERRUPTIBLE);
1869 schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
1874 transaction->t_synchronous_commit = 1;
1877 * If the handle is marked SYNC, we need to set another commit
1878 * going! We also want to force a commit if the transaction is too
1881 if (handle->h_sync ||
1882 time_after_eq(jiffies, transaction->t_expires)) {
1883 /* Do this even for aborted journals: an abort still
1884 * completes the commit thread, it just doesn't write
1885 * anything to disk. */
1887 jbd_debug(2, "transaction too old, requesting commit for "
1888 "handle %p\n", handle);
1889 /* This is non-blocking */
1890 jbd2_log_start_commit(journal, tid);
1893 * Special case: JBD2_SYNC synchronous updates require us
1894 * to wait for the commit to complete.
1896 if (handle->h_sync && !(current->flags & PF_MEMALLOC))
1897 wait_for_commit = 1;
1901 * Once stop_this_handle() drops t_updates, the transaction could start
1902 * committing on us and eventually disappear. So we must not
1903 * dereference transaction pointer again after calling
1904 * stop_this_handle().
1906 stop_this_handle(handle);
1908 if (wait_for_commit)
1909 err = jbd2_log_wait_commit(journal, tid);
1912 if (handle->h_rsv_handle)
1913 jbd2_free_handle(handle->h_rsv_handle);
1914 jbd2_free_handle(handle);
1920 * List management code snippets: various functions for manipulating the
1921 * transaction buffer lists.
1926 * Append a buffer to a transaction list, given the transaction's list head
1929 * j_list_lock is held.
1931 * jh->b_state_lock is held.
1935 __blist_add_buffer(struct journal_head **list, struct journal_head *jh)
1938 jh->b_tnext = jh->b_tprev = jh;
1941 /* Insert at the tail of the list to preserve order */
1942 struct journal_head *first = *list, *last = first->b_tprev;
1944 jh->b_tnext = first;
1945 last->b_tnext = first->b_tprev = jh;
1950 * Remove a buffer from a transaction list, given the transaction's list
1953 * Called with j_list_lock held, and the journal may not be locked.
1955 * jh->b_state_lock is held.
1959 __blist_del_buffer(struct journal_head **list, struct journal_head *jh)
1962 *list = jh->b_tnext;
1966 jh->b_tprev->b_tnext = jh->b_tnext;
1967 jh->b_tnext->b_tprev = jh->b_tprev;
1971 * Remove a buffer from the appropriate transaction list.
1973 * Note that this function can *change* the value of
1974 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1975 * t_reserved_list. If the caller is holding onto a copy of one of these
1976 * pointers, it could go bad. Generally the caller needs to re-read the
1977 * pointer from the transaction_t.
1979 * Called under j_list_lock.
1981 static void __jbd2_journal_temp_unlink_buffer(struct journal_head *jh)
1983 struct journal_head **list = NULL;
1984 transaction_t *transaction;
1985 struct buffer_head *bh = jh2bh(jh);
1987 lockdep_assert_held(&jh->b_state_lock);
1988 transaction = jh->b_transaction;
1990 assert_spin_locked(&transaction->t_journal->j_list_lock);
1992 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
1993 if (jh->b_jlist != BJ_None)
1994 J_ASSERT_JH(jh, transaction != NULL);
1996 switch (jh->b_jlist) {
2000 transaction->t_nr_buffers--;
2001 J_ASSERT_JH(jh, transaction->t_nr_buffers >= 0);
2002 list = &transaction->t_buffers;
2005 list = &transaction->t_forget;
2008 list = &transaction->t_shadow_list;
2011 list = &transaction->t_reserved_list;
2015 __blist_del_buffer(list, jh);
2016 jh->b_jlist = BJ_None;
2017 if (transaction && is_journal_aborted(transaction->t_journal))
2018 clear_buffer_jbddirty(bh);
2019 else if (test_clear_buffer_jbddirty(bh))
2020 mark_buffer_dirty(bh); /* Expose it to the VM */
2024 * Remove buffer from all transactions. The caller is responsible for dropping
2025 * the jh reference that belonged to the transaction.
2027 * Called with bh_state lock and j_list_lock
2029 static void __jbd2_journal_unfile_buffer(struct journal_head *jh)
2031 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2032 J_ASSERT_JH(jh, jh->b_next_transaction == NULL);
2034 __jbd2_journal_temp_unlink_buffer(jh);
2035 jh->b_transaction = NULL;
2038 void jbd2_journal_unfile_buffer(journal_t *journal, struct journal_head *jh)
2040 struct buffer_head *bh = jh2bh(jh);
2042 /* Get reference so that buffer cannot be freed before we unlock it */
2044 spin_lock(&jh->b_state_lock);
2045 spin_lock(&journal->j_list_lock);
2046 __jbd2_journal_unfile_buffer(jh);
2047 spin_unlock(&journal->j_list_lock);
2048 spin_unlock(&jh->b_state_lock);
2049 jbd2_journal_put_journal_head(jh);
2054 * Called from jbd2_journal_try_to_free_buffers().
2056 * Called under jh->b_state_lock
2059 __journal_try_to_free_buffer(journal_t *journal, struct buffer_head *bh)
2061 struct journal_head *jh;
2065 if (buffer_locked(bh) || buffer_dirty(bh))
2068 if (jh->b_next_transaction != NULL || jh->b_transaction != NULL)
2071 spin_lock(&journal->j_list_lock);
2072 if (jh->b_cp_transaction != NULL) {
2073 /* written-back checkpointed metadata buffer */
2074 JBUFFER_TRACE(jh, "remove from checkpoint list");
2075 __jbd2_journal_remove_checkpoint(jh);
2077 spin_unlock(&journal->j_list_lock);
2083 * int jbd2_journal_try_to_free_buffers() - try to free page buffers.
2084 * @journal: journal for operation
2085 * @page: to try and free
2087 * For all the buffers on this page,
2088 * if they are fully written out ordered data, move them onto BUF_CLEAN
2089 * so try_to_free_buffers() can reap them.
2091 * This function returns non-zero if we wish try_to_free_buffers()
2092 * to be called. We do this if the page is releasable by try_to_free_buffers().
2093 * We also do it if the page has locked or dirty buffers and the caller wants
2094 * us to perform sync or async writeout.
2096 * This complicates JBD locking somewhat. We aren't protected by the
2097 * BKL here. We wish to remove the buffer from its committing or
2098 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2100 * This may *change* the value of transaction_t->t_datalist, so anyone
2101 * who looks at t_datalist needs to lock against this function.
2103 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2104 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
2105 * will come out of the lock with the buffer dirty, which makes it
2106 * ineligible for release here.
2108 * Who else is affected by this? hmm... Really the only contender
2109 * is do_get_write_access() - it could be looking at the buffer while
2110 * journal_try_to_free_buffer() is changing its state. But that
2111 * cannot happen because we never reallocate freed data as metadata
2112 * while the data is part of a transaction. Yes?
2114 * Return 0 on failure, 1 on success
2116 int jbd2_journal_try_to_free_buffers(journal_t *journal, struct page *page)
2118 struct buffer_head *head;
2119 struct buffer_head *bh;
2120 bool has_write_io_error = false;
2123 J_ASSERT(PageLocked(page));
2125 head = page_buffers(page);
2128 struct journal_head *jh;
2131 * We take our own ref against the journal_head here to avoid
2132 * having to add tons of locking around each instance of
2133 * jbd2_journal_put_journal_head().
2135 jh = jbd2_journal_grab_journal_head(bh);
2139 spin_lock(&jh->b_state_lock);
2140 __journal_try_to_free_buffer(journal, bh);
2141 spin_unlock(&jh->b_state_lock);
2142 jbd2_journal_put_journal_head(jh);
2147 * If we free a metadata buffer which has been failed to
2148 * write out, the jbd2 checkpoint procedure will not detect
2149 * this failure and may lead to filesystem inconsistency
2150 * after cleanup journal tail.
2152 if (buffer_write_io_error(bh)) {
2153 pr_err("JBD2: Error while async write back metadata bh %llu.",
2154 (unsigned long long)bh->b_blocknr);
2155 has_write_io_error = true;
2157 } while ((bh = bh->b_this_page) != head);
2159 ret = try_to_free_buffers(page);
2162 if (has_write_io_error)
2163 jbd2_journal_abort(journal, -EIO);
2169 * This buffer is no longer needed. If it is on an older transaction's
2170 * checkpoint list we need to record it on this transaction's forget list
2171 * to pin this buffer (and hence its checkpointing transaction) down until
2172 * this transaction commits. If the buffer isn't on a checkpoint list, we
2174 * Returns non-zero if JBD no longer has an interest in the buffer.
2176 * Called under j_list_lock.
2178 * Called under jh->b_state_lock.
2180 static int __dispose_buffer(struct journal_head *jh, transaction_t *transaction)
2183 struct buffer_head *bh = jh2bh(jh);
2185 if (jh->b_cp_transaction) {
2186 JBUFFER_TRACE(jh, "on running+cp transaction");
2187 __jbd2_journal_temp_unlink_buffer(jh);
2189 * We don't want to write the buffer anymore, clear the
2190 * bit so that we don't confuse checks in
2191 * __journal_file_buffer
2193 clear_buffer_dirty(bh);
2194 __jbd2_journal_file_buffer(jh, transaction, BJ_Forget);
2197 JBUFFER_TRACE(jh, "on running transaction");
2198 __jbd2_journal_unfile_buffer(jh);
2199 jbd2_journal_put_journal_head(jh);
2205 * jbd2_journal_invalidatepage
2207 * This code is tricky. It has a number of cases to deal with.
2209 * There are two invariants which this code relies on:
2211 * i_size must be updated on disk before we start calling invalidatepage on the
2214 * This is done in ext3 by defining an ext3_setattr method which
2215 * updates i_size before truncate gets going. By maintaining this
2216 * invariant, we can be sure that it is safe to throw away any buffers
2217 * attached to the current transaction: once the transaction commits,
2218 * we know that the data will not be needed.
2220 * Note however that we can *not* throw away data belonging to the
2221 * previous, committing transaction!
2223 * Any disk blocks which *are* part of the previous, committing
2224 * transaction (and which therefore cannot be discarded immediately) are
2225 * not going to be reused in the new running transaction
2227 * The bitmap committed_data images guarantee this: any block which is
2228 * allocated in one transaction and removed in the next will be marked
2229 * as in-use in the committed_data bitmap, so cannot be reused until
2230 * the next transaction to delete the block commits. This means that
2231 * leaving committing buffers dirty is quite safe: the disk blocks
2232 * cannot be reallocated to a different file and so buffer aliasing is
2236 * The above applies mainly to ordered data mode. In writeback mode we
2237 * don't make guarantees about the order in which data hits disk --- in
2238 * particular we don't guarantee that new dirty data is flushed before
2239 * transaction commit --- so it is always safe just to discard data
2240 * immediately in that mode. --sct
2244 * The journal_unmap_buffer helper function returns zero if the buffer
2245 * concerned remains pinned as an anonymous buffer belonging to an older
2248 * We're outside-transaction here. Either or both of j_running_transaction
2249 * and j_committing_transaction may be NULL.
2251 static int journal_unmap_buffer(journal_t *journal, struct buffer_head *bh,
2254 transaction_t *transaction;
2255 struct journal_head *jh;
2258 BUFFER_TRACE(bh, "entry");
2261 * It is safe to proceed here without the j_list_lock because the
2262 * buffers cannot be stolen by try_to_free_buffers as long as we are
2263 * holding the page lock. --sct
2266 jh = jbd2_journal_grab_journal_head(bh);
2268 goto zap_buffer_unlocked;
2270 /* OK, we have data buffer in journaled mode */
2271 write_lock(&journal->j_state_lock);
2272 spin_lock(&jh->b_state_lock);
2273 spin_lock(&journal->j_list_lock);
2276 * We cannot remove the buffer from checkpoint lists until the
2277 * transaction adding inode to orphan list (let's call it T)
2278 * is committed. Otherwise if the transaction changing the
2279 * buffer would be cleaned from the journal before T is
2280 * committed, a crash will cause that the correct contents of
2281 * the buffer will be lost. On the other hand we have to
2282 * clear the buffer dirty bit at latest at the moment when the
2283 * transaction marking the buffer as freed in the filesystem
2284 * structures is committed because from that moment on the
2285 * block can be reallocated and used by a different page.
2286 * Since the block hasn't been freed yet but the inode has
2287 * already been added to orphan list, it is safe for us to add
2288 * the buffer to BJ_Forget list of the newest transaction.
2290 * Also we have to clear buffer_mapped flag of a truncated buffer
2291 * because the buffer_head may be attached to the page straddling
2292 * i_size (can happen only when blocksize < pagesize) and thus the
2293 * buffer_head can be reused when the file is extended again. So we end
2294 * up keeping around invalidated buffers attached to transactions'
2295 * BJ_Forget list just to stop checkpointing code from cleaning up
2296 * the transaction this buffer was modified in.
2298 transaction = jh->b_transaction;
2299 if (transaction == NULL) {
2300 /* First case: not on any transaction. If it
2301 * has no checkpoint link, then we can zap it:
2302 * it's a writeback-mode buffer so we don't care
2303 * if it hits disk safely. */
2304 if (!jh->b_cp_transaction) {
2305 JBUFFER_TRACE(jh, "not on any transaction: zap");
2309 if (!buffer_dirty(bh)) {
2310 /* bdflush has written it. We can drop it now */
2311 __jbd2_journal_remove_checkpoint(jh);
2315 /* OK, it must be in the journal but still not
2316 * written fully to disk: it's metadata or
2317 * journaled data... */
2319 if (journal->j_running_transaction) {
2320 /* ... and once the current transaction has
2321 * committed, the buffer won't be needed any
2323 JBUFFER_TRACE(jh, "checkpointed: add to BJ_Forget");
2324 may_free = __dispose_buffer(jh,
2325 journal->j_running_transaction);
2328 /* There is no currently-running transaction. So the
2329 * orphan record which we wrote for this file must have
2330 * passed into commit. We must attach this buffer to
2331 * the committing transaction, if it exists. */
2332 if (journal->j_committing_transaction) {
2333 JBUFFER_TRACE(jh, "give to committing trans");
2334 may_free = __dispose_buffer(jh,
2335 journal->j_committing_transaction);
2338 /* The orphan record's transaction has
2339 * committed. We can cleanse this buffer */
2340 clear_buffer_jbddirty(bh);
2341 __jbd2_journal_remove_checkpoint(jh);
2345 } else if (transaction == journal->j_committing_transaction) {
2346 JBUFFER_TRACE(jh, "on committing transaction");
2348 * The buffer is committing, we simply cannot touch
2349 * it. If the page is straddling i_size we have to wait
2350 * for commit and try again.
2353 spin_unlock(&journal->j_list_lock);
2354 spin_unlock(&jh->b_state_lock);
2355 write_unlock(&journal->j_state_lock);
2356 jbd2_journal_put_journal_head(jh);
2360 * OK, buffer won't be reachable after truncate. We just clear
2361 * b_modified to not confuse transaction credit accounting, and
2362 * set j_next_transaction to the running transaction (if there
2363 * is one) and mark buffer as freed so that commit code knows
2364 * it should clear dirty bits when it is done with the buffer.
2366 set_buffer_freed(bh);
2367 if (journal->j_running_transaction && buffer_jbddirty(bh))
2368 jh->b_next_transaction = journal->j_running_transaction;
2370 spin_unlock(&journal->j_list_lock);
2371 spin_unlock(&jh->b_state_lock);
2372 write_unlock(&journal->j_state_lock);
2373 jbd2_journal_put_journal_head(jh);
2376 /* Good, the buffer belongs to the running transaction.
2377 * We are writing our own transaction's data, not any
2378 * previous one's, so it is safe to throw it away
2379 * (remember that we expect the filesystem to have set
2380 * i_size already for this truncate so recovery will not
2381 * expose the disk blocks we are discarding here.) */
2382 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
2383 JBUFFER_TRACE(jh, "on running transaction");
2384 may_free = __dispose_buffer(jh, transaction);
2389 * This is tricky. Although the buffer is truncated, it may be reused
2390 * if blocksize < pagesize and it is attached to the page straddling
2391 * EOF. Since the buffer might have been added to BJ_Forget list of the
2392 * running transaction, journal_get_write_access() won't clear
2393 * b_modified and credit accounting gets confused. So clear b_modified
2397 spin_unlock(&journal->j_list_lock);
2398 spin_unlock(&jh->b_state_lock);
2399 write_unlock(&journal->j_state_lock);
2400 jbd2_journal_put_journal_head(jh);
2401 zap_buffer_unlocked:
2402 clear_buffer_dirty(bh);
2403 J_ASSERT_BH(bh, !buffer_jbddirty(bh));
2404 clear_buffer_mapped(bh);
2405 clear_buffer_req(bh);
2406 clear_buffer_new(bh);
2407 clear_buffer_delay(bh);
2408 clear_buffer_unwritten(bh);
2414 * void jbd2_journal_invalidatepage()
2415 * @journal: journal to use for flush...
2416 * @page: page to flush
2417 * @offset: start of the range to invalidate
2418 * @length: length of the range to invalidate
2420 * Reap page buffers containing data after in the specified range in page.
2421 * Can return -EBUSY if buffers are part of the committing transaction and
2422 * the page is straddling i_size. Caller then has to wait for current commit
2425 int jbd2_journal_invalidatepage(journal_t *journal,
2427 unsigned int offset,
2428 unsigned int length)
2430 struct buffer_head *head, *bh, *next;
2431 unsigned int stop = offset + length;
2432 unsigned int curr_off = 0;
2433 int partial_page = (offset || length < PAGE_SIZE);
2437 if (!PageLocked(page))
2439 if (!page_has_buffers(page))
2442 BUG_ON(stop > PAGE_SIZE || stop < length);
2444 /* We will potentially be playing with lists other than just the
2445 * data lists (especially for journaled data mode), so be
2446 * cautious in our locking. */
2448 head = bh = page_buffers(page);
2450 unsigned int next_off = curr_off + bh->b_size;
2451 next = bh->b_this_page;
2453 if (next_off > stop)
2456 if (offset <= curr_off) {
2457 /* This block is wholly outside the truncation point */
2459 ret = journal_unmap_buffer(journal, bh, partial_page);
2465 curr_off = next_off;
2468 } while (bh != head);
2470 if (!partial_page) {
2471 if (may_free && try_to_free_buffers(page))
2472 J_ASSERT(!page_has_buffers(page));
2478 * File a buffer on the given transaction list.
2480 void __jbd2_journal_file_buffer(struct journal_head *jh,
2481 transaction_t *transaction, int jlist)
2483 struct journal_head **list = NULL;
2485 struct buffer_head *bh = jh2bh(jh);
2487 lockdep_assert_held(&jh->b_state_lock);
2488 assert_spin_locked(&transaction->t_journal->j_list_lock);
2490 J_ASSERT_JH(jh, jh->b_jlist < BJ_Types);
2491 J_ASSERT_JH(jh, jh->b_transaction == transaction ||
2492 jh->b_transaction == NULL);
2494 if (jh->b_transaction && jh->b_jlist == jlist)
2497 if (jlist == BJ_Metadata || jlist == BJ_Reserved ||
2498 jlist == BJ_Shadow || jlist == BJ_Forget) {
2500 * For metadata buffers, we track dirty bit in buffer_jbddirty
2501 * instead of buffer_dirty. We should not see a dirty bit set
2502 * here because we clear it in do_get_write_access but e.g.
2503 * tune2fs can modify the sb and set the dirty bit at any time
2504 * so we try to gracefully handle that.
2506 if (buffer_dirty(bh))
2507 warn_dirty_buffer(bh);
2508 if (test_clear_buffer_dirty(bh) ||
2509 test_clear_buffer_jbddirty(bh))
2513 if (jh->b_transaction)
2514 __jbd2_journal_temp_unlink_buffer(jh);
2516 jbd2_journal_grab_journal_head(bh);
2517 jh->b_transaction = transaction;
2521 J_ASSERT_JH(jh, !jh->b_committed_data);
2522 J_ASSERT_JH(jh, !jh->b_frozen_data);
2525 transaction->t_nr_buffers++;
2526 list = &transaction->t_buffers;
2529 list = &transaction->t_forget;
2532 list = &transaction->t_shadow_list;
2535 list = &transaction->t_reserved_list;
2539 __blist_add_buffer(list, jh);
2540 jh->b_jlist = jlist;
2543 set_buffer_jbddirty(bh);
2546 void jbd2_journal_file_buffer(struct journal_head *jh,
2547 transaction_t *transaction, int jlist)
2549 spin_lock(&jh->b_state_lock);
2550 spin_lock(&transaction->t_journal->j_list_lock);
2551 __jbd2_journal_file_buffer(jh, transaction, jlist);
2552 spin_unlock(&transaction->t_journal->j_list_lock);
2553 spin_unlock(&jh->b_state_lock);
2557 * Remove a buffer from its current buffer list in preparation for
2558 * dropping it from its current transaction entirely. If the buffer has
2559 * already started to be used by a subsequent transaction, refile the
2560 * buffer on that transaction's metadata list.
2562 * Called under j_list_lock
2563 * Called under jh->b_state_lock
2565 * When this function returns true, there's no next transaction to refile to
2566 * and the caller has to drop jh reference through
2567 * jbd2_journal_put_journal_head().
2569 bool __jbd2_journal_refile_buffer(struct journal_head *jh)
2571 int was_dirty, jlist;
2572 struct buffer_head *bh = jh2bh(jh);
2574 lockdep_assert_held(&jh->b_state_lock);
2575 if (jh->b_transaction)
2576 assert_spin_locked(&jh->b_transaction->t_journal->j_list_lock);
2578 /* If the buffer is now unused, just drop it. */
2579 if (jh->b_next_transaction == NULL) {
2580 __jbd2_journal_unfile_buffer(jh);
2585 * It has been modified by a later transaction: add it to the new
2586 * transaction's metadata list.
2589 was_dirty = test_clear_buffer_jbddirty(bh);
2590 __jbd2_journal_temp_unlink_buffer(jh);
2593 * b_transaction must be set, otherwise the new b_transaction won't
2594 * be holding jh reference
2596 J_ASSERT_JH(jh, jh->b_transaction != NULL);
2599 * We set b_transaction here because b_next_transaction will inherit
2600 * our jh reference and thus __jbd2_journal_file_buffer() must not
2603 WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
2604 WRITE_ONCE(jh->b_next_transaction, NULL);
2605 if (buffer_freed(bh))
2607 else if (jh->b_modified)
2608 jlist = BJ_Metadata;
2610 jlist = BJ_Reserved;
2611 __jbd2_journal_file_buffer(jh, jh->b_transaction, jlist);
2612 J_ASSERT_JH(jh, jh->b_transaction->t_state == T_RUNNING);
2615 set_buffer_jbddirty(bh);
2620 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2621 * bh reference so that we can safely unlock bh.
2623 * The jh and bh may be freed by this call.
2625 void jbd2_journal_refile_buffer(journal_t *journal, struct journal_head *jh)
2629 spin_lock(&jh->b_state_lock);
2630 spin_lock(&journal->j_list_lock);
2631 drop = __jbd2_journal_refile_buffer(jh);
2632 spin_unlock(&jh->b_state_lock);
2633 spin_unlock(&journal->j_list_lock);
2635 jbd2_journal_put_journal_head(jh);
2639 * File inode in the inode list of the handle's transaction
2641 static int jbd2_journal_file_inode(handle_t *handle, struct jbd2_inode *jinode,
2642 unsigned long flags, loff_t start_byte, loff_t end_byte)
2644 transaction_t *transaction = handle->h_transaction;
2647 if (is_handle_aborted(handle))
2649 journal = transaction->t_journal;
2651 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode->i_vfs_inode->i_ino,
2652 transaction->t_tid);
2654 spin_lock(&journal->j_list_lock);
2655 jinode->i_flags |= flags;
2657 if (jinode->i_dirty_end) {
2658 jinode->i_dirty_start = min(jinode->i_dirty_start, start_byte);
2659 jinode->i_dirty_end = max(jinode->i_dirty_end, end_byte);
2661 jinode->i_dirty_start = start_byte;
2662 jinode->i_dirty_end = end_byte;
2665 /* Is inode already attached where we need it? */
2666 if (jinode->i_transaction == transaction ||
2667 jinode->i_next_transaction == transaction)
2671 * We only ever set this variable to 1 so the test is safe. Since
2672 * t_need_data_flush is likely to be set, we do the test to save some
2673 * cacheline bouncing
2675 if (!transaction->t_need_data_flush)
2676 transaction->t_need_data_flush = 1;
2677 /* On some different transaction's list - should be
2678 * the committing one */
2679 if (jinode->i_transaction) {
2680 J_ASSERT(jinode->i_next_transaction == NULL);
2681 J_ASSERT(jinode->i_transaction ==
2682 journal->j_committing_transaction);
2683 jinode->i_next_transaction = transaction;
2686 /* Not on any transaction list... */
2687 J_ASSERT(!jinode->i_next_transaction);
2688 jinode->i_transaction = transaction;
2689 list_add(&jinode->i_list, &transaction->t_inode_list);
2691 spin_unlock(&journal->j_list_lock);
2696 int jbd2_journal_inode_ranged_write(handle_t *handle,
2697 struct jbd2_inode *jinode, loff_t start_byte, loff_t length)
2699 return jbd2_journal_file_inode(handle, jinode,
2700 JI_WRITE_DATA | JI_WAIT_DATA, start_byte,
2701 start_byte + length - 1);
2704 int jbd2_journal_inode_ranged_wait(handle_t *handle, struct jbd2_inode *jinode,
2705 loff_t start_byte, loff_t length)
2707 return jbd2_journal_file_inode(handle, jinode, JI_WAIT_DATA,
2708 start_byte, start_byte + length - 1);
2712 * File truncate and transaction commit interact with each other in a
2713 * non-trivial way. If a transaction writing data block A is
2714 * committing, we cannot discard the data by truncate until we have
2715 * written them. Otherwise if we crashed after the transaction with
2716 * write has committed but before the transaction with truncate has
2717 * committed, we could see stale data in block A. This function is a
2718 * helper to solve this problem. It starts writeout of the truncated
2719 * part in case it is in the committing transaction.
2721 * Filesystem code must call this function when inode is journaled in
2722 * ordered mode before truncation happens and after the inode has been
2723 * placed on orphan list with the new inode size. The second condition
2724 * avoids the race that someone writes new data and we start
2725 * committing the transaction after this function has been called but
2726 * before a transaction for truncate is started (and furthermore it
2727 * allows us to optimize the case where the addition to orphan list
2728 * happens in the same transaction as write --- we don't have to write
2729 * any data in such case).
2731 int jbd2_journal_begin_ordered_truncate(journal_t *journal,
2732 struct jbd2_inode *jinode,
2735 transaction_t *inode_trans, *commit_trans;
2738 /* This is a quick check to avoid locking if not necessary */
2739 if (!jinode->i_transaction)
2741 /* Locks are here just to force reading of recent values, it is
2742 * enough that the transaction was not committing before we started
2743 * a transaction adding the inode to orphan list */
2744 read_lock(&journal->j_state_lock);
2745 commit_trans = journal->j_committing_transaction;
2746 read_unlock(&journal->j_state_lock);
2747 spin_lock(&journal->j_list_lock);
2748 inode_trans = jinode->i_transaction;
2749 spin_unlock(&journal->j_list_lock);
2750 if (inode_trans == commit_trans) {
2751 ret = filemap_fdatawrite_range(jinode->i_vfs_inode->i_mapping,
2752 new_size, LLONG_MAX);
2754 jbd2_journal_abort(journal, ret);