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
);
353 * This check is racy but it is just an optimization of allocating new
354 * transaction early if there are high chances we'll need it. If we
355 * guess wrong, we'll retry or free unused transaction.
357 if (!data_race(journal
->j_running_transaction
)) {
359 * If __GFP_FS is not present, then we may be being called from
360 * inside the fs writeback layer, so we MUST NOT fail.
362 if ((gfp_mask
& __GFP_FS
) == 0)
363 gfp_mask
|= __GFP_NOFAIL
;
364 new_transaction
= kmem_cache_zalloc(transaction_cache
,
366 if (!new_transaction
)
370 jbd_debug(3, "New handle %p going live.\n", handle
);
373 * We need to hold j_state_lock until t_updates has been incremented,
374 * for proper journal barrier handling
377 read_lock(&journal
->j_state_lock
);
378 BUG_ON(journal
->j_flags
& JBD2_UNMOUNT
);
379 if (is_journal_aborted(journal
) ||
380 (journal
->j_errno
!= 0 && !(journal
->j_flags
& JBD2_ACK_ERR
))) {
381 read_unlock(&journal
->j_state_lock
);
382 jbd2_journal_free_transaction(new_transaction
);
387 * Wait on the journal's transaction barrier if necessary. Specifically
388 * we allow reserved handles to proceed because otherwise commit could
389 * deadlock on page writeback not being able to complete.
391 if (!handle
->h_reserved
&& journal
->j_barrier_count
) {
392 read_unlock(&journal
->j_state_lock
);
393 wait_event(journal
->j_wait_transaction_locked
,
394 journal
->j_barrier_count
== 0);
398 if (!journal
->j_running_transaction
) {
399 read_unlock(&journal
->j_state_lock
);
400 if (!new_transaction
)
401 goto alloc_transaction
;
402 write_lock(&journal
->j_state_lock
);
403 if (!journal
->j_running_transaction
&&
404 (handle
->h_reserved
|| !journal
->j_barrier_count
)) {
405 jbd2_get_transaction(journal
, new_transaction
);
406 new_transaction
= NULL
;
408 write_unlock(&journal
->j_state_lock
);
412 transaction
= journal
->j_running_transaction
;
414 if (!handle
->h_reserved
) {
415 /* We may have dropped j_state_lock - restart in that case */
416 if (add_transaction_credits(journal
, blocks
, rsv_blocks
))
420 * We have handle reserved so we are allowed to join T_LOCKED
421 * transaction and we don't have to check for transaction size
422 * and journal space. But we still have to wait while running
423 * transaction is being switched to a committing one as it
424 * won't wait for any handles anymore.
426 if (transaction
->t_state
== T_SWITCH
) {
427 wait_transaction_switching(journal
);
430 sub_reserved_credits(journal
, blocks
);
431 handle
->h_reserved
= 0;
434 /* OK, account for the buffers that this operation expects to
435 * use and add the handle to the running transaction.
437 update_t_max_wait(transaction
, ts
);
438 handle
->h_transaction
= transaction
;
439 handle
->h_requested_credits
= blocks
;
440 handle
->h_revoke_credits_requested
= handle
->h_revoke_credits
;
441 handle
->h_start_jiffies
= jiffies
;
442 atomic_inc(&transaction
->t_updates
);
443 atomic_inc(&transaction
->t_handle_count
);
444 jbd_debug(4, "Handle %p given %d credits (total %d, free %lu)\n",
446 atomic_read(&transaction
->t_outstanding_credits
),
447 jbd2_log_space_left(journal
));
448 read_unlock(&journal
->j_state_lock
);
449 current
->journal_info
= handle
;
451 rwsem_acquire_read(&journal
->j_trans_commit_map
, 0, 0, _THIS_IP_
);
452 jbd2_journal_free_transaction(new_transaction
);
454 * Ensure that no allocations done while the transaction is open are
455 * going to recurse back to the fs layer.
457 handle
->saved_alloc_context
= memalloc_nofs_save();
461 /* Allocate a new handle. This should probably be in a slab... */
462 static handle_t
*new_handle(int nblocks
)
464 handle_t
*handle
= jbd2_alloc_handle(GFP_NOFS
);
467 handle
->h_total_credits
= nblocks
;
473 handle_t
*jbd2__journal_start(journal_t
*journal
, int nblocks
, int rsv_blocks
,
474 int revoke_records
, gfp_t gfp_mask
,
475 unsigned int type
, unsigned int line_no
)
477 handle_t
*handle
= journal_current_handle();
481 return ERR_PTR(-EROFS
);
484 J_ASSERT(handle
->h_transaction
->t_journal
== journal
);
489 nblocks
+= DIV_ROUND_UP(revoke_records
,
490 journal
->j_revoke_records_per_block
);
491 handle
= new_handle(nblocks
);
493 return ERR_PTR(-ENOMEM
);
495 handle_t
*rsv_handle
;
497 rsv_handle
= new_handle(rsv_blocks
);
499 jbd2_free_handle(handle
);
500 return ERR_PTR(-ENOMEM
);
502 rsv_handle
->h_reserved
= 1;
503 rsv_handle
->h_journal
= journal
;
504 handle
->h_rsv_handle
= rsv_handle
;
506 handle
->h_revoke_credits
= revoke_records
;
508 err
= start_this_handle(journal
, handle
, gfp_mask
);
510 if (handle
->h_rsv_handle
)
511 jbd2_free_handle(handle
->h_rsv_handle
);
512 jbd2_free_handle(handle
);
515 handle
->h_type
= type
;
516 handle
->h_line_no
= line_no
;
517 trace_jbd2_handle_start(journal
->j_fs_dev
->bd_dev
,
518 handle
->h_transaction
->t_tid
, type
,
523 EXPORT_SYMBOL(jbd2__journal_start
);
527 * jbd2_journal_start() - Obtain a new handle.
528 * @journal: Journal to start transaction on.
529 * @nblocks: number of block buffer we might modify
531 * We make sure that the transaction can guarantee at least nblocks of
532 * modified buffers in the log. We block until the log can guarantee
533 * that much space. Additionally, if rsv_blocks > 0, we also create another
534 * handle with rsv_blocks reserved blocks in the journal. This handle is
535 * stored in h_rsv_handle. It is not attached to any particular transaction
536 * and thus doesn't block transaction commit. If the caller uses this reserved
537 * handle, it has to set h_rsv_handle to NULL as otherwise jbd2_journal_stop()
538 * on the parent handle will dispose the reserved one. Reserved handle has to
539 * be converted to a normal handle using jbd2_journal_start_reserved() before
542 * Return a pointer to a newly allocated handle, or an ERR_PTR() value
545 handle_t
*jbd2_journal_start(journal_t
*journal
, int nblocks
)
547 return jbd2__journal_start(journal
, nblocks
, 0, 0, GFP_NOFS
, 0, 0);
549 EXPORT_SYMBOL(jbd2_journal_start
);
551 static void __jbd2_journal_unreserve_handle(handle_t
*handle
, transaction_t
*t
)
553 journal_t
*journal
= handle
->h_journal
;
555 WARN_ON(!handle
->h_reserved
);
556 sub_reserved_credits(journal
, handle
->h_total_credits
);
558 atomic_sub(handle
->h_total_credits
, &t
->t_outstanding_credits
);
561 void jbd2_journal_free_reserved(handle_t
*handle
)
563 journal_t
*journal
= handle
->h_journal
;
565 /* Get j_state_lock to pin running transaction if it exists */
566 read_lock(&journal
->j_state_lock
);
567 __jbd2_journal_unreserve_handle(handle
, journal
->j_running_transaction
);
568 read_unlock(&journal
->j_state_lock
);
569 jbd2_free_handle(handle
);
571 EXPORT_SYMBOL(jbd2_journal_free_reserved
);
574 * jbd2_journal_start_reserved() - start reserved handle
575 * @handle: handle to start
576 * @type: for handle statistics
577 * @line_no: for handle statistics
579 * Start handle that has been previously reserved with jbd2_journal_reserve().
580 * This attaches @handle to the running transaction (or creates one if there's
581 * not transaction running). Unlike jbd2_journal_start() this function cannot
582 * block on journal commit, checkpointing, or similar stuff. It can block on
583 * memory allocation or frozen journal though.
585 * Return 0 on success, non-zero on error - handle is freed in that case.
587 int jbd2_journal_start_reserved(handle_t
*handle
, unsigned int type
,
588 unsigned int line_no
)
590 journal_t
*journal
= handle
->h_journal
;
593 if (WARN_ON(!handle
->h_reserved
)) {
594 /* Someone passed in normal handle? Just stop it. */
595 jbd2_journal_stop(handle
);
599 * Usefulness of mixing of reserved and unreserved handles is
600 * questionable. So far nobody seems to need it so just error out.
602 if (WARN_ON(current
->journal_info
)) {
603 jbd2_journal_free_reserved(handle
);
607 handle
->h_journal
= NULL
;
609 * GFP_NOFS is here because callers are likely from writeback or
610 * similarly constrained call sites
612 ret
= start_this_handle(journal
, handle
, GFP_NOFS
);
614 handle
->h_journal
= journal
;
615 jbd2_journal_free_reserved(handle
);
618 handle
->h_type
= type
;
619 handle
->h_line_no
= line_no
;
620 trace_jbd2_handle_start(journal
->j_fs_dev
->bd_dev
,
621 handle
->h_transaction
->t_tid
, type
,
622 line_no
, handle
->h_total_credits
);
625 EXPORT_SYMBOL(jbd2_journal_start_reserved
);
628 * jbd2_journal_extend() - extend buffer credits.
629 * @handle: handle to 'extend'
630 * @nblocks: nr blocks to try to extend by.
631 * @revoke_records: number of revoke records to try to extend by.
633 * Some transactions, such as large extends and truncates, can be done
634 * atomically all at once or in several stages. The operation requests
635 * a credit for a number of buffer modifications in advance, but can
636 * extend its credit if it needs more.
638 * jbd2_journal_extend tries to give the running handle more buffer credits.
639 * It does not guarantee that allocation - this is a best-effort only.
640 * The calling process MUST be able to deal cleanly with a failure to
643 * Return 0 on success, non-zero on failure.
645 * return code < 0 implies an error
646 * return code > 0 implies normal transaction-full status.
648 int jbd2_journal_extend(handle_t
*handle
, int nblocks
, int revoke_records
)
650 transaction_t
*transaction
= handle
->h_transaction
;
655 if (is_handle_aborted(handle
))
657 journal
= transaction
->t_journal
;
661 read_lock(&journal
->j_state_lock
);
663 /* Don't extend a locked-down transaction! */
664 if (transaction
->t_state
!= T_RUNNING
) {
665 jbd_debug(3, "denied handle %p %d blocks: "
666 "transaction not running\n", handle
, nblocks
);
670 nblocks
+= DIV_ROUND_UP(
671 handle
->h_revoke_credits_requested
+ revoke_records
,
672 journal
->j_revoke_records_per_block
) -
674 handle
->h_revoke_credits_requested
,
675 journal
->j_revoke_records_per_block
);
676 spin_lock(&transaction
->t_handle_lock
);
677 wanted
= atomic_add_return(nblocks
,
678 &transaction
->t_outstanding_credits
);
680 if (wanted
> journal
->j_max_transaction_buffers
) {
681 jbd_debug(3, "denied handle %p %d blocks: "
682 "transaction too large\n", handle
, nblocks
);
683 atomic_sub(nblocks
, &transaction
->t_outstanding_credits
);
687 trace_jbd2_handle_extend(journal
->j_fs_dev
->bd_dev
,
689 handle
->h_type
, handle
->h_line_no
,
690 handle
->h_total_credits
,
693 handle
->h_total_credits
+= nblocks
;
694 handle
->h_requested_credits
+= nblocks
;
695 handle
->h_revoke_credits
+= revoke_records
;
696 handle
->h_revoke_credits_requested
+= revoke_records
;
699 jbd_debug(3, "extended handle %p by %d\n", handle
, nblocks
);
701 spin_unlock(&transaction
->t_handle_lock
);
703 read_unlock(&journal
->j_state_lock
);
707 static void stop_this_handle(handle_t
*handle
)
709 transaction_t
*transaction
= handle
->h_transaction
;
710 journal_t
*journal
= transaction
->t_journal
;
713 J_ASSERT(journal_current_handle() == handle
);
714 J_ASSERT(atomic_read(&transaction
->t_updates
) > 0);
715 current
->journal_info
= NULL
;
717 * Subtract necessary revoke descriptor blocks from handle credits. We
718 * take care to account only for revoke descriptor blocks the
719 * transaction will really need as large sequences of transactions with
720 * small numbers of revokes are relatively common.
722 revokes
= handle
->h_revoke_credits_requested
- handle
->h_revoke_credits
;
724 int t_revokes
, revoke_descriptors
;
725 int rr_per_blk
= journal
->j_revoke_records_per_block
;
727 WARN_ON_ONCE(DIV_ROUND_UP(revokes
, rr_per_blk
)
728 > handle
->h_total_credits
);
729 t_revokes
= atomic_add_return(revokes
,
730 &transaction
->t_outstanding_revokes
);
732 DIV_ROUND_UP(t_revokes
, rr_per_blk
) -
733 DIV_ROUND_UP(t_revokes
- revokes
, rr_per_blk
);
734 handle
->h_total_credits
-= revoke_descriptors
;
736 atomic_sub(handle
->h_total_credits
,
737 &transaction
->t_outstanding_credits
);
738 if (handle
->h_rsv_handle
)
739 __jbd2_journal_unreserve_handle(handle
->h_rsv_handle
,
741 if (atomic_dec_and_test(&transaction
->t_updates
))
742 wake_up(&journal
->j_wait_updates
);
744 rwsem_release(&journal
->j_trans_commit_map
, _THIS_IP_
);
746 * Scope of the GFP_NOFS context is over here and so we can restore the
747 * original alloc context.
749 memalloc_nofs_restore(handle
->saved_alloc_context
);
753 * jbd2__journal_restart() - restart a handle .
754 * @handle: handle to restart
755 * @nblocks: nr credits requested
756 * @revoke_records: number of revoke record credits requested
757 * @gfp_mask: memory allocation flags (for start_this_handle)
759 * Restart a handle for a multi-transaction filesystem
762 * If the jbd2_journal_extend() call above fails to grant new buffer credits
763 * to a running handle, a call to jbd2_journal_restart will commit the
764 * handle's transaction so far and reattach the handle to a new
765 * transaction capable of guaranteeing the requested number of
766 * credits. We preserve reserved handle if there's any attached to the
769 int jbd2__journal_restart(handle_t
*handle
, int nblocks
, int revoke_records
,
772 transaction_t
*transaction
= handle
->h_transaction
;
778 /* If we've had an abort of any type, don't even think about
779 * actually doing the restart! */
780 if (is_handle_aborted(handle
))
782 journal
= transaction
->t_journal
;
783 tid
= transaction
->t_tid
;
786 * First unlink the handle from its current transaction, and start the
789 jbd_debug(2, "restarting handle %p\n", handle
);
790 stop_this_handle(handle
);
791 handle
->h_transaction
= NULL
;
794 * TODO: If we use READ_ONCE / WRITE_ONCE for j_commit_request we can
795 * get rid of pointless j_state_lock traffic like this.
797 read_lock(&journal
->j_state_lock
);
798 need_to_start
= !tid_geq(journal
->j_commit_request
, tid
);
799 read_unlock(&journal
->j_state_lock
);
801 jbd2_log_start_commit(journal
, tid
);
802 handle
->h_total_credits
= nblocks
+
803 DIV_ROUND_UP(revoke_records
,
804 journal
->j_revoke_records_per_block
);
805 handle
->h_revoke_credits
= revoke_records
;
806 ret
= start_this_handle(journal
, handle
, gfp_mask
);
807 trace_jbd2_handle_restart(journal
->j_fs_dev
->bd_dev
,
808 ret
? 0 : handle
->h_transaction
->t_tid
,
809 handle
->h_type
, handle
->h_line_no
,
810 handle
->h_total_credits
);
813 EXPORT_SYMBOL(jbd2__journal_restart
);
816 int jbd2_journal_restart(handle_t
*handle
, int nblocks
)
818 return jbd2__journal_restart(handle
, nblocks
, 0, GFP_NOFS
);
820 EXPORT_SYMBOL(jbd2_journal_restart
);
823 * jbd2_journal_lock_updates () - establish a transaction barrier.
824 * @journal: Journal to establish a barrier on.
826 * This locks out any further updates from being started, and blocks
827 * until all existing updates have completed, returning only once the
828 * journal is in a quiescent state with no updates running.
830 * The journal lock should not be held on entry.
832 void jbd2_journal_lock_updates(journal_t
*journal
)
836 jbd2_might_wait_for_commit(journal
);
838 write_lock(&journal
->j_state_lock
);
839 ++journal
->j_barrier_count
;
841 /* Wait until there are no reserved handles */
842 if (atomic_read(&journal
->j_reserved_credits
)) {
843 write_unlock(&journal
->j_state_lock
);
844 wait_event(journal
->j_wait_reserved
,
845 atomic_read(&journal
->j_reserved_credits
) == 0);
846 write_lock(&journal
->j_state_lock
);
849 /* Wait until there are no running updates */
851 transaction_t
*transaction
= journal
->j_running_transaction
;
856 spin_lock(&transaction
->t_handle_lock
);
857 prepare_to_wait(&journal
->j_wait_updates
, &wait
,
858 TASK_UNINTERRUPTIBLE
);
859 if (!atomic_read(&transaction
->t_updates
)) {
860 spin_unlock(&transaction
->t_handle_lock
);
861 finish_wait(&journal
->j_wait_updates
, &wait
);
864 spin_unlock(&transaction
->t_handle_lock
);
865 write_unlock(&journal
->j_state_lock
);
867 finish_wait(&journal
->j_wait_updates
, &wait
);
868 write_lock(&journal
->j_state_lock
);
870 write_unlock(&journal
->j_state_lock
);
873 * We have now established a barrier against other normal updates, but
874 * we also need to barrier against other jbd2_journal_lock_updates() calls
875 * to make sure that we serialise special journal-locked operations
878 mutex_lock(&journal
->j_barrier
);
882 * jbd2_journal_unlock_updates () - release barrier
883 * @journal: Journal to release the barrier on.
885 * Release a transaction barrier obtained with jbd2_journal_lock_updates().
887 * Should be called without the journal lock held.
889 void jbd2_journal_unlock_updates (journal_t
*journal
)
891 J_ASSERT(journal
->j_barrier_count
!= 0);
893 mutex_unlock(&journal
->j_barrier
);
894 write_lock(&journal
->j_state_lock
);
895 --journal
->j_barrier_count
;
896 write_unlock(&journal
->j_state_lock
);
897 wake_up(&journal
->j_wait_transaction_locked
);
900 static void warn_dirty_buffer(struct buffer_head
*bh
)
903 "JBD2: Spotted dirty metadata buffer (dev = %pg, blocknr = %llu). "
904 "There's a risk of filesystem corruption in case of system "
906 bh
->b_bdev
, (unsigned long long)bh
->b_blocknr
);
909 /* Call t_frozen trigger and copy buffer data into jh->b_frozen_data. */
910 static void jbd2_freeze_jh_data(struct journal_head
*jh
)
915 struct buffer_head
*bh
= jh2bh(jh
);
917 J_EXPECT_JH(jh
, buffer_uptodate(bh
), "Possible IO failure.\n");
919 offset
= offset_in_page(bh
->b_data
);
920 source
= kmap_atomic(page
);
921 /* Fire data frozen trigger just before we copy the data */
922 jbd2_buffer_frozen_trigger(jh
, source
+ offset
, jh
->b_triggers
);
923 memcpy(jh
->b_frozen_data
, source
+ offset
, bh
->b_size
);
924 kunmap_atomic(source
);
927 * Now that the frozen data is saved off, we need to store any matching
930 jh
->b_frozen_triggers
= jh
->b_triggers
;
934 * If the buffer is already part of the current transaction, then there
935 * is nothing we need to do. If it is already part of a prior
936 * transaction which we are still committing to disk, then we need to
937 * make sure that we do not overwrite the old copy: we do copy-out to
938 * preserve the copy going to disk. We also account the buffer against
939 * the handle's metadata buffer credits (unless the buffer is already
940 * part of the transaction, that is).
944 do_get_write_access(handle_t
*handle
, struct journal_head
*jh
,
947 struct buffer_head
*bh
;
948 transaction_t
*transaction
= handle
->h_transaction
;
951 char *frozen_buffer
= NULL
;
952 unsigned long start_lock
, time_lock
;
954 journal
= transaction
->t_journal
;
956 jbd_debug(5, "journal_head %p, force_copy %d\n", jh
, force_copy
);
958 JBUFFER_TRACE(jh
, "entry");
962 /* @@@ Need to check for errors here at some point. */
964 start_lock
= jiffies
;
966 spin_lock(&jh
->b_state_lock
);
968 /* If it takes too long to lock the buffer, trace it */
969 time_lock
= jbd2_time_diff(start_lock
, jiffies
);
970 if (time_lock
> HZ
/10)
971 trace_jbd2_lock_buffer_stall(bh
->b_bdev
->bd_dev
,
972 jiffies_to_msecs(time_lock
));
974 /* We now hold the buffer lock so it is safe to query the buffer
975 * state. Is the buffer dirty?
977 * If so, there are two possibilities. The buffer may be
978 * non-journaled, and undergoing a quite legitimate writeback.
979 * Otherwise, it is journaled, and we don't expect dirty buffers
980 * in that state (the buffers should be marked JBD_Dirty
981 * instead.) So either the IO is being done under our own
982 * control and this is a bug, or it's a third party IO such as
983 * dump(8) (which may leave the buffer scheduled for read ---
984 * ie. locked but not dirty) or tune2fs (which may actually have
985 * the buffer dirtied, ugh.) */
987 if (buffer_dirty(bh
)) {
989 * First question: is this buffer already part of the current
990 * transaction or the existing committing transaction?
992 if (jh
->b_transaction
) {
994 jh
->b_transaction
== transaction
||
996 journal
->j_committing_transaction
);
997 if (jh
->b_next_transaction
)
998 J_ASSERT_JH(jh
, jh
->b_next_transaction
==
1000 warn_dirty_buffer(bh
);
1003 * In any case we need to clean the dirty flag and we must
1004 * do it under the buffer lock to be sure we don't race
1005 * with running write-out.
1007 JBUFFER_TRACE(jh
, "Journalling dirty buffer");
1008 clear_buffer_dirty(bh
);
1009 set_buffer_jbddirty(bh
);
1015 if (is_handle_aborted(handle
)) {
1016 spin_unlock(&jh
->b_state_lock
);
1022 * The buffer is already part of this transaction if b_transaction or
1023 * b_next_transaction points to it
1025 if (jh
->b_transaction
== transaction
||
1026 jh
->b_next_transaction
== transaction
)
1030 * this is the first time this transaction is touching this buffer,
1031 * reset the modified flag
1036 * If the buffer is not journaled right now, we need to make sure it
1037 * doesn't get written to disk before the caller actually commits the
1040 if (!jh
->b_transaction
) {
1041 JBUFFER_TRACE(jh
, "no transaction");
1042 J_ASSERT_JH(jh
, !jh
->b_next_transaction
);
1043 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
1045 * Make sure all stores to jh (b_modified, b_frozen_data) are
1046 * visible before attaching it to the running transaction.
1047 * Paired with barrier in jbd2_write_access_granted()
1050 spin_lock(&journal
->j_list_lock
);
1051 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Reserved
);
1052 spin_unlock(&journal
->j_list_lock
);
1056 * If there is already a copy-out version of this buffer, then we don't
1057 * need to make another one
1059 if (jh
->b_frozen_data
) {
1060 JBUFFER_TRACE(jh
, "has frozen data");
1061 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
1065 JBUFFER_TRACE(jh
, "owned by older transaction");
1066 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
1067 J_ASSERT_JH(jh
, jh
->b_transaction
== journal
->j_committing_transaction
);
1070 * There is one case we have to be very careful about. If the
1071 * committing transaction is currently writing this buffer out to disk
1072 * and has NOT made a copy-out, then we cannot modify the buffer
1073 * contents at all right now. The essence of copy-out is that it is
1074 * the extra copy, not the primary copy, which gets journaled. If the
1075 * primary copy is already going to disk then we cannot do copy-out
1078 if (buffer_shadow(bh
)) {
1079 JBUFFER_TRACE(jh
, "on shadow: sleep");
1080 spin_unlock(&jh
->b_state_lock
);
1081 wait_on_bit_io(&bh
->b_state
, BH_Shadow
, TASK_UNINTERRUPTIBLE
);
1086 * Only do the copy if the currently-owning transaction still needs it.
1087 * If buffer isn't on BJ_Metadata list, the committing transaction is
1088 * past that stage (here we use the fact that BH_Shadow is set under
1089 * bh_state lock together with refiling to BJ_Shadow list and at this
1090 * point we know the buffer doesn't have BH_Shadow set).
1092 * Subtle point, though: if this is a get_undo_access, then we will be
1093 * relying on the frozen_data to contain the new value of the
1094 * committed_data record after the transaction, so we HAVE to force the
1095 * frozen_data copy in that case.
1097 if (jh
->b_jlist
== BJ_Metadata
|| force_copy
) {
1098 JBUFFER_TRACE(jh
, "generate frozen data");
1099 if (!frozen_buffer
) {
1100 JBUFFER_TRACE(jh
, "allocate memory for buffer");
1101 spin_unlock(&jh
->b_state_lock
);
1102 frozen_buffer
= jbd2_alloc(jh2bh(jh
)->b_size
,
1103 GFP_NOFS
| __GFP_NOFAIL
);
1106 jh
->b_frozen_data
= frozen_buffer
;
1107 frozen_buffer
= NULL
;
1108 jbd2_freeze_jh_data(jh
);
1112 * Make sure all stores to jh (b_modified, b_frozen_data) are visible
1113 * before attaching it to the running transaction. Paired with barrier
1114 * in jbd2_write_access_granted()
1117 jh
->b_next_transaction
= transaction
;
1120 spin_unlock(&jh
->b_state_lock
);
1123 * If we are about to journal a buffer, then any revoke pending on it is
1126 jbd2_journal_cancel_revoke(handle
, jh
);
1129 if (unlikely(frozen_buffer
)) /* It's usually NULL */
1130 jbd2_free(frozen_buffer
, bh
->b_size
);
1132 JBUFFER_TRACE(jh
, "exit");
1136 /* Fast check whether buffer is already attached to the required transaction */
1137 static bool jbd2_write_access_granted(handle_t
*handle
, struct buffer_head
*bh
,
1140 struct journal_head
*jh
;
1143 /* Dirty buffers require special handling... */
1144 if (buffer_dirty(bh
))
1148 * RCU protects us from dereferencing freed pages. So the checks we do
1149 * are guaranteed not to oops. However the jh slab object can get freed
1150 * & reallocated while we work with it. So we have to be careful. When
1151 * we see jh attached to the running transaction, we know it must stay
1152 * so until the transaction is committed. Thus jh won't be freed and
1153 * will be attached to the same bh while we run. However it can
1154 * happen jh gets freed, reallocated, and attached to the transaction
1155 * just after we get pointer to it from bh. So we have to be careful
1156 * and recheck jh still belongs to our bh before we return success.
1159 if (!buffer_jbd(bh
))
1161 /* This should be bh2jh() but that doesn't work with inline functions */
1162 jh
= READ_ONCE(bh
->b_private
);
1165 /* For undo access buffer must have data copied */
1166 if (undo
&& !jh
->b_committed_data
)
1168 if (READ_ONCE(jh
->b_transaction
) != handle
->h_transaction
&&
1169 READ_ONCE(jh
->b_next_transaction
) != handle
->h_transaction
)
1172 * There are two reasons for the barrier here:
1173 * 1) Make sure to fetch b_bh after we did previous checks so that we
1174 * detect when jh went through free, realloc, attach to transaction
1175 * while we were checking. Paired with implicit barrier in that path.
1176 * 2) So that access to bh done after jbd2_write_access_granted()
1177 * doesn't get reordered and see inconsistent state of concurrent
1178 * do_get_write_access().
1181 if (unlikely(jh
->b_bh
!= bh
))
1190 * jbd2_journal_get_write_access() - notify intent to modify a buffer
1191 * for metadata (not data) update.
1192 * @handle: transaction to add buffer modifications to
1193 * @bh: bh to be used for metadata writes
1195 * Returns: error code or 0 on success.
1197 * In full data journalling mode the buffer may be of type BJ_AsyncData,
1198 * because we're ``write()ing`` a buffer which is also part of a shared mapping.
1201 int jbd2_journal_get_write_access(handle_t
*handle
, struct buffer_head
*bh
)
1203 struct journal_head
*jh
;
1206 if (is_handle_aborted(handle
))
1209 if (jbd2_write_access_granted(handle
, bh
, false))
1212 jh
= jbd2_journal_add_journal_head(bh
);
1213 /* We do not want to get caught playing with fields which the
1214 * log thread also manipulates. Make sure that the buffer
1215 * completes any outstanding IO before proceeding. */
1216 rc
= do_get_write_access(handle
, jh
, 0);
1217 jbd2_journal_put_journal_head(jh
);
1223 * When the user wants to journal a newly created buffer_head
1224 * (ie. getblk() returned a new buffer and we are going to populate it
1225 * manually rather than reading off disk), then we need to keep the
1226 * buffer_head locked until it has been completely filled with new
1227 * data. In this case, we should be able to make the assertion that
1228 * the bh is not already part of an existing transaction.
1230 * The buffer should already be locked by the caller by this point.
1231 * There is no lock ranking violation: it was a newly created,
1232 * unlocked buffer beforehand. */
1235 * jbd2_journal_get_create_access () - notify intent to use newly created bh
1236 * @handle: transaction to new buffer to
1239 * Call this if you create a new bh.
1241 int jbd2_journal_get_create_access(handle_t
*handle
, struct buffer_head
*bh
)
1243 transaction_t
*transaction
= handle
->h_transaction
;
1245 struct journal_head
*jh
= jbd2_journal_add_journal_head(bh
);
1248 jbd_debug(5, "journal_head %p\n", jh
);
1250 if (is_handle_aborted(handle
))
1252 journal
= transaction
->t_journal
;
1255 JBUFFER_TRACE(jh
, "entry");
1257 * The buffer may already belong to this transaction due to pre-zeroing
1258 * in the filesystem's new_block code. It may also be on the previous,
1259 * committing transaction's lists, but it HAS to be in Forget state in
1260 * that case: the transaction must have deleted the buffer for it to be
1263 spin_lock(&jh
->b_state_lock
);
1264 J_ASSERT_JH(jh
, (jh
->b_transaction
== transaction
||
1265 jh
->b_transaction
== NULL
||
1266 (jh
->b_transaction
== journal
->j_committing_transaction
&&
1267 jh
->b_jlist
== BJ_Forget
)));
1269 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
1270 J_ASSERT_JH(jh
, buffer_locked(jh2bh(jh
)));
1272 if (jh
->b_transaction
== NULL
) {
1274 * Previous jbd2_journal_forget() could have left the buffer
1275 * with jbddirty bit set because it was being committed. When
1276 * the commit finished, we've filed the buffer for
1277 * checkpointing and marked it dirty. Now we are reallocating
1278 * the buffer so the transaction freeing it must have
1279 * committed and so it's safe to clear the dirty bit.
1281 clear_buffer_dirty(jh2bh(jh
));
1282 /* first access by this transaction */
1285 JBUFFER_TRACE(jh
, "file as BJ_Reserved");
1286 spin_lock(&journal
->j_list_lock
);
1287 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Reserved
);
1288 spin_unlock(&journal
->j_list_lock
);
1289 } else if (jh
->b_transaction
== journal
->j_committing_transaction
) {
1290 /* first access by this transaction */
1293 JBUFFER_TRACE(jh
, "set next transaction");
1294 spin_lock(&journal
->j_list_lock
);
1295 jh
->b_next_transaction
= transaction
;
1296 spin_unlock(&journal
->j_list_lock
);
1298 spin_unlock(&jh
->b_state_lock
);
1301 * akpm: I added this. ext3_alloc_branch can pick up new indirect
1302 * blocks which contain freed but then revoked metadata. We need
1303 * to cancel the revoke in case we end up freeing it yet again
1304 * and the reallocating as data - this would cause a second revoke,
1305 * which hits an assertion error.
1307 JBUFFER_TRACE(jh
, "cancelling revoke");
1308 jbd2_journal_cancel_revoke(handle
, jh
);
1310 jbd2_journal_put_journal_head(jh
);
1315 * jbd2_journal_get_undo_access() - Notify intent to modify metadata with
1316 * non-rewindable consequences
1317 * @handle: transaction
1318 * @bh: buffer to undo
1320 * Sometimes there is a need to distinguish between metadata which has
1321 * been committed to disk and that which has not. The ext3fs code uses
1322 * this for freeing and allocating space, we have to make sure that we
1323 * do not reuse freed space until the deallocation has been committed,
1324 * since if we overwrote that space we would make the delete
1325 * un-rewindable in case of a crash.
1327 * To deal with that, jbd2_journal_get_undo_access requests write access to a
1328 * buffer for parts of non-rewindable operations such as delete
1329 * operations on the bitmaps. The journaling code must keep a copy of
1330 * the buffer's contents prior to the undo_access call until such time
1331 * as we know that the buffer has definitely been committed to disk.
1333 * We never need to know which transaction the committed data is part
1334 * of, buffers touched here are guaranteed to be dirtied later and so
1335 * will be committed to a new transaction in due course, at which point
1336 * we can discard the old committed data pointer.
1338 * Returns error number or 0 on success.
1340 int jbd2_journal_get_undo_access(handle_t
*handle
, struct buffer_head
*bh
)
1343 struct journal_head
*jh
;
1344 char *committed_data
= NULL
;
1346 if (is_handle_aborted(handle
))
1349 if (jbd2_write_access_granted(handle
, bh
, true))
1352 jh
= jbd2_journal_add_journal_head(bh
);
1353 JBUFFER_TRACE(jh
, "entry");
1356 * Do this first --- it can drop the journal lock, so we want to
1357 * make sure that obtaining the committed_data is done
1358 * atomically wrt. completion of any outstanding commits.
1360 err
= do_get_write_access(handle
, jh
, 1);
1365 if (!jh
->b_committed_data
)
1366 committed_data
= jbd2_alloc(jh2bh(jh
)->b_size
,
1367 GFP_NOFS
|__GFP_NOFAIL
);
1369 spin_lock(&jh
->b_state_lock
);
1370 if (!jh
->b_committed_data
) {
1371 /* Copy out the current buffer contents into the
1372 * preserved, committed copy. */
1373 JBUFFER_TRACE(jh
, "generate b_committed data");
1374 if (!committed_data
) {
1375 spin_unlock(&jh
->b_state_lock
);
1379 jh
->b_committed_data
= committed_data
;
1380 committed_data
= NULL
;
1381 memcpy(jh
->b_committed_data
, bh
->b_data
, bh
->b_size
);
1383 spin_unlock(&jh
->b_state_lock
);
1385 jbd2_journal_put_journal_head(jh
);
1386 if (unlikely(committed_data
))
1387 jbd2_free(committed_data
, bh
->b_size
);
1392 * jbd2_journal_set_triggers() - Add triggers for commit writeout
1393 * @bh: buffer to trigger on
1394 * @type: struct jbd2_buffer_trigger_type containing the trigger(s).
1396 * Set any triggers on this journal_head. This is always safe, because
1397 * triggers for a committing buffer will be saved off, and triggers for
1398 * a running transaction will match the buffer in that transaction.
1400 * Call with NULL to clear the triggers.
1402 void jbd2_journal_set_triggers(struct buffer_head
*bh
,
1403 struct jbd2_buffer_trigger_type
*type
)
1405 struct journal_head
*jh
= jbd2_journal_grab_journal_head(bh
);
1409 jh
->b_triggers
= type
;
1410 jbd2_journal_put_journal_head(jh
);
1413 void jbd2_buffer_frozen_trigger(struct journal_head
*jh
, void *mapped_data
,
1414 struct jbd2_buffer_trigger_type
*triggers
)
1416 struct buffer_head
*bh
= jh2bh(jh
);
1418 if (!triggers
|| !triggers
->t_frozen
)
1421 triggers
->t_frozen(triggers
, bh
, mapped_data
, bh
->b_size
);
1424 void jbd2_buffer_abort_trigger(struct journal_head
*jh
,
1425 struct jbd2_buffer_trigger_type
*triggers
)
1427 if (!triggers
|| !triggers
->t_abort
)
1430 triggers
->t_abort(triggers
, jh2bh(jh
));
1434 * jbd2_journal_dirty_metadata() - mark a buffer as containing dirty metadata
1435 * @handle: transaction to add buffer to.
1436 * @bh: buffer to mark
1438 * mark dirty metadata which needs to be journaled as part of the current
1441 * The buffer must have previously had jbd2_journal_get_write_access()
1442 * called so that it has a valid journal_head attached to the buffer
1445 * The buffer is placed on the transaction's metadata list and is marked
1446 * as belonging to the transaction.
1448 * Returns error number or 0 on success.
1450 * Special care needs to be taken if the buffer already belongs to the
1451 * current committing transaction (in which case we should have frozen
1452 * data present for that commit). In that case, we don't relink the
1453 * buffer: that only gets done when the old transaction finally
1454 * completes its commit.
1456 int jbd2_journal_dirty_metadata(handle_t
*handle
, struct buffer_head
*bh
)
1458 transaction_t
*transaction
= handle
->h_transaction
;
1460 struct journal_head
*jh
;
1463 if (is_handle_aborted(handle
))
1465 if (!buffer_jbd(bh
))
1469 * We don't grab jh reference here since the buffer must be part
1470 * of the running transaction.
1473 jbd_debug(5, "journal_head %p\n", jh
);
1474 JBUFFER_TRACE(jh
, "entry");
1477 * This and the following assertions are unreliable since we may see jh
1478 * in inconsistent state unless we grab bh_state lock. But this is
1479 * crucial to catch bugs so let's do a reliable check until the
1480 * lockless handling is fully proven.
1482 if (data_race(jh
->b_transaction
!= transaction
&&
1483 jh
->b_next_transaction
!= transaction
)) {
1484 spin_lock(&jh
->b_state_lock
);
1485 J_ASSERT_JH(jh
, jh
->b_transaction
== transaction
||
1486 jh
->b_next_transaction
== transaction
);
1487 spin_unlock(&jh
->b_state_lock
);
1489 if (jh
->b_modified
== 1) {
1490 /* If it's in our transaction it must be in BJ_Metadata list. */
1491 if (data_race(jh
->b_transaction
== transaction
&&
1492 jh
->b_jlist
!= BJ_Metadata
)) {
1493 spin_lock(&jh
->b_state_lock
);
1494 if (jh
->b_transaction
== transaction
&&
1495 jh
->b_jlist
!= BJ_Metadata
)
1496 pr_err("JBD2: assertion failure: h_type=%u "
1497 "h_line_no=%u block_no=%llu jlist=%u\n",
1498 handle
->h_type
, handle
->h_line_no
,
1499 (unsigned long long) bh
->b_blocknr
,
1501 J_ASSERT_JH(jh
, jh
->b_transaction
!= transaction
||
1502 jh
->b_jlist
== BJ_Metadata
);
1503 spin_unlock(&jh
->b_state_lock
);
1508 journal
= transaction
->t_journal
;
1509 spin_lock(&jh
->b_state_lock
);
1511 if (jh
->b_modified
== 0) {
1513 * This buffer's got modified and becoming part
1514 * of the transaction. This needs to be done
1515 * once a transaction -bzzz
1517 if (WARN_ON_ONCE(jbd2_handle_buffer_credits(handle
) <= 0)) {
1522 handle
->h_total_credits
--;
1526 * fastpath, to avoid expensive locking. If this buffer is already
1527 * on the running transaction's metadata list there is nothing to do.
1528 * Nobody can take it off again because there is a handle open.
1529 * I _think_ we're OK here with SMP barriers - a mistaken decision will
1530 * result in this test being false, so we go in and take the locks.
1532 if (jh
->b_transaction
== transaction
&& jh
->b_jlist
== BJ_Metadata
) {
1533 JBUFFER_TRACE(jh
, "fastpath");
1534 if (unlikely(jh
->b_transaction
!=
1535 journal
->j_running_transaction
)) {
1536 printk(KERN_ERR
"JBD2: %s: "
1537 "jh->b_transaction (%llu, %p, %u) != "
1538 "journal->j_running_transaction (%p, %u)\n",
1540 (unsigned long long) bh
->b_blocknr
,
1542 jh
->b_transaction
? jh
->b_transaction
->t_tid
: 0,
1543 journal
->j_running_transaction
,
1544 journal
->j_running_transaction
?
1545 journal
->j_running_transaction
->t_tid
: 0);
1551 set_buffer_jbddirty(bh
);
1554 * Metadata already on the current transaction list doesn't
1555 * need to be filed. Metadata on another transaction's list must
1556 * be committing, and will be refiled once the commit completes:
1557 * leave it alone for now.
1559 if (jh
->b_transaction
!= transaction
) {
1560 JBUFFER_TRACE(jh
, "already on other transaction");
1561 if (unlikely(((jh
->b_transaction
!=
1562 journal
->j_committing_transaction
)) ||
1563 (jh
->b_next_transaction
!= transaction
))) {
1564 printk(KERN_ERR
"jbd2_journal_dirty_metadata: %s: "
1565 "bad jh for block %llu: "
1566 "transaction (%p, %u), "
1567 "jh->b_transaction (%p, %u), "
1568 "jh->b_next_transaction (%p, %u), jlist %u\n",
1570 (unsigned long long) bh
->b_blocknr
,
1571 transaction
, transaction
->t_tid
,
1574 jh
->b_transaction
->t_tid
: 0,
1575 jh
->b_next_transaction
,
1576 jh
->b_next_transaction
?
1577 jh
->b_next_transaction
->t_tid
: 0,
1582 /* And this case is illegal: we can't reuse another
1583 * transaction's data buffer, ever. */
1587 /* That test should have eliminated the following case: */
1588 J_ASSERT_JH(jh
, jh
->b_frozen_data
== NULL
);
1590 JBUFFER_TRACE(jh
, "file as BJ_Metadata");
1591 spin_lock(&journal
->j_list_lock
);
1592 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Metadata
);
1593 spin_unlock(&journal
->j_list_lock
);
1595 spin_unlock(&jh
->b_state_lock
);
1597 JBUFFER_TRACE(jh
, "exit");
1602 * jbd2_journal_forget() - bforget() for potentially-journaled buffers.
1603 * @handle: transaction handle
1604 * @bh: bh to 'forget'
1606 * We can only do the bforget if there are no commits pending against the
1607 * buffer. If the buffer is dirty in the current running transaction we
1608 * can safely unlink it.
1610 * bh may not be a journalled buffer at all - it may be a non-JBD
1611 * buffer which came off the hashtable. Check for this.
1613 * Decrements bh->b_count by one.
1615 * Allow this call even if the handle has aborted --- it may be part of
1616 * the caller's cleanup after an abort.
1618 int jbd2_journal_forget(handle_t
*handle
, struct buffer_head
*bh
)
1620 transaction_t
*transaction
= handle
->h_transaction
;
1622 struct journal_head
*jh
;
1623 int drop_reserve
= 0;
1625 int was_modified
= 0;
1627 if (is_handle_aborted(handle
))
1629 journal
= transaction
->t_journal
;
1631 BUFFER_TRACE(bh
, "entry");
1633 jh
= jbd2_journal_grab_journal_head(bh
);
1639 spin_lock(&jh
->b_state_lock
);
1641 /* Critical error: attempting to delete a bitmap buffer, maybe?
1642 * Don't do any jbd operations, and return an error. */
1643 if (!J_EXPECT_JH(jh
, !jh
->b_committed_data
,
1644 "inconsistent data on disk")) {
1649 /* keep track of whether or not this transaction modified us */
1650 was_modified
= jh
->b_modified
;
1653 * The buffer's going from the transaction, we must drop
1654 * all references -bzzz
1658 if (jh
->b_transaction
== transaction
) {
1659 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
1661 /* If we are forgetting a buffer which is already part
1662 * of this transaction, then we can just drop it from
1663 * the transaction immediately. */
1664 clear_buffer_dirty(bh
);
1665 clear_buffer_jbddirty(bh
);
1667 JBUFFER_TRACE(jh
, "belongs to current transaction: unfile");
1670 * we only want to drop a reference if this transaction
1671 * modified the buffer
1677 * We are no longer going to journal this buffer.
1678 * However, the commit of this transaction is still
1679 * important to the buffer: the delete that we are now
1680 * processing might obsolete an old log entry, so by
1681 * committing, we can satisfy the buffer's checkpoint.
1683 * So, if we have a checkpoint on the buffer, we should
1684 * now refile the buffer on our BJ_Forget list so that
1685 * we know to remove the checkpoint after we commit.
1688 spin_lock(&journal
->j_list_lock
);
1689 if (jh
->b_cp_transaction
) {
1690 __jbd2_journal_temp_unlink_buffer(jh
);
1691 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Forget
);
1693 __jbd2_journal_unfile_buffer(jh
);
1694 jbd2_journal_put_journal_head(jh
);
1696 spin_unlock(&journal
->j_list_lock
);
1697 } else if (jh
->b_transaction
) {
1698 J_ASSERT_JH(jh
, (jh
->b_transaction
==
1699 journal
->j_committing_transaction
));
1700 /* However, if the buffer is still owned by a prior
1701 * (committing) transaction, we can't drop it yet... */
1702 JBUFFER_TRACE(jh
, "belongs to older transaction");
1703 /* ... but we CAN drop it from the new transaction through
1704 * marking the buffer as freed and set j_next_transaction to
1705 * the new transaction, so that not only the commit code
1706 * knows it should clear dirty bits when it is done with the
1707 * buffer, but also the buffer can be checkpointed only
1708 * after the new transaction commits. */
1710 set_buffer_freed(bh
);
1712 if (!jh
->b_next_transaction
) {
1713 spin_lock(&journal
->j_list_lock
);
1714 jh
->b_next_transaction
= transaction
;
1715 spin_unlock(&journal
->j_list_lock
);
1717 J_ASSERT(jh
->b_next_transaction
== transaction
);
1720 * only drop a reference if this transaction modified
1728 * Finally, if the buffer is not belongs to any
1729 * transaction, we can just drop it now if it has no
1732 spin_lock(&journal
->j_list_lock
);
1733 if (!jh
->b_cp_transaction
) {
1734 JBUFFER_TRACE(jh
, "belongs to none transaction");
1735 spin_unlock(&journal
->j_list_lock
);
1740 * Otherwise, if the buffer has been written to disk,
1741 * it is safe to remove the checkpoint and drop it.
1743 if (!buffer_dirty(bh
)) {
1744 __jbd2_journal_remove_checkpoint(jh
);
1745 spin_unlock(&journal
->j_list_lock
);
1750 * The buffer is still not written to disk, we should
1751 * attach this buffer to current transaction so that the
1752 * buffer can be checkpointed only after the current
1753 * transaction commits.
1755 clear_buffer_dirty(bh
);
1756 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Forget
);
1757 spin_unlock(&journal
->j_list_lock
);
1761 spin_unlock(&jh
->b_state_lock
);
1762 jbd2_journal_put_journal_head(jh
);
1764 /* no need to reserve log space for this block -bzzz */
1765 handle
->h_total_credits
++;
1771 * jbd2_journal_stop() - complete a transaction
1772 * @handle: transaction to complete.
1774 * All done for a particular handle.
1776 * There is not much action needed here. We just return any remaining
1777 * buffer credits to the transaction and remove the handle. The only
1778 * complication is that we need to start a commit operation if the
1779 * filesystem is marked for synchronous update.
1781 * jbd2_journal_stop itself will not usually return an error, but it may
1782 * do so in unusual circumstances. In particular, expect it to
1783 * return -EIO if a jbd2_journal_abort has been executed since the
1784 * transaction began.
1786 int jbd2_journal_stop(handle_t
*handle
)
1788 transaction_t
*transaction
= handle
->h_transaction
;
1790 int err
= 0, wait_for_commit
= 0;
1794 if (--handle
->h_ref
> 0) {
1795 jbd_debug(4, "h_ref %d -> %d\n", handle
->h_ref
+ 1,
1797 if (is_handle_aborted(handle
))
1803 * Handle is already detached from the transaction so there is
1804 * nothing to do other than free the handle.
1806 memalloc_nofs_restore(handle
->saved_alloc_context
);
1809 journal
= transaction
->t_journal
;
1810 tid
= transaction
->t_tid
;
1812 if (is_handle_aborted(handle
))
1815 jbd_debug(4, "Handle %p going down\n", handle
);
1816 trace_jbd2_handle_stats(journal
->j_fs_dev
->bd_dev
,
1817 tid
, handle
->h_type
, handle
->h_line_no
,
1818 jiffies
- handle
->h_start_jiffies
,
1819 handle
->h_sync
, handle
->h_requested_credits
,
1820 (handle
->h_requested_credits
-
1821 handle
->h_total_credits
));
1824 * Implement synchronous transaction batching. If the handle
1825 * was synchronous, don't force a commit immediately. Let's
1826 * yield and let another thread piggyback onto this
1827 * transaction. Keep doing that while new threads continue to
1828 * arrive. It doesn't cost much - we're about to run a commit
1829 * and sleep on IO anyway. Speeds up many-threaded, many-dir
1830 * operations by 30x or more...
1832 * We try and optimize the sleep time against what the
1833 * underlying disk can do, instead of having a static sleep
1834 * time. This is useful for the case where our storage is so
1835 * fast that it is more optimal to go ahead and force a flush
1836 * and wait for the transaction to be committed than it is to
1837 * wait for an arbitrary amount of time for new writers to
1838 * join the transaction. We achieve this by measuring how
1839 * long it takes to commit a transaction, and compare it with
1840 * how long this transaction has been running, and if run time
1841 * < commit time then we sleep for the delta and commit. This
1842 * greatly helps super fast disks that would see slowdowns as
1843 * more threads started doing fsyncs.
1845 * But don't do this if this process was the most recent one
1846 * to perform a synchronous write. We do this to detect the
1847 * case where a single process is doing a stream of sync
1848 * writes. No point in waiting for joiners in that case.
1850 * Setting max_batch_time to 0 disables this completely.
1853 if (handle
->h_sync
&& journal
->j_last_sync_writer
!= pid
&&
1854 journal
->j_max_batch_time
) {
1855 u64 commit_time
, trans_time
;
1857 journal
->j_last_sync_writer
= pid
;
1859 read_lock(&journal
->j_state_lock
);
1860 commit_time
= journal
->j_average_commit_time
;
1861 read_unlock(&journal
->j_state_lock
);
1863 trans_time
= ktime_to_ns(ktime_sub(ktime_get(),
1864 transaction
->t_start_time
));
1866 commit_time
= max_t(u64
, commit_time
,
1867 1000*journal
->j_min_batch_time
);
1868 commit_time
= min_t(u64
, commit_time
,
1869 1000*journal
->j_max_batch_time
);
1871 if (trans_time
< commit_time
) {
1872 ktime_t expires
= ktime_add_ns(ktime_get(),
1874 set_current_state(TASK_UNINTERRUPTIBLE
);
1875 schedule_hrtimeout(&expires
, HRTIMER_MODE_ABS
);
1880 transaction
->t_synchronous_commit
= 1;
1883 * If the handle is marked SYNC, we need to set another commit
1884 * going! We also want to force a commit if the transaction is too
1887 if (handle
->h_sync
||
1888 time_after_eq(jiffies
, transaction
->t_expires
)) {
1889 /* Do this even for aborted journals: an abort still
1890 * completes the commit thread, it just doesn't write
1891 * anything to disk. */
1893 jbd_debug(2, "transaction too old, requesting commit for "
1894 "handle %p\n", handle
);
1895 /* This is non-blocking */
1896 jbd2_log_start_commit(journal
, tid
);
1899 * Special case: JBD2_SYNC synchronous updates require us
1900 * to wait for the commit to complete.
1902 if (handle
->h_sync
&& !(current
->flags
& PF_MEMALLOC
))
1903 wait_for_commit
= 1;
1907 * Once stop_this_handle() drops t_updates, the transaction could start
1908 * committing on us and eventually disappear. So we must not
1909 * dereference transaction pointer again after calling
1910 * stop_this_handle().
1912 stop_this_handle(handle
);
1914 if (wait_for_commit
)
1915 err
= jbd2_log_wait_commit(journal
, tid
);
1918 if (handle
->h_rsv_handle
)
1919 jbd2_free_handle(handle
->h_rsv_handle
);
1920 jbd2_free_handle(handle
);
1926 * List management code snippets: various functions for manipulating the
1927 * transaction buffer lists.
1932 * Append a buffer to a transaction list, given the transaction's list head
1935 * j_list_lock is held.
1937 * jh->b_state_lock is held.
1941 __blist_add_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1944 jh
->b_tnext
= jh
->b_tprev
= jh
;
1947 /* Insert at the tail of the list to preserve order */
1948 struct journal_head
*first
= *list
, *last
= first
->b_tprev
;
1950 jh
->b_tnext
= first
;
1951 last
->b_tnext
= first
->b_tprev
= jh
;
1956 * Remove a buffer from a transaction list, given the transaction's list
1959 * Called with j_list_lock held, and the journal may not be locked.
1961 * jh->b_state_lock is held.
1965 __blist_del_buffer(struct journal_head
**list
, struct journal_head
*jh
)
1968 *list
= jh
->b_tnext
;
1972 jh
->b_tprev
->b_tnext
= jh
->b_tnext
;
1973 jh
->b_tnext
->b_tprev
= jh
->b_tprev
;
1977 * Remove a buffer from the appropriate transaction list.
1979 * Note that this function can *change* the value of
1980 * bh->b_transaction->t_buffers, t_forget, t_shadow_list, t_log_list or
1981 * t_reserved_list. If the caller is holding onto a copy of one of these
1982 * pointers, it could go bad. Generally the caller needs to re-read the
1983 * pointer from the transaction_t.
1985 * Called under j_list_lock.
1987 static void __jbd2_journal_temp_unlink_buffer(struct journal_head
*jh
)
1989 struct journal_head
**list
= NULL
;
1990 transaction_t
*transaction
;
1991 struct buffer_head
*bh
= jh2bh(jh
);
1993 lockdep_assert_held(&jh
->b_state_lock
);
1994 transaction
= jh
->b_transaction
;
1996 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
1998 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
1999 if (jh
->b_jlist
!= BJ_None
)
2000 J_ASSERT_JH(jh
, transaction
!= NULL
);
2002 switch (jh
->b_jlist
) {
2006 transaction
->t_nr_buffers
--;
2007 J_ASSERT_JH(jh
, transaction
->t_nr_buffers
>= 0);
2008 list
= &transaction
->t_buffers
;
2011 list
= &transaction
->t_forget
;
2014 list
= &transaction
->t_shadow_list
;
2017 list
= &transaction
->t_reserved_list
;
2021 __blist_del_buffer(list
, jh
);
2022 jh
->b_jlist
= BJ_None
;
2023 if (transaction
&& is_journal_aborted(transaction
->t_journal
))
2024 clear_buffer_jbddirty(bh
);
2025 else if (test_clear_buffer_jbddirty(bh
))
2026 mark_buffer_dirty(bh
); /* Expose it to the VM */
2030 * Remove buffer from all transactions. The caller is responsible for dropping
2031 * the jh reference that belonged to the transaction.
2033 * Called with bh_state lock and j_list_lock
2035 static void __jbd2_journal_unfile_buffer(struct journal_head
*jh
)
2037 J_ASSERT_JH(jh
, jh
->b_transaction
!= NULL
);
2038 J_ASSERT_JH(jh
, jh
->b_next_transaction
== NULL
);
2040 __jbd2_journal_temp_unlink_buffer(jh
);
2041 jh
->b_transaction
= NULL
;
2044 void jbd2_journal_unfile_buffer(journal_t
*journal
, struct journal_head
*jh
)
2046 struct buffer_head
*bh
= jh2bh(jh
);
2048 /* Get reference so that buffer cannot be freed before we unlock it */
2050 spin_lock(&jh
->b_state_lock
);
2051 spin_lock(&journal
->j_list_lock
);
2052 __jbd2_journal_unfile_buffer(jh
);
2053 spin_unlock(&journal
->j_list_lock
);
2054 spin_unlock(&jh
->b_state_lock
);
2055 jbd2_journal_put_journal_head(jh
);
2060 * Called from jbd2_journal_try_to_free_buffers().
2062 * Called under jh->b_state_lock
2065 __journal_try_to_free_buffer(journal_t
*journal
, struct buffer_head
*bh
)
2067 struct journal_head
*jh
;
2071 if (buffer_locked(bh
) || buffer_dirty(bh
))
2074 if (jh
->b_next_transaction
!= NULL
|| jh
->b_transaction
!= NULL
)
2077 spin_lock(&journal
->j_list_lock
);
2078 if (jh
->b_cp_transaction
!= NULL
) {
2079 /* written-back checkpointed metadata buffer */
2080 JBUFFER_TRACE(jh
, "remove from checkpoint list");
2081 __jbd2_journal_remove_checkpoint(jh
);
2083 spin_unlock(&journal
->j_list_lock
);
2089 * jbd2_journal_try_to_free_buffers() - try to free page buffers.
2090 * @journal: journal for operation
2091 * @page: to try and free
2093 * For all the buffers on this page,
2094 * if they are fully written out ordered data, move them onto BUF_CLEAN
2095 * so try_to_free_buffers() can reap them.
2097 * This function returns non-zero if we wish try_to_free_buffers()
2098 * to be called. We do this if the page is releasable by try_to_free_buffers().
2099 * We also do it if the page has locked or dirty buffers and the caller wants
2100 * us to perform sync or async writeout.
2102 * This complicates JBD locking somewhat. We aren't protected by the
2103 * BKL here. We wish to remove the buffer from its committing or
2104 * running transaction's ->t_datalist via __jbd2_journal_unfile_buffer.
2106 * This may *change* the value of transaction_t->t_datalist, so anyone
2107 * who looks at t_datalist needs to lock against this function.
2109 * Even worse, someone may be doing a jbd2_journal_dirty_data on this
2110 * buffer. So we need to lock against that. jbd2_journal_dirty_data()
2111 * will come out of the lock with the buffer dirty, which makes it
2112 * ineligible for release here.
2114 * Who else is affected by this? hmm... Really the only contender
2115 * is do_get_write_access() - it could be looking at the buffer while
2116 * journal_try_to_free_buffer() is changing its state. But that
2117 * cannot happen because we never reallocate freed data as metadata
2118 * while the data is part of a transaction. Yes?
2120 * Return 0 on failure, 1 on success
2122 int jbd2_journal_try_to_free_buffers(journal_t
*journal
, struct page
*page
)
2124 struct buffer_head
*head
;
2125 struct buffer_head
*bh
;
2126 bool has_write_io_error
= false;
2129 J_ASSERT(PageLocked(page
));
2131 head
= page_buffers(page
);
2134 struct journal_head
*jh
;
2137 * We take our own ref against the journal_head here to avoid
2138 * having to add tons of locking around each instance of
2139 * jbd2_journal_put_journal_head().
2141 jh
= jbd2_journal_grab_journal_head(bh
);
2145 spin_lock(&jh
->b_state_lock
);
2146 __journal_try_to_free_buffer(journal
, bh
);
2147 spin_unlock(&jh
->b_state_lock
);
2148 jbd2_journal_put_journal_head(jh
);
2153 * If we free a metadata buffer which has been failed to
2154 * write out, the jbd2 checkpoint procedure will not detect
2155 * this failure and may lead to filesystem inconsistency
2156 * after cleanup journal tail.
2158 if (buffer_write_io_error(bh
)) {
2159 pr_err("JBD2: Error while async write back metadata bh %llu.",
2160 (unsigned long long)bh
->b_blocknr
);
2161 has_write_io_error
= true;
2163 } while ((bh
= bh
->b_this_page
) != head
);
2165 ret
= try_to_free_buffers(page
);
2168 if (has_write_io_error
)
2169 jbd2_journal_abort(journal
, -EIO
);
2175 * This buffer is no longer needed. If it is on an older transaction's
2176 * checkpoint list we need to record it on this transaction's forget list
2177 * to pin this buffer (and hence its checkpointing transaction) down until
2178 * this transaction commits. If the buffer isn't on a checkpoint list, we
2180 * Returns non-zero if JBD no longer has an interest in the buffer.
2182 * Called under j_list_lock.
2184 * Called under jh->b_state_lock.
2186 static int __dispose_buffer(struct journal_head
*jh
, transaction_t
*transaction
)
2189 struct buffer_head
*bh
= jh2bh(jh
);
2191 if (jh
->b_cp_transaction
) {
2192 JBUFFER_TRACE(jh
, "on running+cp transaction");
2193 __jbd2_journal_temp_unlink_buffer(jh
);
2195 * We don't want to write the buffer anymore, clear the
2196 * bit so that we don't confuse checks in
2197 * __journal_file_buffer
2199 clear_buffer_dirty(bh
);
2200 __jbd2_journal_file_buffer(jh
, transaction
, BJ_Forget
);
2203 JBUFFER_TRACE(jh
, "on running transaction");
2204 __jbd2_journal_unfile_buffer(jh
);
2205 jbd2_journal_put_journal_head(jh
);
2211 * jbd2_journal_invalidatepage
2213 * This code is tricky. It has a number of cases to deal with.
2215 * There are two invariants which this code relies on:
2217 * i_size must be updated on disk before we start calling invalidatepage on the
2220 * This is done in ext3 by defining an ext3_setattr method which
2221 * updates i_size before truncate gets going. By maintaining this
2222 * invariant, we can be sure that it is safe to throw away any buffers
2223 * attached to the current transaction: once the transaction commits,
2224 * we know that the data will not be needed.
2226 * Note however that we can *not* throw away data belonging to the
2227 * previous, committing transaction!
2229 * Any disk blocks which *are* part of the previous, committing
2230 * transaction (and which therefore cannot be discarded immediately) are
2231 * not going to be reused in the new running transaction
2233 * The bitmap committed_data images guarantee this: any block which is
2234 * allocated in one transaction and removed in the next will be marked
2235 * as in-use in the committed_data bitmap, so cannot be reused until
2236 * the next transaction to delete the block commits. This means that
2237 * leaving committing buffers dirty is quite safe: the disk blocks
2238 * cannot be reallocated to a different file and so buffer aliasing is
2242 * The above applies mainly to ordered data mode. In writeback mode we
2243 * don't make guarantees about the order in which data hits disk --- in
2244 * particular we don't guarantee that new dirty data is flushed before
2245 * transaction commit --- so it is always safe just to discard data
2246 * immediately in that mode. --sct
2250 * The journal_unmap_buffer helper function returns zero if the buffer
2251 * concerned remains pinned as an anonymous buffer belonging to an older
2254 * We're outside-transaction here. Either or both of j_running_transaction
2255 * and j_committing_transaction may be NULL.
2257 static int journal_unmap_buffer(journal_t
*journal
, struct buffer_head
*bh
,
2260 transaction_t
*transaction
;
2261 struct journal_head
*jh
;
2264 BUFFER_TRACE(bh
, "entry");
2267 * It is safe to proceed here without the j_list_lock because the
2268 * buffers cannot be stolen by try_to_free_buffers as long as we are
2269 * holding the page lock. --sct
2272 jh
= jbd2_journal_grab_journal_head(bh
);
2274 goto zap_buffer_unlocked
;
2276 /* OK, we have data buffer in journaled mode */
2277 write_lock(&journal
->j_state_lock
);
2278 spin_lock(&jh
->b_state_lock
);
2279 spin_lock(&journal
->j_list_lock
);
2282 * We cannot remove the buffer from checkpoint lists until the
2283 * transaction adding inode to orphan list (let's call it T)
2284 * is committed. Otherwise if the transaction changing the
2285 * buffer would be cleaned from the journal before T is
2286 * committed, a crash will cause that the correct contents of
2287 * the buffer will be lost. On the other hand we have to
2288 * clear the buffer dirty bit at latest at the moment when the
2289 * transaction marking the buffer as freed in the filesystem
2290 * structures is committed because from that moment on the
2291 * block can be reallocated and used by a different page.
2292 * Since the block hasn't been freed yet but the inode has
2293 * already been added to orphan list, it is safe for us to add
2294 * the buffer to BJ_Forget list of the newest transaction.
2296 * Also we have to clear buffer_mapped flag of a truncated buffer
2297 * because the buffer_head may be attached to the page straddling
2298 * i_size (can happen only when blocksize < pagesize) and thus the
2299 * buffer_head can be reused when the file is extended again. So we end
2300 * up keeping around invalidated buffers attached to transactions'
2301 * BJ_Forget list just to stop checkpointing code from cleaning up
2302 * the transaction this buffer was modified in.
2304 transaction
= jh
->b_transaction
;
2305 if (transaction
== NULL
) {
2306 /* First case: not on any transaction. If it
2307 * has no checkpoint link, then we can zap it:
2308 * it's a writeback-mode buffer so we don't care
2309 * if it hits disk safely. */
2310 if (!jh
->b_cp_transaction
) {
2311 JBUFFER_TRACE(jh
, "not on any transaction: zap");
2315 if (!buffer_dirty(bh
)) {
2316 /* bdflush has written it. We can drop it now */
2317 __jbd2_journal_remove_checkpoint(jh
);
2321 /* OK, it must be in the journal but still not
2322 * written fully to disk: it's metadata or
2323 * journaled data... */
2325 if (journal
->j_running_transaction
) {
2326 /* ... and once the current transaction has
2327 * committed, the buffer won't be needed any
2329 JBUFFER_TRACE(jh
, "checkpointed: add to BJ_Forget");
2330 may_free
= __dispose_buffer(jh
,
2331 journal
->j_running_transaction
);
2334 /* There is no currently-running transaction. So the
2335 * orphan record which we wrote for this file must have
2336 * passed into commit. We must attach this buffer to
2337 * the committing transaction, if it exists. */
2338 if (journal
->j_committing_transaction
) {
2339 JBUFFER_TRACE(jh
, "give to committing trans");
2340 may_free
= __dispose_buffer(jh
,
2341 journal
->j_committing_transaction
);
2344 /* The orphan record's transaction has
2345 * committed. We can cleanse this buffer */
2346 clear_buffer_jbddirty(bh
);
2347 __jbd2_journal_remove_checkpoint(jh
);
2351 } else if (transaction
== journal
->j_committing_transaction
) {
2352 JBUFFER_TRACE(jh
, "on committing transaction");
2354 * The buffer is committing, we simply cannot touch
2355 * it. If the page is straddling i_size we have to wait
2356 * for commit and try again.
2359 spin_unlock(&journal
->j_list_lock
);
2360 spin_unlock(&jh
->b_state_lock
);
2361 write_unlock(&journal
->j_state_lock
);
2362 jbd2_journal_put_journal_head(jh
);
2366 * OK, buffer won't be reachable after truncate. We just clear
2367 * b_modified to not confuse transaction credit accounting, and
2368 * set j_next_transaction to the running transaction (if there
2369 * is one) and mark buffer as freed so that commit code knows
2370 * it should clear dirty bits when it is done with the buffer.
2372 set_buffer_freed(bh
);
2373 if (journal
->j_running_transaction
&& buffer_jbddirty(bh
))
2374 jh
->b_next_transaction
= journal
->j_running_transaction
;
2376 spin_unlock(&journal
->j_list_lock
);
2377 spin_unlock(&jh
->b_state_lock
);
2378 write_unlock(&journal
->j_state_lock
);
2379 jbd2_journal_put_journal_head(jh
);
2382 /* Good, the buffer belongs to the running transaction.
2383 * We are writing our own transaction's data, not any
2384 * previous one's, so it is safe to throw it away
2385 * (remember that we expect the filesystem to have set
2386 * i_size already for this truncate so recovery will not
2387 * expose the disk blocks we are discarding here.) */
2388 J_ASSERT_JH(jh
, transaction
== journal
->j_running_transaction
);
2389 JBUFFER_TRACE(jh
, "on running transaction");
2390 may_free
= __dispose_buffer(jh
, transaction
);
2395 * This is tricky. Although the buffer is truncated, it may be reused
2396 * if blocksize < pagesize and it is attached to the page straddling
2397 * EOF. Since the buffer might have been added to BJ_Forget list of the
2398 * running transaction, journal_get_write_access() won't clear
2399 * b_modified and credit accounting gets confused. So clear b_modified
2403 spin_unlock(&journal
->j_list_lock
);
2404 spin_unlock(&jh
->b_state_lock
);
2405 write_unlock(&journal
->j_state_lock
);
2406 jbd2_journal_put_journal_head(jh
);
2407 zap_buffer_unlocked
:
2408 clear_buffer_dirty(bh
);
2409 J_ASSERT_BH(bh
, !buffer_jbddirty(bh
));
2410 clear_buffer_mapped(bh
);
2411 clear_buffer_req(bh
);
2412 clear_buffer_new(bh
);
2413 clear_buffer_delay(bh
);
2414 clear_buffer_unwritten(bh
);
2420 * jbd2_journal_invalidatepage()
2421 * @journal: journal to use for flush...
2422 * @page: page to flush
2423 * @offset: start of the range to invalidate
2424 * @length: length of the range to invalidate
2426 * Reap page buffers containing data after in the specified range in page.
2427 * Can return -EBUSY if buffers are part of the committing transaction and
2428 * the page is straddling i_size. Caller then has to wait for current commit
2431 int jbd2_journal_invalidatepage(journal_t
*journal
,
2433 unsigned int offset
,
2434 unsigned int length
)
2436 struct buffer_head
*head
, *bh
, *next
;
2437 unsigned int stop
= offset
+ length
;
2438 unsigned int curr_off
= 0;
2439 int partial_page
= (offset
|| length
< PAGE_SIZE
);
2443 if (!PageLocked(page
))
2445 if (!page_has_buffers(page
))
2448 BUG_ON(stop
> PAGE_SIZE
|| stop
< length
);
2450 /* We will potentially be playing with lists other than just the
2451 * data lists (especially for journaled data mode), so be
2452 * cautious in our locking. */
2454 head
= bh
= page_buffers(page
);
2456 unsigned int next_off
= curr_off
+ bh
->b_size
;
2457 next
= bh
->b_this_page
;
2459 if (next_off
> stop
)
2462 if (offset
<= curr_off
) {
2463 /* This block is wholly outside the truncation point */
2465 ret
= journal_unmap_buffer(journal
, bh
, partial_page
);
2471 curr_off
= next_off
;
2474 } while (bh
!= head
);
2476 if (!partial_page
) {
2477 if (may_free
&& try_to_free_buffers(page
))
2478 J_ASSERT(!page_has_buffers(page
));
2484 * File a buffer on the given transaction list.
2486 void __jbd2_journal_file_buffer(struct journal_head
*jh
,
2487 transaction_t
*transaction
, int jlist
)
2489 struct journal_head
**list
= NULL
;
2491 struct buffer_head
*bh
= jh2bh(jh
);
2493 lockdep_assert_held(&jh
->b_state_lock
);
2494 assert_spin_locked(&transaction
->t_journal
->j_list_lock
);
2496 J_ASSERT_JH(jh
, jh
->b_jlist
< BJ_Types
);
2497 J_ASSERT_JH(jh
, jh
->b_transaction
== transaction
||
2498 jh
->b_transaction
== NULL
);
2500 if (jh
->b_transaction
&& jh
->b_jlist
== jlist
)
2503 if (jlist
== BJ_Metadata
|| jlist
== BJ_Reserved
||
2504 jlist
== BJ_Shadow
|| jlist
== BJ_Forget
) {
2506 * For metadata buffers, we track dirty bit in buffer_jbddirty
2507 * instead of buffer_dirty. We should not see a dirty bit set
2508 * here because we clear it in do_get_write_access but e.g.
2509 * tune2fs can modify the sb and set the dirty bit at any time
2510 * so we try to gracefully handle that.
2512 if (buffer_dirty(bh
))
2513 warn_dirty_buffer(bh
);
2514 if (test_clear_buffer_dirty(bh
) ||
2515 test_clear_buffer_jbddirty(bh
))
2519 if (jh
->b_transaction
)
2520 __jbd2_journal_temp_unlink_buffer(jh
);
2522 jbd2_journal_grab_journal_head(bh
);
2523 jh
->b_transaction
= transaction
;
2527 J_ASSERT_JH(jh
, !jh
->b_committed_data
);
2528 J_ASSERT_JH(jh
, !jh
->b_frozen_data
);
2531 transaction
->t_nr_buffers
++;
2532 list
= &transaction
->t_buffers
;
2535 list
= &transaction
->t_forget
;
2538 list
= &transaction
->t_shadow_list
;
2541 list
= &transaction
->t_reserved_list
;
2545 __blist_add_buffer(list
, jh
);
2546 jh
->b_jlist
= jlist
;
2549 set_buffer_jbddirty(bh
);
2552 void jbd2_journal_file_buffer(struct journal_head
*jh
,
2553 transaction_t
*transaction
, int jlist
)
2555 spin_lock(&jh
->b_state_lock
);
2556 spin_lock(&transaction
->t_journal
->j_list_lock
);
2557 __jbd2_journal_file_buffer(jh
, transaction
, jlist
);
2558 spin_unlock(&transaction
->t_journal
->j_list_lock
);
2559 spin_unlock(&jh
->b_state_lock
);
2563 * Remove a buffer from its current buffer list in preparation for
2564 * dropping it from its current transaction entirely. If the buffer has
2565 * already started to be used by a subsequent transaction, refile the
2566 * buffer on that transaction's metadata list.
2568 * Called under j_list_lock
2569 * Called under jh->b_state_lock
2571 * When this function returns true, there's no next transaction to refile to
2572 * and the caller has to drop jh reference through
2573 * jbd2_journal_put_journal_head().
2575 bool __jbd2_journal_refile_buffer(struct journal_head
*jh
)
2577 int was_dirty
, jlist
;
2578 struct buffer_head
*bh
= jh2bh(jh
);
2580 lockdep_assert_held(&jh
->b_state_lock
);
2581 if (jh
->b_transaction
)
2582 assert_spin_locked(&jh
->b_transaction
->t_journal
->j_list_lock
);
2584 /* If the buffer is now unused, just drop it. */
2585 if (jh
->b_next_transaction
== NULL
) {
2586 __jbd2_journal_unfile_buffer(jh
);
2591 * It has been modified by a later transaction: add it to the new
2592 * transaction's metadata list.
2595 was_dirty
= test_clear_buffer_jbddirty(bh
);
2596 __jbd2_journal_temp_unlink_buffer(jh
);
2599 * b_transaction must be set, otherwise the new b_transaction won't
2600 * be holding jh reference
2602 J_ASSERT_JH(jh
, jh
->b_transaction
!= NULL
);
2605 * We set b_transaction here because b_next_transaction will inherit
2606 * our jh reference and thus __jbd2_journal_file_buffer() must not
2609 WRITE_ONCE(jh
->b_transaction
, jh
->b_next_transaction
);
2610 WRITE_ONCE(jh
->b_next_transaction
, NULL
);
2611 if (buffer_freed(bh
))
2613 else if (jh
->b_modified
)
2614 jlist
= BJ_Metadata
;
2616 jlist
= BJ_Reserved
;
2617 __jbd2_journal_file_buffer(jh
, jh
->b_transaction
, jlist
);
2618 J_ASSERT_JH(jh
, jh
->b_transaction
->t_state
== T_RUNNING
);
2621 set_buffer_jbddirty(bh
);
2626 * __jbd2_journal_refile_buffer() with necessary locking added. We take our
2627 * bh reference so that we can safely unlock bh.
2629 * The jh and bh may be freed by this call.
2631 void jbd2_journal_refile_buffer(journal_t
*journal
, struct journal_head
*jh
)
2635 spin_lock(&jh
->b_state_lock
);
2636 spin_lock(&journal
->j_list_lock
);
2637 drop
= __jbd2_journal_refile_buffer(jh
);
2638 spin_unlock(&jh
->b_state_lock
);
2639 spin_unlock(&journal
->j_list_lock
);
2641 jbd2_journal_put_journal_head(jh
);
2645 * File inode in the inode list of the handle's transaction
2647 static int jbd2_journal_file_inode(handle_t
*handle
, struct jbd2_inode
*jinode
,
2648 unsigned long flags
, loff_t start_byte
, loff_t end_byte
)
2650 transaction_t
*transaction
= handle
->h_transaction
;
2653 if (is_handle_aborted(handle
))
2655 journal
= transaction
->t_journal
;
2657 jbd_debug(4, "Adding inode %lu, tid:%d\n", jinode
->i_vfs_inode
->i_ino
,
2658 transaction
->t_tid
);
2660 spin_lock(&journal
->j_list_lock
);
2661 jinode
->i_flags
|= flags
;
2663 if (jinode
->i_dirty_end
) {
2664 jinode
->i_dirty_start
= min(jinode
->i_dirty_start
, start_byte
);
2665 jinode
->i_dirty_end
= max(jinode
->i_dirty_end
, end_byte
);
2667 jinode
->i_dirty_start
= start_byte
;
2668 jinode
->i_dirty_end
= end_byte
;
2671 /* Is inode already attached where we need it? */
2672 if (jinode
->i_transaction
== transaction
||
2673 jinode
->i_next_transaction
== transaction
)
2677 * We only ever set this variable to 1 so the test is safe. Since
2678 * t_need_data_flush is likely to be set, we do the test to save some
2679 * cacheline bouncing
2681 if (!transaction
->t_need_data_flush
)
2682 transaction
->t_need_data_flush
= 1;
2683 /* On some different transaction's list - should be
2684 * the committing one */
2685 if (jinode
->i_transaction
) {
2686 J_ASSERT(jinode
->i_next_transaction
== NULL
);
2687 J_ASSERT(jinode
->i_transaction
==
2688 journal
->j_committing_transaction
);
2689 jinode
->i_next_transaction
= transaction
;
2692 /* Not on any transaction list... */
2693 J_ASSERT(!jinode
->i_next_transaction
);
2694 jinode
->i_transaction
= transaction
;
2695 list_add(&jinode
->i_list
, &transaction
->t_inode_list
);
2697 spin_unlock(&journal
->j_list_lock
);
2702 int jbd2_journal_inode_ranged_write(handle_t
*handle
,
2703 struct jbd2_inode
*jinode
, loff_t start_byte
, loff_t length
)
2705 return jbd2_journal_file_inode(handle
, jinode
,
2706 JI_WRITE_DATA
| JI_WAIT_DATA
, start_byte
,
2707 start_byte
+ length
- 1);
2710 int jbd2_journal_inode_ranged_wait(handle_t
*handle
, struct jbd2_inode
*jinode
,
2711 loff_t start_byte
, loff_t length
)
2713 return jbd2_journal_file_inode(handle
, jinode
, JI_WAIT_DATA
,
2714 start_byte
, start_byte
+ length
- 1);
2718 * File truncate and transaction commit interact with each other in a
2719 * non-trivial way. If a transaction writing data block A is
2720 * committing, we cannot discard the data by truncate until we have
2721 * written them. Otherwise if we crashed after the transaction with
2722 * write has committed but before the transaction with truncate has
2723 * committed, we could see stale data in block A. This function is a
2724 * helper to solve this problem. It starts writeout of the truncated
2725 * part in case it is in the committing transaction.
2727 * Filesystem code must call this function when inode is journaled in
2728 * ordered mode before truncation happens and after the inode has been
2729 * placed on orphan list with the new inode size. The second condition
2730 * avoids the race that someone writes new data and we start
2731 * committing the transaction after this function has been called but
2732 * before a transaction for truncate is started (and furthermore it
2733 * allows us to optimize the case where the addition to orphan list
2734 * happens in the same transaction as write --- we don't have to write
2735 * any data in such case).
2737 int jbd2_journal_begin_ordered_truncate(journal_t
*journal
,
2738 struct jbd2_inode
*jinode
,
2741 transaction_t
*inode_trans
, *commit_trans
;
2744 /* This is a quick check to avoid locking if not necessary */
2745 if (!jinode
->i_transaction
)
2747 /* Locks are here just to force reading of recent values, it is
2748 * enough that the transaction was not committing before we started
2749 * a transaction adding the inode to orphan list */
2750 read_lock(&journal
->j_state_lock
);
2751 commit_trans
= journal
->j_committing_transaction
;
2752 read_unlock(&journal
->j_state_lock
);
2753 spin_lock(&journal
->j_list_lock
);
2754 inode_trans
= jinode
->i_transaction
;
2755 spin_unlock(&journal
->j_list_lock
);
2756 if (inode_trans
== commit_trans
) {
2757 ret
= filemap_fdatawrite_range(jinode
->i_vfs_inode
->i_mapping
,
2758 new_size
, LLONG_MAX
);
2760 jbd2_journal_abort(journal
, ret
);