2 * linux/fs/jbd2/commit.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6 * Copyright 1998 Red Hat corp --- All Rights Reserved
8 * This file is part of the Linux kernel and is made available under
9 * the terms of the GNU General Public License, version 2, or at your
10 * option, any later version, incorporated herein by reference.
12 * Journal commit routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
16 #include <linux/time.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
22 #include <linux/pagemap.h>
23 #include <linux/jiffies.h>
24 #include <linux/crc32.h>
25 #include <linux/writeback.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bio.h>
28 #include <linux/blkdev.h>
29 #include <linux/bitops.h>
30 #include <trace/events/jbd2.h>
31 #include <asm/system.h>
34 * Default IO end handler for temporary BJ_IO buffer_heads.
36 static void journal_end_buffer_io_sync(struct buffer_head
*bh
, int uptodate
)
40 set_buffer_uptodate(bh
);
42 clear_buffer_uptodate(bh
);
47 * When an ext4 file is truncated, it is possible that some pages are not
48 * successfully freed, because they are attached to a committing transaction.
49 * After the transaction commits, these pages are left on the LRU, with no
50 * ->mapping, and with attached buffers. These pages are trivially reclaimable
51 * by the VM, but their apparent absence upsets the VM accounting, and it makes
52 * the numbers in /proc/meminfo look odd.
54 * So here, we have a buffer which has just come off the forget list. Look to
55 * see if we can strip all buffers from the backing page.
57 * Called under lock_journal(), and possibly under journal_datalist_lock. The
58 * caller provided us with a ref against the buffer, and we drop that here.
60 static void release_buffer_page(struct buffer_head
*bh
)
66 if (atomic_read(&bh
->b_count
) != 1)
74 /* OK, it's a truncated page */
75 if (!trylock_page(page
))
80 try_to_free_buffers(page
);
82 page_cache_release(page
);
90 * Done it all: now submit the commit record. We should have
91 * cleaned up our previous buffers by now, so if we are in abort
92 * mode we can now just skip the rest of the journal write
95 * Returns 1 if the journal needs to be aborted or 0 on success
97 static int journal_submit_commit_record(journal_t
*journal
,
98 transaction_t
*commit_transaction
,
99 struct buffer_head
**cbh
,
102 struct journal_head
*descriptor
;
103 struct commit_header
*tmp
;
104 struct buffer_head
*bh
;
106 struct timespec now
= current_kernel_time();
110 if (is_journal_aborted(journal
))
113 descriptor
= jbd2_journal_get_descriptor_buffer(journal
);
117 bh
= jh2bh(descriptor
);
119 tmp
= (struct commit_header
*)bh
->b_data
;
120 tmp
->h_magic
= cpu_to_be32(JBD2_MAGIC_NUMBER
);
121 tmp
->h_blocktype
= cpu_to_be32(JBD2_COMMIT_BLOCK
);
122 tmp
->h_sequence
= cpu_to_be32(commit_transaction
->t_tid
);
123 tmp
->h_commit_sec
= cpu_to_be64(now
.tv_sec
);
124 tmp
->h_commit_nsec
= cpu_to_be32(now
.tv_nsec
);
126 if (JBD2_HAS_COMPAT_FEATURE(journal
,
127 JBD2_FEATURE_COMPAT_CHECKSUM
)) {
128 tmp
->h_chksum_type
= JBD2_CRC32_CHKSUM
;
129 tmp
->h_chksum_size
= JBD2_CRC32_CHKSUM_SIZE
;
130 tmp
->h_chksum
[0] = cpu_to_be32(crc32_sum
);
133 JBUFFER_TRACE(descriptor
, "submit commit block");
135 clear_buffer_dirty(bh
);
136 set_buffer_uptodate(bh
);
137 bh
->b_end_io
= journal_end_buffer_io_sync
;
139 if (journal
->j_flags
& JBD2_BARRIER
&&
140 !JBD2_HAS_INCOMPAT_FEATURE(journal
,
141 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
))
142 ret
= submit_bh(WRITE_SYNC
| WRITE_FLUSH_FUA
, bh
);
144 ret
= submit_bh(WRITE_SYNC
, bh
);
151 * This function along with journal_submit_commit_record
152 * allows to write the commit record asynchronously.
154 static int journal_wait_on_commit_record(journal_t
*journal
,
155 struct buffer_head
*bh
)
159 clear_buffer_dirty(bh
);
162 if (unlikely(!buffer_uptodate(bh
)))
164 put_bh(bh
); /* One for getblk() */
165 jbd2_journal_put_journal_head(bh2jh(bh
));
171 * write the filemap data using writepage() address_space_operations.
172 * We don't do block allocation here even for delalloc. We don't
173 * use writepages() because with dealyed allocation we may be doing
174 * block allocation in writepages().
176 static int journal_submit_inode_data_buffers(struct address_space
*mapping
)
179 struct writeback_control wbc
= {
180 .sync_mode
= WB_SYNC_ALL
,
181 .nr_to_write
= mapping
->nrpages
* 2,
183 .range_end
= i_size_read(mapping
->host
),
186 ret
= generic_writepages(mapping
, &wbc
);
191 * Submit all the data buffers of inode associated with the transaction to
194 * We are in a committing transaction. Therefore no new inode can be added to
195 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
196 * operate on from being released while we write out pages.
198 static int journal_submit_data_buffers(journal_t
*journal
,
199 transaction_t
*commit_transaction
)
201 struct jbd2_inode
*jinode
;
203 struct address_space
*mapping
;
205 spin_lock(&journal
->j_list_lock
);
206 list_for_each_entry(jinode
, &commit_transaction
->t_inode_list
, i_list
) {
207 mapping
= jinode
->i_vfs_inode
->i_mapping
;
208 set_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
209 spin_unlock(&journal
->j_list_lock
);
211 * submit the inode data buffers. We use writepage
212 * instead of writepages. Because writepages can do
213 * block allocation with delalloc. We need to write
214 * only allocated blocks here.
216 trace_jbd2_submit_inode_data(jinode
->i_vfs_inode
);
217 err
= journal_submit_inode_data_buffers(mapping
);
220 spin_lock(&journal
->j_list_lock
);
221 J_ASSERT(jinode
->i_transaction
== commit_transaction
);
222 clear_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
223 smp_mb__after_clear_bit();
224 wake_up_bit(&jinode
->i_flags
, __JI_COMMIT_RUNNING
);
226 spin_unlock(&journal
->j_list_lock
);
231 * Wait for data submitted for writeout, refile inodes to proper
232 * transaction if needed.
235 static int journal_finish_inode_data_buffers(journal_t
*journal
,
236 transaction_t
*commit_transaction
)
238 struct jbd2_inode
*jinode
, *next_i
;
241 /* For locking, see the comment in journal_submit_data_buffers() */
242 spin_lock(&journal
->j_list_lock
);
243 list_for_each_entry(jinode
, &commit_transaction
->t_inode_list
, i_list
) {
244 set_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
245 spin_unlock(&journal
->j_list_lock
);
246 err
= filemap_fdatawait(jinode
->i_vfs_inode
->i_mapping
);
249 * Because AS_EIO is cleared by
250 * filemap_fdatawait_range(), set it again so
251 * that user process can get -EIO from fsync().
254 &jinode
->i_vfs_inode
->i_mapping
->flags
);
259 spin_lock(&journal
->j_list_lock
);
260 clear_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
261 smp_mb__after_clear_bit();
262 wake_up_bit(&jinode
->i_flags
, __JI_COMMIT_RUNNING
);
265 /* Now refile inode to proper lists */
266 list_for_each_entry_safe(jinode
, next_i
,
267 &commit_transaction
->t_inode_list
, i_list
) {
268 list_del(&jinode
->i_list
);
269 if (jinode
->i_next_transaction
) {
270 jinode
->i_transaction
= jinode
->i_next_transaction
;
271 jinode
->i_next_transaction
= NULL
;
272 list_add(&jinode
->i_list
,
273 &jinode
->i_transaction
->t_inode_list
);
275 jinode
->i_transaction
= NULL
;
278 spin_unlock(&journal
->j_list_lock
);
283 static __u32
jbd2_checksum_data(__u32 crc32_sum
, struct buffer_head
*bh
)
285 struct page
*page
= bh
->b_page
;
289 addr
= kmap_atomic(page
, KM_USER0
);
290 checksum
= crc32_be(crc32_sum
,
291 (void *)(addr
+ offset_in_page(bh
->b_data
)), bh
->b_size
);
292 kunmap_atomic(addr
, KM_USER0
);
297 static void write_tag_block(int tag_bytes
, journal_block_tag_t
*tag
,
298 unsigned long long block
)
300 tag
->t_blocknr
= cpu_to_be32(block
& (u32
)~0);
301 if (tag_bytes
> JBD2_TAG_SIZE32
)
302 tag
->t_blocknr_high
= cpu_to_be32((block
>> 31) >> 1);
306 * jbd2_journal_commit_transaction
308 * The primary function for committing a transaction to the log. This
309 * function is called by the journal thread to begin a complete commit.
311 void jbd2_journal_commit_transaction(journal_t
*journal
)
313 struct transaction_stats_s stats
;
314 transaction_t
*commit_transaction
;
315 struct journal_head
*jh
, *new_jh
, *descriptor
;
316 struct buffer_head
**wbuf
= journal
->j_wbuf
;
320 unsigned long long blocknr
;
324 journal_header_t
*header
;
325 journal_block_tag_t
*tag
= NULL
;
330 int tag_bytes
= journal_tag_bytes(journal
);
331 struct buffer_head
*cbh
= NULL
; /* For transactional checksums */
332 __u32 crc32_sum
= ~0;
333 struct blk_plug plug
;
336 * First job: lock down the current transaction and wait for
337 * all outstanding updates to complete.
340 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
341 if (journal
->j_flags
& JBD2_FLUSHED
) {
342 jbd_debug(3, "super block updated\n");
343 mutex_lock(&journal
->j_checkpoint_mutex
);
345 * We hold j_checkpoint_mutex so tail cannot change under us.
346 * We don't need any special data guarantees for writing sb
347 * since journal is empty and it is ok for write to be
348 * flushed only with transaction commit.
350 jbd2_journal_update_sb_log_tail(journal
,
351 journal
->j_tail_sequence
,
354 mutex_unlock(&journal
->j_checkpoint_mutex
);
356 jbd_debug(3, "superblock not updated\n");
359 J_ASSERT(journal
->j_running_transaction
!= NULL
);
360 J_ASSERT(journal
->j_committing_transaction
== NULL
);
362 commit_transaction
= journal
->j_running_transaction
;
363 J_ASSERT(commit_transaction
->t_state
== T_RUNNING
);
365 trace_jbd2_start_commit(journal
, commit_transaction
);
366 jbd_debug(1, "JBD2: starting commit of transaction %d\n",
367 commit_transaction
->t_tid
);
369 write_lock(&journal
->j_state_lock
);
370 commit_transaction
->t_state
= T_LOCKED
;
372 trace_jbd2_commit_locking(journal
, commit_transaction
);
373 stats
.run
.rs_wait
= commit_transaction
->t_max_wait
;
374 stats
.run
.rs_locked
= jiffies
;
375 stats
.run
.rs_running
= jbd2_time_diff(commit_transaction
->t_start
,
376 stats
.run
.rs_locked
);
378 spin_lock(&commit_transaction
->t_handle_lock
);
379 while (atomic_read(&commit_transaction
->t_updates
)) {
382 prepare_to_wait(&journal
->j_wait_updates
, &wait
,
383 TASK_UNINTERRUPTIBLE
);
384 if (atomic_read(&commit_transaction
->t_updates
)) {
385 spin_unlock(&commit_transaction
->t_handle_lock
);
386 write_unlock(&journal
->j_state_lock
);
388 write_lock(&journal
->j_state_lock
);
389 spin_lock(&commit_transaction
->t_handle_lock
);
391 finish_wait(&journal
->j_wait_updates
, &wait
);
393 spin_unlock(&commit_transaction
->t_handle_lock
);
395 J_ASSERT (atomic_read(&commit_transaction
->t_outstanding_credits
) <=
396 journal
->j_max_transaction_buffers
);
399 * First thing we are allowed to do is to discard any remaining
400 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
401 * that there are no such buffers: if a large filesystem
402 * operation like a truncate needs to split itself over multiple
403 * transactions, then it may try to do a jbd2_journal_restart() while
404 * there are still BJ_Reserved buffers outstanding. These must
405 * be released cleanly from the current transaction.
407 * In this case, the filesystem must still reserve write access
408 * again before modifying the buffer in the new transaction, but
409 * we do not require it to remember exactly which old buffers it
410 * has reserved. This is consistent with the existing behaviour
411 * that multiple jbd2_journal_get_write_access() calls to the same
412 * buffer are perfectly permissible.
414 while (commit_transaction
->t_reserved_list
) {
415 jh
= commit_transaction
->t_reserved_list
;
416 JBUFFER_TRACE(jh
, "reserved, unused: refile");
418 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
419 * leave undo-committed data.
421 if (jh
->b_committed_data
) {
422 struct buffer_head
*bh
= jh2bh(jh
);
424 jbd_lock_bh_state(bh
);
425 jbd2_free(jh
->b_committed_data
, bh
->b_size
);
426 jh
->b_committed_data
= NULL
;
427 jbd_unlock_bh_state(bh
);
429 jbd2_journal_refile_buffer(journal
, jh
);
433 * Now try to drop any written-back buffers from the journal's
434 * checkpoint lists. We do this *before* commit because it potentially
437 spin_lock(&journal
->j_list_lock
);
438 __jbd2_journal_clean_checkpoint_list(journal
);
439 spin_unlock(&journal
->j_list_lock
);
441 jbd_debug(3, "JBD2: commit phase 1\n");
444 * Clear revoked flag to reflect there is no revoked buffers
445 * in the next transaction which is going to be started.
447 jbd2_clear_buffer_revoked_flags(journal
);
450 * Switch to a new revoke table.
452 jbd2_journal_switch_revoke_table(journal
);
454 trace_jbd2_commit_flushing(journal
, commit_transaction
);
455 stats
.run
.rs_flushing
= jiffies
;
456 stats
.run
.rs_locked
= jbd2_time_diff(stats
.run
.rs_locked
,
457 stats
.run
.rs_flushing
);
459 commit_transaction
->t_state
= T_FLUSH
;
460 journal
->j_committing_transaction
= commit_transaction
;
461 journal
->j_running_transaction
= NULL
;
462 start_time
= ktime_get();
463 commit_transaction
->t_log_start
= journal
->j_head
;
464 wake_up(&journal
->j_wait_transaction_locked
);
465 write_unlock(&journal
->j_state_lock
);
467 jbd_debug(3, "JBD2: commit phase 2\n");
470 * Now start flushing things to disk, in the order they appear
471 * on the transaction lists. Data blocks go first.
473 err
= journal_submit_data_buffers(journal
, commit_transaction
);
475 jbd2_journal_abort(journal
, err
);
477 blk_start_plug(&plug
);
478 jbd2_journal_write_revoke_records(journal
, commit_transaction
,
480 blk_finish_plug(&plug
);
482 jbd_debug(3, "JBD2: commit phase 2\n");
485 * Way to go: we have now written out all of the data for a
486 * transaction! Now comes the tricky part: we need to write out
487 * metadata. Loop over the transaction's entire buffer list:
489 write_lock(&journal
->j_state_lock
);
490 commit_transaction
->t_state
= T_COMMIT
;
491 write_unlock(&journal
->j_state_lock
);
493 trace_jbd2_commit_logging(journal
, commit_transaction
);
494 stats
.run
.rs_logging
= jiffies
;
495 stats
.run
.rs_flushing
= jbd2_time_diff(stats
.run
.rs_flushing
,
496 stats
.run
.rs_logging
);
497 stats
.run
.rs_blocks
=
498 atomic_read(&commit_transaction
->t_outstanding_credits
);
499 stats
.run
.rs_blocks_logged
= 0;
501 J_ASSERT(commit_transaction
->t_nr_buffers
<=
502 atomic_read(&commit_transaction
->t_outstanding_credits
));
507 blk_start_plug(&plug
);
508 while (commit_transaction
->t_buffers
) {
510 /* Find the next buffer to be journaled... */
512 jh
= commit_transaction
->t_buffers
;
514 /* If we're in abort mode, we just un-journal the buffer and
517 if (is_journal_aborted(journal
)) {
518 clear_buffer_jbddirty(jh2bh(jh
));
519 JBUFFER_TRACE(jh
, "journal is aborting: refile");
520 jbd2_buffer_abort_trigger(jh
,
522 jh
->b_frozen_triggers
:
524 jbd2_journal_refile_buffer(journal
, jh
);
525 /* If that was the last one, we need to clean up
526 * any descriptor buffers which may have been
527 * already allocated, even if we are now
529 if (!commit_transaction
->t_buffers
)
530 goto start_journal_io
;
534 /* Make sure we have a descriptor block in which to
535 record the metadata buffer. */
538 struct buffer_head
*bh
;
540 J_ASSERT (bufs
== 0);
542 jbd_debug(4, "JBD2: get descriptor\n");
544 descriptor
= jbd2_journal_get_descriptor_buffer(journal
);
546 jbd2_journal_abort(journal
, -EIO
);
550 bh
= jh2bh(descriptor
);
551 jbd_debug(4, "JBD2: got buffer %llu (%p)\n",
552 (unsigned long long)bh
->b_blocknr
, bh
->b_data
);
553 header
= (journal_header_t
*)&bh
->b_data
[0];
554 header
->h_magic
= cpu_to_be32(JBD2_MAGIC_NUMBER
);
555 header
->h_blocktype
= cpu_to_be32(JBD2_DESCRIPTOR_BLOCK
);
556 header
->h_sequence
= cpu_to_be32(commit_transaction
->t_tid
);
558 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
559 space_left
= bh
->b_size
- sizeof(journal_header_t
);
561 set_buffer_jwrite(bh
);
562 set_buffer_dirty(bh
);
565 /* Record it so that we can wait for IO
567 BUFFER_TRACE(bh
, "ph3: file as descriptor");
568 jbd2_journal_file_buffer(descriptor
, commit_transaction
,
572 /* Where is the buffer to be written? */
574 err
= jbd2_journal_next_log_block(journal
, &blocknr
);
575 /* If the block mapping failed, just abandon the buffer
576 and repeat this loop: we'll fall into the
577 refile-on-abort condition above. */
579 jbd2_journal_abort(journal
, err
);
584 * start_this_handle() uses t_outstanding_credits to determine
585 * the free space in the log, but this counter is changed
586 * by jbd2_journal_next_log_block() also.
588 atomic_dec(&commit_transaction
->t_outstanding_credits
);
590 /* Bump b_count to prevent truncate from stumbling over
591 the shadowed buffer! @@@ This can go if we ever get
592 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
593 atomic_inc(&jh2bh(jh
)->b_count
);
595 /* Make a temporary IO buffer with which to write it out
596 (this will requeue both the metadata buffer and the
597 temporary IO buffer). new_bh goes on BJ_IO*/
599 set_bit(BH_JWrite
, &jh2bh(jh
)->b_state
);
601 * akpm: jbd2_journal_write_metadata_buffer() sets
602 * new_bh->b_transaction to commit_transaction.
603 * We need to clean this up before we release new_bh
604 * (which is of type BJ_IO)
606 JBUFFER_TRACE(jh
, "ph3: write metadata");
607 flags
= jbd2_journal_write_metadata_buffer(commit_transaction
,
608 jh
, &new_jh
, blocknr
);
610 jbd2_journal_abort(journal
, flags
);
613 set_bit(BH_JWrite
, &jh2bh(new_jh
)->b_state
);
614 wbuf
[bufs
++] = jh2bh(new_jh
);
616 /* Record the new block's tag in the current descriptor
621 tag_flag
|= JBD2_FLAG_ESCAPE
;
623 tag_flag
|= JBD2_FLAG_SAME_UUID
;
625 tag
= (journal_block_tag_t
*) tagp
;
626 write_tag_block(tag_bytes
, tag
, jh2bh(jh
)->b_blocknr
);
627 tag
->t_flags
= cpu_to_be32(tag_flag
);
629 space_left
-= tag_bytes
;
632 memcpy (tagp
, journal
->j_uuid
, 16);
638 /* If there's no more to do, or if the descriptor is full,
641 if (bufs
== journal
->j_wbufsize
||
642 commit_transaction
->t_buffers
== NULL
||
643 space_left
< tag_bytes
+ 16) {
645 jbd_debug(4, "JBD2: Submit %d IOs\n", bufs
);
647 /* Write an end-of-descriptor marker before
648 submitting the IOs. "tag" still points to
649 the last tag we set up. */
651 tag
->t_flags
|= cpu_to_be32(JBD2_FLAG_LAST_TAG
);
654 for (i
= 0; i
< bufs
; i
++) {
655 struct buffer_head
*bh
= wbuf
[i
];
659 if (JBD2_HAS_COMPAT_FEATURE(journal
,
660 JBD2_FEATURE_COMPAT_CHECKSUM
)) {
662 jbd2_checksum_data(crc32_sum
, bh
);
666 clear_buffer_dirty(bh
);
667 set_buffer_uptodate(bh
);
668 bh
->b_end_io
= journal_end_buffer_io_sync
;
669 submit_bh(WRITE_SYNC
, bh
);
672 stats
.run
.rs_blocks_logged
+= bufs
;
674 /* Force a new descriptor to be generated next
675 time round the loop. */
681 err
= journal_finish_inode_data_buffers(journal
, commit_transaction
);
684 "JBD2: Detected IO errors while flushing file data "
685 "on %s\n", journal
->j_devname
);
686 if (journal
->j_flags
& JBD2_ABORT_ON_SYNCDATA_ERR
)
687 jbd2_journal_abort(journal
, err
);
691 write_lock(&journal
->j_state_lock
);
692 J_ASSERT(commit_transaction
->t_state
== T_COMMIT
);
693 commit_transaction
->t_state
= T_COMMIT_DFLUSH
;
694 write_unlock(&journal
->j_state_lock
);
696 * If the journal is not located on the file system device,
697 * then we must flush the file system device before we issue
700 if (commit_transaction
->t_need_data_flush
&&
701 (journal
->j_fs_dev
!= journal
->j_dev
) &&
702 (journal
->j_flags
& JBD2_BARRIER
))
703 blkdev_issue_flush(journal
->j_fs_dev
, GFP_KERNEL
, NULL
);
705 /* Done it all: now write the commit record asynchronously. */
706 if (JBD2_HAS_INCOMPAT_FEATURE(journal
,
707 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
)) {
708 err
= journal_submit_commit_record(journal
, commit_transaction
,
711 __jbd2_journal_abort_hard(journal
);
714 blk_finish_plug(&plug
);
716 /* Lo and behold: we have just managed to send a transaction to
717 the log. Before we can commit it, wait for the IO so far to
718 complete. Control buffers being written are on the
719 transaction's t_log_list queue, and metadata buffers are on
720 the t_iobuf_list queue.
722 Wait for the buffers in reverse order. That way we are
723 less likely to be woken up until all IOs have completed, and
724 so we incur less scheduling load.
727 jbd_debug(3, "JBD2: commit phase 3\n");
730 * akpm: these are BJ_IO, and j_list_lock is not needed.
731 * See __journal_try_to_free_buffer.
734 while (commit_transaction
->t_iobuf_list
!= NULL
) {
735 struct buffer_head
*bh
;
737 jh
= commit_transaction
->t_iobuf_list
->b_tprev
;
739 if (buffer_locked(bh
)) {
746 if (unlikely(!buffer_uptodate(bh
)))
749 clear_buffer_jwrite(bh
);
751 JBUFFER_TRACE(jh
, "ph4: unfile after journal write");
752 jbd2_journal_unfile_buffer(journal
, jh
);
755 * ->t_iobuf_list should contain only dummy buffer_heads
756 * which were created by jbd2_journal_write_metadata_buffer().
758 BUFFER_TRACE(bh
, "dumping temporary bh");
759 jbd2_journal_put_journal_head(jh
);
761 J_ASSERT_BH(bh
, atomic_read(&bh
->b_count
) == 0);
762 free_buffer_head(bh
);
764 /* We also have to unlock and free the corresponding
766 jh
= commit_transaction
->t_shadow_list
->b_tprev
;
768 clear_bit(BH_JWrite
, &bh
->b_state
);
769 J_ASSERT_BH(bh
, buffer_jbddirty(bh
));
771 /* The metadata is now released for reuse, but we need
772 to remember it against this transaction so that when
773 we finally commit, we can do any checkpointing
775 JBUFFER_TRACE(jh
, "file as BJ_Forget");
776 jbd2_journal_file_buffer(jh
, commit_transaction
, BJ_Forget
);
778 * Wake up any transactions which were waiting for this IO to
779 * complete. The barrier must be here so that changes by
780 * jbd2_journal_file_buffer() take effect before wake_up_bit()
781 * does the waitqueue check.
784 wake_up_bit(&bh
->b_state
, BH_Unshadow
);
785 JBUFFER_TRACE(jh
, "brelse shadowed buffer");
789 J_ASSERT (commit_transaction
->t_shadow_list
== NULL
);
791 jbd_debug(3, "JBD2: commit phase 4\n");
793 /* Here we wait for the revoke record and descriptor record buffers */
795 while (commit_transaction
->t_log_list
!= NULL
) {
796 struct buffer_head
*bh
;
798 jh
= commit_transaction
->t_log_list
->b_tprev
;
800 if (buffer_locked(bh
)) {
802 goto wait_for_ctlbuf
;
805 goto wait_for_ctlbuf
;
807 if (unlikely(!buffer_uptodate(bh
)))
810 BUFFER_TRACE(bh
, "ph5: control buffer writeout done: unfile");
811 clear_buffer_jwrite(bh
);
812 jbd2_journal_unfile_buffer(journal
, jh
);
813 jbd2_journal_put_journal_head(jh
);
814 __brelse(bh
); /* One for getblk */
815 /* AKPM: bforget here */
819 jbd2_journal_abort(journal
, err
);
821 jbd_debug(3, "JBD2: commit phase 5\n");
822 write_lock(&journal
->j_state_lock
);
823 J_ASSERT(commit_transaction
->t_state
== T_COMMIT_DFLUSH
);
824 commit_transaction
->t_state
= T_COMMIT_JFLUSH
;
825 write_unlock(&journal
->j_state_lock
);
827 if (!JBD2_HAS_INCOMPAT_FEATURE(journal
,
828 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
)) {
829 err
= journal_submit_commit_record(journal
, commit_transaction
,
832 __jbd2_journal_abort_hard(journal
);
835 err
= journal_wait_on_commit_record(journal
, cbh
);
836 if (JBD2_HAS_INCOMPAT_FEATURE(journal
,
837 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
) &&
838 journal
->j_flags
& JBD2_BARRIER
) {
839 blkdev_issue_flush(journal
->j_dev
, GFP_KERNEL
, NULL
);
843 jbd2_journal_abort(journal
, err
);
845 /* End of a transaction! Finally, we can do checkpoint
846 processing: any buffers committed as a result of this
847 transaction can be removed from any checkpoint list it was on
850 jbd_debug(3, "JBD2: commit phase 6\n");
852 J_ASSERT(list_empty(&commit_transaction
->t_inode_list
));
853 J_ASSERT(commit_transaction
->t_buffers
== NULL
);
854 J_ASSERT(commit_transaction
->t_checkpoint_list
== NULL
);
855 J_ASSERT(commit_transaction
->t_iobuf_list
== NULL
);
856 J_ASSERT(commit_transaction
->t_shadow_list
== NULL
);
857 J_ASSERT(commit_transaction
->t_log_list
== NULL
);
861 * As there are other places (journal_unmap_buffer()) adding buffers
862 * to this list we have to be careful and hold the j_list_lock.
864 spin_lock(&journal
->j_list_lock
);
865 while (commit_transaction
->t_forget
) {
866 transaction_t
*cp_transaction
;
867 struct buffer_head
*bh
;
870 jh
= commit_transaction
->t_forget
;
871 spin_unlock(&journal
->j_list_lock
);
874 * Get a reference so that bh cannot be freed before we are
878 jbd_lock_bh_state(bh
);
879 J_ASSERT_JH(jh
, jh
->b_transaction
== commit_transaction
);
882 * If there is undo-protected committed data against
883 * this buffer, then we can remove it now. If it is a
884 * buffer needing such protection, the old frozen_data
885 * field now points to a committed version of the
886 * buffer, so rotate that field to the new committed
889 * Otherwise, we can just throw away the frozen data now.
891 * We also know that the frozen data has already fired
892 * its triggers if they exist, so we can clear that too.
894 if (jh
->b_committed_data
) {
895 jbd2_free(jh
->b_committed_data
, bh
->b_size
);
896 jh
->b_committed_data
= NULL
;
897 if (jh
->b_frozen_data
) {
898 jh
->b_committed_data
= jh
->b_frozen_data
;
899 jh
->b_frozen_data
= NULL
;
900 jh
->b_frozen_triggers
= NULL
;
902 } else if (jh
->b_frozen_data
) {
903 jbd2_free(jh
->b_frozen_data
, bh
->b_size
);
904 jh
->b_frozen_data
= NULL
;
905 jh
->b_frozen_triggers
= NULL
;
908 spin_lock(&journal
->j_list_lock
);
909 cp_transaction
= jh
->b_cp_transaction
;
910 if (cp_transaction
) {
911 JBUFFER_TRACE(jh
, "remove from old cp transaction");
912 cp_transaction
->t_chp_stats
.cs_dropped
++;
913 __jbd2_journal_remove_checkpoint(jh
);
916 /* Only re-checkpoint the buffer_head if it is marked
917 * dirty. If the buffer was added to the BJ_Forget list
918 * by jbd2_journal_forget, it may no longer be dirty and
919 * there's no point in keeping a checkpoint record for
922 /* A buffer which has been freed while still being
923 * journaled by a previous transaction may end up still
924 * being dirty here, but we want to avoid writing back
925 * that buffer in the future after the "add to orphan"
926 * operation been committed, That's not only a performance
927 * gain, it also stops aliasing problems if the buffer is
928 * left behind for writeback and gets reallocated for another
929 * use in a different page. */
930 if (buffer_freed(bh
) && !jh
->b_next_transaction
) {
931 clear_buffer_freed(bh
);
932 clear_buffer_jbddirty(bh
);
935 if (buffer_jbddirty(bh
)) {
936 JBUFFER_TRACE(jh
, "add to new checkpointing trans");
937 __jbd2_journal_insert_checkpoint(jh
, commit_transaction
);
938 if (is_journal_aborted(journal
))
939 clear_buffer_jbddirty(bh
);
941 J_ASSERT_BH(bh
, !buffer_dirty(bh
));
943 * The buffer on BJ_Forget list and not jbddirty means
944 * it has been freed by this transaction and hence it
945 * could not have been reallocated until this
946 * transaction has committed. *BUT* it could be
947 * reallocated once we have written all the data to
948 * disk and before we process the buffer on BJ_Forget
951 if (!jh
->b_next_transaction
)
954 JBUFFER_TRACE(jh
, "refile or unfile buffer");
955 __jbd2_journal_refile_buffer(jh
);
956 jbd_unlock_bh_state(bh
);
958 release_buffer_page(bh
); /* Drops bh reference */
961 cond_resched_lock(&journal
->j_list_lock
);
963 spin_unlock(&journal
->j_list_lock
);
965 * This is a bit sleazy. We use j_list_lock to protect transition
966 * of a transaction into T_FINISHED state and calling
967 * __jbd2_journal_drop_transaction(). Otherwise we could race with
968 * other checkpointing code processing the transaction...
970 write_lock(&journal
->j_state_lock
);
971 spin_lock(&journal
->j_list_lock
);
973 * Now recheck if some buffers did not get attached to the transaction
974 * while the lock was dropped...
976 if (commit_transaction
->t_forget
) {
977 spin_unlock(&journal
->j_list_lock
);
978 write_unlock(&journal
->j_state_lock
);
982 /* Done with this transaction! */
984 jbd_debug(3, "JBD2: commit phase 7\n");
986 J_ASSERT(commit_transaction
->t_state
== T_COMMIT_JFLUSH
);
988 commit_transaction
->t_start
= jiffies
;
989 stats
.run
.rs_logging
= jbd2_time_diff(stats
.run
.rs_logging
,
990 commit_transaction
->t_start
);
993 * File the transaction statistics
995 stats
.ts_tid
= commit_transaction
->t_tid
;
996 stats
.run
.rs_handle_count
=
997 atomic_read(&commit_transaction
->t_handle_count
);
998 trace_jbd2_run_stats(journal
->j_fs_dev
->bd_dev
,
999 commit_transaction
->t_tid
, &stats
.run
);
1002 * Calculate overall stats
1004 spin_lock(&journal
->j_history_lock
);
1005 journal
->j_stats
.ts_tid
++;
1006 journal
->j_stats
.run
.rs_wait
+= stats
.run
.rs_wait
;
1007 journal
->j_stats
.run
.rs_running
+= stats
.run
.rs_running
;
1008 journal
->j_stats
.run
.rs_locked
+= stats
.run
.rs_locked
;
1009 journal
->j_stats
.run
.rs_flushing
+= stats
.run
.rs_flushing
;
1010 journal
->j_stats
.run
.rs_logging
+= stats
.run
.rs_logging
;
1011 journal
->j_stats
.run
.rs_handle_count
+= stats
.run
.rs_handle_count
;
1012 journal
->j_stats
.run
.rs_blocks
+= stats
.run
.rs_blocks
;
1013 journal
->j_stats
.run
.rs_blocks_logged
+= stats
.run
.rs_blocks_logged
;
1014 spin_unlock(&journal
->j_history_lock
);
1016 commit_transaction
->t_state
= T_FINISHED
;
1017 J_ASSERT(commit_transaction
== journal
->j_committing_transaction
);
1018 journal
->j_commit_sequence
= commit_transaction
->t_tid
;
1019 journal
->j_committing_transaction
= NULL
;
1020 commit_time
= ktime_to_ns(ktime_sub(ktime_get(), start_time
));
1023 * weight the commit time higher than the average time so we don't
1024 * react too strongly to vast changes in the commit time
1026 if (likely(journal
->j_average_commit_time
))
1027 journal
->j_average_commit_time
= (commit_time
+
1028 journal
->j_average_commit_time
*3) / 4;
1030 journal
->j_average_commit_time
= commit_time
;
1031 write_unlock(&journal
->j_state_lock
);
1033 if (commit_transaction
->t_checkpoint_list
== NULL
&&
1034 commit_transaction
->t_checkpoint_io_list
== NULL
) {
1035 __jbd2_journal_drop_transaction(journal
, commit_transaction
);
1038 if (journal
->j_checkpoint_transactions
== NULL
) {
1039 journal
->j_checkpoint_transactions
= commit_transaction
;
1040 commit_transaction
->t_cpnext
= commit_transaction
;
1041 commit_transaction
->t_cpprev
= commit_transaction
;
1043 commit_transaction
->t_cpnext
=
1044 journal
->j_checkpoint_transactions
;
1045 commit_transaction
->t_cpprev
=
1046 commit_transaction
->t_cpnext
->t_cpprev
;
1047 commit_transaction
->t_cpnext
->t_cpprev
=
1049 commit_transaction
->t_cpprev
->t_cpnext
=
1053 spin_unlock(&journal
->j_list_lock
);
1055 if (journal
->j_commit_callback
)
1056 journal
->j_commit_callback(journal
, commit_transaction
);
1058 trace_jbd2_end_commit(journal
, commit_transaction
);
1059 jbd_debug(1, "JBD2: commit %d complete, head %d\n",
1060 journal
->j_commit_sequence
, journal
->j_tail_sequence
);
1062 jbd2_journal_free_transaction(commit_transaction
);
1064 wake_up(&journal
->j_wait_done_commit
);