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 commit_transaction
->t_flushed_data_blocks
= 1;
223 clear_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
224 smp_mb__after_clear_bit();
225 wake_up_bit(&jinode
->i_flags
, __JI_COMMIT_RUNNING
);
227 spin_unlock(&journal
->j_list_lock
);
232 * Wait for data submitted for writeout, refile inodes to proper
233 * transaction if needed.
236 static int journal_finish_inode_data_buffers(journal_t
*journal
,
237 transaction_t
*commit_transaction
)
239 struct jbd2_inode
*jinode
, *next_i
;
242 /* For locking, see the comment in journal_submit_data_buffers() */
243 spin_lock(&journal
->j_list_lock
);
244 list_for_each_entry(jinode
, &commit_transaction
->t_inode_list
, i_list
) {
245 set_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
246 spin_unlock(&journal
->j_list_lock
);
247 err
= filemap_fdatawait(jinode
->i_vfs_inode
->i_mapping
);
250 * Because AS_EIO is cleared by
251 * filemap_fdatawait_range(), set it again so
252 * that user process can get -EIO from fsync().
255 &jinode
->i_vfs_inode
->i_mapping
->flags
);
260 spin_lock(&journal
->j_list_lock
);
261 clear_bit(__JI_COMMIT_RUNNING
, &jinode
->i_flags
);
262 smp_mb__after_clear_bit();
263 wake_up_bit(&jinode
->i_flags
, __JI_COMMIT_RUNNING
);
266 /* Now refile inode to proper lists */
267 list_for_each_entry_safe(jinode
, next_i
,
268 &commit_transaction
->t_inode_list
, i_list
) {
269 list_del(&jinode
->i_list
);
270 if (jinode
->i_next_transaction
) {
271 jinode
->i_transaction
= jinode
->i_next_transaction
;
272 jinode
->i_next_transaction
= NULL
;
273 list_add(&jinode
->i_list
,
274 &jinode
->i_transaction
->t_inode_list
);
276 jinode
->i_transaction
= NULL
;
279 spin_unlock(&journal
->j_list_lock
);
284 static __u32
jbd2_checksum_data(__u32 crc32_sum
, struct buffer_head
*bh
)
286 struct page
*page
= bh
->b_page
;
290 addr
= kmap_atomic(page
, KM_USER0
);
291 checksum
= crc32_be(crc32_sum
,
292 (void *)(addr
+ offset_in_page(bh
->b_data
)), bh
->b_size
);
293 kunmap_atomic(addr
, KM_USER0
);
298 static void write_tag_block(int tag_bytes
, journal_block_tag_t
*tag
,
299 unsigned long long block
)
301 tag
->t_blocknr
= cpu_to_be32(block
& (u32
)~0);
302 if (tag_bytes
> JBD2_TAG_SIZE32
)
303 tag
->t_blocknr_high
= cpu_to_be32((block
>> 31) >> 1);
307 * jbd2_journal_commit_transaction
309 * The primary function for committing a transaction to the log. This
310 * function is called by the journal thread to begin a complete commit.
312 void jbd2_journal_commit_transaction(journal_t
*journal
)
314 struct transaction_stats_s stats
;
315 transaction_t
*commit_transaction
;
316 struct journal_head
*jh
, *new_jh
, *descriptor
;
317 struct buffer_head
**wbuf
= journal
->j_wbuf
;
321 unsigned long long blocknr
;
325 journal_header_t
*header
;
326 journal_block_tag_t
*tag
= NULL
;
331 int tag_bytes
= journal_tag_bytes(journal
);
332 struct buffer_head
*cbh
= NULL
; /* For transactional checksums */
333 __u32 crc32_sum
= ~0;
334 struct blk_plug plug
;
337 * First job: lock down the current transaction and wait for
338 * all outstanding updates to complete.
342 spin_lock(&journal
->j_list_lock
);
343 summarise_journal_usage(journal
);
344 spin_unlock(&journal
->j_list_lock
);
347 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
348 if (journal
->j_flags
& JBD2_FLUSHED
) {
349 jbd_debug(3, "super block updated\n");
350 jbd2_journal_update_superblock(journal
, 1);
352 jbd_debug(3, "superblock not updated\n");
355 J_ASSERT(journal
->j_running_transaction
!= NULL
);
356 J_ASSERT(journal
->j_committing_transaction
== NULL
);
358 commit_transaction
= journal
->j_running_transaction
;
359 J_ASSERT(commit_transaction
->t_state
== T_RUNNING
);
361 trace_jbd2_start_commit(journal
, commit_transaction
);
362 jbd_debug(1, "JBD: starting commit of transaction %d\n",
363 commit_transaction
->t_tid
);
365 write_lock(&journal
->j_state_lock
);
366 commit_transaction
->t_state
= T_LOCKED
;
368 trace_jbd2_commit_locking(journal
, commit_transaction
);
369 stats
.run
.rs_wait
= commit_transaction
->t_max_wait
;
370 stats
.run
.rs_locked
= jiffies
;
371 stats
.run
.rs_running
= jbd2_time_diff(commit_transaction
->t_start
,
372 stats
.run
.rs_locked
);
374 spin_lock(&commit_transaction
->t_handle_lock
);
375 while (atomic_read(&commit_transaction
->t_updates
)) {
378 prepare_to_wait(&journal
->j_wait_updates
, &wait
,
379 TASK_UNINTERRUPTIBLE
);
380 if (atomic_read(&commit_transaction
->t_updates
)) {
381 spin_unlock(&commit_transaction
->t_handle_lock
);
382 write_unlock(&journal
->j_state_lock
);
384 write_lock(&journal
->j_state_lock
);
385 spin_lock(&commit_transaction
->t_handle_lock
);
387 finish_wait(&journal
->j_wait_updates
, &wait
);
389 spin_unlock(&commit_transaction
->t_handle_lock
);
391 J_ASSERT (atomic_read(&commit_transaction
->t_outstanding_credits
) <=
392 journal
->j_max_transaction_buffers
);
395 * First thing we are allowed to do is to discard any remaining
396 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
397 * that there are no such buffers: if a large filesystem
398 * operation like a truncate needs to split itself over multiple
399 * transactions, then it may try to do a jbd2_journal_restart() while
400 * there are still BJ_Reserved buffers outstanding. These must
401 * be released cleanly from the current transaction.
403 * In this case, the filesystem must still reserve write access
404 * again before modifying the buffer in the new transaction, but
405 * we do not require it to remember exactly which old buffers it
406 * has reserved. This is consistent with the existing behaviour
407 * that multiple jbd2_journal_get_write_access() calls to the same
408 * buffer are perfectly permissible.
410 while (commit_transaction
->t_reserved_list
) {
411 jh
= commit_transaction
->t_reserved_list
;
412 JBUFFER_TRACE(jh
, "reserved, unused: refile");
414 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
415 * leave undo-committed data.
417 if (jh
->b_committed_data
) {
418 struct buffer_head
*bh
= jh2bh(jh
);
420 jbd_lock_bh_state(bh
);
421 jbd2_free(jh
->b_committed_data
, bh
->b_size
);
422 jh
->b_committed_data
= NULL
;
423 jbd_unlock_bh_state(bh
);
425 jbd2_journal_refile_buffer(journal
, jh
);
429 * Now try to drop any written-back buffers from the journal's
430 * checkpoint lists. We do this *before* commit because it potentially
433 spin_lock(&journal
->j_list_lock
);
434 __jbd2_journal_clean_checkpoint_list(journal
);
435 spin_unlock(&journal
->j_list_lock
);
437 jbd_debug (3, "JBD: commit phase 1\n");
440 * Switch to a new revoke table.
442 jbd2_journal_switch_revoke_table(journal
);
444 trace_jbd2_commit_flushing(journal
, commit_transaction
);
445 stats
.run
.rs_flushing
= jiffies
;
446 stats
.run
.rs_locked
= jbd2_time_diff(stats
.run
.rs_locked
,
447 stats
.run
.rs_flushing
);
449 commit_transaction
->t_state
= T_FLUSH
;
450 journal
->j_committing_transaction
= commit_transaction
;
451 journal
->j_running_transaction
= NULL
;
452 start_time
= ktime_get();
453 commit_transaction
->t_log_start
= journal
->j_head
;
454 wake_up(&journal
->j_wait_transaction_locked
);
455 write_unlock(&journal
->j_state_lock
);
457 jbd_debug (3, "JBD: commit phase 2\n");
460 * Now start flushing things to disk, in the order they appear
461 * on the transaction lists. Data blocks go first.
463 err
= journal_submit_data_buffers(journal
, commit_transaction
);
465 jbd2_journal_abort(journal
, err
);
467 blk_start_plug(&plug
);
468 jbd2_journal_write_revoke_records(journal
, commit_transaction
,
470 blk_finish_plug(&plug
);
472 jbd_debug(3, "JBD: commit phase 2\n");
475 * Way to go: we have now written out all of the data for a
476 * transaction! Now comes the tricky part: we need to write out
477 * metadata. Loop over the transaction's entire buffer list:
479 write_lock(&journal
->j_state_lock
);
480 commit_transaction
->t_state
= T_COMMIT
;
481 write_unlock(&journal
->j_state_lock
);
483 trace_jbd2_commit_logging(journal
, commit_transaction
);
484 stats
.run
.rs_logging
= jiffies
;
485 stats
.run
.rs_flushing
= jbd2_time_diff(stats
.run
.rs_flushing
,
486 stats
.run
.rs_logging
);
487 stats
.run
.rs_blocks
=
488 atomic_read(&commit_transaction
->t_outstanding_credits
);
489 stats
.run
.rs_blocks_logged
= 0;
491 J_ASSERT(commit_transaction
->t_nr_buffers
<=
492 atomic_read(&commit_transaction
->t_outstanding_credits
));
497 blk_start_plug(&plug
);
498 while (commit_transaction
->t_buffers
) {
500 /* Find the next buffer to be journaled... */
502 jh
= commit_transaction
->t_buffers
;
504 /* If we're in abort mode, we just un-journal the buffer and
507 if (is_journal_aborted(journal
)) {
508 clear_buffer_jbddirty(jh2bh(jh
));
509 JBUFFER_TRACE(jh
, "journal is aborting: refile");
510 jbd2_buffer_abort_trigger(jh
,
512 jh
->b_frozen_triggers
:
514 jbd2_journal_refile_buffer(journal
, jh
);
515 /* If that was the last one, we need to clean up
516 * any descriptor buffers which may have been
517 * already allocated, even if we are now
519 if (!commit_transaction
->t_buffers
)
520 goto start_journal_io
;
524 /* Make sure we have a descriptor block in which to
525 record the metadata buffer. */
528 struct buffer_head
*bh
;
530 J_ASSERT (bufs
== 0);
532 jbd_debug(4, "JBD: get descriptor\n");
534 descriptor
= jbd2_journal_get_descriptor_buffer(journal
);
536 jbd2_journal_abort(journal
, -EIO
);
540 bh
= jh2bh(descriptor
);
541 jbd_debug(4, "JBD: got buffer %llu (%p)\n",
542 (unsigned long long)bh
->b_blocknr
, bh
->b_data
);
543 header
= (journal_header_t
*)&bh
->b_data
[0];
544 header
->h_magic
= cpu_to_be32(JBD2_MAGIC_NUMBER
);
545 header
->h_blocktype
= cpu_to_be32(JBD2_DESCRIPTOR_BLOCK
);
546 header
->h_sequence
= cpu_to_be32(commit_transaction
->t_tid
);
548 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
549 space_left
= bh
->b_size
- sizeof(journal_header_t
);
551 set_buffer_jwrite(bh
);
552 set_buffer_dirty(bh
);
555 /* Record it so that we can wait for IO
557 BUFFER_TRACE(bh
, "ph3: file as descriptor");
558 jbd2_journal_file_buffer(descriptor
, commit_transaction
,
562 /* Where is the buffer to be written? */
564 err
= jbd2_journal_next_log_block(journal
, &blocknr
);
565 /* If the block mapping failed, just abandon the buffer
566 and repeat this loop: we'll fall into the
567 refile-on-abort condition above. */
569 jbd2_journal_abort(journal
, err
);
574 * start_this_handle() uses t_outstanding_credits to determine
575 * the free space in the log, but this counter is changed
576 * by jbd2_journal_next_log_block() also.
578 atomic_dec(&commit_transaction
->t_outstanding_credits
);
580 /* Bump b_count to prevent truncate from stumbling over
581 the shadowed buffer! @@@ This can go if we ever get
582 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
583 atomic_inc(&jh2bh(jh
)->b_count
);
585 /* Make a temporary IO buffer with which to write it out
586 (this will requeue both the metadata buffer and the
587 temporary IO buffer). new_bh goes on BJ_IO*/
589 set_bit(BH_JWrite
, &jh2bh(jh
)->b_state
);
591 * akpm: jbd2_journal_write_metadata_buffer() sets
592 * new_bh->b_transaction to commit_transaction.
593 * We need to clean this up before we release new_bh
594 * (which is of type BJ_IO)
596 JBUFFER_TRACE(jh
, "ph3: write metadata");
597 flags
= jbd2_journal_write_metadata_buffer(commit_transaction
,
598 jh
, &new_jh
, blocknr
);
600 jbd2_journal_abort(journal
, flags
);
603 set_bit(BH_JWrite
, &jh2bh(new_jh
)->b_state
);
604 wbuf
[bufs
++] = jh2bh(new_jh
);
606 /* Record the new block's tag in the current descriptor
611 tag_flag
|= JBD2_FLAG_ESCAPE
;
613 tag_flag
|= JBD2_FLAG_SAME_UUID
;
615 tag
= (journal_block_tag_t
*) tagp
;
616 write_tag_block(tag_bytes
, tag
, jh2bh(jh
)->b_blocknr
);
617 tag
->t_flags
= cpu_to_be32(tag_flag
);
619 space_left
-= tag_bytes
;
622 memcpy (tagp
, journal
->j_uuid
, 16);
628 /* If there's no more to do, or if the descriptor is full,
631 if (bufs
== journal
->j_wbufsize
||
632 commit_transaction
->t_buffers
== NULL
||
633 space_left
< tag_bytes
+ 16) {
635 jbd_debug(4, "JBD: Submit %d IOs\n", bufs
);
637 /* Write an end-of-descriptor marker before
638 submitting the IOs. "tag" still points to
639 the last tag we set up. */
641 tag
->t_flags
|= cpu_to_be32(JBD2_FLAG_LAST_TAG
);
644 for (i
= 0; i
< bufs
; i
++) {
645 struct buffer_head
*bh
= wbuf
[i
];
649 if (JBD2_HAS_COMPAT_FEATURE(journal
,
650 JBD2_FEATURE_COMPAT_CHECKSUM
)) {
652 jbd2_checksum_data(crc32_sum
, bh
);
656 clear_buffer_dirty(bh
);
657 set_buffer_uptodate(bh
);
658 bh
->b_end_io
= journal_end_buffer_io_sync
;
659 submit_bh(WRITE_SYNC
, bh
);
662 stats
.run
.rs_blocks_logged
+= bufs
;
664 /* Force a new descriptor to be generated next
665 time round the loop. */
671 err
= journal_finish_inode_data_buffers(journal
, commit_transaction
);
674 "JBD2: Detected IO errors while flushing file data "
675 "on %s\n", journal
->j_devname
);
676 if (journal
->j_flags
& JBD2_ABORT_ON_SYNCDATA_ERR
)
677 jbd2_journal_abort(journal
, err
);
682 * If the journal is not located on the file system device,
683 * then we must flush the file system device before we issue
686 if (commit_transaction
->t_flushed_data_blocks
&&
687 (journal
->j_fs_dev
!= journal
->j_dev
) &&
688 (journal
->j_flags
& JBD2_BARRIER
))
689 blkdev_issue_flush(journal
->j_fs_dev
, GFP_KERNEL
, NULL
);
691 /* Done it all: now write the commit record asynchronously. */
692 if (JBD2_HAS_INCOMPAT_FEATURE(journal
,
693 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
)) {
694 err
= journal_submit_commit_record(journal
, commit_transaction
,
697 __jbd2_journal_abort_hard(journal
);
700 blk_finish_plug(&plug
);
702 /* Lo and behold: we have just managed to send a transaction to
703 the log. Before we can commit it, wait for the IO so far to
704 complete. Control buffers being written are on the
705 transaction's t_log_list queue, and metadata buffers are on
706 the t_iobuf_list queue.
708 Wait for the buffers in reverse order. That way we are
709 less likely to be woken up until all IOs have completed, and
710 so we incur less scheduling load.
713 jbd_debug(3, "JBD: commit phase 3\n");
716 * akpm: these are BJ_IO, and j_list_lock is not needed.
717 * See __journal_try_to_free_buffer.
720 while (commit_transaction
->t_iobuf_list
!= NULL
) {
721 struct buffer_head
*bh
;
723 jh
= commit_transaction
->t_iobuf_list
->b_tprev
;
725 if (buffer_locked(bh
)) {
732 if (unlikely(!buffer_uptodate(bh
)))
735 clear_buffer_jwrite(bh
);
737 JBUFFER_TRACE(jh
, "ph4: unfile after journal write");
738 jbd2_journal_unfile_buffer(journal
, jh
);
741 * ->t_iobuf_list should contain only dummy buffer_heads
742 * which were created by jbd2_journal_write_metadata_buffer().
744 BUFFER_TRACE(bh
, "dumping temporary bh");
745 jbd2_journal_put_journal_head(jh
);
747 J_ASSERT_BH(bh
, atomic_read(&bh
->b_count
) == 0);
748 free_buffer_head(bh
);
750 /* We also have to unlock and free the corresponding
752 jh
= commit_transaction
->t_shadow_list
->b_tprev
;
754 clear_bit(BH_JWrite
, &bh
->b_state
);
755 J_ASSERT_BH(bh
, buffer_jbddirty(bh
));
757 /* The metadata is now released for reuse, but we need
758 to remember it against this transaction so that when
759 we finally commit, we can do any checkpointing
761 JBUFFER_TRACE(jh
, "file as BJ_Forget");
762 jbd2_journal_file_buffer(jh
, commit_transaction
, BJ_Forget
);
764 * Wake up any transactions which were waiting for this IO to
765 * complete. The barrier must be here so that changes by
766 * jbd2_journal_file_buffer() take effect before wake_up_bit()
767 * does the waitqueue check.
770 wake_up_bit(&bh
->b_state
, BH_Unshadow
);
771 JBUFFER_TRACE(jh
, "brelse shadowed buffer");
775 J_ASSERT (commit_transaction
->t_shadow_list
== NULL
);
777 jbd_debug(3, "JBD: commit phase 4\n");
779 /* Here we wait for the revoke record and descriptor record buffers */
781 while (commit_transaction
->t_log_list
!= NULL
) {
782 struct buffer_head
*bh
;
784 jh
= commit_transaction
->t_log_list
->b_tprev
;
786 if (buffer_locked(bh
)) {
788 goto wait_for_ctlbuf
;
791 goto wait_for_ctlbuf
;
793 if (unlikely(!buffer_uptodate(bh
)))
796 BUFFER_TRACE(bh
, "ph5: control buffer writeout done: unfile");
797 clear_buffer_jwrite(bh
);
798 jbd2_journal_unfile_buffer(journal
, jh
);
799 jbd2_journal_put_journal_head(jh
);
800 __brelse(bh
); /* One for getblk */
801 /* AKPM: bforget here */
805 jbd2_journal_abort(journal
, err
);
807 jbd_debug(3, "JBD: commit phase 5\n");
809 if (!JBD2_HAS_INCOMPAT_FEATURE(journal
,
810 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
)) {
811 err
= journal_submit_commit_record(journal
, commit_transaction
,
814 __jbd2_journal_abort_hard(journal
);
817 err
= journal_wait_on_commit_record(journal
, cbh
);
818 if (JBD2_HAS_INCOMPAT_FEATURE(journal
,
819 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT
) &&
820 journal
->j_flags
& JBD2_BARRIER
) {
821 blkdev_issue_flush(journal
->j_dev
, GFP_KERNEL
, NULL
);
825 jbd2_journal_abort(journal
, err
);
827 /* End of a transaction! Finally, we can do checkpoint
828 processing: any buffers committed as a result of this
829 transaction can be removed from any checkpoint list it was on
832 jbd_debug(3, "JBD: commit phase 6\n");
834 J_ASSERT(list_empty(&commit_transaction
->t_inode_list
));
835 J_ASSERT(commit_transaction
->t_buffers
== NULL
);
836 J_ASSERT(commit_transaction
->t_checkpoint_list
== NULL
);
837 J_ASSERT(commit_transaction
->t_iobuf_list
== NULL
);
838 J_ASSERT(commit_transaction
->t_shadow_list
== NULL
);
839 J_ASSERT(commit_transaction
->t_log_list
== NULL
);
843 * As there are other places (journal_unmap_buffer()) adding buffers
844 * to this list we have to be careful and hold the j_list_lock.
846 spin_lock(&journal
->j_list_lock
);
847 while (commit_transaction
->t_forget
) {
848 transaction_t
*cp_transaction
;
849 struct buffer_head
*bh
;
851 jh
= commit_transaction
->t_forget
;
852 spin_unlock(&journal
->j_list_lock
);
854 jbd_lock_bh_state(bh
);
855 J_ASSERT_JH(jh
, jh
->b_transaction
== commit_transaction
);
858 * If there is undo-protected committed data against
859 * this buffer, then we can remove it now. If it is a
860 * buffer needing such protection, the old frozen_data
861 * field now points to a committed version of the
862 * buffer, so rotate that field to the new committed
865 * Otherwise, we can just throw away the frozen data now.
867 * We also know that the frozen data has already fired
868 * its triggers if they exist, so we can clear that too.
870 if (jh
->b_committed_data
) {
871 jbd2_free(jh
->b_committed_data
, bh
->b_size
);
872 jh
->b_committed_data
= NULL
;
873 if (jh
->b_frozen_data
) {
874 jh
->b_committed_data
= jh
->b_frozen_data
;
875 jh
->b_frozen_data
= NULL
;
876 jh
->b_frozen_triggers
= NULL
;
878 } else if (jh
->b_frozen_data
) {
879 jbd2_free(jh
->b_frozen_data
, bh
->b_size
);
880 jh
->b_frozen_data
= NULL
;
881 jh
->b_frozen_triggers
= NULL
;
884 spin_lock(&journal
->j_list_lock
);
885 cp_transaction
= jh
->b_cp_transaction
;
886 if (cp_transaction
) {
887 JBUFFER_TRACE(jh
, "remove from old cp transaction");
888 cp_transaction
->t_chp_stats
.cs_dropped
++;
889 __jbd2_journal_remove_checkpoint(jh
);
892 /* Only re-checkpoint the buffer_head if it is marked
893 * dirty. If the buffer was added to the BJ_Forget list
894 * by jbd2_journal_forget, it may no longer be dirty and
895 * there's no point in keeping a checkpoint record for
898 /* A buffer which has been freed while still being
899 * journaled by a previous transaction may end up still
900 * being dirty here, but we want to avoid writing back
901 * that buffer in the future after the "add to orphan"
902 * operation been committed, That's not only a performance
903 * gain, it also stops aliasing problems if the buffer is
904 * left behind for writeback and gets reallocated for another
905 * use in a different page. */
906 if (buffer_freed(bh
) && !jh
->b_next_transaction
) {
907 clear_buffer_freed(bh
);
908 clear_buffer_jbddirty(bh
);
911 if (buffer_jbddirty(bh
)) {
912 JBUFFER_TRACE(jh
, "add to new checkpointing trans");
913 __jbd2_journal_insert_checkpoint(jh
, commit_transaction
);
914 if (is_journal_aborted(journal
))
915 clear_buffer_jbddirty(bh
);
916 JBUFFER_TRACE(jh
, "refile for checkpoint writeback");
917 __jbd2_journal_refile_buffer(jh
);
918 jbd_unlock_bh_state(bh
);
920 J_ASSERT_BH(bh
, !buffer_dirty(bh
));
921 /* The buffer on BJ_Forget list and not jbddirty means
922 * it has been freed by this transaction and hence it
923 * could not have been reallocated until this
924 * transaction has committed. *BUT* it could be
925 * reallocated once we have written all the data to
926 * disk and before we process the buffer on BJ_Forget
928 JBUFFER_TRACE(jh
, "refile or unfile freed buffer");
929 __jbd2_journal_refile_buffer(jh
);
930 if (!jh
->b_transaction
) {
931 jbd_unlock_bh_state(bh
);
933 jbd2_journal_remove_journal_head(bh
);
934 release_buffer_page(bh
);
936 jbd_unlock_bh_state(bh
);
938 cond_resched_lock(&journal
->j_list_lock
);
940 spin_unlock(&journal
->j_list_lock
);
942 * This is a bit sleazy. We use j_list_lock to protect transition
943 * of a transaction into T_FINISHED state and calling
944 * __jbd2_journal_drop_transaction(). Otherwise we could race with
945 * other checkpointing code processing the transaction...
947 write_lock(&journal
->j_state_lock
);
948 spin_lock(&journal
->j_list_lock
);
950 * Now recheck if some buffers did not get attached to the transaction
951 * while the lock was dropped...
953 if (commit_transaction
->t_forget
) {
954 spin_unlock(&journal
->j_list_lock
);
955 write_unlock(&journal
->j_state_lock
);
959 /* Done with this transaction! */
961 jbd_debug(3, "JBD: commit phase 7\n");
963 J_ASSERT(commit_transaction
->t_state
== T_COMMIT
);
965 commit_transaction
->t_start
= jiffies
;
966 stats
.run
.rs_logging
= jbd2_time_diff(stats
.run
.rs_logging
,
967 commit_transaction
->t_start
);
970 * File the transaction statistics
972 stats
.ts_tid
= commit_transaction
->t_tid
;
973 stats
.run
.rs_handle_count
=
974 atomic_read(&commit_transaction
->t_handle_count
);
975 trace_jbd2_run_stats(journal
->j_fs_dev
->bd_dev
,
976 commit_transaction
->t_tid
, &stats
.run
);
979 * Calculate overall stats
981 spin_lock(&journal
->j_history_lock
);
982 journal
->j_stats
.ts_tid
++;
983 journal
->j_stats
.run
.rs_wait
+= stats
.run
.rs_wait
;
984 journal
->j_stats
.run
.rs_running
+= stats
.run
.rs_running
;
985 journal
->j_stats
.run
.rs_locked
+= stats
.run
.rs_locked
;
986 journal
->j_stats
.run
.rs_flushing
+= stats
.run
.rs_flushing
;
987 journal
->j_stats
.run
.rs_logging
+= stats
.run
.rs_logging
;
988 journal
->j_stats
.run
.rs_handle_count
+= stats
.run
.rs_handle_count
;
989 journal
->j_stats
.run
.rs_blocks
+= stats
.run
.rs_blocks
;
990 journal
->j_stats
.run
.rs_blocks_logged
+= stats
.run
.rs_blocks_logged
;
991 spin_unlock(&journal
->j_history_lock
);
993 commit_transaction
->t_state
= T_FINISHED
;
994 J_ASSERT(commit_transaction
== journal
->j_committing_transaction
);
995 journal
->j_commit_sequence
= commit_transaction
->t_tid
;
996 journal
->j_committing_transaction
= NULL
;
997 commit_time
= ktime_to_ns(ktime_sub(ktime_get(), start_time
));
1000 * weight the commit time higher than the average time so we don't
1001 * react too strongly to vast changes in the commit time
1003 if (likely(journal
->j_average_commit_time
))
1004 journal
->j_average_commit_time
= (commit_time
+
1005 journal
->j_average_commit_time
*3) / 4;
1007 journal
->j_average_commit_time
= commit_time
;
1008 write_unlock(&journal
->j_state_lock
);
1010 if (commit_transaction
->t_checkpoint_list
== NULL
&&
1011 commit_transaction
->t_checkpoint_io_list
== NULL
) {
1012 __jbd2_journal_drop_transaction(journal
, commit_transaction
);
1015 if (journal
->j_checkpoint_transactions
== NULL
) {
1016 journal
->j_checkpoint_transactions
= commit_transaction
;
1017 commit_transaction
->t_cpnext
= commit_transaction
;
1018 commit_transaction
->t_cpprev
= commit_transaction
;
1020 commit_transaction
->t_cpnext
=
1021 journal
->j_checkpoint_transactions
;
1022 commit_transaction
->t_cpprev
=
1023 commit_transaction
->t_cpnext
->t_cpprev
;
1024 commit_transaction
->t_cpnext
->t_cpprev
=
1026 commit_transaction
->t_cpprev
->t_cpnext
=
1030 spin_unlock(&journal
->j_list_lock
);
1032 if (journal
->j_commit_callback
)
1033 journal
->j_commit_callback(journal
, commit_transaction
);
1035 trace_jbd2_end_commit(journal
, commit_transaction
);
1036 jbd_debug(1, "JBD: commit %d complete, head %d\n",
1037 journal
->j_commit_sequence
, journal
->j_tail_sequence
);
1039 kfree(commit_transaction
);
1041 wake_up(&journal
->j_wait_done_commit
);