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();
108 if (is_journal_aborted(journal))
111 descriptor = jbd2_journal_get_descriptor_buffer(journal);
115 bh = jh2bh(descriptor);
117 tmp = (struct commit_header *)bh->b_data;
118 tmp->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
119 tmp->h_blocktype = cpu_to_be32(JBD2_COMMIT_BLOCK);
120 tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
121 tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
122 tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
124 if (JBD2_HAS_COMPAT_FEATURE(journal,
125 JBD2_FEATURE_COMPAT_CHECKSUM)) {
126 tmp->h_chksum_type = JBD2_CRC32_CHKSUM;
127 tmp->h_chksum_size = JBD2_CRC32_CHKSUM_SIZE;
128 tmp->h_chksum[0] = cpu_to_be32(crc32_sum);
131 JBUFFER_TRACE(descriptor, "submit commit block");
133 clear_buffer_dirty(bh);
134 set_buffer_uptodate(bh);
135 bh->b_end_io = journal_end_buffer_io_sync;
137 if (journal->j_flags & JBD2_BARRIER &&
138 !JBD2_HAS_INCOMPAT_FEATURE(journal,
139 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
140 ret = submit_bh(WRITE_SYNC_PLUG | WRITE_FLUSH_FUA, bh);
142 ret = submit_bh(WRITE_SYNC_PLUG, bh);
149 * This function along with journal_submit_commit_record
150 * allows to write the commit record asynchronously.
152 static int journal_wait_on_commit_record(journal_t *journal,
153 struct buffer_head *bh)
157 clear_buffer_dirty(bh);
160 if (unlikely(!buffer_uptodate(bh)))
162 put_bh(bh); /* One for getblk() */
163 jbd2_journal_put_journal_head(bh2jh(bh));
169 * write the filemap data using writepage() address_space_operations.
170 * We don't do block allocation here even for delalloc. We don't
171 * use writepages() because with dealyed allocation we may be doing
172 * block allocation in writepages().
174 static int journal_submit_inode_data_buffers(struct address_space *mapping)
177 struct writeback_control wbc = {
178 .sync_mode = WB_SYNC_ALL,
179 .nr_to_write = mapping->nrpages * 2,
181 .range_end = i_size_read(mapping->host),
184 ret = generic_writepages(mapping, &wbc);
189 * Submit all the data buffers of inode associated with the transaction to
192 * We are in a committing transaction. Therefore no new inode can be added to
193 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
194 * operate on from being released while we write out pages.
196 static int journal_submit_data_buffers(journal_t *journal,
197 transaction_t *commit_transaction)
199 struct jbd2_inode *jinode;
201 struct address_space *mapping;
203 spin_lock(&journal->j_list_lock);
204 list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
205 mapping = jinode->i_vfs_inode->i_mapping;
206 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
207 spin_unlock(&journal->j_list_lock);
209 * submit the inode data buffers. We use writepage
210 * instead of writepages. Because writepages can do
211 * block allocation with delalloc. We need to write
212 * only allocated blocks here.
214 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
215 err = journal_submit_inode_data_buffers(mapping);
218 spin_lock(&journal->j_list_lock);
219 J_ASSERT(jinode->i_transaction == commit_transaction);
220 commit_transaction->t_flushed_data_blocks = 1;
221 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
222 smp_mb__after_clear_bit();
223 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
225 spin_unlock(&journal->j_list_lock);
230 * Wait for data submitted for writeout, refile inodes to proper
231 * transaction if needed.
234 static int journal_finish_inode_data_buffers(journal_t *journal,
235 transaction_t *commit_transaction)
237 struct jbd2_inode *jinode, *next_i;
240 /* For locking, see the comment in journal_submit_data_buffers() */
241 spin_lock(&journal->j_list_lock);
242 list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
243 set_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
244 spin_unlock(&journal->j_list_lock);
245 err = filemap_fdatawait(jinode->i_vfs_inode->i_mapping);
248 * Because AS_EIO is cleared by
249 * filemap_fdatawait_range(), set it again so
250 * that user process can get -EIO from fsync().
253 &jinode->i_vfs_inode->i_mapping->flags);
258 spin_lock(&journal->j_list_lock);
259 clear_bit(__JI_COMMIT_RUNNING, &jinode->i_flags);
260 smp_mb__after_clear_bit();
261 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
264 /* Now refile inode to proper lists */
265 list_for_each_entry_safe(jinode, next_i,
266 &commit_transaction->t_inode_list, i_list) {
267 list_del(&jinode->i_list);
268 if (jinode->i_next_transaction) {
269 jinode->i_transaction = jinode->i_next_transaction;
270 jinode->i_next_transaction = NULL;
271 list_add(&jinode->i_list,
272 &jinode->i_transaction->t_inode_list);
274 jinode->i_transaction = NULL;
277 spin_unlock(&journal->j_list_lock);
282 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
284 struct page *page = bh->b_page;
288 addr = kmap_atomic(page, KM_USER0);
289 checksum = crc32_be(crc32_sum,
290 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
291 kunmap_atomic(addr, KM_USER0);
296 static void write_tag_block(int tag_bytes, journal_block_tag_t *tag,
297 unsigned long long block)
299 tag->t_blocknr = cpu_to_be32(block & (u32)~0);
300 if (tag_bytes > JBD2_TAG_SIZE32)
301 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
305 * jbd2_journal_commit_transaction
307 * The primary function for committing a transaction to the log. This
308 * function is called by the journal thread to begin a complete commit.
310 void jbd2_journal_commit_transaction(journal_t *journal)
312 struct transaction_stats_s stats;
313 transaction_t *commit_transaction;
314 struct journal_head *jh, *new_jh, *descriptor;
315 struct buffer_head **wbuf = journal->j_wbuf;
319 unsigned long long blocknr;
323 journal_header_t *header;
324 journal_block_tag_t *tag = NULL;
329 int tag_bytes = journal_tag_bytes(journal);
330 struct buffer_head *cbh = NULL; /* For transactional checksums */
331 __u32 crc32_sum = ~0;
332 int write_op = WRITE_SYNC;
335 * First job: lock down the current transaction and wait for
336 * all outstanding updates to complete.
340 spin_lock(&journal->j_list_lock);
341 summarise_journal_usage(journal);
342 spin_unlock(&journal->j_list_lock);
345 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
346 if (journal->j_flags & JBD2_FLUSHED) {
347 jbd_debug(3, "super block updated\n");
348 jbd2_journal_update_superblock(journal, 1);
350 jbd_debug(3, "superblock not updated\n");
353 J_ASSERT(journal->j_running_transaction != NULL);
354 J_ASSERT(journal->j_committing_transaction == NULL);
356 commit_transaction = journal->j_running_transaction;
357 J_ASSERT(commit_transaction->t_state == T_RUNNING);
359 trace_jbd2_start_commit(journal, commit_transaction);
360 jbd_debug(1, "JBD: starting commit of transaction %d\n",
361 commit_transaction->t_tid);
363 write_lock(&journal->j_state_lock);
364 commit_transaction->t_state = T_LOCKED;
367 * Use plugged writes here, since we want to submit several before
368 * we unplug the device. We don't do explicit unplugging in here,
369 * instead we rely on sync_buffer() doing the unplug for us.
371 if (commit_transaction->t_synchronous_commit)
372 write_op = WRITE_SYNC_PLUG;
373 trace_jbd2_commit_locking(journal, commit_transaction);
374 stats.run.rs_wait = commit_transaction->t_max_wait;
375 stats.run.rs_locked = jiffies;
376 stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
377 stats.run.rs_locked);
379 spin_lock(&commit_transaction->t_handle_lock);
380 while (atomic_read(&commit_transaction->t_updates)) {
383 prepare_to_wait(&journal->j_wait_updates, &wait,
384 TASK_UNINTERRUPTIBLE);
385 if (atomic_read(&commit_transaction->t_updates)) {
386 spin_unlock(&commit_transaction->t_handle_lock);
387 write_unlock(&journal->j_state_lock);
389 write_lock(&journal->j_state_lock);
390 spin_lock(&commit_transaction->t_handle_lock);
392 finish_wait(&journal->j_wait_updates, &wait);
394 spin_unlock(&commit_transaction->t_handle_lock);
396 J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
397 journal->j_max_transaction_buffers);
400 * First thing we are allowed to do is to discard any remaining
401 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
402 * that there are no such buffers: if a large filesystem
403 * operation like a truncate needs to split itself over multiple
404 * transactions, then it may try to do a jbd2_journal_restart() while
405 * there are still BJ_Reserved buffers outstanding. These must
406 * be released cleanly from the current transaction.
408 * In this case, the filesystem must still reserve write access
409 * again before modifying the buffer in the new transaction, but
410 * we do not require it to remember exactly which old buffers it
411 * has reserved. This is consistent with the existing behaviour
412 * that multiple jbd2_journal_get_write_access() calls to the same
413 * buffer are perfectly permissable.
415 while (commit_transaction->t_reserved_list) {
416 jh = commit_transaction->t_reserved_list;
417 JBUFFER_TRACE(jh, "reserved, unused: refile");
419 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
420 * leave undo-committed data.
422 if (jh->b_committed_data) {
423 struct buffer_head *bh = jh2bh(jh);
425 jbd_lock_bh_state(bh);
426 jbd2_free(jh->b_committed_data, bh->b_size);
427 jh->b_committed_data = NULL;
428 jbd_unlock_bh_state(bh);
430 jbd2_journal_refile_buffer(journal, jh);
434 * Now try to drop any written-back buffers from the journal's
435 * checkpoint lists. We do this *before* commit because it potentially
438 spin_lock(&journal->j_list_lock);
439 __jbd2_journal_clean_checkpoint_list(journal);
440 spin_unlock(&journal->j_list_lock);
442 jbd_debug (3, "JBD: commit phase 1\n");
445 * Switch to a new revoke table.
447 jbd2_journal_switch_revoke_table(journal);
449 trace_jbd2_commit_flushing(journal, commit_transaction);
450 stats.run.rs_flushing = jiffies;
451 stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
452 stats.run.rs_flushing);
454 commit_transaction->t_state = T_FLUSH;
455 journal->j_committing_transaction = commit_transaction;
456 journal->j_running_transaction = NULL;
457 start_time = ktime_get();
458 commit_transaction->t_log_start = journal->j_head;
459 wake_up(&journal->j_wait_transaction_locked);
460 write_unlock(&journal->j_state_lock);
462 jbd_debug (3, "JBD: commit phase 2\n");
465 * Now start flushing things to disk, in the order they appear
466 * on the transaction lists. Data blocks go first.
468 err = journal_submit_data_buffers(journal, commit_transaction);
470 jbd2_journal_abort(journal, err);
472 jbd2_journal_write_revoke_records(journal, commit_transaction,
475 jbd_debug(3, "JBD: commit phase 2\n");
478 * Way to go: we have now written out all of the data for a
479 * transaction! Now comes the tricky part: we need to write out
480 * metadata. Loop over the transaction's entire buffer list:
482 write_lock(&journal->j_state_lock);
483 commit_transaction->t_state = T_COMMIT;
484 write_unlock(&journal->j_state_lock);
486 trace_jbd2_commit_logging(journal, commit_transaction);
487 stats.run.rs_logging = jiffies;
488 stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
489 stats.run.rs_logging);
490 stats.run.rs_blocks =
491 atomic_read(&commit_transaction->t_outstanding_credits);
492 stats.run.rs_blocks_logged = 0;
494 J_ASSERT(commit_transaction->t_nr_buffers <=
495 atomic_read(&commit_transaction->t_outstanding_credits));
500 while (commit_transaction->t_buffers) {
502 /* Find the next buffer to be journaled... */
504 jh = commit_transaction->t_buffers;
506 /* If we're in abort mode, we just un-journal the buffer and
509 if (is_journal_aborted(journal)) {
510 clear_buffer_jbddirty(jh2bh(jh));
511 JBUFFER_TRACE(jh, "journal is aborting: refile");
512 jbd2_buffer_abort_trigger(jh,
514 jh->b_frozen_triggers :
516 jbd2_journal_refile_buffer(journal, jh);
517 /* If that was the last one, we need to clean up
518 * any descriptor buffers which may have been
519 * already allocated, even if we are now
521 if (!commit_transaction->t_buffers)
522 goto start_journal_io;
526 /* Make sure we have a descriptor block in which to
527 record the metadata buffer. */
530 struct buffer_head *bh;
532 J_ASSERT (bufs == 0);
534 jbd_debug(4, "JBD: get descriptor\n");
536 descriptor = jbd2_journal_get_descriptor_buffer(journal);
538 jbd2_journal_abort(journal, -EIO);
542 bh = jh2bh(descriptor);
543 jbd_debug(4, "JBD: got buffer %llu (%p)\n",
544 (unsigned long long)bh->b_blocknr, bh->b_data);
545 header = (journal_header_t *)&bh->b_data[0];
546 header->h_magic = cpu_to_be32(JBD2_MAGIC_NUMBER);
547 header->h_blocktype = cpu_to_be32(JBD2_DESCRIPTOR_BLOCK);
548 header->h_sequence = cpu_to_be32(commit_transaction->t_tid);
550 tagp = &bh->b_data[sizeof(journal_header_t)];
551 space_left = bh->b_size - sizeof(journal_header_t);
553 set_buffer_jwrite(bh);
554 set_buffer_dirty(bh);
557 /* Record it so that we can wait for IO
559 BUFFER_TRACE(bh, "ph3: file as descriptor");
560 jbd2_journal_file_buffer(descriptor, commit_transaction,
564 /* Where is the buffer to be written? */
566 err = jbd2_journal_next_log_block(journal, &blocknr);
567 /* If the block mapping failed, just abandon the buffer
568 and repeat this loop: we'll fall into the
569 refile-on-abort condition above. */
571 jbd2_journal_abort(journal, err);
576 * start_this_handle() uses t_outstanding_credits to determine
577 * the free space in the log, but this counter is changed
578 * by jbd2_journal_next_log_block() also.
580 atomic_dec(&commit_transaction->t_outstanding_credits);
582 /* Bump b_count to prevent truncate from stumbling over
583 the shadowed buffer! @@@ This can go if we ever get
584 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
585 atomic_inc(&jh2bh(jh)->b_count);
587 /* Make a temporary IO buffer with which to write it out
588 (this will requeue both the metadata buffer and the
589 temporary IO buffer). new_bh goes on BJ_IO*/
591 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
593 * akpm: jbd2_journal_write_metadata_buffer() sets
594 * new_bh->b_transaction to commit_transaction.
595 * We need to clean this up before we release new_bh
596 * (which is of type BJ_IO)
598 JBUFFER_TRACE(jh, "ph3: write metadata");
599 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
600 jh, &new_jh, blocknr);
602 jbd2_journal_abort(journal, flags);
605 set_bit(BH_JWrite, &jh2bh(new_jh)->b_state);
606 wbuf[bufs++] = jh2bh(new_jh);
608 /* Record the new block's tag in the current descriptor
613 tag_flag |= JBD2_FLAG_ESCAPE;
615 tag_flag |= JBD2_FLAG_SAME_UUID;
617 tag = (journal_block_tag_t *) tagp;
618 write_tag_block(tag_bytes, tag, jh2bh(jh)->b_blocknr);
619 tag->t_flags = cpu_to_be32(tag_flag);
621 space_left -= tag_bytes;
624 memcpy (tagp, journal->j_uuid, 16);
630 /* If there's no more to do, or if the descriptor is full,
633 if (bufs == journal->j_wbufsize ||
634 commit_transaction->t_buffers == NULL ||
635 space_left < tag_bytes + 16) {
637 jbd_debug(4, "JBD: Submit %d IOs\n", bufs);
639 /* Write an end-of-descriptor marker before
640 submitting the IOs. "tag" still points to
641 the last tag we set up. */
643 tag->t_flags |= cpu_to_be32(JBD2_FLAG_LAST_TAG);
646 for (i = 0; i < bufs; i++) {
647 struct buffer_head *bh = wbuf[i];
651 if (JBD2_HAS_COMPAT_FEATURE(journal,
652 JBD2_FEATURE_COMPAT_CHECKSUM)) {
654 jbd2_checksum_data(crc32_sum, bh);
658 clear_buffer_dirty(bh);
659 set_buffer_uptodate(bh);
660 bh->b_end_io = journal_end_buffer_io_sync;
661 submit_bh(write_op, bh);
664 stats.run.rs_blocks_logged += bufs;
666 /* Force a new descriptor to be generated next
667 time round the loop. */
673 err = journal_finish_inode_data_buffers(journal, commit_transaction);
676 "JBD2: Detected IO errors while flushing file data "
677 "on %s\n", journal->j_devname);
678 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
679 jbd2_journal_abort(journal, err);
684 * If the journal is not located on the file system device,
685 * then we must flush the file system device before we issue
688 if (commit_transaction->t_flushed_data_blocks &&
689 (journal->j_fs_dev != journal->j_dev) &&
690 (journal->j_flags & JBD2_BARRIER))
691 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
693 /* Done it all: now write the commit record asynchronously. */
694 if (JBD2_HAS_INCOMPAT_FEATURE(journal,
695 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
696 err = journal_submit_commit_record(journal, commit_transaction,
699 __jbd2_journal_abort_hard(journal);
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);
763 /* Wake up any transactions which were waiting for this
765 wake_up_bit(&bh->b_state, BH_Unshadow);
766 JBUFFER_TRACE(jh, "brelse shadowed buffer");
770 J_ASSERT (commit_transaction->t_shadow_list == NULL);
772 jbd_debug(3, "JBD: commit phase 4\n");
774 /* Here we wait for the revoke record and descriptor record buffers */
776 while (commit_transaction->t_log_list != NULL) {
777 struct buffer_head *bh;
779 jh = commit_transaction->t_log_list->b_tprev;
781 if (buffer_locked(bh)) {
783 goto wait_for_ctlbuf;
786 goto wait_for_ctlbuf;
788 if (unlikely(!buffer_uptodate(bh)))
791 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
792 clear_buffer_jwrite(bh);
793 jbd2_journal_unfile_buffer(journal, jh);
794 jbd2_journal_put_journal_head(jh);
795 __brelse(bh); /* One for getblk */
796 /* AKPM: bforget here */
800 jbd2_journal_abort(journal, err);
802 jbd_debug(3, "JBD: commit phase 5\n");
804 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
805 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
806 err = journal_submit_commit_record(journal, commit_transaction,
809 __jbd2_journal_abort_hard(journal);
811 if (!err && !is_journal_aborted(journal))
812 err = journal_wait_on_commit_record(journal, cbh);
813 if (JBD2_HAS_INCOMPAT_FEATURE(journal,
814 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
815 journal->j_flags & JBD2_BARRIER) {
816 blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL);
820 jbd2_journal_abort(journal, err);
822 /* End of a transaction! Finally, we can do checkpoint
823 processing: any buffers committed as a result of this
824 transaction can be removed from any checkpoint list it was on
827 jbd_debug(3, "JBD: commit phase 6\n");
829 J_ASSERT(list_empty(&commit_transaction->t_inode_list));
830 J_ASSERT(commit_transaction->t_buffers == NULL);
831 J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
832 J_ASSERT(commit_transaction->t_iobuf_list == NULL);
833 J_ASSERT(commit_transaction->t_shadow_list == NULL);
834 J_ASSERT(commit_transaction->t_log_list == NULL);
838 * As there are other places (journal_unmap_buffer()) adding buffers
839 * to this list we have to be careful and hold the j_list_lock.
841 spin_lock(&journal->j_list_lock);
842 while (commit_transaction->t_forget) {
843 transaction_t *cp_transaction;
844 struct buffer_head *bh;
846 jh = commit_transaction->t_forget;
847 spin_unlock(&journal->j_list_lock);
849 jbd_lock_bh_state(bh);
850 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
853 * If there is undo-protected committed data against
854 * this buffer, then we can remove it now. If it is a
855 * buffer needing such protection, the old frozen_data
856 * field now points to a committed version of the
857 * buffer, so rotate that field to the new committed
860 * Otherwise, we can just throw away the frozen data now.
862 * We also know that the frozen data has already fired
863 * its triggers if they exist, so we can clear that too.
865 if (jh->b_committed_data) {
866 jbd2_free(jh->b_committed_data, bh->b_size);
867 jh->b_committed_data = NULL;
868 if (jh->b_frozen_data) {
869 jh->b_committed_data = jh->b_frozen_data;
870 jh->b_frozen_data = NULL;
871 jh->b_frozen_triggers = NULL;
873 } else if (jh->b_frozen_data) {
874 jbd2_free(jh->b_frozen_data, bh->b_size);
875 jh->b_frozen_data = NULL;
876 jh->b_frozen_triggers = NULL;
879 spin_lock(&journal->j_list_lock);
880 cp_transaction = jh->b_cp_transaction;
881 if (cp_transaction) {
882 JBUFFER_TRACE(jh, "remove from old cp transaction");
883 cp_transaction->t_chp_stats.cs_dropped++;
884 __jbd2_journal_remove_checkpoint(jh);
887 /* Only re-checkpoint the buffer_head if it is marked
888 * dirty. If the buffer was added to the BJ_Forget list
889 * by jbd2_journal_forget, it may no longer be dirty and
890 * there's no point in keeping a checkpoint record for
893 /* A buffer which has been freed while still being
894 * journaled by a previous transaction may end up still
895 * being dirty here, but we want to avoid writing back
896 * that buffer in the future after the "add to orphan"
897 * operation been committed, That's not only a performance
898 * gain, it also stops aliasing problems if the buffer is
899 * left behind for writeback and gets reallocated for another
900 * use in a different page. */
901 if (buffer_freed(bh) && !jh->b_next_transaction) {
902 clear_buffer_freed(bh);
903 clear_buffer_jbddirty(bh);
906 if (buffer_jbddirty(bh)) {
907 JBUFFER_TRACE(jh, "add to new checkpointing trans");
908 __jbd2_journal_insert_checkpoint(jh, commit_transaction);
909 if (is_journal_aborted(journal))
910 clear_buffer_jbddirty(bh);
911 JBUFFER_TRACE(jh, "refile for checkpoint writeback");
912 __jbd2_journal_refile_buffer(jh);
913 jbd_unlock_bh_state(bh);
915 J_ASSERT_BH(bh, !buffer_dirty(bh));
916 /* The buffer on BJ_Forget list and not jbddirty means
917 * it has been freed by this transaction and hence it
918 * could not have been reallocated until this
919 * transaction has committed. *BUT* it could be
920 * reallocated once we have written all the data to
921 * disk and before we process the buffer on BJ_Forget
923 JBUFFER_TRACE(jh, "refile or unfile freed buffer");
924 __jbd2_journal_refile_buffer(jh);
925 if (!jh->b_transaction) {
926 jbd_unlock_bh_state(bh);
928 jbd2_journal_remove_journal_head(bh);
929 release_buffer_page(bh);
931 jbd_unlock_bh_state(bh);
933 cond_resched_lock(&journal->j_list_lock);
935 spin_unlock(&journal->j_list_lock);
937 * This is a bit sleazy. We use j_list_lock to protect transition
938 * of a transaction into T_FINISHED state and calling
939 * __jbd2_journal_drop_transaction(). Otherwise we could race with
940 * other checkpointing code processing the transaction...
942 write_lock(&journal->j_state_lock);
943 spin_lock(&journal->j_list_lock);
945 * Now recheck if some buffers did not get attached to the transaction
946 * while the lock was dropped...
948 if (commit_transaction->t_forget) {
949 spin_unlock(&journal->j_list_lock);
950 write_unlock(&journal->j_state_lock);
954 /* Done with this transaction! */
956 jbd_debug(3, "JBD: commit phase 7\n");
958 J_ASSERT(commit_transaction->t_state == T_COMMIT);
960 commit_transaction->t_start = jiffies;
961 stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
962 commit_transaction->t_start);
965 * File the transaction statistics
967 stats.ts_tid = commit_transaction->t_tid;
968 stats.run.rs_handle_count =
969 atomic_read(&commit_transaction->t_handle_count);
970 trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
971 commit_transaction->t_tid, &stats.run);
974 * Calculate overall stats
976 spin_lock(&journal->j_history_lock);
977 journal->j_stats.ts_tid++;
978 journal->j_stats.run.rs_wait += stats.run.rs_wait;
979 journal->j_stats.run.rs_running += stats.run.rs_running;
980 journal->j_stats.run.rs_locked += stats.run.rs_locked;
981 journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
982 journal->j_stats.run.rs_logging += stats.run.rs_logging;
983 journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
984 journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
985 journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
986 spin_unlock(&journal->j_history_lock);
988 commit_transaction->t_state = T_FINISHED;
989 J_ASSERT(commit_transaction == journal->j_committing_transaction);
990 journal->j_commit_sequence = commit_transaction->t_tid;
991 journal->j_committing_transaction = NULL;
992 commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
995 * weight the commit time higher than the average time so we don't
996 * react too strongly to vast changes in the commit time
998 if (likely(journal->j_average_commit_time))
999 journal->j_average_commit_time = (commit_time +
1000 journal->j_average_commit_time*3) / 4;
1002 journal->j_average_commit_time = commit_time;
1003 write_unlock(&journal->j_state_lock);
1005 if (commit_transaction->t_checkpoint_list == NULL &&
1006 commit_transaction->t_checkpoint_io_list == NULL) {
1007 __jbd2_journal_drop_transaction(journal, commit_transaction);
1010 if (journal->j_checkpoint_transactions == NULL) {
1011 journal->j_checkpoint_transactions = commit_transaction;
1012 commit_transaction->t_cpnext = commit_transaction;
1013 commit_transaction->t_cpprev = commit_transaction;
1015 commit_transaction->t_cpnext =
1016 journal->j_checkpoint_transactions;
1017 commit_transaction->t_cpprev =
1018 commit_transaction->t_cpnext->t_cpprev;
1019 commit_transaction->t_cpnext->t_cpprev =
1021 commit_transaction->t_cpprev->t_cpnext =
1025 spin_unlock(&journal->j_list_lock);
1027 if (journal->j_commit_callback)
1028 journal->j_commit_callback(journal, commit_transaction);
1030 trace_jbd2_end_commit(journal, commit_transaction);
1031 jbd_debug(1, "JBD: commit %d complete, head %d\n",
1032 journal->j_commit_sequence, journal->j_tail_sequence);
1034 kfree(commit_transaction);
1036 wake_up(&journal->j_wait_done_commit);