1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/commit.c
7 * Copyright 1998 Red Hat corp --- All Rights Reserved
9 * Journal commit routines for the generic filesystem journaling code;
10 * part of the ext2fs journaling system.
13 #include <linux/time.h>
15 #include <linux/jbd2.h>
16 #include <linux/errno.h>
17 #include <linux/slab.h>
19 #include <linux/pagemap.h>
20 #include <linux/jiffies.h>
21 #include <linux/crc32.h>
22 #include <linux/writeback.h>
23 #include <linux/backing-dev.h>
24 #include <linux/bio.h>
25 #include <linux/blkdev.h>
26 #include <linux/bitops.h>
27 #include <trace/events/jbd2.h>
30 * IO end handler for temporary buffer_heads handling writes to the journal.
32 static void journal_end_buffer_io_sync(struct buffer_head *bh, int uptodate)
34 struct buffer_head *orig_bh = bh->b_private;
38 set_buffer_uptodate(bh);
40 clear_buffer_uptodate(bh);
42 clear_bit_unlock(BH_Shadow, &orig_bh->b_state);
43 smp_mb__after_atomic();
44 wake_up_bit(&orig_bh->b_state, BH_Shadow);
50 * When an ext4 file is truncated, it is possible that some pages are not
51 * successfully freed, because they are attached to a committing transaction.
52 * After the transaction commits, these pages are left on the LRU, with no
53 * ->mapping, and with attached buffers. These pages are trivially reclaimable
54 * by the VM, but their apparent absence upsets the VM accounting, and it makes
55 * the numbers in /proc/meminfo look odd.
57 * So here, we have a buffer which has just come off the forget list. Look to
58 * see if we can strip all buffers from the backing page.
60 * Called under lock_journal(), and possibly under journal_datalist_lock. The
61 * caller provided us with a ref against the buffer, and we drop that here.
63 static void release_buffer_page(struct buffer_head *bh)
70 if (atomic_read(&bh->b_count) != 1)
75 folio = page_folio(page);
79 /* OK, it's a truncated page */
80 if (!folio_trylock(folio))
85 try_to_free_buffers(folio);
94 static void jbd2_commit_block_csum_set(journal_t *j, struct buffer_head *bh)
96 struct commit_header *h;
99 if (!jbd2_journal_has_csum_v2or3(j))
102 h = (struct commit_header *)(bh->b_data);
103 h->h_chksum_type = 0;
104 h->h_chksum_size = 0;
106 csum = jbd2_chksum(j, j->j_csum_seed, bh->b_data, j->j_blocksize);
107 h->h_chksum[0] = cpu_to_be32(csum);
111 * Done it all: now submit the commit record. We should have
112 * cleaned up our previous buffers by now, so if we are in abort
113 * mode we can now just skip the rest of the journal write
116 * Returns 1 if the journal needs to be aborted or 0 on success
118 static int journal_submit_commit_record(journal_t *journal,
119 transaction_t *commit_transaction,
120 struct buffer_head **cbh,
123 struct commit_header *tmp;
124 struct buffer_head *bh;
125 struct timespec64 now;
126 blk_opf_t write_flags = REQ_OP_WRITE | REQ_SYNC;
130 if (is_journal_aborted(journal))
133 bh = jbd2_journal_get_descriptor_buffer(commit_transaction,
138 tmp = (struct commit_header *)bh->b_data;
139 ktime_get_coarse_real_ts64(&now);
140 tmp->h_commit_sec = cpu_to_be64(now.tv_sec);
141 tmp->h_commit_nsec = cpu_to_be32(now.tv_nsec);
143 if (jbd2_has_feature_checksum(journal)) {
144 tmp->h_chksum_type = JBD2_CRC32_CHKSUM;
145 tmp->h_chksum_size = JBD2_CRC32_CHKSUM_SIZE;
146 tmp->h_chksum[0] = cpu_to_be32(crc32_sum);
148 jbd2_commit_block_csum_set(journal, bh);
150 BUFFER_TRACE(bh, "submit commit block");
152 clear_buffer_dirty(bh);
153 set_buffer_uptodate(bh);
154 bh->b_end_io = journal_end_buffer_io_sync;
156 if (journal->j_flags & JBD2_BARRIER &&
157 !jbd2_has_feature_async_commit(journal))
158 write_flags |= REQ_PREFLUSH | REQ_FUA;
160 submit_bh(write_flags, bh);
166 * This function along with journal_submit_commit_record
167 * allows to write the commit record asynchronously.
169 static int journal_wait_on_commit_record(journal_t *journal,
170 struct buffer_head *bh)
174 clear_buffer_dirty(bh);
177 if (unlikely(!buffer_uptodate(bh)))
179 put_bh(bh); /* One for getblk() */
185 * write the filemap data using writepage() address_space_operations.
186 * We don't do block allocation here even for delalloc. We don't
187 * use writepages() because with delayed allocation we may be doing
188 * block allocation in writepages().
190 int jbd2_journal_submit_inode_data_buffers(struct jbd2_inode *jinode)
192 struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
193 struct writeback_control wbc = {
194 .sync_mode = WB_SYNC_ALL,
195 .nr_to_write = mapping->nrpages * 2,
196 .range_start = jinode->i_dirty_start,
197 .range_end = jinode->i_dirty_end,
201 * submit the inode data buffers. We use writepage
202 * instead of writepages. Because writepages can do
203 * block allocation with delalloc. We need to write
204 * only allocated blocks here.
206 return generic_writepages(mapping, &wbc);
209 /* Send all the data buffers related to an inode */
210 int jbd2_submit_inode_data(journal_t *journal, struct jbd2_inode *jinode)
212 if (!jinode || !(jinode->i_flags & JI_WRITE_DATA))
215 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
216 return journal->j_submit_inode_data_buffers(jinode);
219 EXPORT_SYMBOL(jbd2_submit_inode_data);
221 int jbd2_wait_inode_data(journal_t *journal, struct jbd2_inode *jinode)
223 if (!jinode || !(jinode->i_flags & JI_WAIT_DATA) ||
224 !jinode->i_vfs_inode || !jinode->i_vfs_inode->i_mapping)
226 return filemap_fdatawait_range_keep_errors(
227 jinode->i_vfs_inode->i_mapping, jinode->i_dirty_start,
228 jinode->i_dirty_end);
230 EXPORT_SYMBOL(jbd2_wait_inode_data);
233 * Submit all the data buffers of inode associated with the transaction to
236 * We are in a committing transaction. Therefore no new inode can be added to
237 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
238 * operate on from being released while we write out pages.
240 static int journal_submit_data_buffers(journal_t *journal,
241 transaction_t *commit_transaction)
243 struct jbd2_inode *jinode;
246 spin_lock(&journal->j_list_lock);
247 list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
248 if (!(jinode->i_flags & JI_WRITE_DATA))
250 jinode->i_flags |= JI_COMMIT_RUNNING;
251 spin_unlock(&journal->j_list_lock);
252 /* submit the inode data buffers. */
253 trace_jbd2_submit_inode_data(jinode->i_vfs_inode);
254 if (journal->j_submit_inode_data_buffers) {
255 err = journal->j_submit_inode_data_buffers(jinode);
259 spin_lock(&journal->j_list_lock);
260 J_ASSERT(jinode->i_transaction == commit_transaction);
261 jinode->i_flags &= ~JI_COMMIT_RUNNING;
263 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
265 spin_unlock(&journal->j_list_lock);
269 int jbd2_journal_finish_inode_data_buffers(struct jbd2_inode *jinode)
271 struct address_space *mapping = jinode->i_vfs_inode->i_mapping;
273 return filemap_fdatawait_range_keep_errors(mapping,
274 jinode->i_dirty_start,
275 jinode->i_dirty_end);
279 * Wait for data submitted for writeout, refile inodes to proper
280 * transaction if needed.
283 static int journal_finish_inode_data_buffers(journal_t *journal,
284 transaction_t *commit_transaction)
286 struct jbd2_inode *jinode, *next_i;
289 /* For locking, see the comment in journal_submit_data_buffers() */
290 spin_lock(&journal->j_list_lock);
291 list_for_each_entry(jinode, &commit_transaction->t_inode_list, i_list) {
292 if (!(jinode->i_flags & JI_WAIT_DATA))
294 jinode->i_flags |= JI_COMMIT_RUNNING;
295 spin_unlock(&journal->j_list_lock);
296 /* wait for the inode data buffers writeout. */
297 if (journal->j_finish_inode_data_buffers) {
298 err = journal->j_finish_inode_data_buffers(jinode);
302 spin_lock(&journal->j_list_lock);
303 jinode->i_flags &= ~JI_COMMIT_RUNNING;
305 wake_up_bit(&jinode->i_flags, __JI_COMMIT_RUNNING);
308 /* Now refile inode to proper lists */
309 list_for_each_entry_safe(jinode, next_i,
310 &commit_transaction->t_inode_list, i_list) {
311 list_del(&jinode->i_list);
312 if (jinode->i_next_transaction) {
313 jinode->i_transaction = jinode->i_next_transaction;
314 jinode->i_next_transaction = NULL;
315 list_add(&jinode->i_list,
316 &jinode->i_transaction->t_inode_list);
318 jinode->i_transaction = NULL;
319 jinode->i_dirty_start = 0;
320 jinode->i_dirty_end = 0;
323 spin_unlock(&journal->j_list_lock);
328 static __u32 jbd2_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
330 struct page *page = bh->b_page;
334 addr = kmap_atomic(page);
335 checksum = crc32_be(crc32_sum,
336 (void *)(addr + offset_in_page(bh->b_data)), bh->b_size);
342 static void write_tag_block(journal_t *j, journal_block_tag_t *tag,
343 unsigned long long block)
345 tag->t_blocknr = cpu_to_be32(block & (u32)~0);
346 if (jbd2_has_feature_64bit(j))
347 tag->t_blocknr_high = cpu_to_be32((block >> 31) >> 1);
350 static void jbd2_block_tag_csum_set(journal_t *j, journal_block_tag_t *tag,
351 struct buffer_head *bh, __u32 sequence)
353 journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
354 struct page *page = bh->b_page;
359 if (!jbd2_journal_has_csum_v2or3(j))
362 seq = cpu_to_be32(sequence);
363 addr = kmap_atomic(page);
364 csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
365 csum32 = jbd2_chksum(j, csum32, addr + offset_in_page(bh->b_data),
369 if (jbd2_has_feature_csum3(j))
370 tag3->t_checksum = cpu_to_be32(csum32);
372 tag->t_checksum = cpu_to_be16(csum32);
375 * jbd2_journal_commit_transaction
377 * The primary function for committing a transaction to the log. This
378 * function is called by the journal thread to begin a complete commit.
380 void jbd2_journal_commit_transaction(journal_t *journal)
382 struct transaction_stats_s stats;
383 transaction_t *commit_transaction;
384 struct journal_head *jh;
385 struct buffer_head *descriptor;
386 struct buffer_head **wbuf = journal->j_wbuf;
390 unsigned long long blocknr;
394 journal_block_tag_t *tag = NULL;
399 int tag_bytes = journal_tag_bytes(journal);
400 struct buffer_head *cbh = NULL; /* For transactional checksums */
401 __u32 crc32_sum = ~0;
402 struct blk_plug plug;
403 /* Tail of the journal */
404 unsigned long first_block;
411 if (jbd2_journal_has_csum_v2or3(journal))
412 csum_size = sizeof(struct jbd2_journal_block_tail);
415 * First job: lock down the current transaction and wait for
416 * all outstanding updates to complete.
419 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
420 if (journal->j_flags & JBD2_FLUSHED) {
421 jbd2_debug(3, "super block updated\n");
422 mutex_lock_io(&journal->j_checkpoint_mutex);
424 * We hold j_checkpoint_mutex so tail cannot change under us.
425 * We don't need any special data guarantees for writing sb
426 * since journal is empty and it is ok for write to be
427 * flushed only with transaction commit.
429 jbd2_journal_update_sb_log_tail(journal,
430 journal->j_tail_sequence,
433 mutex_unlock(&journal->j_checkpoint_mutex);
435 jbd2_debug(3, "superblock not updated\n");
438 J_ASSERT(journal->j_running_transaction != NULL);
439 J_ASSERT(journal->j_committing_transaction == NULL);
441 write_lock(&journal->j_state_lock);
442 journal->j_flags |= JBD2_FULL_COMMIT_ONGOING;
443 while (journal->j_flags & JBD2_FAST_COMMIT_ONGOING) {
446 prepare_to_wait(&journal->j_fc_wait, &wait,
447 TASK_UNINTERRUPTIBLE);
448 write_unlock(&journal->j_state_lock);
450 write_lock(&journal->j_state_lock);
451 finish_wait(&journal->j_fc_wait, &wait);
453 * TODO: by blocking fast commits here, we are increasing
454 * fsync() latency slightly. Strictly speaking, we don't need
455 * to block fast commits until the transaction enters T_FLUSH
456 * state. So an optimization is possible where we block new fast
457 * commits here and wait for existing ones to complete
458 * just before we enter T_FLUSH. That way, the existing fast
459 * commits and this full commit can proceed parallely.
462 write_unlock(&journal->j_state_lock);
464 commit_transaction = journal->j_running_transaction;
466 trace_jbd2_start_commit(journal, commit_transaction);
467 jbd2_debug(1, "JBD2: starting commit of transaction %d\n",
468 commit_transaction->t_tid);
470 write_lock(&journal->j_state_lock);
471 journal->j_fc_off = 0;
472 J_ASSERT(commit_transaction->t_state == T_RUNNING);
473 commit_transaction->t_state = T_LOCKED;
475 trace_jbd2_commit_locking(journal, commit_transaction);
476 stats.run.rs_wait = commit_transaction->t_max_wait;
477 stats.run.rs_request_delay = 0;
478 stats.run.rs_locked = jiffies;
479 if (commit_transaction->t_requested)
480 stats.run.rs_request_delay =
481 jbd2_time_diff(commit_transaction->t_requested,
482 stats.run.rs_locked);
483 stats.run.rs_running = jbd2_time_diff(commit_transaction->t_start,
484 stats.run.rs_locked);
486 // waits for any t_updates to finish
487 jbd2_journal_wait_updates(journal);
489 commit_transaction->t_state = T_SWITCH;
491 J_ASSERT (atomic_read(&commit_transaction->t_outstanding_credits) <=
492 journal->j_max_transaction_buffers);
495 * First thing we are allowed to do is to discard any remaining
496 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
497 * that there are no such buffers: if a large filesystem
498 * operation like a truncate needs to split itself over multiple
499 * transactions, then it may try to do a jbd2_journal_restart() while
500 * there are still BJ_Reserved buffers outstanding. These must
501 * be released cleanly from the current transaction.
503 * In this case, the filesystem must still reserve write access
504 * again before modifying the buffer in the new transaction, but
505 * we do not require it to remember exactly which old buffers it
506 * has reserved. This is consistent with the existing behaviour
507 * that multiple jbd2_journal_get_write_access() calls to the same
508 * buffer are perfectly permissible.
509 * We use journal->j_state_lock here to serialize processing of
510 * t_reserved_list with eviction of buffers from journal_unmap_buffer().
512 while (commit_transaction->t_reserved_list) {
513 jh = commit_transaction->t_reserved_list;
514 JBUFFER_TRACE(jh, "reserved, unused: refile");
516 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
517 * leave undo-committed data.
519 if (jh->b_committed_data) {
520 struct buffer_head *bh = jh2bh(jh);
522 spin_lock(&jh->b_state_lock);
523 jbd2_free(jh->b_committed_data, bh->b_size);
524 jh->b_committed_data = NULL;
525 spin_unlock(&jh->b_state_lock);
527 jbd2_journal_refile_buffer(journal, jh);
530 write_unlock(&journal->j_state_lock);
532 * Now try to drop any written-back buffers from the journal's
533 * checkpoint lists. We do this *before* commit because it potentially
536 spin_lock(&journal->j_list_lock);
537 __jbd2_journal_clean_checkpoint_list(journal, false);
538 spin_unlock(&journal->j_list_lock);
540 jbd2_debug(3, "JBD2: commit phase 1\n");
543 * Clear revoked flag to reflect there is no revoked buffers
544 * in the next transaction which is going to be started.
546 jbd2_clear_buffer_revoked_flags(journal);
549 * Switch to a new revoke table.
551 jbd2_journal_switch_revoke_table(journal);
553 write_lock(&journal->j_state_lock);
555 * Reserved credits cannot be claimed anymore, free them
557 atomic_sub(atomic_read(&journal->j_reserved_credits),
558 &commit_transaction->t_outstanding_credits);
560 trace_jbd2_commit_flushing(journal, commit_transaction);
561 stats.run.rs_flushing = jiffies;
562 stats.run.rs_locked = jbd2_time_diff(stats.run.rs_locked,
563 stats.run.rs_flushing);
565 commit_transaction->t_state = T_FLUSH;
566 journal->j_committing_transaction = commit_transaction;
567 journal->j_running_transaction = NULL;
568 start_time = ktime_get();
569 commit_transaction->t_log_start = journal->j_head;
570 wake_up_all(&journal->j_wait_transaction_locked);
571 write_unlock(&journal->j_state_lock);
573 jbd2_debug(3, "JBD2: commit phase 2a\n");
576 * Now start flushing things to disk, in the order they appear
577 * on the transaction lists. Data blocks go first.
579 err = journal_submit_data_buffers(journal, commit_transaction);
581 jbd2_journal_abort(journal, err);
583 blk_start_plug(&plug);
584 jbd2_journal_write_revoke_records(commit_transaction, &log_bufs);
586 jbd2_debug(3, "JBD2: commit phase 2b\n");
589 * Way to go: we have now written out all of the data for a
590 * transaction! Now comes the tricky part: we need to write out
591 * metadata. Loop over the transaction's entire buffer list:
593 write_lock(&journal->j_state_lock);
594 commit_transaction->t_state = T_COMMIT;
595 write_unlock(&journal->j_state_lock);
597 trace_jbd2_commit_logging(journal, commit_transaction);
598 stats.run.rs_logging = jiffies;
599 stats.run.rs_flushing = jbd2_time_diff(stats.run.rs_flushing,
600 stats.run.rs_logging);
601 stats.run.rs_blocks = commit_transaction->t_nr_buffers;
602 stats.run.rs_blocks_logged = 0;
604 J_ASSERT(commit_transaction->t_nr_buffers <=
605 atomic_read(&commit_transaction->t_outstanding_credits));
610 while (commit_transaction->t_buffers) {
612 /* Find the next buffer to be journaled... */
614 jh = commit_transaction->t_buffers;
616 /* If we're in abort mode, we just un-journal the buffer and
619 if (is_journal_aborted(journal)) {
620 clear_buffer_jbddirty(jh2bh(jh));
621 JBUFFER_TRACE(jh, "journal is aborting: refile");
622 jbd2_buffer_abort_trigger(jh,
624 jh->b_frozen_triggers :
626 jbd2_journal_refile_buffer(journal, jh);
627 /* If that was the last one, we need to clean up
628 * any descriptor buffers which may have been
629 * already allocated, even if we are now
631 if (!commit_transaction->t_buffers)
632 goto start_journal_io;
636 /* Make sure we have a descriptor block in which to
637 record the metadata buffer. */
640 J_ASSERT (bufs == 0);
642 jbd2_debug(4, "JBD2: get descriptor\n");
644 descriptor = jbd2_journal_get_descriptor_buffer(
646 JBD2_DESCRIPTOR_BLOCK);
648 jbd2_journal_abort(journal, -EIO);
652 jbd2_debug(4, "JBD2: got buffer %llu (%p)\n",
653 (unsigned long long)descriptor->b_blocknr,
655 tagp = &descriptor->b_data[sizeof(journal_header_t)];
656 space_left = descriptor->b_size -
657 sizeof(journal_header_t);
659 set_buffer_jwrite(descriptor);
660 set_buffer_dirty(descriptor);
661 wbuf[bufs++] = descriptor;
663 /* Record it so that we can wait for IO
665 BUFFER_TRACE(descriptor, "ph3: file as descriptor");
666 jbd2_file_log_bh(&log_bufs, descriptor);
669 /* Where is the buffer to be written? */
671 err = jbd2_journal_next_log_block(journal, &blocknr);
672 /* If the block mapping failed, just abandon the buffer
673 and repeat this loop: we'll fall into the
674 refile-on-abort condition above. */
676 jbd2_journal_abort(journal, err);
681 * start_this_handle() uses t_outstanding_credits to determine
682 * the free space in the log.
684 atomic_dec(&commit_transaction->t_outstanding_credits);
686 /* Bump b_count to prevent truncate from stumbling over
687 the shadowed buffer! @@@ This can go if we ever get
688 rid of the shadow pairing of buffers. */
689 atomic_inc(&jh2bh(jh)->b_count);
692 * Make a temporary IO buffer with which to write it out
693 * (this will requeue the metadata buffer to BJ_Shadow).
695 set_bit(BH_JWrite, &jh2bh(jh)->b_state);
696 JBUFFER_TRACE(jh, "ph3: write metadata");
697 flags = jbd2_journal_write_metadata_buffer(commit_transaction,
698 jh, &wbuf[bufs], blocknr);
700 jbd2_journal_abort(journal, flags);
703 jbd2_file_log_bh(&io_bufs, wbuf[bufs]);
705 /* Record the new block's tag in the current descriptor
710 tag_flag |= JBD2_FLAG_ESCAPE;
712 tag_flag |= JBD2_FLAG_SAME_UUID;
714 tag = (journal_block_tag_t *) tagp;
715 write_tag_block(journal, tag, jh2bh(jh)->b_blocknr);
716 tag->t_flags = cpu_to_be16(tag_flag);
717 jbd2_block_tag_csum_set(journal, tag, wbuf[bufs],
718 commit_transaction->t_tid);
720 space_left -= tag_bytes;
724 memcpy (tagp, journal->j_uuid, 16);
730 /* If there's no more to do, or if the descriptor is full,
733 if (bufs == journal->j_wbufsize ||
734 commit_transaction->t_buffers == NULL ||
735 space_left < tag_bytes + 16 + csum_size) {
737 jbd2_debug(4, "JBD2: Submit %d IOs\n", bufs);
739 /* Write an end-of-descriptor marker before
740 submitting the IOs. "tag" still points to
741 the last tag we set up. */
743 tag->t_flags |= cpu_to_be16(JBD2_FLAG_LAST_TAG);
746 jbd2_descriptor_block_csum_set(journal,
749 for (i = 0; i < bufs; i++) {
750 struct buffer_head *bh = wbuf[i];
754 if (jbd2_has_feature_checksum(journal)) {
756 jbd2_checksum_data(crc32_sum, bh);
760 clear_buffer_dirty(bh);
761 set_buffer_uptodate(bh);
762 bh->b_end_io = journal_end_buffer_io_sync;
763 submit_bh(REQ_OP_WRITE | REQ_SYNC, bh);
767 /* Force a new descriptor to be generated next
768 time round the loop. */
774 err = journal_finish_inode_data_buffers(journal, commit_transaction);
777 "JBD2: Detected IO errors while flushing file data "
778 "on %s\n", journal->j_devname);
779 if (journal->j_flags & JBD2_ABORT_ON_SYNCDATA_ERR)
780 jbd2_journal_abort(journal, err);
785 * Get current oldest transaction in the log before we issue flush
786 * to the filesystem device. After the flush we can be sure that
787 * blocks of all older transactions are checkpointed to persistent
788 * storage and we will be safe to update journal start in the
789 * superblock with the numbers we get here.
792 jbd2_journal_get_log_tail(journal, &first_tid, &first_block);
794 write_lock(&journal->j_state_lock);
796 long freed = first_block - journal->j_tail;
798 if (first_block < journal->j_tail)
799 freed += journal->j_last - journal->j_first;
800 /* Update tail only if we free significant amount of space */
801 if (freed < jbd2_journal_get_max_txn_bufs(journal))
804 J_ASSERT(commit_transaction->t_state == T_COMMIT);
805 commit_transaction->t_state = T_COMMIT_DFLUSH;
806 write_unlock(&journal->j_state_lock);
809 * If the journal is not located on the file system device,
810 * then we must flush the file system device before we issue
813 if (commit_transaction->t_need_data_flush &&
814 (journal->j_fs_dev != journal->j_dev) &&
815 (journal->j_flags & JBD2_BARRIER))
816 blkdev_issue_flush(journal->j_fs_dev);
818 /* Done it all: now write the commit record asynchronously. */
819 if (jbd2_has_feature_async_commit(journal)) {
820 err = journal_submit_commit_record(journal, commit_transaction,
823 jbd2_journal_abort(journal, err);
826 blk_finish_plug(&plug);
828 /* Lo and behold: we have just managed to send a transaction to
829 the log. Before we can commit it, wait for the IO so far to
830 complete. Control buffers being written are on the
831 transaction's t_log_list queue, and metadata buffers are on
834 Wait for the buffers in reverse order. That way we are
835 less likely to be woken up until all IOs have completed, and
836 so we incur less scheduling load.
839 jbd2_debug(3, "JBD2: commit phase 3\n");
841 while (!list_empty(&io_bufs)) {
842 struct buffer_head *bh = list_entry(io_bufs.prev,
849 if (unlikely(!buffer_uptodate(bh)))
851 jbd2_unfile_log_bh(bh);
852 stats.run.rs_blocks_logged++;
855 * The list contains temporary buffer heads created by
856 * jbd2_journal_write_metadata_buffer().
858 BUFFER_TRACE(bh, "dumping temporary bh");
860 J_ASSERT_BH(bh, atomic_read(&bh->b_count) == 0);
861 free_buffer_head(bh);
863 /* We also have to refile the corresponding shadowed buffer */
864 jh = commit_transaction->t_shadow_list->b_tprev;
866 clear_buffer_jwrite(bh);
867 J_ASSERT_BH(bh, buffer_jbddirty(bh));
868 J_ASSERT_BH(bh, !buffer_shadow(bh));
870 /* The metadata is now released for reuse, but we need
871 to remember it against this transaction so that when
872 we finally commit, we can do any checkpointing
874 JBUFFER_TRACE(jh, "file as BJ_Forget");
875 jbd2_journal_file_buffer(jh, commit_transaction, BJ_Forget);
876 JBUFFER_TRACE(jh, "brelse shadowed buffer");
880 J_ASSERT (commit_transaction->t_shadow_list == NULL);
882 jbd2_debug(3, "JBD2: commit phase 4\n");
884 /* Here we wait for the revoke record and descriptor record buffers */
885 while (!list_empty(&log_bufs)) {
886 struct buffer_head *bh;
888 bh = list_entry(log_bufs.prev, struct buffer_head, b_assoc_buffers);
892 if (unlikely(!buffer_uptodate(bh)))
895 BUFFER_TRACE(bh, "ph5: control buffer writeout done: unfile");
896 clear_buffer_jwrite(bh);
897 jbd2_unfile_log_bh(bh);
898 stats.run.rs_blocks_logged++;
899 __brelse(bh); /* One for getblk */
900 /* AKPM: bforget here */
904 jbd2_journal_abort(journal, err);
906 jbd2_debug(3, "JBD2: commit phase 5\n");
907 write_lock(&journal->j_state_lock);
908 J_ASSERT(commit_transaction->t_state == T_COMMIT_DFLUSH);
909 commit_transaction->t_state = T_COMMIT_JFLUSH;
910 write_unlock(&journal->j_state_lock);
912 if (!jbd2_has_feature_async_commit(journal)) {
913 err = journal_submit_commit_record(journal, commit_transaction,
916 jbd2_journal_abort(journal, err);
919 err = journal_wait_on_commit_record(journal, cbh);
920 stats.run.rs_blocks_logged++;
921 if (jbd2_has_feature_async_commit(journal) &&
922 journal->j_flags & JBD2_BARRIER) {
923 blkdev_issue_flush(journal->j_dev);
927 jbd2_journal_abort(journal, err);
930 atomic_read(&commit_transaction->t_outstanding_credits) < 0);
933 * Now disk caches for filesystem device are flushed so we are safe to
934 * erase checkpointed transactions from the log by updating journal
938 jbd2_update_log_tail(journal, first_tid, first_block);
940 /* End of a transaction! Finally, we can do checkpoint
941 processing: any buffers committed as a result of this
942 transaction can be removed from any checkpoint list it was on
945 jbd2_debug(3, "JBD2: commit phase 6\n");
947 J_ASSERT(list_empty(&commit_transaction->t_inode_list));
948 J_ASSERT(commit_transaction->t_buffers == NULL);
949 J_ASSERT(commit_transaction->t_checkpoint_list == NULL);
950 J_ASSERT(commit_transaction->t_shadow_list == NULL);
954 * As there are other places (journal_unmap_buffer()) adding buffers
955 * to this list we have to be careful and hold the j_list_lock.
957 spin_lock(&journal->j_list_lock);
958 while (commit_transaction->t_forget) {
959 transaction_t *cp_transaction;
960 struct buffer_head *bh;
964 jh = commit_transaction->t_forget;
965 spin_unlock(&journal->j_list_lock);
968 * Get a reference so that bh cannot be freed before we are
972 spin_lock(&jh->b_state_lock);
973 J_ASSERT_JH(jh, jh->b_transaction == commit_transaction);
976 * If there is undo-protected committed data against
977 * this buffer, then we can remove it now. If it is a
978 * buffer needing such protection, the old frozen_data
979 * field now points to a committed version of the
980 * buffer, so rotate that field to the new committed
983 * Otherwise, we can just throw away the frozen data now.
985 * We also know that the frozen data has already fired
986 * its triggers if they exist, so we can clear that too.
988 if (jh->b_committed_data) {
989 jbd2_free(jh->b_committed_data, bh->b_size);
990 jh->b_committed_data = NULL;
991 if (jh->b_frozen_data) {
992 jh->b_committed_data = jh->b_frozen_data;
993 jh->b_frozen_data = NULL;
994 jh->b_frozen_triggers = NULL;
996 } else if (jh->b_frozen_data) {
997 jbd2_free(jh->b_frozen_data, bh->b_size);
998 jh->b_frozen_data = NULL;
999 jh->b_frozen_triggers = NULL;
1002 spin_lock(&journal->j_list_lock);
1003 cp_transaction = jh->b_cp_transaction;
1004 if (cp_transaction) {
1005 JBUFFER_TRACE(jh, "remove from old cp transaction");
1006 cp_transaction->t_chp_stats.cs_dropped++;
1007 __jbd2_journal_remove_checkpoint(jh);
1010 /* Only re-checkpoint the buffer_head if it is marked
1011 * dirty. If the buffer was added to the BJ_Forget list
1012 * by jbd2_journal_forget, it may no longer be dirty and
1013 * there's no point in keeping a checkpoint record for
1017 * A buffer which has been freed while still being journaled
1018 * by a previous transaction, refile the buffer to BJ_Forget of
1019 * the running transaction. If the just committed transaction
1020 * contains "add to orphan" operation, we can completely
1021 * invalidate the buffer now. We are rather through in that
1022 * since the buffer may be still accessible when blocksize <
1023 * pagesize and it is attached to the last partial page.
1025 if (buffer_freed(bh) && !jh->b_next_transaction) {
1026 struct address_space *mapping;
1028 clear_buffer_freed(bh);
1029 clear_buffer_jbddirty(bh);
1032 * Block device buffers need to stay mapped all the
1033 * time, so it is enough to clear buffer_jbddirty and
1034 * buffer_freed bits. For the file mapping buffers (i.e.
1035 * journalled data) we need to unmap buffer and clear
1036 * more bits. We also need to be careful about the check
1037 * because the data page mapping can get cleared under
1038 * our hands. Note that if mapping == NULL, we don't
1039 * need to make buffer unmapped because the page is
1040 * already detached from the mapping and buffers cannot
1043 mapping = READ_ONCE(bh->b_page->mapping);
1044 if (mapping && !sb_is_blkdev_sb(mapping->host->i_sb)) {
1045 clear_buffer_mapped(bh);
1046 clear_buffer_new(bh);
1047 clear_buffer_req(bh);
1052 if (buffer_jbddirty(bh)) {
1053 JBUFFER_TRACE(jh, "add to new checkpointing trans");
1054 __jbd2_journal_insert_checkpoint(jh, commit_transaction);
1055 if (is_journal_aborted(journal))
1056 clear_buffer_jbddirty(bh);
1058 J_ASSERT_BH(bh, !buffer_dirty(bh));
1060 * The buffer on BJ_Forget list and not jbddirty means
1061 * it has been freed by this transaction and hence it
1062 * could not have been reallocated until this
1063 * transaction has committed. *BUT* it could be
1064 * reallocated once we have written all the data to
1065 * disk and before we process the buffer on BJ_Forget
1068 if (!jh->b_next_transaction)
1071 JBUFFER_TRACE(jh, "refile or unfile buffer");
1072 drop_ref = __jbd2_journal_refile_buffer(jh);
1073 spin_unlock(&jh->b_state_lock);
1075 jbd2_journal_put_journal_head(jh);
1077 release_buffer_page(bh); /* Drops bh reference */
1080 cond_resched_lock(&journal->j_list_lock);
1082 spin_unlock(&journal->j_list_lock);
1084 * This is a bit sleazy. We use j_list_lock to protect transition
1085 * of a transaction into T_FINISHED state and calling
1086 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1087 * other checkpointing code processing the transaction...
1089 write_lock(&journal->j_state_lock);
1090 spin_lock(&journal->j_list_lock);
1092 * Now recheck if some buffers did not get attached to the transaction
1093 * while the lock was dropped...
1095 if (commit_transaction->t_forget) {
1096 spin_unlock(&journal->j_list_lock);
1097 write_unlock(&journal->j_state_lock);
1101 /* Add the transaction to the checkpoint list
1102 * __journal_remove_checkpoint() can not destroy transaction
1103 * under us because it is not marked as T_FINISHED yet */
1104 if (journal->j_checkpoint_transactions == NULL) {
1105 journal->j_checkpoint_transactions = commit_transaction;
1106 commit_transaction->t_cpnext = commit_transaction;
1107 commit_transaction->t_cpprev = commit_transaction;
1109 commit_transaction->t_cpnext =
1110 journal->j_checkpoint_transactions;
1111 commit_transaction->t_cpprev =
1112 commit_transaction->t_cpnext->t_cpprev;
1113 commit_transaction->t_cpnext->t_cpprev =
1115 commit_transaction->t_cpprev->t_cpnext =
1118 spin_unlock(&journal->j_list_lock);
1120 /* Done with this transaction! */
1122 jbd2_debug(3, "JBD2: commit phase 7\n");
1124 J_ASSERT(commit_transaction->t_state == T_COMMIT_JFLUSH);
1126 commit_transaction->t_start = jiffies;
1127 stats.run.rs_logging = jbd2_time_diff(stats.run.rs_logging,
1128 commit_transaction->t_start);
1131 * File the transaction statistics
1133 stats.ts_tid = commit_transaction->t_tid;
1134 stats.run.rs_handle_count =
1135 atomic_read(&commit_transaction->t_handle_count);
1136 trace_jbd2_run_stats(journal->j_fs_dev->bd_dev,
1137 commit_transaction->t_tid, &stats.run);
1138 stats.ts_requested = (commit_transaction->t_requested) ? 1 : 0;
1140 commit_transaction->t_state = T_COMMIT_CALLBACK;
1141 J_ASSERT(commit_transaction == journal->j_committing_transaction);
1142 journal->j_commit_sequence = commit_transaction->t_tid;
1143 journal->j_committing_transaction = NULL;
1144 commit_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1147 * weight the commit time higher than the average time so we don't
1148 * react too strongly to vast changes in the commit time
1150 if (likely(journal->j_average_commit_time))
1151 journal->j_average_commit_time = (commit_time +
1152 journal->j_average_commit_time*3) / 4;
1154 journal->j_average_commit_time = commit_time;
1156 write_unlock(&journal->j_state_lock);
1158 if (journal->j_commit_callback)
1159 journal->j_commit_callback(journal, commit_transaction);
1160 if (journal->j_fc_cleanup_callback)
1161 journal->j_fc_cleanup_callback(journal, 1, commit_transaction->t_tid);
1163 trace_jbd2_end_commit(journal, commit_transaction);
1164 jbd2_debug(1, "JBD2: commit %d complete, head %d\n",
1165 journal->j_commit_sequence, journal->j_tail_sequence);
1167 write_lock(&journal->j_state_lock);
1168 journal->j_flags &= ~JBD2_FULL_COMMIT_ONGOING;
1169 journal->j_flags &= ~JBD2_FAST_COMMIT_ONGOING;
1170 spin_lock(&journal->j_list_lock);
1171 commit_transaction->t_state = T_FINISHED;
1172 /* Check if the transaction can be dropped now that we are finished */
1173 if (commit_transaction->t_checkpoint_list == NULL &&
1174 commit_transaction->t_checkpoint_io_list == NULL) {
1175 __jbd2_journal_drop_transaction(journal, commit_transaction);
1176 jbd2_journal_free_transaction(commit_transaction);
1178 spin_unlock(&journal->j_list_lock);
1179 write_unlock(&journal->j_state_lock);
1180 wake_up(&journal->j_wait_done_commit);
1181 wake_up(&journal->j_fc_wait);
1184 * Calculate overall stats
1186 spin_lock(&journal->j_history_lock);
1187 journal->j_stats.ts_tid++;
1188 journal->j_stats.ts_requested += stats.ts_requested;
1189 journal->j_stats.run.rs_wait += stats.run.rs_wait;
1190 journal->j_stats.run.rs_request_delay += stats.run.rs_request_delay;
1191 journal->j_stats.run.rs_running += stats.run.rs_running;
1192 journal->j_stats.run.rs_locked += stats.run.rs_locked;
1193 journal->j_stats.run.rs_flushing += stats.run.rs_flushing;
1194 journal->j_stats.run.rs_logging += stats.run.rs_logging;
1195 journal->j_stats.run.rs_handle_count += stats.run.rs_handle_count;
1196 journal->j_stats.run.rs_blocks += stats.run.rs_blocks;
1197 journal->j_stats.run.rs_blocks_logged += stats.run.rs_blocks_logged;
1198 spin_unlock(&journal->j_history_lock);