| blob | b121d7d434c67510a58c7d962610abc68f3423cc |
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * linux/fs/jbd2/commit.c
4 *
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
6 *
7 * Copyright 1998 Red Hat corp --- All Rights Reserved
8 *
9 * Journal commit routines for the generic filesystem journaling code;
10 * part of the ext2fs journaling system.
11 */
13 #include <linux/time.h>
14 #include <linux/fs.h>
15 #include <linux/jbd2.h>
16 #include <linux/errno.h>
17 #include <linux/slab.h>
18 #include <linux/mm.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>
29 /*
30 * IO end handler for temporary buffer_heads handling writes to the journal.
31 */
33 {
39 else
45 }
47 }
49 /*
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.
56 *
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.
59 *
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.
62 */
64 {
77 /* OK, it's a truncated page */
88 nope:
90 }
93 {
95 __u32 csum;
106 }
108 /*
109 * Done it all: now submit the commit record. We should have
110 * cleaned up our previous buffers by now, so if we are in abort
111 * mode we can now just skip the rest of the journal write
112 * entirely.
113 *
114 * Returns 1 if the journal needs to be aborted or 0 on success
115 */
119 __u32 crc32_sum)
120 {
132 JBD2_COMMIT_BLOCK);
145 }
158 else
163 }
165 /*
166 * This function along with journal_submit_commit_record
167 * allows to write the commit record asynchronously.
168 */
171 {
182 }
184 /*
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().
189 */
191 {
198 };
200 /*
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.
205 */
207 }
209 /* Send all the data buffers related to an inode */
211 {
219 }
223 {
230 }
233 /*
234 * Submit all the data buffers of inode associated with the transaction to
235 * disk.
236 *
237 * We are in a committing transaction. Therefore no new inode can be added to
238 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
239 * operate on from being released while we write out pages.
240 */
243 {
253 /* submit the inode data buffers. */
259 }
265 }
268 }
271 {
277 }
279 /*
280 * Wait for data submitted for writeout, refile inodes to proper
281 * transaction if needed.
282 *
283 */
286 {
290 /* For locking, see the comment in journal_submit_data_buffers() */
297 /* wait for the inode data buffers writeout. */
302 }
307 }
309 /* Now refile inode to proper lists */
322 }
323 }
327 }
330 {
333 __u32 checksum;
341 }
345 {
349 }
353 {
357 __u32 csum32;
358 __be32 seq;
372 else
374 }
375 /*
376 * jbd2_journal_commit_transaction
377 *
378 * The primary function for committing a transaction to the log. This
379 * function is called by the journal thread to begin a complete commit.
380 */
382 {
392 ktime_t start_time;
393 u64 commit_time;
404 /* Tail of the journal */
406 tid_t first_tid;
415 /*
416 * First job: lock down the current transaction and wait for
417 * all outstanding updates to complete.
418 */
420 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
424 /*
425 * We hold j_checkpoint_mutex so tail cannot change under us.
426 * We don't need any special data guarantees for writing sb
427 * since journal is empty and it is ok for write to be
428 * flushed only with transaction commit.
429 */
433 REQ_SYNC);
437 }
448 TASK_UNINTERRUPTIBLE);
453 /*
454 * TODO: by blocking fast commits here, we are increasing
455 * fsync() latency slightly. Strictly speaking, we don't need
456 * to block fast commits until the transaction enters T_FLUSH
457 * state. So an optimization is possible where we block new fast
458 * commits here and wait for existing ones to complete
459 * just before we enter T_FLUSH. That way, the existing fast
460 * commits and this full commit can proceed parallely.
461 */
462 }
492 TASK_UNINTERRUPTIBLE);
499 }
501 }
509 /*
510 * First thing we are allowed to do is to discard any remaining
511 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
512 * that there are no such buffers: if a large filesystem
513 * operation like a truncate needs to split itself over multiple
514 * transactions, then it may try to do a jbd2_journal_restart() while
515 * there are still BJ_Reserved buffers outstanding. These must
516 * be released cleanly from the current transaction.
517 *
518 * In this case, the filesystem must still reserve write access
519 * again before modifying the buffer in the new transaction, but
520 * we do not require it to remember exactly which old buffers it
521 * has reserved. This is consistent with the existing behaviour
522 * that multiple jbd2_journal_get_write_access() calls to the same
523 * buffer are perfectly permissible.
524 */
528 /*
529 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
530 * leave undo-committed data.
531 */
539 }
541 }
543 /*
544 * Now try to drop any written-back buffers from the journal's
545 * checkpoint lists. We do this *before* commit because it potentially
546 * frees some memory
547 */
554 /*
555 * Clear revoked flag to reflect there is no revoked buffers
556 * in the next transaction which is going to be started.
557 */
560 /*
561 * Switch to a new revoke table.
562 */
565 /*
566 * Reserved credits cannot be claimed anymore, free them
567 */
587 /*
588 * Now start flushing things to disk, in the order they appear
589 * on the transaction lists. Data blocks go first.
590 */
600 /*
601 * Way to go: we have now written out all of the data for a
602 * transaction! Now comes the tricky part: we need to write out
603 * metadata. Loop over the transaction's entire buffer list:
604 */
624 /* Find the next buffer to be journaled... */
628 /* If we're in abort mode, we just un-journal the buffer and
629 release it. */
639 /* If that was the last one, we need to clean up
640 * any descriptor buffers which may have been
641 * already allocated, even if we are now
642 * aborting. */
646 }
648 /* Make sure we have a descriptor block in which to
649 record the metadata buffer. */
657 commit_transaction,
658 JBD2_DESCRIPTOR_BLOCK);
662 }
675 /* Record it so that we can wait for IO
676 completion later */
679 }
681 /* Where is the buffer to be written? */
684 /* If the block mapping failed, just abandon the buffer
685 and repeat this loop: we'll fall into the
686 refile-on-abort condition above. */
690 }
692 /*
693 * start_this_handle() uses t_outstanding_credits to determine
694 * the free space in the log.
695 */
698 /* Bump b_count to prevent truncate from stumbling over
699 the shadowed buffer! @@@ This can go if we ever get
700 rid of the shadow pairing of buffers. */
703 /*
704 * Make a temporary IO buffer with which to write it out
705 * (this will requeue the metadata buffer to BJ_Shadow).
706 */
714 }
717 /* Record the new block's tag in the current descriptor
718 buffer */
733 bufs++;
740 }
742 /* If there's no more to do, or if the descriptor is full,
743 let the IO rip! */
751 /* Write an end-of-descriptor marker before
752 submitting the IOs. "tag" still points to
753 the last tag we set up. */
756 start_journal_io:
759 descriptor);
763 /*
764 * Compute checksum.
765 */
767 crc32_sum =
769 }
776 }
779 /* Force a new descriptor to be generated next
780 time round the loop. */
783 }
784 }
789 "JBD2: Detected IO errors while flushing file data "
794 }
796 /*
797 * Get current oldest transaction in the log before we issue flush
798 * to the filesystem device. After the flush we can be sure that
799 * blocks of all older transactions are checkpointed to persistent
800 * storage and we will be safe to update journal start in the
801 * superblock with the numbers we get here.
802 */
803 update_tail =
812 /* Update tail only if we free significant amount of space */
815 }
820 /*
821 * If the journal is not located on the file system device,
822 * then we must flush the file system device before we issue
823 * the commit record
824 */
830 /* Done it all: now write the commit record asynchronously. */
836 }
840 /* Lo and behold: we have just managed to send a transaction to
841 the log. Before we can commit it, wait for the IO so far to
842 complete. Control buffers being written are on the
843 transaction's t_log_list queue, and metadata buffers are on
844 the io_bufs list.
846 Wait for the buffers in reverse order. That way we are
847 less likely to be woken up until all IOs have completed, and
848 so we incur less scheduling load.
849 */
856 b_assoc_buffers);
866 /*
867 * The list contains temporary buffer heads created by
868 * jbd2_journal_write_metadata_buffer().
869 */
875 /* We also have to refile the corresponding shadowed buffer */
882 /* The metadata is now released for reuse, but we need
883 to remember it against this transaction so that when
884 we finally commit, we can do any checkpointing
885 required. */
890 }
896 /* Here we wait for the revoke record and descriptor record buffers */
912 /* AKPM: bforget here */
913 }
929 }
936 }
944 /*
945 * Now disk caches for filesystem device are flushed so we are safe to
946 * erase checkpointed transactions from the log by updating journal
947 * superblock.
948 */
952 /* End of a transaction! Finally, we can do checkpoint
953 processing: any buffers committed as a result of this
954 transaction can be removed from any checkpoint list it was on
955 before. */
964 restart_loop:
965 /*
966 * As there are other places (journal_unmap_buffer()) adding buffers
967 * to this list we have to be careful and hold the j_list_lock.
968 */
979 /*
980 * Get a reference so that bh cannot be freed before we are
981 * done with it.
982 */
987 /*
988 * If there is undo-protected committed data against
989 * this buffer, then we can remove it now. If it is a
990 * buffer needing such protection, the old frozen_data
991 * field now points to a committed version of the
992 * buffer, so rotate that field to the new committed
993 * data.
994 *
995 * Otherwise, we can just throw away the frozen data now.
996 *
997 * We also know that the frozen data has already fired
998 * its triggers if they exist, so we can clear that too.
999 */
1007 }
1012 }
1020 }
1022 /* Only re-checkpoint the buffer_head if it is marked
1023 * dirty. If the buffer was added to the BJ_Forget list
1024 * by jbd2_journal_forget, it may no longer be dirty and
1025 * there's no point in keeping a checkpoint record for
1026 * it. */
1028 /*
1029 * A buffer which has been freed while still being journaled
1030 * by a previous transaction, refile the buffer to BJ_Forget of
1031 * the running transaction. If the just committed transaction
1032 * contains "add to orphan" operation, we can completely
1033 * invalidate the buffer now. We are rather through in that
1034 * since the buffer may be still accessible when blocksize <
1035 * pagesize and it is attached to the last partial page.
1036 */
1043 /*
1044 * Block device buffers need to stay mapped all the
1045 * time, so it is enough to clear buffer_jbddirty and
1046 * buffer_freed bits. For the file mapping buffers (i.e.
1047 * journalled data) we need to unmap buffer and clear
1048 * more bits. We also need to be careful about the check
1049 * because the data page mapping can get cleared under
1050 * our hands. Note that if mapping == NULL, we don't
1051 * need to make buffer unmapped because the page is
1052 * already detached from the mapping and buffers cannot
1053 * get reused.
1054 */
1061 }
1062 }
1071 /*
1072 * The buffer on BJ_Forget list and not jbddirty means
1073 * it has been freed by this transaction and hence it
1074 * could not have been reallocated until this
1075 * transaction has committed. *BUT* it could be
1076 * reallocated once we have written all the data to
1077 * disk and before we process the buffer on BJ_Forget
1078 * list.
1079 */
1082 }
1090 else
1093 }
1095 /*
1096 * This is a bit sleazy. We use j_list_lock to protect transition
1097 * of a transaction into T_FINISHED state and calling
1098 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1099 * other checkpointing code processing the transaction...
1100 */
1103 /*
1104 * Now recheck if some buffers did not get attached to the transaction
1105 * while the lock was dropped...
1106 */
1111 }
1113 /* Add the transaction to the checkpoint list
1114 * __journal_remove_checkpoint() can not destroy transaction
1115 * under us because it is not marked as T_FINISHED yet */
1126 commit_transaction;
1128 commit_transaction;
1129 }
1132 /* Done with this transaction! */
1142 /*
1143 * File the transaction statistics
1144 */
1158 /*
1159 * weight the commit time higher than the average time so we don't
1160 * react too strongly to vast changes in the commit time
1161 */
1165 else
1184 /* Check if the transaction can be dropped now that we are finished */
1189 }
1195 /*
1196 * Calculate overall stats
1197 */
1211 }