| blob | c321fa06081ce6cadafa253bc8257f5bf0da8497 |
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 dealyed allocation we may be doing
188 * block allocation in writepages().
189 */
192 {
199 };
203 }
205 /*
206 * Submit all the data buffers of inode associated with the transaction to
207 * disk.
208 *
209 * We are in a committing transaction. Therefore no new inode can be added to
210 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
211 * operate on from being released while we write out pages.
212 */
215 {
230 /*
231 * submit the inode data buffers. We use writepage
232 * instead of writepages. Because writepages can do
233 * block allocation with delalloc. We need to write
234 * only allocated blocks here.
235 */
238 dirty_end);
246 }
249 }
251 /*
252 * Wait for data submitted for writeout, refile inodes to proper
253 * transaction if needed.
254 *
255 */
258 {
262 /* For locking, see the comment in journal_submit_data_buffers() */
274 dirty_end);
281 }
283 /* Now refile inode to proper lists */
296 }
297 }
301 }
304 {
307 __u32 checksum;
315 }
319 {
323 }
327 {
331 __u32 csum32;
332 __be32 seq;
346 else
348 }
349 /*
350 * jbd2_journal_commit_transaction
351 *
352 * The primary function for committing a transaction to the log. This
353 * function is called by the journal thread to begin a complete commit.
354 */
356 {
366 ktime_t start_time;
367 u64 commit_time;
378 /* Tail of the journal */
380 tid_t first_tid;
389 /*
390 * First job: lock down the current transaction and wait for
391 * all outstanding updates to complete.
392 */
394 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
398 /*
399 * We hold j_checkpoint_mutex so tail cannot change under us.
400 * We don't need any special data guarantees for writing sb
401 * since journal is empty and it is ok for write to be
402 * flushed only with transaction commit.
403 */
407 REQ_SYNC);
411 }
442 TASK_UNINTERRUPTIBLE);
449 }
451 }
457 /*
458 * First thing we are allowed to do is to discard any remaining
459 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
460 * that there are no such buffers: if a large filesystem
461 * operation like a truncate needs to split itself over multiple
462 * transactions, then it may try to do a jbd2_journal_restart() while
463 * there are still BJ_Reserved buffers outstanding. These must
464 * be released cleanly from the current transaction.
465 *
466 * In this case, the filesystem must still reserve write access
467 * again before modifying the buffer in the new transaction, but
468 * we do not require it to remember exactly which old buffers it
469 * has reserved. This is consistent with the existing behaviour
470 * that multiple jbd2_journal_get_write_access() calls to the same
471 * buffer are perfectly permissible.
472 */
476 /*
477 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
478 * leave undo-committed data.
479 */
487 }
489 }
491 /*
492 * Now try to drop any written-back buffers from the journal's
493 * checkpoint lists. We do this *before* commit because it potentially
494 * frees some memory
495 */
502 /*
503 * Clear revoked flag to reflect there is no revoked buffers
504 * in the next transaction which is going to be started.
505 */
508 /*
509 * Switch to a new revoke table.
510 */
513 /*
514 * Reserved credits cannot be claimed anymore, free them
515 */
534 /*
535 * Now start flushing things to disk, in the order they appear
536 * on the transaction lists. Data blocks go first.
537 */
547 /*
548 * Way to go: we have now written out all of the data for a
549 * transaction! Now comes the tricky part: we need to write out
550 * metadata. Loop over the transaction's entire buffer list:
551 */
572 /* Find the next buffer to be journaled... */
576 /* If we're in abort mode, we just un-journal the buffer and
577 release it. */
587 /* If that was the last one, we need to clean up
588 * any descriptor buffers which may have been
589 * already allocated, even if we are now
590 * aborting. */
594 }
596 /* Make sure we have a descriptor block in which to
597 record the metadata buffer. */
605 commit_transaction,
606 JBD2_DESCRIPTOR_BLOCK);
610 }
623 /* Record it so that we can wait for IO
624 completion later */
627 }
629 /* Where is the buffer to be written? */
632 /* If the block mapping failed, just abandon the buffer
633 and repeat this loop: we'll fall into the
634 refile-on-abort condition above. */
638 }
640 /*
641 * start_this_handle() uses t_outstanding_credits to determine
642 * the free space in the log, but this counter is changed
643 * by jbd2_journal_next_log_block() also.
644 */
647 /* Bump b_count to prevent truncate from stumbling over
648 the shadowed buffer! @@@ This can go if we ever get
649 rid of the shadow pairing of buffers. */
652 /*
653 * Make a temporary IO buffer with which to write it out
654 * (this will requeue the metadata buffer to BJ_Shadow).
655 */
663 }
666 /* Record the new block's tag in the current descriptor
667 buffer */
682 bufs++;
689 }
691 /* If there's no more to do, or if the descriptor is full,
692 let the IO rip! */
700 /* Write an end-of-descriptor marker before
701 submitting the IOs. "tag" still points to
702 the last tag we set up. */
705 start_journal_io:
708 descriptor);
712 /*
713 * Compute checksum.
714 */
716 crc32_sum =
718 }
725 }
728 /* Force a new descriptor to be generated next
729 time round the loop. */
732 }
733 }
738 "JBD2: Detected IO errors while flushing file data "
743 }
745 /*
746 * Get current oldest transaction in the log before we issue flush
747 * to the filesystem device. After the flush we can be sure that
748 * blocks of all older transactions are checkpointed to persistent
749 * storage and we will be safe to update journal start in the
750 * superblock with the numbers we get here.
751 */
752 update_tail =
761 /* Update tail only if we free significant amount of space */
764 }
769 /*
770 * If the journal is not located on the file system device,
771 * then we must flush the file system device before we issue
772 * the commit record
773 */
779 /* Done it all: now write the commit record asynchronously. */
785 }
789 /* Lo and behold: we have just managed to send a transaction to
790 the log. Before we can commit it, wait for the IO so far to
791 complete. Control buffers being written are on the
792 transaction's t_log_list queue, and metadata buffers are on
793 the io_bufs list.
795 Wait for the buffers in reverse order. That way we are
796 less likely to be woken up until all IOs have completed, and
797 so we incur less scheduling load.
798 */
805 b_assoc_buffers);
815 /*
816 * The list contains temporary buffer heads created by
817 * jbd2_journal_write_metadata_buffer().
818 */
824 /* We also have to refile the corresponding shadowed buffer */
831 /* The metadata is now released for reuse, but we need
832 to remember it against this transaction so that when
833 we finally commit, we can do any checkpointing
834 required. */
839 }
845 /* Here we wait for the revoke record and descriptor record buffers */
861 /* AKPM: bforget here */
862 }
878 }
885 }
890 /*
891 * Now disk caches for filesystem device are flushed so we are safe to
892 * erase checkpointed transactions from the log by updating journal
893 * superblock.
894 */
898 /* End of a transaction! Finally, we can do checkpoint
899 processing: any buffers committed as a result of this
900 transaction can be removed from any checkpoint list it was on
901 before. */
910 restart_loop:
911 /*
912 * As there are other places (journal_unmap_buffer()) adding buffers
913 * to this list we have to be careful and hold the j_list_lock.
914 */
924 /*
925 * Get a reference so that bh cannot be freed before we are
926 * done with it.
927 */
932 /*
933 * If there is undo-protected committed data against
934 * this buffer, then we can remove it now. If it is a
935 * buffer needing such protection, the old frozen_data
936 * field now points to a committed version of the
937 * buffer, so rotate that field to the new committed
938 * data.
939 *
940 * Otherwise, we can just throw away the frozen data now.
941 *
942 * We also know that the frozen data has already fired
943 * its triggers if they exist, so we can clear that too.
944 */
952 }
957 }
965 }
967 /* Only re-checkpoint the buffer_head if it is marked
968 * dirty. If the buffer was added to the BJ_Forget list
969 * by jbd2_journal_forget, it may no longer be dirty and
970 * there's no point in keeping a checkpoint record for
971 * it. */
973 /*
974 * A buffer which has been freed while still being journaled
975 * by a previous transaction, refile the buffer to BJ_Forget of
976 * the running transaction. If the just committed transaction
977 * contains "add to orphan" operation, we can completely
978 * invalidate the buffer now. We are rather through in that
979 * since the buffer may be still accessible when blocksize <
980 * pagesize and it is attached to the last partial page.
981 */
988 /*
989 * Block device buffers need to stay mapped all the
990 * time, so it is enough to clear buffer_jbddirty and
991 * buffer_freed bits. For the file mapping buffers (i.e.
992 * journalled data) we need to unmap buffer and clear
993 * more bits. We also need to be careful about the check
994 * because the data page mapping can get cleared under
995 * out hands, which alse need not to clear more bits
996 * because the page and buffers will be freed and can
997 * never be reused once we are done with them.
998 */
1005 }
1006 }
1015 /*
1016 * The buffer on BJ_Forget list and not jbddirty means
1017 * it has been freed by this transaction and hence it
1018 * could not have been reallocated until this
1019 * transaction has committed. *BUT* it could be
1020 * reallocated once we have written all the data to
1021 * disk and before we process the buffer on BJ_Forget
1022 * list.
1023 */
1026 }
1032 else
1035 }
1037 /*
1038 * This is a bit sleazy. We use j_list_lock to protect transition
1039 * of a transaction into T_FINISHED state and calling
1040 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1041 * other checkpointing code processing the transaction...
1042 */
1045 /*
1046 * Now recheck if some buffers did not get attached to the transaction
1047 * while the lock was dropped...
1048 */
1053 }
1055 /* Add the transaction to the checkpoint list
1056 * __journal_remove_checkpoint() can not destroy transaction
1057 * under us because it is not marked as T_FINISHED yet */
1068 commit_transaction;
1070 commit_transaction;
1071 }
1074 /* Done with this transaction! */
1084 /*
1085 * File the transaction statistics
1086 */
1100 /*
1101 * weight the commit time higher than the average time so we don't
1102 * react too strongly to vast changes in the commit time
1103 */
1107 else
1122 /* Check if the transaction can be dropped now that we are finished */
1127 }
1132 /*
1133 * Calculate overall stats
1134 */
1148 }