| blob | 9153ff3a08e76c8b8fce12c90e25a4acc99139f5 |
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 {
75 /* OK, it's a truncated page */
86 nope:
88 }
91 {
93 __u32 csum;
104 }
106 /*
107 * Done it all: now submit the commit record. We should have
108 * cleaned up our previous buffers by now, so if we are in abort
109 * mode we can now just skip the rest of the journal write
110 * entirely.
111 *
112 * Returns 1 if the journal needs to be aborted or 0 on success
113 */
117 __u32 crc32_sum)
118 {
130 JBD2_COMMIT_BLOCK);
143 }
159 }
161 /*
162 * This function along with journal_submit_commit_record
163 * allows to write the commit record asynchronously.
164 */
167 {
178 }
180 /* Send all the data buffers related to an inode */
182 {
189 }
193 {
200 }
203 /*
204 * Submit all the data buffers of inode associated with the transaction to
205 * disk.
206 *
207 * We are in a committing transaction. Therefore no new inode can be added to
208 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
209 * operate on from being released while we write out pages.
210 */
213 {
223 /* submit the inode data buffers. */
229 }
235 }
238 }
241 {
247 }
249 /*
250 * Wait for data submitted for writeout, refile inodes to proper
251 * transaction if needed.
252 *
253 */
256 {
260 /* For locking, see the comment in journal_submit_data_buffers() */
267 /* wait for the inode data buffers writeout. */
272 }
278 }
280 /* Now refile inode to proper lists */
293 }
294 }
298 }
301 {
303 __u32 checksum;
310 }
314 {
318 }
322 {
325 __u32 csum32;
326 __be32 seq;
339 else
341 }
342 /*
343 * jbd2_journal_commit_transaction
344 *
345 * The primary function for committing a transaction to the log. This
346 * function is called by the journal thread to begin a complete commit.
347 */
349 {
359 ktime_t start_time;
360 u64 commit_time;
371 /* Tail of the journal */
373 tid_t first_tid;
382 /*
383 * First job: lock down the current transaction and wait for
384 * all outstanding updates to complete.
385 */
387 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
391 /*
392 * We hold j_checkpoint_mutex so tail cannot change under us.
393 * We don't need any special data guarantees for writing sb
394 * since journal is empty and it is ok for write to be
395 * flushed only with transaction commit.
396 */
403 }
414 TASK_UNINTERRUPTIBLE);
419 /*
420 * TODO: by blocking fast commits here, we are increasing
421 * fsync() latency slightly. Strictly speaking, we don't need
422 * to block fast commits until the transaction enters T_FLUSH
423 * state. So an optimization is possible where we block new fast
424 * commits here and wait for existing ones to complete
425 * just before we enter T_FLUSH. That way, the existing fast
426 * commits and this full commit can proceed parallely.
427 */
428 }
453 // waits for any t_updates to finish
461 /*
462 * First thing we are allowed to do is to discard any remaining
463 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
464 * that there are no such buffers: if a large filesystem
465 * operation like a truncate needs to split itself over multiple
466 * transactions, then it may try to do a jbd2_journal_restart() while
467 * there are still BJ_Reserved buffers outstanding. These must
468 * be released cleanly from the current transaction.
469 *
470 * In this case, the filesystem must still reserve write access
471 * again before modifying the buffer in the new transaction, but
472 * we do not require it to remember exactly which old buffers it
473 * has reserved. This is consistent with the existing behaviour
474 * that multiple jbd2_journal_get_write_access() calls to the same
475 * buffer are perfectly permissible.
476 * We use journal->j_state_lock here to serialize processing of
477 * t_reserved_list with eviction of buffers from journal_unmap_buffer().
478 */
482 /*
483 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
484 * leave undo-committed data.
485 */
493 }
495 }
498 /*
499 * Now try to drop any written-back buffers from the journal's
500 * checkpoint lists. We do this *before* commit because it potentially
501 * frees some memory
502 */
509 /*
510 * Clear revoked flag to reflect there is no revoked buffers
511 * in the next transaction which is going to be started.
512 */
515 /*
516 * Switch to a new revoke table.
517 */
521 /*
522 * Reserved credits cannot be claimed anymore, free them
523 */
542 /*
543 * Now start flushing things to disk, in the order they appear
544 * on the transaction lists. Data blocks go first.
545 */
555 /*
556 * Way to go: we have now written out all of the data for a
557 * transaction! Now comes the tricky part: we need to write out
558 * metadata. Loop over the transaction's entire buffer list:
559 */
578 /* Find the next buffer to be journaled... */
582 /* If we're in abort mode, we just un-journal the buffer and
583 release it. */
593 /* If that was the last one, we need to clean up
594 * any descriptor buffers which may have been
595 * already allocated, even if we are now
596 * aborting. */
600 }
602 /* Make sure we have a descriptor block in which to
603 record the metadata buffer. */
611 commit_transaction,
612 JBD2_DESCRIPTOR_BLOCK);
616 }
629 /* Record it so that we can wait for IO
630 completion later */
633 }
635 /* Where is the buffer to be written? */
638 /* If the block mapping failed, just abandon the buffer
639 and repeat this loop: we'll fall into the
640 refile-on-abort condition above. */
644 }
646 /*
647 * start_this_handle() uses t_outstanding_credits to determine
648 * the free space in the log.
649 */
652 /* Bump b_count to prevent truncate from stumbling over
653 the shadowed buffer! @@@ This can go if we ever get
654 rid of the shadow pairing of buffers. */
657 /*
658 * Make a temporary IO buffer with which to write it out
659 * (this will requeue the metadata buffer to BJ_Shadow).
660 */
667 /* Record the new block's tag in the current descriptor
668 buffer */
683 bufs++;
690 }
692 /* If there's no more to do, or if the descriptor is full,
693 let the IO rip! */
701 /* Write an end-of-descriptor marker before
702 submitting the IOs. "tag" still points to
703 the last tag we set up. */
706 start_journal_io:
709 descriptor);
714 /*
715 * Compute checksum.
716 */
718 crc32_sum =
720 }
727 bh);
728 }
731 /* Force a new descriptor to be generated next
732 time round the loop. */
735 }
736 }
741 "JBD2: Detected IO errors while flushing file data "
746 }
748 /*
749 * Get current oldest transaction in the log before we issue flush
750 * to the filesystem device. After the flush we can be sure that
751 * blocks of all older transactions are checkpointed to persistent
752 * storage and we will be safe to update journal start in the
753 * superblock with the numbers we get here.
754 */
755 update_tail =
764 /* Update tail only if we free significant amount of space */
767 }
772 /*
773 * If the journal is not located on the file system device,
774 * then we must flush the file system device before we issue
775 * the commit record
776 */
782 /* Done it all: now write the commit record asynchronously. */
788 }
792 /* Lo and behold: we have just managed to send a transaction to
793 the log. Before we can commit it, wait for the IO so far to
794 complete. Control buffers being written are on the
795 transaction's t_log_list queue, and metadata buffers are on
796 the io_bufs list.
798 Wait for the buffers in reverse order. That way we are
799 less likely to be woken up until all IOs have completed, and
800 so we incur less scheduling load.
801 */
808 b_assoc_buffers);
818 /*
819 * The list contains temporary buffer heads created by
820 * jbd2_journal_write_metadata_buffer().
821 */
827 /* We also have to refile the corresponding shadowed buffer */
834 /* The metadata is now released for reuse, but we need
835 to remember it against this transaction so that when
836 we finally commit, we can do any checkpointing
837 required. */
842 }
848 /* Here we wait for the revoke record and descriptor record buffers */
864 /* AKPM: bforget here */
865 }
881 }
888 }
896 /*
897 * Now disk caches for filesystem device are flushed so we are safe to
898 * erase checkpointed transactions from the log by updating journal
899 * superblock.
900 */
904 /* End of a transaction! Finally, we can do checkpoint
905 processing: any buffers committed as a result of this
906 transaction can be removed from any checkpoint list it was on
907 before. */
916 restart_loop:
917 /*
918 * As there are other places (journal_unmap_buffer()) adding buffers
919 * to this list we have to be careful and hold the j_list_lock.
920 */
931 /*
932 * Get a reference so that bh cannot be freed before we are
933 * done with it.
934 */
939 /*
940 * If there is undo-protected committed data against
941 * this buffer, then we can remove it now. If it is a
942 * buffer needing such protection, the old frozen_data
943 * field now points to a committed version of the
944 * buffer, so rotate that field to the new committed
945 * data.
946 *
947 * Otherwise, we can just throw away the frozen data now.
948 *
949 * We also know that the frozen data has already fired
950 * its triggers if they exist, so we can clear that too.
951 */
959 }
964 }
972 }
974 /* Only re-checkpoint the buffer_head if it is marked
975 * dirty. If the buffer was added to the BJ_Forget list
976 * by jbd2_journal_forget, it may no longer be dirty and
977 * there's no point in keeping a checkpoint record for
978 * it. */
980 /*
981 * A buffer which has been freed while still being journaled
982 * by a previous transaction, refile the buffer to BJ_Forget of
983 * the running transaction. If the just committed transaction
984 * contains "add to orphan" operation, we can completely
985 * invalidate the buffer now. We are rather through in that
986 * since the buffer may be still accessible when blocksize <
987 * pagesize and it is attached to the last partial page.
988 */
995 /*
996 * Block device buffers need to stay mapped all the
997 * time, so it is enough to clear buffer_jbddirty and
998 * buffer_freed bits. For the file mapping buffers (i.e.
999 * journalled data) we need to unmap buffer and clear
1000 * more bits. We also need to be careful about the check
1001 * because the data page mapping can get cleared under
1002 * our hands. Note that if mapping == NULL, we don't
1003 * need to make buffer unmapped because the page is
1004 * already detached from the mapping and buffers cannot
1005 * get reused.
1006 */
1013 }
1014 }
1023 /*
1024 * The buffer on BJ_Forget list and not jbddirty means
1025 * it has been freed by this transaction and hence it
1026 * could not have been reallocated until this
1027 * transaction has committed. *BUT* it could be
1028 * reallocated once we have written all the data to
1029 * disk and before we process the buffer on BJ_Forget
1030 * list.
1031 */
1034 }
1042 else
1045 }
1047 /*
1048 * This is a bit sleazy. We use j_list_lock to protect transition
1049 * of a transaction into T_FINISHED state and calling
1050 * __jbd2_journal_drop_transaction(). Otherwise we could race with
1051 * other checkpointing code processing the transaction...
1052 */
1055 /*
1056 * Now recheck if some buffers did not get attached to the transaction
1057 * while the lock was dropped...
1058 */
1063 }
1065 /* Add the transaction to the checkpoint list
1066 * __journal_remove_checkpoint() can not destroy transaction
1067 * under us because it is not marked as T_FINISHED yet */
1078 commit_transaction;
1080 commit_transaction;
1081 }
1084 /* Done with this transaction! */
1094 /*
1095 * File the transaction statistics
1096 */
1110 /*
1111 * weight the commit time higher than the average time so we don't
1112 * react too strongly to vast changes in the commit time
1113 */
1117 else
1136 /* Check if the transaction can be dropped now that we are finished */
1140 }
1146 /*
1147 * Calculate overall stats
1148 */
1162 }