| blob | 36345fefa3ffee1f66e56c908cca50b282abeba5 |
1 /*
2 * linux/fs/jbd2/commit.c
3 *
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1998
5 *
6 * Copyright 1998 Red Hat corp --- All Rights Reserved
7 *
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.
11 *
12 * Journal commit routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
14 */
16 #include <linux/time.h>
17 #include <linux/fs.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/slab.h>
21 #include <linux/mm.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>
32 /*
33 * IO end handler for temporary buffer_heads handling writes to the journal.
34 */
36 {
42 else
48 }
50 }
52 /*
53 * When an ext4 file is truncated, it is possible that some pages are not
54 * successfully freed, because they are attached to a committing transaction.
55 * After the transaction commits, these pages are left on the LRU, with no
56 * ->mapping, and with attached buffers. These pages are trivially reclaimable
57 * by the VM, but their apparent absence upsets the VM accounting, and it makes
58 * the numbers in /proc/meminfo look odd.
59 *
60 * So here, we have a buffer which has just come off the forget list. Look to
61 * see if we can strip all buffers from the backing page.
62 *
63 * Called under lock_journal(), and possibly under journal_datalist_lock. The
64 * caller provided us with a ref against the buffer, and we drop that here.
65 */
67 {
80 /* OK, it's a truncated page */
91 nope:
93 }
96 {
98 __u32 csum;
109 }
111 /*
112 * Done it all: now submit the commit record. We should have
113 * cleaned up our previous buffers by now, so if we are in abort
114 * mode we can now just skip the rest of the journal write
115 * entirely.
116 *
117 * Returns 1 if the journal needs to be aborted or 0 on success
118 */
122 __u32 crc32_sum)
123 {
149 }
161 else
166 }
168 /*
169 * This function along with journal_submit_commit_record
170 * allows to write the commit record asynchronously.
171 */
174 {
185 }
187 /*
188 * write the filemap data using writepage() address_space_operations.
189 * We don't do block allocation here even for delalloc. We don't
190 * use writepages() because with dealyed allocation we may be doing
191 * block allocation in writepages().
192 */
194 {
201 };
205 }
207 /*
208 * Submit all the data buffers of inode associated with the transaction to
209 * disk.
210 *
211 * We are in a committing transaction. Therefore no new inode can be added to
212 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
213 * operate on from being released while we write out pages.
214 */
217 {
227 /*
228 * submit the inode data buffers. We use writepage
229 * instead of writepages. Because writepages can do
230 * block allocation with delalloc. We need to write
231 * only allocated blocks here.
232 */
242 }
245 }
247 /*
248 * Wait for data submitted for writeout, refile inodes to proper
249 * transaction if needed.
250 *
251 */
254 {
258 /* For locking, see the comment in journal_submit_data_buffers() */
265 /*
266 * Because AS_EIO is cleared by
267 * filemap_fdatawait_range(), set it again so
268 * that user process can get -EIO from fsync().
269 */
275 }
280 }
282 /* Now refile inode to proper lists */
293 }
294 }
298 }
301 {
304 __u32 checksum;
312 }
316 {
320 }
324 {
326 __u32 csum;
336 }
340 {
344 __u32 csum32;
345 __be32 seq;
359 else
361 }
362 /*
363 * jbd2_journal_commit_transaction
364 *
365 * The primary function for committing a transaction to the log. This
366 * function is called by the journal thread to begin a complete commit.
367 */
369 {
379 ktime_t start_time;
380 u64 commit_time;
392 /* Tail of the journal */
394 tid_t first_tid;
403 /*
404 * First job: lock down the current transaction and wait for
405 * all outstanding updates to complete.
406 */
408 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
412 /*
413 * We hold j_checkpoint_mutex so tail cannot change under us.
414 * We don't need any special data guarantees for writing sb
415 * since journal is empty and it is ok for write to be
416 * flushed only with transaction commit.
417 */
421 WRITE_SYNC);
425 }
456 TASK_UNINTERRUPTIBLE);
463 }
465 }
471 /*
472 * First thing we are allowed to do is to discard any remaining
473 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
474 * that there are no such buffers: if a large filesystem
475 * operation like a truncate needs to split itself over multiple
476 * transactions, then it may try to do a jbd2_journal_restart() while
477 * there are still BJ_Reserved buffers outstanding. These must
478 * be released cleanly from the current transaction.
479 *
480 * In this case, the filesystem must still reserve write access
481 * again before modifying the buffer in the new transaction, but
482 * we do not require it to remember exactly which old buffers it
483 * has reserved. This is consistent with the existing behaviour
484 * that multiple jbd2_journal_get_write_access() calls to the same
485 * buffer are perfectly permissible.
486 */
490 /*
491 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
492 * leave undo-committed data.
493 */
501 }
503 }
505 /*
506 * Now try to drop any written-back buffers from the journal's
507 * checkpoint lists. We do this *before* commit because it potentially
508 * frees some memory
509 */
516 /*
517 * Clear revoked flag to reflect there is no revoked buffers
518 * in the next transaction which is going to be started.
519 */
522 /*
523 * Switch to a new revoke table.
524 */
527 /*
528 * Reserved credits cannot be claimed anymore, free them
529 */
548 /*
549 * Now start flushing things to disk, in the order they appear
550 * on the transaction lists. Data blocks go first.
551 */
562 /*
563 * Way to go: we have now written out all of the data for a
564 * transaction! Now comes the tricky part: we need to write out
565 * metadata. Loop over the transaction's entire buffer list:
566 */
587 /* Find the next buffer to be journaled... */
591 /* If we're in abort mode, we just un-journal the buffer and
592 release it. */
602 /* If that was the last one, we need to clean up
603 * any descriptor buffers which may have been
604 * already allocated, even if we are now
605 * aborting. */
609 }
611 /* Make sure we have a descriptor block in which to
612 record the metadata buffer. */
623 }
641 /* Record it so that we can wait for IO
642 completion later */
645 }
647 /* Where is the buffer to be written? */
650 /* If the block mapping failed, just abandon the buffer
651 and repeat this loop: we'll fall into the
652 refile-on-abort condition above. */
656 }
658 /*
659 * start_this_handle() uses t_outstanding_credits to determine
660 * the free space in the log, but this counter is changed
661 * by jbd2_journal_next_log_block() also.
662 */
665 /* Bump b_count to prevent truncate from stumbling over
666 the shadowed buffer! @@@ This can go if we ever get
667 rid of the shadow pairing of buffers. */
670 /*
671 * Make a temporary IO buffer with which to write it out
672 * (this will requeue the metadata buffer to BJ_Shadow).
673 */
681 }
684 /* Record the new block's tag in the current descriptor
685 buffer */
700 bufs++;
707 }
709 /* If there's no more to do, or if the descriptor is full,
710 let the IO rip! */
718 /* Write an end-of-descriptor marker before
719 submitting the IOs. "tag" still points to
720 the last tag we set up. */
725 start_journal_io:
728 /*
729 * Compute checksum.
730 */
732 crc32_sum =
734 }
741 }
745 /* Force a new descriptor to be generated next
746 time round the loop. */
749 }
750 }
755 "JBD2: Detected IO errors while flushing file data "
760 }
762 /*
763 * Get current oldest transaction in the log before we issue flush
764 * to the filesystem device. After the flush we can be sure that
765 * blocks of all older transactions are checkpointed to persistent
766 * storage and we will be safe to update journal start in the
767 * superblock with the numbers we get here.
768 */
769 update_tail =
778 /* Update tail only if we free significant amount of space */
781 }
786 /*
787 * If the journal is not located on the file system device,
788 * then we must flush the file system device before we issue
789 * the commit record
790 */
796 /* Done it all: now write the commit record asynchronously. */
802 }
806 /* Lo and behold: we have just managed to send a transaction to
807 the log. Before we can commit it, wait for the IO so far to
808 complete. Control buffers being written are on the
809 transaction's t_log_list queue, and metadata buffers are on
810 the io_bufs list.
812 Wait for the buffers in reverse order. That way we are
813 less likely to be woken up until all IOs have completed, and
814 so we incur less scheduling load.
815 */
822 b_assoc_buffers);
831 /*
832 * The list contains temporary buffer heads created by
833 * jbd2_journal_write_metadata_buffer().
834 */
840 /* We also have to refile the corresponding shadowed buffer */
847 /* The metadata is now released for reuse, but we need
848 to remember it against this transaction so that when
849 we finally commit, we can do any checkpointing
850 required. */
855 }
861 /* Here we wait for the revoke record and descriptor record buffers */
876 /* AKPM: bforget here */
877 }
893 }
899 }
904 /*
905 * Now disk caches for filesystem device are flushed so we are safe to
906 * erase checkpointed transactions from the log by updating journal
907 * superblock.
908 */
912 /* End of a transaction! Finally, we can do checkpoint
913 processing: any buffers committed as a result of this
914 transaction can be removed from any checkpoint list it was on
915 before. */
924 restart_loop:
925 /*
926 * As there are other places (journal_unmap_buffer()) adding buffers
927 * to this list we have to be careful and hold the j_list_lock.
928 */
938 /*
939 * Get a reference so that bh cannot be freed before we are
940 * done with it.
941 */
946 /*
947 * If there is undo-protected committed data against
948 * this buffer, then we can remove it now. If it is a
949 * buffer needing such protection, the old frozen_data
950 * field now points to a committed version of the
951 * buffer, so rotate that field to the new committed
952 * data.
953 *
954 * Otherwise, we can just throw away the frozen data now.
955 *
956 * We also know that the frozen data has already fired
957 * its triggers if they exist, so we can clear that too.
958 */
966 }
971 }
979 }
981 /* Only re-checkpoint the buffer_head if it is marked
982 * dirty. If the buffer was added to the BJ_Forget list
983 * by jbd2_journal_forget, it may no longer be dirty and
984 * there's no point in keeping a checkpoint record for
985 * it. */
987 /*
988 * A buffer which has been freed while still being journaled by
989 * a previous transaction.
990 */
992 /*
993 * If the running transaction is the one containing
994 * "add to orphan" operation (b_next_transaction !=
995 * NULL), we have to wait for that transaction to
996 * commit before we can really get rid of the buffer.
997 * So just clear b_modified to not confuse transaction
998 * credit accounting and refile the buffer to
999 * BJ_Forget of the running transaction. If the just
1000 * committed transaction contains "add to orphan"
1001 * operation, we can completely invalidate the buffer
1002 * now. We are rather through in that since the
1003 * buffer may be still accessible when blocksize <
1004 * pagesize and it is attached to the last partial
1005 * page.
1006 */
1015 }
1016 }
1025 /*
1026 * The buffer on BJ_Forget list and not jbddirty means
1027 * it has been freed by this transaction and hence it
1028 * could not have been reallocated until this
1029 * transaction has committed. *BUT* it could be
1030 * reallocated once we have written all the data to
1031 * disk and before we process the buffer on BJ_Forget
1032 * list.
1033 */
1036 }
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
1132 /* Check if the transaction can be dropped now that we are finished */
1137 }
1142 /*
1143 * Calculate overall stats
1144 */
1158 }