| blob | eef6979821a4c8db91f48854c46f81c291ef4c27 |
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>
31 #include <asm/system.h>
33 /*
34 * Default IO end handler for temporary BJ_IO buffer_heads.
35 */
37 {
41 else
44 }
46 /*
47 * When an ext4 file is truncated, it is possible that some pages are not
48 * successfully freed, because they are attached to a committing transaction.
49 * After the transaction commits, these pages are left on the LRU, with no
50 * ->mapping, and with attached buffers. These pages are trivially reclaimable
51 * by the VM, but their apparent absence upsets the VM accounting, and it makes
52 * the numbers in /proc/meminfo look odd.
53 *
54 * So here, we have a buffer which has just come off the forget list. Look to
55 * see if we can strip all buffers from the backing page.
56 *
57 * Called under lock_journal(), and possibly under journal_datalist_lock. The
58 * caller provided us with a ref against the buffer, and we drop that here.
59 */
61 {
74 /* OK, it's a truncated page */
85 nope:
87 }
89 /*
90 * Done it all: now submit the commit record. We should have
91 * cleaned up our previous buffers by now, so if we are in abort
92 * mode we can now just skip the rest of the journal write
93 * entirely.
94 *
95 * Returns 1 if the journal needs to be aborted or 0 on success
96 */
100 __u32 crc32_sum)
101 {
127 JBD2_FEATURE_COMPAT_CHECKSUM)) {
131 }
141 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT))
143 else
148 }
150 /*
151 * This function along with journal_submit_commit_record
152 * allows to write the commit record asynchronously.
153 */
156 {
168 }
170 /*
171 * write the filemap data using writepage() address_space_operations.
172 * We don't do block allocation here even for delalloc. We don't
173 * use writepages() because with dealyed allocation we may be doing
174 * block allocation in writepages().
175 */
177 {
184 };
188 }
190 /*
191 * Submit all the data buffers of inode associated with the transaction to
192 * disk.
193 *
194 * We are in a committing transaction. Therefore no new inode can be added to
195 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
196 * operate on from being released while we write out pages.
197 */
200 {
210 /*
211 * submit the inode data buffers. We use writepage
212 * instead of writepages. Because writepages can do
213 * block allocation with delalloc. We need to write
214 * only allocated blocks here.
215 */
225 }
228 }
230 /*
231 * Wait for data submitted for writeout, refile inodes to proper
232 * transaction if needed.
233 *
234 */
237 {
241 /* For locking, see the comment in journal_submit_data_buffers() */
248 /*
249 * Because AS_EIO is cleared by
250 * filemap_fdatawait_range(), set it again so
251 * that user process can get -EIO from fsync().
252 */
258 }
263 }
265 /* Now refile inode to proper lists */
276 }
277 }
281 }
284 {
287 __u32 checksum;
295 }
299 {
303 }
305 /*
306 * jbd2_journal_commit_transaction
307 *
308 * The primary function for committing a transaction to the log. This
309 * function is called by the journal thread to begin a complete commit.
310 */
312 {
321 ktime_t start_time;
322 u64 commit_time;
335 /*
336 * First job: lock down the current transaction and wait for
337 * all outstanding updates to complete.
338 */
340 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
346 }
372 TASK_UNINTERRUPTIBLE);
379 }
381 }
387 /*
388 * First thing we are allowed to do is to discard any remaining
389 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
390 * that there are no such buffers: if a large filesystem
391 * operation like a truncate needs to split itself over multiple
392 * transactions, then it may try to do a jbd2_journal_restart() while
393 * there are still BJ_Reserved buffers outstanding. These must
394 * be released cleanly from the current transaction.
395 *
396 * In this case, the filesystem must still reserve write access
397 * again before modifying the buffer in the new transaction, but
398 * we do not require it to remember exactly which old buffers it
399 * has reserved. This is consistent with the existing behaviour
400 * that multiple jbd2_journal_get_write_access() calls to the same
401 * buffer are perfectly permissible.
402 */
406 /*
407 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
408 * leave undo-committed data.
409 */
417 }
419 }
421 /*
422 * Now try to drop any written-back buffers from the journal's
423 * checkpoint lists. We do this *before* commit because it potentially
424 * frees some memory
425 */
432 /*
433 * Switch to a new revoke table.
434 */
452 /*
453 * Now start flushing things to disk, in the order they appear
454 * on the transaction lists. Data blocks go first.
455 */
462 WRITE_SYNC);
467 /*
468 * Way to go: we have now written out all of the data for a
469 * transaction! Now comes the tricky part: we need to write out
470 * metadata. Loop over the transaction's entire buffer list:
471 */
493 /* Find the next buffer to be journaled... */
497 /* If we're in abort mode, we just un-journal the buffer and
498 release it. */
508 /* If that was the last one, we need to clean up
509 * any descriptor buffers which may have been
510 * already allocated, even if we are now
511 * aborting. */
515 }
517 /* Make sure we have a descriptor block in which to
518 record the metadata buffer. */
531 }
548 /* Record it so that we can wait for IO
549 completion later */
552 BJ_LogCtl);
553 }
555 /* Where is the buffer to be written? */
558 /* If the block mapping failed, just abandon the buffer
559 and repeat this loop: we'll fall into the
560 refile-on-abort condition above. */
564 }
566 /*
567 * start_this_handle() uses t_outstanding_credits to determine
568 * the free space in the log, but this counter is changed
569 * by jbd2_journal_next_log_block() also.
570 */
573 /* Bump b_count to prevent truncate from stumbling over
574 the shadowed buffer! @@@ This can go if we ever get
575 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
578 /* Make a temporary IO buffer with which to write it out
579 (this will requeue both the metadata buffer and the
580 temporary IO buffer). new_bh goes on BJ_IO*/
583 /*
584 * akpm: jbd2_journal_write_metadata_buffer() sets
585 * new_bh->b_transaction to commit_transaction.
586 * We need to clean this up before we release new_bh
587 * (which is of type BJ_IO)
588 */
595 }
599 /* Record the new block's tag in the current descriptor
600 buffer */
619 }
621 /* If there's no more to do, or if the descriptor is full,
622 let the IO rip! */
630 /* Write an end-of-descriptor marker before
631 submitting the IOs. "tag" still points to
632 the last tag we set up. */
636 start_journal_io:
639 /*
640 * Compute checksum.
641 */
643 JBD2_FEATURE_COMPAT_CHECKSUM)) {
644 crc32_sum =
646 }
653 }
657 /* Force a new descriptor to be generated next
658 time round the loop. */
661 }
662 }
667 "JBD2: Detected IO errors while flushing file data "
672 }
678 /*
679 * If the journal is not located on the file system device,
680 * then we must flush the file system device before we issue
681 * the commit record
682 */
688 /* Done it all: now write the commit record asynchronously. */
690 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
695 }
699 /* Lo and behold: we have just managed to send a transaction to
700 the log. Before we can commit it, wait for the IO so far to
701 complete. Control buffers being written are on the
702 transaction's t_log_list queue, and metadata buffers are on
703 the t_iobuf_list queue.
705 Wait for the buffers in reverse order. That way we are
706 less likely to be woken up until all IOs have completed, and
707 so we incur less scheduling load.
708 */
712 /*
713 * akpm: these are BJ_IO, and j_list_lock is not needed.
714 * See __journal_try_to_free_buffer.
715 */
716 wait_for_iobuf:
725 }
737 /*
738 * ->t_iobuf_list should contain only dummy buffer_heads
739 * which were created by jbd2_journal_write_metadata_buffer().
740 */
747 /* We also have to unlock and free the corresponding
748 shadowed buffer */
754 /* The metadata is now released for reuse, but we need
755 to remember it against this transaction so that when
756 we finally commit, we can do any checkpointing
757 required. */
760 /*
761 * Wake up any transactions which were waiting for this IO to
762 * complete. The barrier must be here so that changes by
763 * jbd2_journal_file_buffer() take effect before wake_up_bit()
764 * does the waitqueue check.
765 */
770 }
776 /* Here we wait for the revoke record and descriptor record buffers */
777 wait_for_ctlbuf:
786 }
798 /* AKPM: bforget here */
799 }
811 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
816 }
820 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT) &&
823 }
828 /* End of a transaction! Finally, we can do checkpoint
829 processing: any buffers committed as a result of this
830 transaction can be removed from any checkpoint list it was on
831 before. */
842 restart_loop:
843 /*
844 * As there are other places (journal_unmap_buffer()) adding buffers
845 * to this list we have to be careful and hold the j_list_lock.
846 */
856 /*
857 * Get a reference so that bh cannot be freed before we are
858 * done with it.
859 */
864 /*
865 * If there is undo-protected committed data against
866 * this buffer, then we can remove it now. If it is a
867 * buffer needing such protection, the old frozen_data
868 * field now points to a committed version of the
869 * buffer, so rotate that field to the new committed
870 * data.
871 *
872 * Otherwise, we can just throw away the frozen data now.
873 *
874 * We also know that the frozen data has already fired
875 * its triggers if they exist, so we can clear that too.
876 */
884 }
889 }
897 }
899 /* Only re-checkpoint the buffer_head if it is marked
900 * dirty. If the buffer was added to the BJ_Forget list
901 * by jbd2_journal_forget, it may no longer be dirty and
902 * there's no point in keeping a checkpoint record for
903 * it. */
905 /* A buffer which has been freed while still being
906 * journaled by a previous transaction may end up still
907 * being dirty here, but we want to avoid writing back
908 * that buffer in the future after the "add to orphan"
909 * operation been committed, That's not only a performance
910 * gain, it also stops aliasing problems if the buffer is
911 * left behind for writeback and gets reallocated for another
912 * use in a different page. */
916 }
925 /*
926 * The buffer on BJ_Forget list and not jbddirty means
927 * it has been freed by this transaction and hence it
928 * could not have been reallocated until this
929 * transaction has committed. *BUT* it could be
930 * reallocated once we have written all the data to
931 * disk and before we process the buffer on BJ_Forget
932 * list.
933 */
936 }
942 else
945 }
947 /*
948 * This is a bit sleazy. We use j_list_lock to protect transition
949 * of a transaction into T_FINISHED state and calling
950 * __jbd2_journal_drop_transaction(). Otherwise we could race with
951 * other checkpointing code processing the transaction...
952 */
955 /*
956 * Now recheck if some buffers did not get attached to the transaction
957 * while the lock was dropped...
958 */
963 }
965 /* Done with this transaction! */
975 /*
976 * File the transaction statistics
977 */
984 /*
985 * Calculate overall stats
986 */
1005 /*
1006 * weight the commit time higher than the average time so we don't
1007 * react too strongly to vast changes in the commit time
1008 */
1012 else
1031 commit_transaction;
1033 commit_transaction;
1034 }
1035 }
1048 }