| blob | d4cfd6d2779e07fe1d5381975840ccb3e3486566 |
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 <trace/events/jbd2.h>
31 /*
32 * Default IO end handler for temporary BJ_IO buffer_heads.
33 */
35 {
39 else
42 }
44 /*
45 * When an ext4 file is truncated, it is possible that some pages are not
46 * successfully freed, because they are attached to a committing transaction.
47 * After the transaction commits, these pages are left on the LRU, with no
48 * ->mapping, and with attached buffers. These pages are trivially reclaimable
49 * by the VM, but their apparent absence upsets the VM accounting, and it makes
50 * the numbers in /proc/meminfo look odd.
51 *
52 * So here, we have a buffer which has just come off the forget list. Look to
53 * see if we can strip all buffers from the backing page.
54 *
55 * Called under lock_journal(), and possibly under journal_datalist_lock. The
56 * caller provided us with a ref against the buffer, and we drop that here.
57 */
59 {
72 /* OK, it's a truncated page */
83 nope:
85 }
87 /*
88 * Done it all: now submit the commit record. We should have
89 * cleaned up our previous buffers by now, so if we are in abort
90 * mode we can now just skip the rest of the journal write
91 * entirely.
92 *
93 * Returns 1 if the journal needs to be aborted or 0 on success
94 */
98 __u32 crc32_sum)
99 {
124 JBD2_FEATURE_COMPAT_CHECKSUM)) {
128 }
138 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
141 }
146 /* is it possible for another commit to fail at roughly
147 * the same time as this one? If so, we don't want to
148 * trust the barrier flag in the super, but instead want
149 * to remember if we sent a barrier request
150 */
153 "JBD: barrier-based sync failed on %s - "
159 /* And try again, without the barrier */
164 }
167 }
169 /*
170 * This function along with journal_submit_commit_record
171 * allows to write the commit record asynchronously.
172 */
175 {
178 retry:
183 "JBD2: wait_on_commit_record: sync failed on %s - "
198 }
200 }
208 }
210 /*
211 * write the filemap data using writepage() address_space_operations.
212 * We don't do block allocation here even for delalloc. We don't
213 * use writepages() because with dealyed allocation we may be doing
214 * block allocation in writepages().
215 */
217 {
224 };
228 }
230 /*
231 * Submit all the data buffers of inode associated with the transaction to
232 * disk.
233 *
234 * We are in a committing transaction. Therefore no new inode can be added to
235 * our inode list. We use JI_COMMIT_RUNNING flag to protect inode we currently
236 * operate on from being released while we write out pages.
237 */
240 {
250 /*
251 * submit the inode data buffers. We use writepage
252 * instead of writepages. Because writepages can do
253 * block allocation with delalloc. We need to write
254 * only allocated blocks here.
255 */
264 }
267 }
269 /*
270 * Wait for data submitted for writeout, refile inodes to proper
271 * transaction if needed.
272 *
273 */
276 {
280 /* For locking, see the comment in journal_submit_data_buffers() */
287 /*
288 * Because AS_EIO is cleared by
289 * wait_on_page_writeback_range(), set it again so
290 * that user process can get -EIO from fsync().
291 */
297 }
301 }
303 /* Now refile inode to proper lists */
314 }
315 }
319 }
322 {
325 __u32 checksum;
333 }
337 {
341 }
343 /*
344 * jbd2_journal_commit_transaction
345 *
346 * The primary function for committing a transaction to the log. This
347 * function is called by the journal thread to begin a complete commit.
348 */
350 {
359 ktime_t start_time;
360 u64 commit_time;
373 /*
374 * First job: lock down the current transaction and wait for
375 * all outstanding updates to complete.
376 */
378 #ifdef COMMIT_STATS
382 #endif
384 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
390 }
405 /*
406 * Use plugged writes here, since we want to submit several before
407 * we unplug the device. We don't do explicit unplugging in here,
408 * instead we rely on sync_buffer() doing the unplug for us.
409 */
423 TASK_UNINTERRUPTIBLE);
430 }
432 }
438 /*
439 * First thing we are allowed to do is to discard any remaining
440 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
441 * that there are no such buffers: if a large filesystem
442 * operation like a truncate needs to split itself over multiple
443 * transactions, then it may try to do a jbd2_journal_restart() while
444 * there are still BJ_Reserved buffers outstanding. These must
445 * be released cleanly from the current transaction.
446 *
447 * In this case, the filesystem must still reserve write access
448 * again before modifying the buffer in the new transaction, but
449 * we do not require it to remember exactly which old buffers it
450 * has reserved. This is consistent with the existing behaviour
451 * that multiple jbd2_journal_get_write_access() calls to the same
452 * buffer are perfectly permissable.
453 */
457 /*
458 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
459 * leave undo-committed data.
460 */
468 }
470 }
472 /*
473 * Now try to drop any written-back buffers from the journal's
474 * checkpoint lists. We do this *before* commit because it potentially
475 * frees some memory
476 */
483 /*
484 * Switch to a new revoke table.
485 */
503 /*
504 * Now start flushing things to disk, in the order they appear
505 * on the transaction lists. Data blocks go first.
506 */
512 write_op);
516 /*
517 * Way to go: we have now written out all of the data for a
518 * transaction! Now comes the tricky part: we need to write out
519 * metadata. Loop over the transaction's entire buffer list:
520 */
540 /* Find the next buffer to be journaled... */
544 /* If we're in abort mode, we just un-journal the buffer and
545 release it. */
555 /* If that was the last one, we need to clean up
556 * any descriptor buffers which may have been
557 * already allocated, even if we are now
558 * aborting. */
562 }
564 /* Make sure we have a descriptor block in which to
565 record the metadata buffer. */
578 }
595 /* Record it so that we can wait for IO
596 completion later */
599 BJ_LogCtl);
600 }
602 /* Where is the buffer to be written? */
605 /* If the block mapping failed, just abandon the buffer
606 and repeat this loop: we'll fall into the
607 refile-on-abort condition above. */
611 }
613 /*
614 * start_this_handle() uses t_outstanding_credits to determine
615 * the free space in the log, but this counter is changed
616 * by jbd2_journal_next_log_block() also.
617 */
620 /* Bump b_count to prevent truncate from stumbling over
621 the shadowed buffer! @@@ This can go if we ever get
622 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
625 /* Make a temporary IO buffer with which to write it out
626 (this will requeue both the metadata buffer and the
627 temporary IO buffer). new_bh goes on BJ_IO*/
630 /*
631 * akpm: jbd2_journal_write_metadata_buffer() sets
632 * new_bh->b_transaction to commit_transaction.
633 * We need to clean this up before we release new_bh
634 * (which is of type BJ_IO)
635 */
642 /* Record the new block's tag in the current descriptor
643 buffer */
662 }
664 /* If there's no more to do, or if the descriptor is full,
665 let the IO rip! */
673 /* Write an end-of-descriptor marker before
674 submitting the IOs. "tag" still points to
675 the last tag we set up. */
679 start_journal_io:
682 /*
683 * Compute checksum.
684 */
686 JBD2_FEATURE_COMPAT_CHECKSUM)) {
687 crc32_sum =
689 }
696 }
700 /* Force a new descriptor to be generated next
701 time round the loop. */
704 }
705 }
707 /* Done it all: now write the commit record asynchronously. */
710 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
717 }
719 /*
720 * This is the right place to wait for data buffers both for ASYNC
721 * and !ASYNC commit. If commit is ASYNC, we need to wait only after
722 * the commit block went to disk (which happens above). If commit is
723 * SYNC, we need to wait for data buffers before we start writing
724 * commit block, which happens below in such setting.
725 */
729 "JBD2: Detected IO errors while flushing file data "
734 }
736 /* Lo and behold: we have just managed to send a transaction to
737 the log. Before we can commit it, wait for the IO so far to
738 complete. Control buffers being written are on the
739 transaction's t_log_list queue, and metadata buffers are on
740 the t_iobuf_list queue.
742 Wait for the buffers in reverse order. That way we are
743 less likely to be woken up until all IOs have completed, and
744 so we incur less scheduling load.
745 */
749 /*
750 * akpm: these are BJ_IO, and j_list_lock is not needed.
751 * See __journal_try_to_free_buffer.
752 */
753 wait_for_iobuf:
762 }
774 /*
775 * ->t_iobuf_list should contain only dummy buffer_heads
776 * which were created by jbd2_journal_write_metadata_buffer().
777 */
784 /* We also have to unlock and free the corresponding
785 shadowed buffer */
791 /* The metadata is now released for reuse, but we need
792 to remember it against this transaction so that when
793 we finally commit, we can do any checkpointing
794 required. */
797 /* Wake up any transactions which were waiting for this
798 IO to complete */
802 }
808 /* Here we wait for the revoke record and descriptor record buffers */
809 wait_for_ctlbuf:
818 }
830 /* AKPM: bforget here */
831 }
839 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
844 }
851 /* End of a transaction! Finally, we can do checkpoint
852 processing: any buffers committed as a result of this
853 transaction can be removed from any checkpoint list it was on
854 before. */
865 restart_loop:
866 /*
867 * As there are other places (journal_unmap_buffer()) adding buffers
868 * to this list we have to be careful and hold the j_list_lock.
869 */
882 /*
883 * If there is undo-protected committed data against
884 * this buffer, then we can remove it now. If it is a
885 * buffer needing such protection, the old frozen_data
886 * field now points to a committed version of the
887 * buffer, so rotate that field to the new committed
888 * data.
889 *
890 * Otherwise, we can just throw away the frozen data now.
891 *
892 * We also know that the frozen data has already fired
893 * its triggers if they exist, so we can clear that too.
894 */
902 }
907 }
915 }
917 /* Only re-checkpoint the buffer_head if it is marked
918 * dirty. If the buffer was added to the BJ_Forget list
919 * by jbd2_journal_forget, it may no longer be dirty and
920 * there's no point in keeping a checkpoint record for
921 * it. */
923 /* A buffer which has been freed while still being
924 * journaled by a previous transaction may end up still
925 * being dirty here, but we want to avoid writing back
926 * that buffer in the future now that the last use has
927 * been committed. That's not only a performance gain,
928 * it also stops aliasing problems if the buffer is left
929 * behind for writeback and gets reallocated for another
930 * use in a different page. */
934 }
946 /* The buffer on BJ_Forget list and not jbddirty means
947 * it has been freed by this transaction and hence it
948 * could not have been reallocated until this
949 * transaction has committed. *BUT* it could be
950 * reallocated once we have written all the data to
951 * disk and before we process the buffer on BJ_Forget
952 * list. */
957 /* needs a brelse */
962 }
964 }
966 /*
967 * This is a bit sleazy. We use j_list_lock to protect transition
968 * of a transaction into T_FINISHED state and calling
969 * __jbd2_journal_drop_transaction(). Otherwise we could race with
970 * other checkpointing code processing the transaction...
971 */
974 /*
975 * Now recheck if some buffers did not get attached to the transaction
976 * while the lock was dropped...
977 */
982 }
984 /* Done with this transaction! */
994 /*
995 * File the transaction statistics
996 */
1002 /*
1003 * Calculate overall stats
1004 */
1023 /*
1024 * weight the commit time higher than the average time so we don't
1025 * react too strongly to vast changes in the commit time
1026 */
1030 else
1049 commit_transaction;
1051 commit_transaction;
1052 }
1053 }
1066 }