| blob | 0b7d3b8226fdb01725b76b6ba18643a4fbd41381 |
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/marker.h>
20 #include <linux/errno.h>
21 #include <linux/slab.h>
22 #include <linux/mm.h>
23 #include <linux/pagemap.h>
24 #include <linux/jiffies.h>
25 #include <linux/crc32.h>
26 #include <linux/writeback.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bio.h>
30 /*
31 * Default IO end handler for temporary BJ_IO buffer_heads.
32 */
34 {
38 else
41 }
43 /*
44 * When an ext4 file is truncated, it is possible that some pages are not
45 * successfully freed, because they are attached to a committing transaction.
46 * After the transaction commits, these pages are left on the LRU, with no
47 * ->mapping, and with attached buffers. These pages are trivially reclaimable
48 * by the VM, but their apparent absence upsets the VM accounting, and it makes
49 * the numbers in /proc/meminfo look odd.
50 *
51 * So here, we have a buffer which has just come off the forget list. Look to
52 * see if we can strip all buffers from the backing page.
53 *
54 * Called under lock_journal(), and possibly under journal_datalist_lock. The
55 * caller provided us with a ref against the buffer, and we drop that here.
56 */
58 {
71 /* OK, it's a truncated page */
82 nope:
84 }
86 /*
87 * Done it all: now submit the commit record. We should have
88 * cleaned up our previous buffers by now, so if we are in abort
89 * mode we can now just skip the rest of the journal write
90 * entirely.
91 *
92 * Returns 1 if the journal needs to be aborted or 0 on success
93 */
97 __u32 crc32_sum)
98 {
123 JBD2_FEATURE_COMPAT_CHECKSUM)) {
127 }
137 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
140 }
145 /* is it possible for another commit to fail at roughly
146 * the same time as this one? If so, we don't want to
147 * trust the barrier flag in the super, but instead want
148 * to remember if we sent a barrier request
149 */
152 "JBD: barrier-based sync failed on %s - "
158 /* And try again, without the barrier */
163 }
166 }
168 /*
169 * This function along with journal_submit_commit_record
170 * allows to write the commit record asynchronously.
171 */
174 {
177 retry:
182 "JBD2: wait_on_commit_record: sync failed on %s - "
197 }
199 }
207 }
209 /*
210 * write the filemap data using writepage() address_space_operations.
211 * We don't do block allocation here even for delalloc. We don't
212 * use writepages() because with dealyed allocation we may be doing
213 * block allocation in writepages().
214 */
216 {
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 */
263 }
266 }
268 /*
269 * Wait for data submitted for writeout, refile inodes to proper
270 * transaction if needed.
271 *
272 */
275 {
279 /* For locking, see the comment in journal_submit_data_buffers() */
286 /*
287 * Because AS_EIO is cleared by
288 * wait_on_page_writeback_range(), set it again so
289 * that user process can get -EIO from fsync().
290 */
296 }
300 }
302 /* Now refile inode to proper lists */
313 }
314 }
318 }
321 {
324 __u32 checksum;
332 }
336 {
340 }
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 {
358 ktime_t start_time;
359 u64 commit_time;
372 /*
373 * First job: lock down the current transaction and wait for
374 * all outstanding updates to complete.
375 */
377 #ifdef COMMIT_STATS
381 #endif
383 /* Do we need to erase the effects of a prior jbd2_journal_flush? */
389 }
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 */
422 TASK_UNINTERRUPTIBLE);
429 }
431 }
437 /*
438 * First thing we are allowed to do is to discard any remaining
439 * BJ_Reserved buffers. Note, it is _not_ permissible to assume
440 * that there are no such buffers: if a large filesystem
441 * operation like a truncate needs to split itself over multiple
442 * transactions, then it may try to do a jbd2_journal_restart() while
443 * there are still BJ_Reserved buffers outstanding. These must
444 * be released cleanly from the current transaction.
445 *
446 * In this case, the filesystem must still reserve write access
447 * again before modifying the buffer in the new transaction, but
448 * we do not require it to remember exactly which old buffers it
449 * has reserved. This is consistent with the existing behaviour
450 * that multiple jbd2_journal_get_write_access() calls to the same
451 * buffer are perfectly permissable.
452 */
456 /*
457 * A jbd2_journal_get_undo_access()+jbd2_journal_release_buffer() may
458 * leave undo-committed data.
459 */
467 }
469 }
471 /*
472 * Now try to drop any written-back buffers from the journal's
473 * checkpoint lists. We do this *before* commit because it potentially
474 * frees some memory
475 */
482 /*
483 * Switch to a new revoke table.
484 */
501 /*
502 * Now start flushing things to disk, in the order they appear
503 * on the transaction lists. Data blocks go first.
504 */
510 write_op);
514 /*
515 * Way to go: we have now written out all of the data for a
516 * transaction! Now comes the tricky part: we need to write out
517 * metadata. Loop over the transaction's entire buffer list:
518 */
537 /* Find the next buffer to be journaled... */
541 /* If we're in abort mode, we just un-journal the buffer and
542 release it. */
552 /* If that was the last one, we need to clean up
553 * any descriptor buffers which may have been
554 * already allocated, even if we are now
555 * aborting. */
559 }
561 /* Make sure we have a descriptor block in which to
562 record the metadata buffer. */
575 }
592 /* Record it so that we can wait for IO
593 completion later */
596 BJ_LogCtl);
597 }
599 /* Where is the buffer to be written? */
602 /* If the block mapping failed, just abandon the buffer
603 and repeat this loop: we'll fall into the
604 refile-on-abort condition above. */
608 }
610 /*
611 * start_this_handle() uses t_outstanding_credits to determine
612 * the free space in the log, but this counter is changed
613 * by jbd2_journal_next_log_block() also.
614 */
617 /* Bump b_count to prevent truncate from stumbling over
618 the shadowed buffer! @@@ This can go if we ever get
619 rid of the BJ_IO/BJ_Shadow pairing of buffers. */
622 /* Make a temporary IO buffer with which to write it out
623 (this will requeue both the metadata buffer and the
624 temporary IO buffer). new_bh goes on BJ_IO*/
627 /*
628 * akpm: jbd2_journal_write_metadata_buffer() sets
629 * new_bh->b_transaction to commit_transaction.
630 * We need to clean this up before we release new_bh
631 * (which is of type BJ_IO)
632 */
639 /* Record the new block's tag in the current descriptor
640 buffer */
659 }
661 /* If there's no more to do, or if the descriptor is full,
662 let the IO rip! */
670 /* Write an end-of-descriptor marker before
671 submitting the IOs. "tag" still points to
672 the last tag we set up. */
676 start_journal_io:
679 /*
680 * Compute checksum.
681 */
683 JBD2_FEATURE_COMPAT_CHECKSUM)) {
684 crc32_sum =
686 }
693 }
697 /* Force a new descriptor to be generated next
698 time round the loop. */
701 }
702 }
704 /* Done it all: now write the commit record asynchronously. */
707 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
712 }
714 /*
715 * This is the right place to wait for data buffers both for ASYNC
716 * and !ASYNC commit. If commit is ASYNC, we need to wait only after
717 * the commit block went to disk (which happens above). If commit is
718 * SYNC, we need to wait for data buffers before we start writing
719 * commit block, which happens below in such setting.
720 */
724 "JBD2: Detected IO errors while flushing file data "
729 }
731 /* Lo and behold: we have just managed to send a transaction to
732 the log. Before we can commit it, wait for the IO so far to
733 complete. Control buffers being written are on the
734 transaction's t_log_list queue, and metadata buffers are on
735 the t_iobuf_list queue.
737 Wait for the buffers in reverse order. That way we are
738 less likely to be woken up until all IOs have completed, and
739 so we incur less scheduling load.
740 */
744 /*
745 * akpm: these are BJ_IO, and j_list_lock is not needed.
746 * See __journal_try_to_free_buffer.
747 */
748 wait_for_iobuf:
757 }
769 /*
770 * ->t_iobuf_list should contain only dummy buffer_heads
771 * which were created by jbd2_journal_write_metadata_buffer().
772 */
779 /* We also have to unlock and free the corresponding
780 shadowed buffer */
786 /* The metadata is now released for reuse, but we need
787 to remember it against this transaction so that when
788 we finally commit, we can do any checkpointing
789 required. */
792 /* Wake up any transactions which were waiting for this
793 IO to complete */
797 }
803 /* Here we wait for the revoke record and descriptor record buffers */
804 wait_for_ctlbuf:
813 }
825 /* AKPM: bforget here */
826 }
834 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
839 }
846 /* End of a transaction! Finally, we can do checkpoint
847 processing: any buffers committed as a result of this
848 transaction can be removed from any checkpoint list it was on
849 before. */
860 restart_loop:
861 /*
862 * As there are other places (journal_unmap_buffer()) adding buffers
863 * to this list we have to be careful and hold the j_list_lock.
864 */
877 /*
878 * If there is undo-protected committed data against
879 * this buffer, then we can remove it now. If it is a
880 * buffer needing such protection, the old frozen_data
881 * field now points to a committed version of the
882 * buffer, so rotate that field to the new committed
883 * data.
884 *
885 * Otherwise, we can just throw away the frozen data now.
886 *
887 * We also know that the frozen data has already fired
888 * its triggers if they exist, so we can clear that too.
889 */
897 }
902 }
910 }
912 /* Only re-checkpoint the buffer_head if it is marked
913 * dirty. If the buffer was added to the BJ_Forget list
914 * by jbd2_journal_forget, it may no longer be dirty and
915 * there's no point in keeping a checkpoint record for
916 * it. */
918 /* A buffer which has been freed while still being
919 * journaled by a previous transaction may end up still
920 * being dirty here, but we want to avoid writing back
921 * that buffer in the future now that the last use has
922 * been committed. That's not only a performance gain,
923 * it also stops aliasing problems if the buffer is left
924 * behind for writeback and gets reallocated for another
925 * use in a different page. */
929 }
941 /* The buffer on BJ_Forget list and not jbddirty means
942 * it has been freed by this transaction and hence it
943 * could not have been reallocated until this
944 * transaction has committed. *BUT* it could be
945 * reallocated once we have written all the data to
946 * disk and before we process the buffer on BJ_Forget
947 * list. */
952 /* needs a brelse */
957 }
959 }
961 /*
962 * This is a bit sleazy. We use j_list_lock to protect transition
963 * of a transaction into T_FINISHED state and calling
964 * __jbd2_journal_drop_transaction(). Otherwise we could race with
965 * other checkpointing code processing the transaction...
966 */
969 /*
970 * Now recheck if some buffers did not get attached to the transaction
971 * while the lock was dropped...
972 */
977 }
979 /* Done with this transaction! */
989 /*
990 * File the transaction for history
991 */
1001 /*
1002 * Calculate overall stats
1003 */
1021 /*
1022 * weight the commit time higher than the average time so we don't
1023 * react too strongly to vast changes in the commit time
1024 */
1028 else
1047 commit_transaction;
1049 commit_transaction;
1050 }
1051 }
1066 }