1 // SPDX-License-Identifier: GPL-2.0+
3 * linux/fs/jbd2/recovery.c
5 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
7 * Copyright 1999-2000 Red Hat Software --- All Rights Reserved
9 * Journal recovery routines for the generic filesystem journaling code;
10 * part of the ext2fs journaling system.
16 #include <linux/time.h>
18 #include <linux/jbd2.h>
19 #include <linux/errno.h>
20 #include <linux/crc32.h>
21 #include <linux/blkdev.h>
25 * Maintain information about the progress of the recovery job, so that
26 * the different passes can carry information between them.
30 tid_t start_transaction
;
31 tid_t end_transaction
;
38 enum passtype
{PASS_SCAN
, PASS_REVOKE
, PASS_REPLAY
};
39 static int do_one_pass(journal_t
*journal
,
40 struct recovery_info
*info
, enum passtype pass
);
41 static int scan_revoke_records(journal_t
*, struct buffer_head
*,
42 tid_t
, struct recovery_info
*);
46 /* Release readahead buffers after use */
47 static void journal_brelse_array(struct buffer_head
*b
[], int n
)
55 * When reading from the journal, we are going through the block device
56 * layer directly and so there is no readahead being done for us. We
57 * need to implement any readahead ourselves if we want it to happen at
58 * all. Recovery is basically one long sequential read, so make sure we
59 * do the IO in reasonably large chunks.
61 * This is not so critical that we need to be enormously clever about
62 * the readahead size, though. 128K is a purely arbitrary, good-enough
67 static int do_readahead(journal_t
*journal
, unsigned int start
)
70 unsigned int max
, nbufs
, next
;
71 unsigned long long blocknr
;
72 struct buffer_head
*bh
;
74 struct buffer_head
* bufs
[MAXBUF
];
76 /* Do up to 128K of readahead */
77 max
= start
+ (128 * 1024 / journal
->j_blocksize
);
78 if (max
> journal
->j_maxlen
)
79 max
= journal
->j_maxlen
;
81 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
82 * a time to the block device IO layer. */
86 for (next
= start
; next
< max
; next
++) {
87 err
= jbd2_journal_bmap(journal
, next
, &blocknr
);
90 printk(KERN_ERR
"JBD2: bad block at offset %u\n",
95 bh
= __getblk(journal
->j_dev
, blocknr
, journal
->j_blocksize
);
101 if (!buffer_uptodate(bh
) && !buffer_locked(bh
)) {
103 if (nbufs
== MAXBUF
) {
104 ll_rw_block(REQ_OP_READ
, 0, nbufs
, bufs
);
105 journal_brelse_array(bufs
, nbufs
);
113 ll_rw_block(REQ_OP_READ
, 0, nbufs
, bufs
);
118 journal_brelse_array(bufs
, nbufs
);
122 #endif /* __KERNEL__ */
126 * Read a block from the journal
129 static int jread(struct buffer_head
**bhp
, journal_t
*journal
,
133 unsigned long long blocknr
;
134 struct buffer_head
*bh
;
138 if (offset
>= journal
->j_maxlen
) {
139 printk(KERN_ERR
"JBD2: corrupted journal superblock\n");
140 return -EFSCORRUPTED
;
143 err
= jbd2_journal_bmap(journal
, offset
, &blocknr
);
146 printk(KERN_ERR
"JBD2: bad block at offset %u\n",
151 bh
= __getblk(journal
->j_dev
, blocknr
, journal
->j_blocksize
);
155 if (!buffer_uptodate(bh
)) {
156 /* If this is a brand new buffer, start readahead.
157 Otherwise, we assume we are already reading it. */
159 do_readahead(journal
, offset
);
163 if (!buffer_uptodate(bh
)) {
164 printk(KERN_ERR
"JBD2: Failed to read block at offset %u\n",
174 static int jbd2_descriptor_block_csum_verify(journal_t
*j
, void *buf
)
176 struct jbd2_journal_block_tail
*tail
;
180 if (!jbd2_journal_has_csum_v2or3(j
))
183 tail
= (struct jbd2_journal_block_tail
*)(buf
+ j
->j_blocksize
-
184 sizeof(struct jbd2_journal_block_tail
));
185 provided
= tail
->t_checksum
;
186 tail
->t_checksum
= 0;
187 calculated
= jbd2_chksum(j
, j
->j_csum_seed
, buf
, j
->j_blocksize
);
188 tail
->t_checksum
= provided
;
190 return provided
== cpu_to_be32(calculated
);
194 * Count the number of in-use tags in a journal descriptor block.
197 static int count_tags(journal_t
*journal
, struct buffer_head
*bh
)
200 journal_block_tag_t
* tag
;
201 int nr
= 0, size
= journal
->j_blocksize
;
202 int tag_bytes
= journal_tag_bytes(journal
);
204 if (jbd2_journal_has_csum_v2or3(journal
))
205 size
-= sizeof(struct jbd2_journal_block_tail
);
207 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
209 while ((tagp
- bh
->b_data
+ tag_bytes
) <= size
) {
210 tag
= (journal_block_tag_t
*) tagp
;
214 if (!(tag
->t_flags
& cpu_to_be16(JBD2_FLAG_SAME_UUID
)))
217 if (tag
->t_flags
& cpu_to_be16(JBD2_FLAG_LAST_TAG
))
225 /* Make sure we wrap around the log correctly! */
226 #define wrap(journal, var) \
228 if (var >= (journal)->j_last) \
229 var -= ((journal)->j_last - (journal)->j_first); \
233 * jbd2_journal_recover - recovers a on-disk journal
234 * @journal: the journal to recover
236 * The primary function for recovering the log contents when mounting a
239 * Recovery is done in three passes. In the first pass, we look for the
240 * end of the log. In the second, we assemble the list of revoke
241 * blocks. In the third and final pass, we replay any un-revoked blocks
244 int jbd2_journal_recover(journal_t
*journal
)
247 journal_superblock_t
* sb
;
249 struct recovery_info info
;
251 memset(&info
, 0, sizeof(info
));
252 sb
= journal
->j_superblock
;
255 * The journal superblock's s_start field (the current log head)
256 * is always zero if, and only if, the journal was cleanly
261 jbd_debug(1, "No recovery required, last transaction %d\n",
262 be32_to_cpu(sb
->s_sequence
));
263 journal
->j_transaction_sequence
= be32_to_cpu(sb
->s_sequence
) + 1;
267 err
= do_one_pass(journal
, &info
, PASS_SCAN
);
269 err
= do_one_pass(journal
, &info
, PASS_REVOKE
);
271 err
= do_one_pass(journal
, &info
, PASS_REPLAY
);
273 jbd_debug(1, "JBD2: recovery, exit status %d, "
274 "recovered transactions %u to %u\n",
275 err
, info
.start_transaction
, info
.end_transaction
);
276 jbd_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
277 info
.nr_replays
, info
.nr_revoke_hits
, info
.nr_revokes
);
279 /* Restart the log at the next transaction ID, thus invalidating
280 * any existing commit records in the log. */
281 journal
->j_transaction_sequence
= ++info
.end_transaction
;
283 jbd2_journal_clear_revoke(journal
);
284 err2
= sync_blockdev(journal
->j_fs_dev
);
287 /* Make sure all replayed data is on permanent storage */
288 if (journal
->j_flags
& JBD2_BARRIER
) {
289 err2
= blkdev_issue_flush(journal
->j_fs_dev
, GFP_KERNEL
);
297 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
298 * @journal: journal to startup
300 * Locate any valid recovery information from the journal and set up the
301 * journal structures in memory to ignore it (presumably because the
302 * caller has evidence that it is out of date).
303 * This function doesn't appear to be exported..
305 * We perform one pass over the journal to allow us to tell the user how
306 * much recovery information is being erased, and to let us initialise
307 * the journal transaction sequence numbers to the next unused ID.
309 int jbd2_journal_skip_recovery(journal_t
*journal
)
313 struct recovery_info info
;
315 memset (&info
, 0, sizeof(info
));
317 err
= do_one_pass(journal
, &info
, PASS_SCAN
);
320 printk(KERN_ERR
"JBD2: error %d scanning journal\n", err
);
321 ++journal
->j_transaction_sequence
;
323 #ifdef CONFIG_JBD2_DEBUG
324 int dropped
= info
.end_transaction
-
325 be32_to_cpu(journal
->j_superblock
->s_sequence
);
327 "JBD2: ignoring %d transaction%s from the journal.\n",
328 dropped
, (dropped
== 1) ? "" : "s");
330 journal
->j_transaction_sequence
= ++info
.end_transaction
;
337 static inline unsigned long long read_tag_block(journal_t
*journal
,
338 journal_block_tag_t
*tag
)
340 unsigned long long block
= be32_to_cpu(tag
->t_blocknr
);
341 if (jbd2_has_feature_64bit(journal
))
342 block
|= (u64
)be32_to_cpu(tag
->t_blocknr_high
) << 32;
347 * calc_chksums calculates the checksums for the blocks described in the
350 static int calc_chksums(journal_t
*journal
, struct buffer_head
*bh
,
351 unsigned long *next_log_block
, __u32
*crc32_sum
)
353 int i
, num_blks
, err
;
354 unsigned long io_block
;
355 struct buffer_head
*obh
;
357 num_blks
= count_tags(journal
, bh
);
358 /* Calculate checksum of the descriptor block. */
359 *crc32_sum
= crc32_be(*crc32_sum
, (void *)bh
->b_data
, bh
->b_size
);
361 for (i
= 0; i
< num_blks
; i
++) {
362 io_block
= (*next_log_block
)++;
363 wrap(journal
, *next_log_block
);
364 err
= jread(&obh
, journal
, io_block
);
366 printk(KERN_ERR
"JBD2: IO error %d recovering block "
367 "%lu in log\n", err
, io_block
);
370 *crc32_sum
= crc32_be(*crc32_sum
, (void *)obh
->b_data
,
378 static int jbd2_commit_block_csum_verify(journal_t
*j
, void *buf
)
380 struct commit_header
*h
;
384 if (!jbd2_journal_has_csum_v2or3(j
))
388 provided
= h
->h_chksum
[0];
390 calculated
= jbd2_chksum(j
, j
->j_csum_seed
, buf
, j
->j_blocksize
);
391 h
->h_chksum
[0] = provided
;
393 return provided
== cpu_to_be32(calculated
);
396 static int jbd2_block_tag_csum_verify(journal_t
*j
, journal_block_tag_t
*tag
,
397 void *buf
, __u32 sequence
)
399 journal_block_tag3_t
*tag3
= (journal_block_tag3_t
*)tag
;
403 if (!jbd2_journal_has_csum_v2or3(j
))
406 seq
= cpu_to_be32(sequence
);
407 csum32
= jbd2_chksum(j
, j
->j_csum_seed
, (__u8
*)&seq
, sizeof(seq
));
408 csum32
= jbd2_chksum(j
, csum32
, buf
, j
->j_blocksize
);
410 if (jbd2_has_feature_csum3(j
))
411 return tag3
->t_checksum
== cpu_to_be32(csum32
);
413 return tag
->t_checksum
== cpu_to_be16(csum32
);
416 static int do_one_pass(journal_t
*journal
,
417 struct recovery_info
*info
, enum passtype pass
)
419 unsigned int first_commit_ID
, next_commit_ID
;
420 unsigned long next_log_block
;
421 int err
, success
= 0;
422 journal_superblock_t
* sb
;
423 journal_header_t
* tmp
;
424 struct buffer_head
* bh
;
425 unsigned int sequence
;
427 int tag_bytes
= journal_tag_bytes(journal
);
428 __u32 crc32_sum
= ~0; /* Transactional Checksums */
429 int descr_csum_size
= 0;
431 bool need_check_commit_time
= false;
432 __u64 last_trans_commit_time
= 0, commit_time
;
435 * First thing is to establish what we expect to find in the log
436 * (in terms of transaction IDs), and where (in terms of log
437 * block offsets): query the superblock.
440 sb
= journal
->j_superblock
;
441 next_commit_ID
= be32_to_cpu(sb
->s_sequence
);
442 next_log_block
= be32_to_cpu(sb
->s_start
);
444 first_commit_ID
= next_commit_ID
;
445 if (pass
== PASS_SCAN
)
446 info
->start_transaction
= first_commit_ID
;
448 jbd_debug(1, "Starting recovery pass %d\n", pass
);
451 * Now we walk through the log, transaction by transaction,
452 * making sure that each transaction has a commit block in the
453 * expected place. Each complete transaction gets replayed back
454 * into the main filesystem.
460 journal_block_tag_t
* tag
;
461 struct buffer_head
* obh
;
462 struct buffer_head
* nbh
;
466 /* If we already know where to stop the log traversal,
467 * check right now that we haven't gone past the end of
470 if (pass
!= PASS_SCAN
)
471 if (tid_geq(next_commit_ID
, info
->end_transaction
))
474 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
475 next_commit_ID
, next_log_block
, journal
->j_last
);
477 /* Skip over each chunk of the transaction looking
478 * either the next descriptor block or the final commit
481 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block
);
482 err
= jread(&bh
, journal
, next_log_block
);
487 wrap(journal
, next_log_block
);
489 /* What kind of buffer is it?
491 * If it is a descriptor block, check that it has the
492 * expected sequence number. Otherwise, we're all done
495 tmp
= (journal_header_t
*)bh
->b_data
;
497 if (tmp
->h_magic
!= cpu_to_be32(JBD2_MAGIC_NUMBER
)) {
502 blocktype
= be32_to_cpu(tmp
->h_blocktype
);
503 sequence
= be32_to_cpu(tmp
->h_sequence
);
504 jbd_debug(3, "Found magic %d, sequence %d\n",
505 blocktype
, sequence
);
507 if (sequence
!= next_commit_ID
) {
512 /* OK, we have a valid descriptor block which matches
513 * all of the sequence number checks. What are we going
514 * to do with it? That depends on the pass... */
517 case JBD2_DESCRIPTOR_BLOCK
:
518 /* Verify checksum first */
519 if (jbd2_journal_has_csum_v2or3(journal
))
521 sizeof(struct jbd2_journal_block_tail
);
522 if (descr_csum_size
> 0 &&
523 !jbd2_descriptor_block_csum_verify(journal
,
526 * PASS_SCAN can see stale blocks due to lazy
527 * journal init. Don't error out on those yet.
529 if (pass
!= PASS_SCAN
) {
530 pr_err("JBD2: Invalid checksum recovering block %lu in log\n",
536 need_check_commit_time
= true;
538 "invalid descriptor block found in %lu\n",
542 /* If it is a valid descriptor block, replay it
543 * in pass REPLAY; if journal_checksums enabled, then
544 * calculate checksums in PASS_SCAN, otherwise,
545 * just skip over the blocks it describes. */
546 if (pass
!= PASS_REPLAY
) {
547 if (pass
== PASS_SCAN
&&
548 jbd2_has_feature_checksum(journal
) &&
549 !need_check_commit_time
&&
550 !info
->end_transaction
) {
551 if (calc_chksums(journal
, bh
,
560 next_log_block
+= count_tags(journal
, bh
);
561 wrap(journal
, next_log_block
);
566 /* A descriptor block: we can now write all of
567 * the data blocks. Yay, useful work is finally
568 * getting done here! */
570 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
571 while ((tagp
- bh
->b_data
+ tag_bytes
)
572 <= journal
->j_blocksize
- descr_csum_size
) {
573 unsigned long io_block
;
575 tag
= (journal_block_tag_t
*) tagp
;
576 flags
= be16_to_cpu(tag
->t_flags
);
578 io_block
= next_log_block
++;
579 wrap(journal
, next_log_block
);
580 err
= jread(&obh
, journal
, io_block
);
582 /* Recover what we can, but
583 * report failure at the end. */
586 "JBD2: IO error %d recovering "
587 "block %ld in log\n",
590 unsigned long long blocknr
;
592 J_ASSERT(obh
!= NULL
);
593 blocknr
= read_tag_block(journal
,
596 /* If the block has been
597 * revoked, then we're all done
599 if (jbd2_journal_test_revoke
603 ++info
->nr_revoke_hits
;
607 /* Look for block corruption */
608 if (!jbd2_block_tag_csum_verify(
609 journal
, tag
, obh
->b_data
,
610 be32_to_cpu(tmp
->h_sequence
))) {
612 success
= -EFSBADCRC
;
613 printk(KERN_ERR
"JBD2: Invalid "
614 "checksum recovering "
615 "data block %llu in "
621 /* Find a buffer for the new
622 * data being restored */
623 nbh
= __getblk(journal
->j_fs_dev
,
625 journal
->j_blocksize
);
628 "JBD2: Out of memory "
629 "during recovery.\n");
637 memcpy(nbh
->b_data
, obh
->b_data
,
638 journal
->j_blocksize
);
639 if (flags
& JBD2_FLAG_ESCAPE
) {
640 *((__be32
*)nbh
->b_data
) =
641 cpu_to_be32(JBD2_MAGIC_NUMBER
);
644 BUFFER_TRACE(nbh
, "marking dirty");
645 set_buffer_uptodate(nbh
);
646 mark_buffer_dirty(nbh
);
647 BUFFER_TRACE(nbh
, "marking uptodate");
649 /* ll_rw_block(WRITE, 1, &nbh); */
657 if (!(flags
& JBD2_FLAG_SAME_UUID
))
660 if (flags
& JBD2_FLAG_LAST_TAG
)
667 case JBD2_COMMIT_BLOCK
:
668 /* How to differentiate between interrupted commit
669 * and journal corruption ?
672 * Checksum Verification Failed
674 * ____________________
676 * async_commit sync_commit
678 * | GO TO NEXT "Journal Corruption"
681 * {(n+1)th transanction}
683 * _______|______________
685 * Commit block found Commit block not found
687 * "Journal Corruption" |
688 * _____________|_________
690 * nth trans corrupt OR nth trans
691 * and (n+1)th interrupted interrupted
692 * before commit block
693 * could reach the disk.
694 * (Cannot find the difference in above
695 * mentioned conditions. Hence assume
696 * "Interrupted Commit".)
698 commit_time
= be64_to_cpu(
699 ((struct commit_header
*)bh
->b_data
)->h_commit_sec
);
701 * If need_check_commit_time is set, it means we are in
702 * PASS_SCAN and csum verify failed before. If
703 * commit_time is increasing, it's the same journal,
704 * otherwise it is stale journal block, just end this
707 if (need_check_commit_time
) {
708 if (commit_time
>= last_trans_commit_time
) {
709 pr_err("JBD2: Invalid checksum found in transaction %u\n",
717 * It likely does not belong to same journal,
718 * just end this recovery with success.
720 jbd_debug(1, "JBD2: Invalid checksum ignored in transaction %u, likely stale data\n",
728 * Found an expected commit block: if checksums
729 * are present, verify them in PASS_SCAN; else not
730 * much to do other than move on to the next sequence
733 if (pass
== PASS_SCAN
&&
734 jbd2_has_feature_checksum(journal
)) {
735 struct commit_header
*cbh
=
736 (struct commit_header
*)bh
->b_data
;
737 unsigned found_chksum
=
738 be32_to_cpu(cbh
->h_chksum
[0]);
740 if (info
->end_transaction
) {
741 journal
->j_failed_commit
=
742 info
->end_transaction
;
747 /* Neither checksum match nor unused? */
748 if (!((crc32_sum
== found_chksum
&&
749 cbh
->h_chksum_type
==
751 cbh
->h_chksum_size
==
752 JBD2_CRC32_CHKSUM_SIZE
) ||
753 (cbh
->h_chksum_type
== 0 &&
754 cbh
->h_chksum_size
== 0 &&
760 if (pass
== PASS_SCAN
&&
761 !jbd2_commit_block_csum_verify(journal
,
764 if (commit_time
< last_trans_commit_time
)
765 goto ignore_crc_mismatch
;
766 info
->end_transaction
= next_commit_ID
;
768 if (!jbd2_has_feature_async_commit(journal
)) {
769 journal
->j_failed_commit
=
775 if (pass
== PASS_SCAN
)
776 last_trans_commit_time
= commit_time
;
781 case JBD2_REVOKE_BLOCK
:
783 * Check revoke block crc in pass_scan, if csum verify
784 * failed, check commit block time later.
786 if (pass
== PASS_SCAN
&&
787 !jbd2_descriptor_block_csum_verify(journal
,
789 jbd_debug(1, "JBD2: invalid revoke block found in %lu\n",
791 need_check_commit_time
= true;
793 /* If we aren't in the REVOKE pass, then we can
794 * just skip over this block. */
795 if (pass
!= PASS_REVOKE
) {
800 err
= scan_revoke_records(journal
, bh
,
801 next_commit_ID
, info
);
808 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
817 * We broke out of the log scan loop: either we came to the
818 * known end of the log or we found an unexpected block in the
819 * log. If the latter happened, then we know that the "current"
820 * transaction marks the end of the valid log.
823 if (pass
== PASS_SCAN
) {
824 if (!info
->end_transaction
)
825 info
->end_transaction
= next_commit_ID
;
827 /* It's really bad news if different passes end up at
828 * different places (but possible due to IO errors). */
829 if (info
->end_transaction
!= next_commit_ID
) {
830 printk(KERN_ERR
"JBD2: recovery pass %d ended at "
831 "transaction %u, expected %u\n",
832 pass
, next_commit_ID
, info
->end_transaction
);
837 if (block_error
&& success
== 0)
845 /* Scan a revoke record, marking all blocks mentioned as revoked. */
847 static int scan_revoke_records(journal_t
*journal
, struct buffer_head
*bh
,
848 tid_t sequence
, struct recovery_info
*info
)
850 jbd2_journal_revoke_header_t
*header
;
856 header
= (jbd2_journal_revoke_header_t
*) bh
->b_data
;
857 offset
= sizeof(jbd2_journal_revoke_header_t
);
858 rcount
= be32_to_cpu(header
->r_count
);
860 if (jbd2_journal_has_csum_v2or3(journal
))
861 csum_size
= sizeof(struct jbd2_journal_block_tail
);
862 if (rcount
> journal
->j_blocksize
- csum_size
)
866 if (jbd2_has_feature_64bit(journal
))
869 while (offset
+ record_len
<= max
) {
870 unsigned long long blocknr
;
874 blocknr
= be32_to_cpu(* ((__be32
*) (bh
->b_data
+offset
)));
876 blocknr
= be64_to_cpu(* ((__be64
*) (bh
->b_data
+offset
)));
877 offset
+= record_len
;
878 err
= jbd2_journal_set_revoke(journal
, blocknr
, sequence
);