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 static int do_one_pass(journal_t
*journal
,
39 struct recovery_info
*info
, enum passtype pass
);
40 static int scan_revoke_records(journal_t
*, struct buffer_head
*,
41 tid_t
, struct recovery_info
*);
45 /* Release readahead buffers after use */
46 static void journal_brelse_array(struct buffer_head
*b
[], int n
)
54 * When reading from the journal, we are going through the block device
55 * layer directly and so there is no readahead being done for us. We
56 * need to implement any readahead ourselves if we want it to happen at
57 * all. Recovery is basically one long sequential read, so make sure we
58 * do the IO in reasonably large chunks.
60 * This is not so critical that we need to be enormously clever about
61 * the readahead size, though. 128K is a purely arbitrary, good-enough
66 static int do_readahead(journal_t
*journal
, unsigned int start
)
69 unsigned int max
, nbufs
, next
;
70 unsigned long long blocknr
;
71 struct buffer_head
*bh
;
73 struct buffer_head
* bufs
[MAXBUF
];
75 /* Do up to 128K of readahead */
76 max
= start
+ (128 * 1024 / journal
->j_blocksize
);
77 if (max
> journal
->j_total_len
)
78 max
= journal
->j_total_len
;
80 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
81 * a time to the block device IO layer. */
85 for (next
= start
; next
< max
; next
++) {
86 err
= jbd2_journal_bmap(journal
, next
, &blocknr
);
89 printk(KERN_ERR
"JBD2: bad block at offset %u\n",
94 bh
= __getblk(journal
->j_dev
, blocknr
, journal
->j_blocksize
);
100 if (!buffer_uptodate(bh
) && !buffer_locked(bh
)) {
102 if (nbufs
== MAXBUF
) {
103 ll_rw_block(REQ_OP_READ
, 0, nbufs
, bufs
);
104 journal_brelse_array(bufs
, nbufs
);
112 ll_rw_block(REQ_OP_READ
, 0, nbufs
, bufs
);
117 journal_brelse_array(bufs
, nbufs
);
121 #endif /* __KERNEL__ */
125 * Read a block from the journal
128 static int jread(struct buffer_head
**bhp
, journal_t
*journal
,
132 unsigned long long blocknr
;
133 struct buffer_head
*bh
;
137 if (offset
>= journal
->j_total_len
) {
138 printk(KERN_ERR
"JBD2: corrupted journal superblock\n");
139 return -EFSCORRUPTED
;
142 err
= jbd2_journal_bmap(journal
, offset
, &blocknr
);
145 printk(KERN_ERR
"JBD2: bad block at offset %u\n",
150 bh
= __getblk(journal
->j_dev
, blocknr
, journal
->j_blocksize
);
154 if (!buffer_uptodate(bh
)) {
155 /* If this is a brand new buffer, start readahead.
156 Otherwise, we assume we are already reading it. */
158 do_readahead(journal
, offset
);
162 if (!buffer_uptodate(bh
)) {
163 printk(KERN_ERR
"JBD2: Failed to read block at offset %u\n",
173 static int jbd2_descriptor_block_csum_verify(journal_t
*j
, void *buf
)
175 struct jbd2_journal_block_tail
*tail
;
179 if (!jbd2_journal_has_csum_v2or3(j
))
182 tail
= (struct jbd2_journal_block_tail
*)(buf
+ j
->j_blocksize
-
183 sizeof(struct jbd2_journal_block_tail
));
184 provided
= tail
->t_checksum
;
185 tail
->t_checksum
= 0;
186 calculated
= jbd2_chksum(j
, j
->j_csum_seed
, buf
, j
->j_blocksize
);
187 tail
->t_checksum
= provided
;
189 return provided
== cpu_to_be32(calculated
);
193 * Count the number of in-use tags in a journal descriptor block.
196 static int count_tags(journal_t
*journal
, struct buffer_head
*bh
)
199 journal_block_tag_t
* tag
;
200 int nr
= 0, size
= journal
->j_blocksize
;
201 int tag_bytes
= journal_tag_bytes(journal
);
203 if (jbd2_journal_has_csum_v2or3(journal
))
204 size
-= sizeof(struct jbd2_journal_block_tail
);
206 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
208 while ((tagp
- bh
->b_data
+ tag_bytes
) <= size
) {
209 tag
= (journal_block_tag_t
*) tagp
;
213 if (!(tag
->t_flags
& cpu_to_be16(JBD2_FLAG_SAME_UUID
)))
216 if (tag
->t_flags
& cpu_to_be16(JBD2_FLAG_LAST_TAG
))
224 /* Make sure we wrap around the log correctly! */
225 #define wrap(journal, var) \
227 unsigned long _wrap_last = \
228 jbd2_has_feature_fast_commit(journal) ? \
229 (journal)->j_fc_last : (journal)->j_last; \
231 if (var >= _wrap_last) \
232 var -= (_wrap_last - (journal)->j_first); \
235 static int fc_do_one_pass(journal_t
*journal
,
236 struct recovery_info
*info
, enum passtype pass
)
238 unsigned int expected_commit_id
= info
->end_transaction
;
239 unsigned long next_fc_block
;
240 struct buffer_head
*bh
;
243 next_fc_block
= journal
->j_fc_first
;
244 if (!journal
->j_fc_replay_callback
)
247 while (next_fc_block
<= journal
->j_fc_last
) {
248 jbd_debug(3, "Fast commit replay: next block %ld\n",
250 err
= jread(&bh
, journal
, next_fc_block
);
252 jbd_debug(3, "Fast commit replay: read error\n");
256 err
= journal
->j_fc_replay_callback(journal
, bh
, pass
,
257 next_fc_block
- journal
->j_fc_first
,
260 if (err
< 0 || err
== JBD2_FC_REPLAY_STOP
)
266 jbd_debug(3, "Fast commit replay failed, err = %d\n", err
);
272 * jbd2_journal_recover - recovers a on-disk journal
273 * @journal: the journal to recover
275 * The primary function for recovering the log contents when mounting a
278 * Recovery is done in three passes. In the first pass, we look for the
279 * end of the log. In the second, we assemble the list of revoke
280 * blocks. In the third and final pass, we replay any un-revoked blocks
283 int jbd2_journal_recover(journal_t
*journal
)
286 journal_superblock_t
* sb
;
288 struct recovery_info info
;
290 memset(&info
, 0, sizeof(info
));
291 sb
= journal
->j_superblock
;
294 * The journal superblock's s_start field (the current log head)
295 * is always zero if, and only if, the journal was cleanly
300 jbd_debug(1, "No recovery required, last transaction %d\n",
301 be32_to_cpu(sb
->s_sequence
));
302 journal
->j_transaction_sequence
= be32_to_cpu(sb
->s_sequence
) + 1;
306 err
= do_one_pass(journal
, &info
, PASS_SCAN
);
308 err
= do_one_pass(journal
, &info
, PASS_REVOKE
);
310 err
= do_one_pass(journal
, &info
, PASS_REPLAY
);
312 jbd_debug(1, "JBD2: recovery, exit status %d, "
313 "recovered transactions %u to %u\n",
314 err
, info
.start_transaction
, info
.end_transaction
);
315 jbd_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
316 info
.nr_replays
, info
.nr_revoke_hits
, info
.nr_revokes
);
318 /* Restart the log at the next transaction ID, thus invalidating
319 * any existing commit records in the log. */
320 journal
->j_transaction_sequence
= ++info
.end_transaction
;
322 jbd2_journal_clear_revoke(journal
);
323 err2
= sync_blockdev(journal
->j_fs_dev
);
326 /* Make sure all replayed data is on permanent storage */
327 if (journal
->j_flags
& JBD2_BARRIER
) {
328 err2
= blkdev_issue_flush(journal
->j_fs_dev
, GFP_KERNEL
);
336 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
337 * @journal: journal to startup
339 * Locate any valid recovery information from the journal and set up the
340 * journal structures in memory to ignore it (presumably because the
341 * caller has evidence that it is out of date).
342 * This function doesn't appear to be exported..
344 * We perform one pass over the journal to allow us to tell the user how
345 * much recovery information is being erased, and to let us initialise
346 * the journal transaction sequence numbers to the next unused ID.
348 int jbd2_journal_skip_recovery(journal_t
*journal
)
352 struct recovery_info info
;
354 memset (&info
, 0, sizeof(info
));
356 err
= do_one_pass(journal
, &info
, PASS_SCAN
);
359 printk(KERN_ERR
"JBD2: error %d scanning journal\n", err
);
360 ++journal
->j_transaction_sequence
;
362 #ifdef CONFIG_JBD2_DEBUG
363 int dropped
= info
.end_transaction
-
364 be32_to_cpu(journal
->j_superblock
->s_sequence
);
366 "JBD2: ignoring %d transaction%s from the journal.\n",
367 dropped
, (dropped
== 1) ? "" : "s");
369 journal
->j_transaction_sequence
= ++info
.end_transaction
;
376 static inline unsigned long long read_tag_block(journal_t
*journal
,
377 journal_block_tag_t
*tag
)
379 unsigned long long block
= be32_to_cpu(tag
->t_blocknr
);
380 if (jbd2_has_feature_64bit(journal
))
381 block
|= (u64
)be32_to_cpu(tag
->t_blocknr_high
) << 32;
386 * calc_chksums calculates the checksums for the blocks described in the
389 static int calc_chksums(journal_t
*journal
, struct buffer_head
*bh
,
390 unsigned long *next_log_block
, __u32
*crc32_sum
)
392 int i
, num_blks
, err
;
393 unsigned long io_block
;
394 struct buffer_head
*obh
;
396 num_blks
= count_tags(journal
, bh
);
397 /* Calculate checksum of the descriptor block. */
398 *crc32_sum
= crc32_be(*crc32_sum
, (void *)bh
->b_data
, bh
->b_size
);
400 for (i
= 0; i
< num_blks
; i
++) {
401 io_block
= (*next_log_block
)++;
402 wrap(journal
, *next_log_block
);
403 err
= jread(&obh
, journal
, io_block
);
405 printk(KERN_ERR
"JBD2: IO error %d recovering block "
406 "%lu in log\n", err
, io_block
);
409 *crc32_sum
= crc32_be(*crc32_sum
, (void *)obh
->b_data
,
417 static int jbd2_commit_block_csum_verify(journal_t
*j
, void *buf
)
419 struct commit_header
*h
;
423 if (!jbd2_journal_has_csum_v2or3(j
))
427 provided
= h
->h_chksum
[0];
429 calculated
= jbd2_chksum(j
, j
->j_csum_seed
, buf
, j
->j_blocksize
);
430 h
->h_chksum
[0] = provided
;
432 return provided
== cpu_to_be32(calculated
);
435 static int jbd2_block_tag_csum_verify(journal_t
*j
, journal_block_tag_t
*tag
,
436 void *buf
, __u32 sequence
)
438 journal_block_tag3_t
*tag3
= (journal_block_tag3_t
*)tag
;
442 if (!jbd2_journal_has_csum_v2or3(j
))
445 seq
= cpu_to_be32(sequence
);
446 csum32
= jbd2_chksum(j
, j
->j_csum_seed
, (__u8
*)&seq
, sizeof(seq
));
447 csum32
= jbd2_chksum(j
, csum32
, buf
, j
->j_blocksize
);
449 if (jbd2_has_feature_csum3(j
))
450 return tag3
->t_checksum
== cpu_to_be32(csum32
);
452 return tag
->t_checksum
== cpu_to_be16(csum32
);
455 static int do_one_pass(journal_t
*journal
,
456 struct recovery_info
*info
, enum passtype pass
)
458 unsigned int first_commit_ID
, next_commit_ID
;
459 unsigned long next_log_block
;
460 int err
, success
= 0;
461 journal_superblock_t
* sb
;
462 journal_header_t
* tmp
;
463 struct buffer_head
* bh
;
464 unsigned int sequence
;
466 int tag_bytes
= journal_tag_bytes(journal
);
467 __u32 crc32_sum
= ~0; /* Transactional Checksums */
468 int descr_csum_size
= 0;
470 bool need_check_commit_time
= false;
471 __u64 last_trans_commit_time
= 0, commit_time
;
474 * First thing is to establish what we expect to find in the log
475 * (in terms of transaction IDs), and where (in terms of log
476 * block offsets): query the superblock.
479 sb
= journal
->j_superblock
;
480 next_commit_ID
= be32_to_cpu(sb
->s_sequence
);
481 next_log_block
= be32_to_cpu(sb
->s_start
);
483 first_commit_ID
= next_commit_ID
;
484 if (pass
== PASS_SCAN
)
485 info
->start_transaction
= first_commit_ID
;
487 jbd_debug(1, "Starting recovery pass %d\n", pass
);
490 * Now we walk through the log, transaction by transaction,
491 * making sure that each transaction has a commit block in the
492 * expected place. Each complete transaction gets replayed back
493 * into the main filesystem.
499 journal_block_tag_t
* tag
;
500 struct buffer_head
* obh
;
501 struct buffer_head
* nbh
;
505 /* If we already know where to stop the log traversal,
506 * check right now that we haven't gone past the end of
509 if (pass
!= PASS_SCAN
)
510 if (tid_geq(next_commit_ID
, info
->end_transaction
))
513 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
514 next_commit_ID
, next_log_block
,
515 jbd2_has_feature_fast_commit(journal
) ?
516 journal
->j_fc_last
: journal
->j_last
);
518 /* Skip over each chunk of the transaction looking
519 * either the next descriptor block or the final commit
522 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block
);
523 err
= jread(&bh
, journal
, next_log_block
);
528 wrap(journal
, next_log_block
);
530 /* What kind of buffer is it?
532 * If it is a descriptor block, check that it has the
533 * expected sequence number. Otherwise, we're all done
536 tmp
= (journal_header_t
*)bh
->b_data
;
538 if (tmp
->h_magic
!= cpu_to_be32(JBD2_MAGIC_NUMBER
)) {
543 blocktype
= be32_to_cpu(tmp
->h_blocktype
);
544 sequence
= be32_to_cpu(tmp
->h_sequence
);
545 jbd_debug(3, "Found magic %d, sequence %d\n",
546 blocktype
, sequence
);
548 if (sequence
!= next_commit_ID
) {
553 /* OK, we have a valid descriptor block which matches
554 * all of the sequence number checks. What are we going
555 * to do with it? That depends on the pass... */
558 case JBD2_DESCRIPTOR_BLOCK
:
559 /* Verify checksum first */
560 if (jbd2_journal_has_csum_v2or3(journal
))
562 sizeof(struct jbd2_journal_block_tail
);
563 if (descr_csum_size
> 0 &&
564 !jbd2_descriptor_block_csum_verify(journal
,
567 * PASS_SCAN can see stale blocks due to lazy
568 * journal init. Don't error out on those yet.
570 if (pass
!= PASS_SCAN
) {
571 pr_err("JBD2: Invalid checksum recovering block %lu in log\n",
577 need_check_commit_time
= true;
579 "invalid descriptor block found in %lu\n",
583 /* If it is a valid descriptor block, replay it
584 * in pass REPLAY; if journal_checksums enabled, then
585 * calculate checksums in PASS_SCAN, otherwise,
586 * just skip over the blocks it describes. */
587 if (pass
!= PASS_REPLAY
) {
588 if (pass
== PASS_SCAN
&&
589 jbd2_has_feature_checksum(journal
) &&
590 !need_check_commit_time
&&
591 !info
->end_transaction
) {
592 if (calc_chksums(journal
, bh
,
601 next_log_block
+= count_tags(journal
, bh
);
602 wrap(journal
, next_log_block
);
607 /* A descriptor block: we can now write all of
608 * the data blocks. Yay, useful work is finally
609 * getting done here! */
611 tagp
= &bh
->b_data
[sizeof(journal_header_t
)];
612 while ((tagp
- bh
->b_data
+ tag_bytes
)
613 <= journal
->j_blocksize
- descr_csum_size
) {
614 unsigned long io_block
;
616 tag
= (journal_block_tag_t
*) tagp
;
617 flags
= be16_to_cpu(tag
->t_flags
);
619 io_block
= next_log_block
++;
620 wrap(journal
, next_log_block
);
621 err
= jread(&obh
, journal
, io_block
);
623 /* Recover what we can, but
624 * report failure at the end. */
627 "JBD2: IO error %d recovering "
628 "block %ld in log\n",
631 unsigned long long blocknr
;
633 J_ASSERT(obh
!= NULL
);
634 blocknr
= read_tag_block(journal
,
637 /* If the block has been
638 * revoked, then we're all done
640 if (jbd2_journal_test_revoke
644 ++info
->nr_revoke_hits
;
648 /* Look for block corruption */
649 if (!jbd2_block_tag_csum_verify(
650 journal
, tag
, obh
->b_data
,
651 be32_to_cpu(tmp
->h_sequence
))) {
653 success
= -EFSBADCRC
;
654 printk(KERN_ERR
"JBD2: Invalid "
655 "checksum recovering "
656 "data block %llu in "
662 /* Find a buffer for the new
663 * data being restored */
664 nbh
= __getblk(journal
->j_fs_dev
,
666 journal
->j_blocksize
);
669 "JBD2: Out of memory "
670 "during recovery.\n");
678 memcpy(nbh
->b_data
, obh
->b_data
,
679 journal
->j_blocksize
);
680 if (flags
& JBD2_FLAG_ESCAPE
) {
681 *((__be32
*)nbh
->b_data
) =
682 cpu_to_be32(JBD2_MAGIC_NUMBER
);
685 BUFFER_TRACE(nbh
, "marking dirty");
686 set_buffer_uptodate(nbh
);
687 mark_buffer_dirty(nbh
);
688 BUFFER_TRACE(nbh
, "marking uptodate");
690 /* ll_rw_block(WRITE, 1, &nbh); */
698 if (!(flags
& JBD2_FLAG_SAME_UUID
))
701 if (flags
& JBD2_FLAG_LAST_TAG
)
708 case JBD2_COMMIT_BLOCK
:
709 /* How to differentiate between interrupted commit
710 * and journal corruption ?
713 * Checksum Verification Failed
715 * ____________________
717 * async_commit sync_commit
719 * | GO TO NEXT "Journal Corruption"
722 * {(n+1)th transanction}
724 * _______|______________
726 * Commit block found Commit block not found
728 * "Journal Corruption" |
729 * _____________|_________
731 * nth trans corrupt OR nth trans
732 * and (n+1)th interrupted interrupted
733 * before commit block
734 * could reach the disk.
735 * (Cannot find the difference in above
736 * mentioned conditions. Hence assume
737 * "Interrupted Commit".)
739 commit_time
= be64_to_cpu(
740 ((struct commit_header
*)bh
->b_data
)->h_commit_sec
);
742 * If need_check_commit_time is set, it means we are in
743 * PASS_SCAN and csum verify failed before. If
744 * commit_time is increasing, it's the same journal,
745 * otherwise it is stale journal block, just end this
748 if (need_check_commit_time
) {
749 if (commit_time
>= last_trans_commit_time
) {
750 pr_err("JBD2: Invalid checksum found in transaction %u\n",
758 * It likely does not belong to same journal,
759 * just end this recovery with success.
761 jbd_debug(1, "JBD2: Invalid checksum ignored in transaction %u, likely stale data\n",
769 * Found an expected commit block: if checksums
770 * are present, verify them in PASS_SCAN; else not
771 * much to do other than move on to the next sequence
774 if (pass
== PASS_SCAN
&&
775 jbd2_has_feature_checksum(journal
)) {
776 struct commit_header
*cbh
=
777 (struct commit_header
*)bh
->b_data
;
778 unsigned found_chksum
=
779 be32_to_cpu(cbh
->h_chksum
[0]);
781 if (info
->end_transaction
) {
782 journal
->j_failed_commit
=
783 info
->end_transaction
;
788 /* Neither checksum match nor unused? */
789 if (!((crc32_sum
== found_chksum
&&
790 cbh
->h_chksum_type
==
792 cbh
->h_chksum_size
==
793 JBD2_CRC32_CHKSUM_SIZE
) ||
794 (cbh
->h_chksum_type
== 0 &&
795 cbh
->h_chksum_size
== 0 &&
801 if (pass
== PASS_SCAN
&&
802 !jbd2_commit_block_csum_verify(journal
,
805 if (commit_time
< last_trans_commit_time
)
806 goto ignore_crc_mismatch
;
807 info
->end_transaction
= next_commit_ID
;
809 if (!jbd2_has_feature_async_commit(journal
)) {
810 journal
->j_failed_commit
=
816 if (pass
== PASS_SCAN
)
817 last_trans_commit_time
= commit_time
;
822 case JBD2_REVOKE_BLOCK
:
824 * Check revoke block crc in pass_scan, if csum verify
825 * failed, check commit block time later.
827 if (pass
== PASS_SCAN
&&
828 !jbd2_descriptor_block_csum_verify(journal
,
830 jbd_debug(1, "JBD2: invalid revoke block found in %lu\n",
832 need_check_commit_time
= true;
834 /* If we aren't in the REVOKE pass, then we can
835 * just skip over this block. */
836 if (pass
!= PASS_REVOKE
) {
841 err
= scan_revoke_records(journal
, bh
,
842 next_commit_ID
, info
);
849 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
858 * We broke out of the log scan loop: either we came to the
859 * known end of the log or we found an unexpected block in the
860 * log. If the latter happened, then we know that the "current"
861 * transaction marks the end of the valid log.
864 if (pass
== PASS_SCAN
) {
865 if (!info
->end_transaction
)
866 info
->end_transaction
= next_commit_ID
;
868 /* It's really bad news if different passes end up at
869 * different places (but possible due to IO errors). */
870 if (info
->end_transaction
!= next_commit_ID
) {
871 printk(KERN_ERR
"JBD2: recovery pass %d ended at "
872 "transaction %u, expected %u\n",
873 pass
, next_commit_ID
, info
->end_transaction
);
879 if (jbd2_has_feature_fast_commit(journal
) && pass
!= PASS_REVOKE
) {
880 err
= fc_do_one_pass(journal
, info
, pass
);
885 if (block_error
&& success
== 0)
893 /* Scan a revoke record, marking all blocks mentioned as revoked. */
895 static int scan_revoke_records(journal_t
*journal
, struct buffer_head
*bh
,
896 tid_t sequence
, struct recovery_info
*info
)
898 jbd2_journal_revoke_header_t
*header
;
904 header
= (jbd2_journal_revoke_header_t
*) bh
->b_data
;
905 offset
= sizeof(jbd2_journal_revoke_header_t
);
906 rcount
= be32_to_cpu(header
->r_count
);
908 if (jbd2_journal_has_csum_v2or3(journal
))
909 csum_size
= sizeof(struct jbd2_journal_block_tail
);
910 if (rcount
> journal
->j_blocksize
- csum_size
)
914 if (jbd2_has_feature_64bit(journal
))
917 while (offset
+ record_len
<= max
) {
918 unsigned long long blocknr
;
922 blocknr
= be32_to_cpu(* ((__be32
*) (bh
->b_data
+offset
)));
924 blocknr
= be64_to_cpu(* ((__be64
*) (bh
->b_data
+offset
)));
925 offset
+= record_len
;
926 err
= jbd2_journal_set_revoke(journal
, blocknr
, sequence
);