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 bh_readahead_batch(nbufs, bufs, 0);
104 journal_brelse_array(bufs, nbufs);
112 bh_readahead_batch(nbufs, bufs, 0);
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)) {
156 * If this is a brand new buffer, start readahead.
157 * Otherwise, we assume we are already reading it.
159 bool need_readahead = !buffer_req(bh);
161 bh_read_nowait(bh, 0);
163 do_readahead(journal, offset);
167 if (!buffer_uptodate(bh)) {
168 printk(KERN_ERR "JBD2: Failed to read block at offset %u\n",
178 static int jbd2_descriptor_block_csum_verify(journal_t *j, void *buf)
180 struct jbd2_journal_block_tail *tail;
184 if (!jbd2_journal_has_csum_v2or3(j))
187 tail = (struct jbd2_journal_block_tail *)((char *)buf +
188 j->j_blocksize - sizeof(struct jbd2_journal_block_tail));
189 provided = tail->t_checksum;
190 tail->t_checksum = 0;
191 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
192 tail->t_checksum = provided;
194 return provided == cpu_to_be32(calculated);
198 * Count the number of in-use tags in a journal descriptor block.
201 static int count_tags(journal_t *journal, struct buffer_head *bh)
204 journal_block_tag_t tag;
205 int nr = 0, size = journal->j_blocksize;
206 int tag_bytes = journal_tag_bytes(journal);
208 if (jbd2_journal_has_csum_v2or3(journal))
209 size -= sizeof(struct jbd2_journal_block_tail);
211 tagp = &bh->b_data[sizeof(journal_header_t)];
213 while ((tagp - bh->b_data + tag_bytes) <= size) {
214 memcpy(&tag, tagp, sizeof(tag));
218 if (!(tag.t_flags & cpu_to_be16(JBD2_FLAG_SAME_UUID)))
221 if (tag.t_flags & cpu_to_be16(JBD2_FLAG_LAST_TAG))
229 /* Make sure we wrap around the log correctly! */
230 #define wrap(journal, var) \
232 if (var >= (journal)->j_last) \
233 var -= ((journal)->j_last - (journal)->j_first); \
236 static int fc_do_one_pass(journal_t *journal,
237 struct recovery_info *info, enum passtype pass)
239 unsigned int expected_commit_id = info->end_transaction;
240 unsigned long next_fc_block;
241 struct buffer_head *bh;
244 next_fc_block = journal->j_fc_first;
245 if (!journal->j_fc_replay_callback)
248 while (next_fc_block <= journal->j_fc_last) {
249 jbd2_debug(3, "Fast commit replay: next block %ld\n",
251 err = jread(&bh, journal, next_fc_block);
253 jbd2_debug(3, "Fast commit replay: read error\n");
257 err = journal->j_fc_replay_callback(journal, bh, pass,
258 next_fc_block - journal->j_fc_first,
262 if (err < 0 || err == JBD2_FC_REPLAY_STOP)
268 jbd2_debug(3, "Fast commit replay failed, err = %d\n", err);
274 * jbd2_journal_recover - recovers a on-disk journal
275 * @journal: the journal to recover
277 * The primary function for recovering the log contents when mounting a
280 * Recovery is done in three passes. In the first pass, we look for the
281 * end of the log. In the second, we assemble the list of revoke
282 * blocks. In the third and final pass, we replay any un-revoked blocks
285 int jbd2_journal_recover(journal_t *journal)
288 journal_superblock_t * sb;
290 struct recovery_info info;
292 struct address_space *mapping;
294 memset(&info, 0, sizeof(info));
295 sb = journal->j_superblock;
298 * The journal superblock's s_start field (the current log head)
299 * is always zero if, and only if, the journal was cleanly
304 jbd2_debug(1, "No recovery required, last transaction %d\n",
305 be32_to_cpu(sb->s_sequence));
306 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
311 mapping = journal->j_fs_dev->bd_inode->i_mapping;
312 errseq_check_and_advance(&mapping->wb_err, &wb_err);
313 err = do_one_pass(journal, &info, PASS_SCAN);
315 err = do_one_pass(journal, &info, PASS_REVOKE);
317 err = do_one_pass(journal, &info, PASS_REPLAY);
319 jbd2_debug(1, "JBD2: recovery, exit status %d, "
320 "recovered transactions %u to %u\n",
321 err, info.start_transaction, info.end_transaction);
322 jbd2_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
323 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
325 /* Restart the log at the next transaction ID, thus invalidating
326 * any existing commit records in the log. */
327 journal->j_transaction_sequence = ++info.end_transaction;
329 jbd2_journal_clear_revoke(journal);
330 err2 = sync_blockdev(journal->j_fs_dev);
333 err2 = errseq_check_and_advance(&mapping->wb_err, &wb_err);
336 /* Make sure all replayed data is on permanent storage */
337 if (journal->j_flags & JBD2_BARRIER) {
338 err2 = blkdev_issue_flush(journal->j_fs_dev);
346 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
347 * @journal: journal to startup
349 * Locate any valid recovery information from the journal and set up the
350 * journal structures in memory to ignore it (presumably because the
351 * caller has evidence that it is out of date).
352 * This function doesn't appear to be exported..
354 * We perform one pass over the journal to allow us to tell the user how
355 * much recovery information is being erased, and to let us initialise
356 * the journal transaction sequence numbers to the next unused ID.
358 int jbd2_journal_skip_recovery(journal_t *journal)
362 struct recovery_info info;
364 memset (&info, 0, sizeof(info));
366 err = do_one_pass(journal, &info, PASS_SCAN);
369 printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
370 ++journal->j_transaction_sequence;
372 #ifdef CONFIG_JBD2_DEBUG
373 int dropped = info.end_transaction -
374 be32_to_cpu(journal->j_superblock->s_sequence);
376 "JBD2: ignoring %d transaction%s from the journal.\n",
377 dropped, (dropped == 1) ? "" : "s");
379 journal->j_transaction_sequence = ++info.end_transaction;
386 static inline unsigned long long read_tag_block(journal_t *journal,
387 journal_block_tag_t *tag)
389 unsigned long long block = be32_to_cpu(tag->t_blocknr);
390 if (jbd2_has_feature_64bit(journal))
391 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
396 * calc_chksums calculates the checksums for the blocks described in the
399 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
400 unsigned long *next_log_block, __u32 *crc32_sum)
402 int i, num_blks, err;
403 unsigned long io_block;
404 struct buffer_head *obh;
406 num_blks = count_tags(journal, bh);
407 /* Calculate checksum of the descriptor block. */
408 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
410 for (i = 0; i < num_blks; i++) {
411 io_block = (*next_log_block)++;
412 wrap(journal, *next_log_block);
413 err = jread(&obh, journal, io_block);
415 printk(KERN_ERR "JBD2: IO error %d recovering block "
416 "%lu in log\n", err, io_block);
419 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
427 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
429 struct commit_header *h;
433 if (!jbd2_journal_has_csum_v2or3(j))
437 provided = h->h_chksum[0];
439 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
440 h->h_chksum[0] = provided;
442 return provided == cpu_to_be32(calculated);
445 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
446 journal_block_tag3_t *tag3,
447 void *buf, __u32 sequence)
452 if (!jbd2_journal_has_csum_v2or3(j))
455 seq = cpu_to_be32(sequence);
456 csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
457 csum32 = jbd2_chksum(j, csum32, buf, j->j_blocksize);
459 if (jbd2_has_feature_csum3(j))
460 return tag3->t_checksum == cpu_to_be32(csum32);
462 return tag->t_checksum == cpu_to_be16(csum32);
465 static int do_one_pass(journal_t *journal,
466 struct recovery_info *info, enum passtype pass)
468 unsigned int first_commit_ID, next_commit_ID;
469 unsigned long next_log_block;
470 int err, success = 0;
471 journal_superblock_t * sb;
472 journal_header_t * tmp;
473 struct buffer_head * bh;
474 unsigned int sequence;
476 int tag_bytes = journal_tag_bytes(journal);
477 __u32 crc32_sum = ~0; /* Transactional Checksums */
478 int descr_csum_size = 0;
480 bool need_check_commit_time = false;
481 __u64 last_trans_commit_time = 0, commit_time;
484 * First thing is to establish what we expect to find in the log
485 * (in terms of transaction IDs), and where (in terms of log
486 * block offsets): query the superblock.
489 sb = journal->j_superblock;
490 next_commit_ID = be32_to_cpu(sb->s_sequence);
491 next_log_block = be32_to_cpu(sb->s_start);
493 first_commit_ID = next_commit_ID;
494 if (pass == PASS_SCAN)
495 info->start_transaction = first_commit_ID;
497 jbd2_debug(1, "Starting recovery pass %d\n", pass);
500 * Now we walk through the log, transaction by transaction,
501 * making sure that each transaction has a commit block in the
502 * expected place. Each complete transaction gets replayed back
503 * into the main filesystem.
509 journal_block_tag_t tag;
510 struct buffer_head * obh;
511 struct buffer_head * nbh;
515 /* If we already know where to stop the log traversal,
516 * check right now that we haven't gone past the end of
519 if (pass != PASS_SCAN)
520 if (tid_geq(next_commit_ID, info->end_transaction))
523 jbd2_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
524 next_commit_ID, next_log_block, journal->j_last);
526 /* Skip over each chunk of the transaction looking
527 * either the next descriptor block or the final commit
530 jbd2_debug(3, "JBD2: checking block %ld\n", next_log_block);
531 err = jread(&bh, journal, next_log_block);
536 wrap(journal, next_log_block);
538 /* What kind of buffer is it?
540 * If it is a descriptor block, check that it has the
541 * expected sequence number. Otherwise, we're all done
544 tmp = (journal_header_t *)bh->b_data;
546 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
551 blocktype = be32_to_cpu(tmp->h_blocktype);
552 sequence = be32_to_cpu(tmp->h_sequence);
553 jbd2_debug(3, "Found magic %d, sequence %d\n",
554 blocktype, sequence);
556 if (sequence != next_commit_ID) {
561 /* OK, we have a valid descriptor block which matches
562 * all of the sequence number checks. What are we going
563 * to do with it? That depends on the pass... */
566 case JBD2_DESCRIPTOR_BLOCK:
567 /* Verify checksum first */
568 if (jbd2_journal_has_csum_v2or3(journal))
570 sizeof(struct jbd2_journal_block_tail);
571 if (descr_csum_size > 0 &&
572 !jbd2_descriptor_block_csum_verify(journal,
575 * PASS_SCAN can see stale blocks due to lazy
576 * journal init. Don't error out on those yet.
578 if (pass != PASS_SCAN) {
579 pr_err("JBD2: Invalid checksum recovering block %lu in log\n",
585 need_check_commit_time = true;
587 "invalid descriptor block found in %lu\n",
591 /* If it is a valid descriptor block, replay it
592 * in pass REPLAY; if journal_checksums enabled, then
593 * calculate checksums in PASS_SCAN, otherwise,
594 * just skip over the blocks it describes. */
595 if (pass != PASS_REPLAY) {
596 if (pass == PASS_SCAN &&
597 jbd2_has_feature_checksum(journal) &&
598 !need_check_commit_time &&
599 !info->end_transaction) {
600 if (calc_chksums(journal, bh,
609 next_log_block += count_tags(journal, bh);
610 wrap(journal, next_log_block);
615 /* A descriptor block: we can now write all of
616 * the data blocks. Yay, useful work is finally
617 * getting done here! */
619 tagp = &bh->b_data[sizeof(journal_header_t)];
620 while ((tagp - bh->b_data + tag_bytes)
621 <= journal->j_blocksize - descr_csum_size) {
622 unsigned long io_block;
624 memcpy(&tag, tagp, sizeof(tag));
625 flags = be16_to_cpu(tag.t_flags);
627 io_block = next_log_block++;
628 wrap(journal, next_log_block);
629 err = jread(&obh, journal, io_block);
631 /* Recover what we can, but
632 * report failure at the end. */
635 "JBD2: IO error %d recovering "
636 "block %ld in log\n",
639 unsigned long long blocknr;
641 J_ASSERT(obh != NULL);
642 blocknr = read_tag_block(journal,
645 /* If the block has been
646 * revoked, then we're all done
648 if (jbd2_journal_test_revoke
652 ++info->nr_revoke_hits;
656 /* Look for block corruption */
657 if (!jbd2_block_tag_csum_verify(
658 journal, &tag, (journal_block_tag3_t *)tagp,
659 obh->b_data, be32_to_cpu(tmp->h_sequence))) {
661 success = -EFSBADCRC;
662 printk(KERN_ERR "JBD2: Invalid "
663 "checksum recovering "
664 "data block %llu in "
670 /* Find a buffer for the new
671 * data being restored */
672 nbh = __getblk(journal->j_fs_dev,
674 journal->j_blocksize);
677 "JBD2: Out of memory "
678 "during recovery.\n");
686 memcpy(nbh->b_data, obh->b_data,
687 journal->j_blocksize);
688 if (flags & JBD2_FLAG_ESCAPE) {
689 *((__be32 *)nbh->b_data) =
690 cpu_to_be32(JBD2_MAGIC_NUMBER);
693 BUFFER_TRACE(nbh, "marking dirty");
694 set_buffer_uptodate(nbh);
695 mark_buffer_dirty(nbh);
696 BUFFER_TRACE(nbh, "marking uptodate");
705 if (!(flags & JBD2_FLAG_SAME_UUID))
708 if (flags & JBD2_FLAG_LAST_TAG)
715 case JBD2_COMMIT_BLOCK:
716 /* How to differentiate between interrupted commit
717 * and journal corruption ?
720 * Checksum Verification Failed
722 * ____________________
724 * async_commit sync_commit
726 * | GO TO NEXT "Journal Corruption"
729 * {(n+1)th transanction}
731 * _______|______________
733 * Commit block found Commit block not found
735 * "Journal Corruption" |
736 * _____________|_________
738 * nth trans corrupt OR nth trans
739 * and (n+1)th interrupted interrupted
740 * before commit block
741 * could reach the disk.
742 * (Cannot find the difference in above
743 * mentioned conditions. Hence assume
744 * "Interrupted Commit".)
746 commit_time = be64_to_cpu(
747 ((struct commit_header *)bh->b_data)->h_commit_sec);
749 * If need_check_commit_time is set, it means we are in
750 * PASS_SCAN and csum verify failed before. If
751 * commit_time is increasing, it's the same journal,
752 * otherwise it is stale journal block, just end this
755 if (need_check_commit_time) {
756 if (commit_time >= last_trans_commit_time) {
757 pr_err("JBD2: Invalid checksum found in transaction %u\n",
765 * It likely does not belong to same journal,
766 * just end this recovery with success.
768 jbd2_debug(1, "JBD2: Invalid checksum ignored in transaction %u, likely stale data\n",
775 * Found an expected commit block: if checksums
776 * are present, verify them in PASS_SCAN; else not
777 * much to do other than move on to the next sequence
780 if (pass == PASS_SCAN &&
781 jbd2_has_feature_checksum(journal)) {
782 struct commit_header *cbh =
783 (struct commit_header *)bh->b_data;
784 unsigned found_chksum =
785 be32_to_cpu(cbh->h_chksum[0]);
787 if (info->end_transaction) {
788 journal->j_failed_commit =
789 info->end_transaction;
794 /* Neither checksum match nor unused? */
795 if (!((crc32_sum == found_chksum &&
796 cbh->h_chksum_type ==
798 cbh->h_chksum_size ==
799 JBD2_CRC32_CHKSUM_SIZE) ||
800 (cbh->h_chksum_type == 0 &&
801 cbh->h_chksum_size == 0 &&
807 if (pass == PASS_SCAN &&
808 !jbd2_commit_block_csum_verify(journal,
811 if (commit_time < last_trans_commit_time)
812 goto ignore_crc_mismatch;
813 info->end_transaction = next_commit_ID;
815 if (!jbd2_has_feature_async_commit(journal)) {
816 journal->j_failed_commit =
822 if (pass == PASS_SCAN)
823 last_trans_commit_time = commit_time;
828 case JBD2_REVOKE_BLOCK:
830 * Check revoke block crc in pass_scan, if csum verify
831 * failed, check commit block time later.
833 if (pass == PASS_SCAN &&
834 !jbd2_descriptor_block_csum_verify(journal,
836 jbd2_debug(1, "JBD2: invalid revoke block found in %lu\n",
838 need_check_commit_time = true;
840 /* If we aren't in the REVOKE pass, then we can
841 * just skip over this block. */
842 if (pass != PASS_REVOKE) {
847 err = scan_revoke_records(journal, bh,
848 next_commit_ID, info);
855 jbd2_debug(3, "Unrecognised magic %d, end of scan.\n",
864 * We broke out of the log scan loop: either we came to the
865 * known end of the log or we found an unexpected block in the
866 * log. If the latter happened, then we know that the "current"
867 * transaction marks the end of the valid log.
870 if (pass == PASS_SCAN) {
871 if (!info->end_transaction)
872 info->end_transaction = next_commit_ID;
874 /* It's really bad news if different passes end up at
875 * different places (but possible due to IO errors). */
876 if (info->end_transaction != next_commit_ID) {
877 printk(KERN_ERR "JBD2: recovery pass %d ended at "
878 "transaction %u, expected %u\n",
879 pass, next_commit_ID, info->end_transaction);
885 if (jbd2_has_feature_fast_commit(journal) && pass != PASS_REVOKE) {
886 err = fc_do_one_pass(journal, info, pass);
891 if (block_error && success == 0)
899 /* Scan a revoke record, marking all blocks mentioned as revoked. */
901 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
902 tid_t sequence, struct recovery_info *info)
904 jbd2_journal_revoke_header_t *header;
906 unsigned csum_size = 0;
910 header = (jbd2_journal_revoke_header_t *) bh->b_data;
911 offset = sizeof(jbd2_journal_revoke_header_t);
912 rcount = be32_to_cpu(header->r_count);
914 if (jbd2_journal_has_csum_v2or3(journal))
915 csum_size = sizeof(struct jbd2_journal_block_tail);
916 if (rcount > journal->j_blocksize - csum_size)
920 if (jbd2_has_feature_64bit(journal))
923 while (offset + record_len <= max) {
924 unsigned long long blocknr;
928 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
930 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
931 offset += record_len;
932 err = jbd2_journal_set_revoke(journal, blocknr, sequence);