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",
250 err = jread(&bh, journal, next_fc_block);
252 jbd_debug(3, "Fast commit replay: read error");
256 jbd_debug(3, "Processing fast commit blk with seq %d");
257 err = journal->j_fc_replay_callback(journal, bh, pass,
258 next_fc_block - journal->j_fc_first,
261 if (err < 0 || err == JBD2_FC_REPLAY_STOP)
267 jbd_debug(3, "Fast commit replay failed, err = %d\n", err);
273 * jbd2_journal_recover - recovers a on-disk journal
274 * @journal: the journal to recover
276 * The primary function for recovering the log contents when mounting a
279 * Recovery is done in three passes. In the first pass, we look for the
280 * end of the log. In the second, we assemble the list of revoke
281 * blocks. In the third and final pass, we replay any un-revoked blocks
284 int jbd2_journal_recover(journal_t *journal)
287 journal_superblock_t * sb;
289 struct recovery_info info;
291 memset(&info, 0, sizeof(info));
292 sb = journal->j_superblock;
295 * The journal superblock's s_start field (the current log head)
296 * is always zero if, and only if, the journal was cleanly
301 jbd_debug(1, "No recovery required, last transaction %d\n",
302 be32_to_cpu(sb->s_sequence));
303 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
307 err = do_one_pass(journal, &info, PASS_SCAN);
309 err = do_one_pass(journal, &info, PASS_REVOKE);
311 err = do_one_pass(journal, &info, PASS_REPLAY);
313 jbd_debug(1, "JBD2: recovery, exit status %d, "
314 "recovered transactions %u to %u\n",
315 err, info.start_transaction, info.end_transaction);
316 jbd_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
317 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
319 /* Restart the log at the next transaction ID, thus invalidating
320 * any existing commit records in the log. */
321 journal->j_transaction_sequence = ++info.end_transaction;
323 jbd2_journal_clear_revoke(journal);
324 err2 = sync_blockdev(journal->j_fs_dev);
327 /* Make sure all replayed data is on permanent storage */
328 if (journal->j_flags & JBD2_BARRIER) {
329 err2 = blkdev_issue_flush(journal->j_fs_dev);
337 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
338 * @journal: journal to startup
340 * Locate any valid recovery information from the journal and set up the
341 * journal structures in memory to ignore it (presumably because the
342 * caller has evidence that it is out of date).
343 * This function doesn't appear to be exported..
345 * We perform one pass over the journal to allow us to tell the user how
346 * much recovery information is being erased, and to let us initialise
347 * the journal transaction sequence numbers to the next unused ID.
349 int jbd2_journal_skip_recovery(journal_t *journal)
353 struct recovery_info info;
355 memset (&info, 0, sizeof(info));
357 err = do_one_pass(journal, &info, PASS_SCAN);
360 printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
361 ++journal->j_transaction_sequence;
363 #ifdef CONFIG_JBD2_DEBUG
364 int dropped = info.end_transaction -
365 be32_to_cpu(journal->j_superblock->s_sequence);
367 "JBD2: ignoring %d transaction%s from the journal.\n",
368 dropped, (dropped == 1) ? "" : "s");
370 journal->j_transaction_sequence = ++info.end_transaction;
377 static inline unsigned long long read_tag_block(journal_t *journal,
378 journal_block_tag_t *tag)
380 unsigned long long block = be32_to_cpu(tag->t_blocknr);
381 if (jbd2_has_feature_64bit(journal))
382 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
387 * calc_chksums calculates the checksums for the blocks described in the
390 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
391 unsigned long *next_log_block, __u32 *crc32_sum)
393 int i, num_blks, err;
394 unsigned long io_block;
395 struct buffer_head *obh;
397 num_blks = count_tags(journal, bh);
398 /* Calculate checksum of the descriptor block. */
399 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
401 for (i = 0; i < num_blks; i++) {
402 io_block = (*next_log_block)++;
403 wrap(journal, *next_log_block);
404 err = jread(&obh, journal, io_block);
406 printk(KERN_ERR "JBD2: IO error %d recovering block "
407 "%lu in log\n", err, io_block);
410 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
418 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
420 struct commit_header *h;
424 if (!jbd2_journal_has_csum_v2or3(j))
428 provided = h->h_chksum[0];
430 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
431 h->h_chksum[0] = provided;
433 return provided == cpu_to_be32(calculated);
436 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
437 void *buf, __u32 sequence)
439 journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
443 if (!jbd2_journal_has_csum_v2or3(j))
446 seq = cpu_to_be32(sequence);
447 csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
448 csum32 = jbd2_chksum(j, csum32, buf, j->j_blocksize);
450 if (jbd2_has_feature_csum3(j))
451 return tag3->t_checksum == cpu_to_be32(csum32);
453 return tag->t_checksum == cpu_to_be16(csum32);
456 static int do_one_pass(journal_t *journal,
457 struct recovery_info *info, enum passtype pass)
459 unsigned int first_commit_ID, next_commit_ID;
460 unsigned long next_log_block;
461 int err, success = 0;
462 journal_superblock_t * sb;
463 journal_header_t * tmp;
464 struct buffer_head * bh;
465 unsigned int sequence;
467 int tag_bytes = journal_tag_bytes(journal);
468 __u32 crc32_sum = ~0; /* Transactional Checksums */
469 int descr_csum_size = 0;
471 bool need_check_commit_time = false;
472 __u64 last_trans_commit_time = 0, commit_time;
475 * First thing is to establish what we expect to find in the log
476 * (in terms of transaction IDs), and where (in terms of log
477 * block offsets): query the superblock.
480 sb = journal->j_superblock;
481 next_commit_ID = be32_to_cpu(sb->s_sequence);
482 next_log_block = be32_to_cpu(sb->s_start);
484 first_commit_ID = next_commit_ID;
485 if (pass == PASS_SCAN)
486 info->start_transaction = first_commit_ID;
488 jbd_debug(1, "Starting recovery pass %d\n", pass);
491 * Now we walk through the log, transaction by transaction,
492 * making sure that each transaction has a commit block in the
493 * expected place. Each complete transaction gets replayed back
494 * into the main filesystem.
500 journal_block_tag_t * tag;
501 struct buffer_head * obh;
502 struct buffer_head * nbh;
506 /* If we already know where to stop the log traversal,
507 * check right now that we haven't gone past the end of
510 if (pass != PASS_SCAN)
511 if (tid_geq(next_commit_ID, info->end_transaction))
514 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
515 next_commit_ID, next_log_block,
516 jbd2_has_feature_fast_commit(journal) ?
517 journal->j_fc_last : journal->j_last);
519 /* Skip over each chunk of the transaction looking
520 * either the next descriptor block or the final commit
523 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
524 err = jread(&bh, journal, next_log_block);
529 wrap(journal, next_log_block);
531 /* What kind of buffer is it?
533 * If it is a descriptor block, check that it has the
534 * expected sequence number. Otherwise, we're all done
537 tmp = (journal_header_t *)bh->b_data;
539 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
544 blocktype = be32_to_cpu(tmp->h_blocktype);
545 sequence = be32_to_cpu(tmp->h_sequence);
546 jbd_debug(3, "Found magic %d, sequence %d\n",
547 blocktype, sequence);
549 if (sequence != next_commit_ID) {
554 /* OK, we have a valid descriptor block which matches
555 * all of the sequence number checks. What are we going
556 * to do with it? That depends on the pass... */
559 case JBD2_DESCRIPTOR_BLOCK:
560 /* Verify checksum first */
561 if (jbd2_journal_has_csum_v2or3(journal))
563 sizeof(struct jbd2_journal_block_tail);
564 if (descr_csum_size > 0 &&
565 !jbd2_descriptor_block_csum_verify(journal,
568 * PASS_SCAN can see stale blocks due to lazy
569 * journal init. Don't error out on those yet.
571 if (pass != PASS_SCAN) {
572 pr_err("JBD2: Invalid checksum recovering block %lu in log\n",
578 need_check_commit_time = true;
580 "invalid descriptor block found in %lu\n",
584 /* If it is a valid descriptor block, replay it
585 * in pass REPLAY; if journal_checksums enabled, then
586 * calculate checksums in PASS_SCAN, otherwise,
587 * just skip over the blocks it describes. */
588 if (pass != PASS_REPLAY) {
589 if (pass == PASS_SCAN &&
590 jbd2_has_feature_checksum(journal) &&
591 !need_check_commit_time &&
592 !info->end_transaction) {
593 if (calc_chksums(journal, bh,
602 next_log_block += count_tags(journal, bh);
603 wrap(journal, next_log_block);
608 /* A descriptor block: we can now write all of
609 * the data blocks. Yay, useful work is finally
610 * getting done here! */
612 tagp = &bh->b_data[sizeof(journal_header_t)];
613 while ((tagp - bh->b_data + tag_bytes)
614 <= journal->j_blocksize - descr_csum_size) {
615 unsigned long io_block;
617 tag = (journal_block_tag_t *) tagp;
618 flags = be16_to_cpu(tag->t_flags);
620 io_block = next_log_block++;
621 wrap(journal, next_log_block);
622 err = jread(&obh, journal, io_block);
624 /* Recover what we can, but
625 * report failure at the end. */
628 "JBD2: IO error %d recovering "
629 "block %ld in log\n",
632 unsigned long long blocknr;
634 J_ASSERT(obh != NULL);
635 blocknr = read_tag_block(journal,
638 /* If the block has been
639 * revoked, then we're all done
641 if (jbd2_journal_test_revoke
645 ++info->nr_revoke_hits;
649 /* Look for block corruption */
650 if (!jbd2_block_tag_csum_verify(
651 journal, tag, obh->b_data,
652 be32_to_cpu(tmp->h_sequence))) {
654 success = -EFSBADCRC;
655 printk(KERN_ERR "JBD2: Invalid "
656 "checksum recovering "
657 "data block %llu in "
663 /* Find a buffer for the new
664 * data being restored */
665 nbh = __getblk(journal->j_fs_dev,
667 journal->j_blocksize);
670 "JBD2: Out of memory "
671 "during recovery.\n");
679 memcpy(nbh->b_data, obh->b_data,
680 journal->j_blocksize);
681 if (flags & JBD2_FLAG_ESCAPE) {
682 *((__be32 *)nbh->b_data) =
683 cpu_to_be32(JBD2_MAGIC_NUMBER);
686 BUFFER_TRACE(nbh, "marking dirty");
687 set_buffer_uptodate(nbh);
688 mark_buffer_dirty(nbh);
689 BUFFER_TRACE(nbh, "marking uptodate");
691 /* ll_rw_block(WRITE, 1, &nbh); */
699 if (!(flags & JBD2_FLAG_SAME_UUID))
702 if (flags & JBD2_FLAG_LAST_TAG)
709 case JBD2_COMMIT_BLOCK:
710 /* How to differentiate between interrupted commit
711 * and journal corruption ?
714 * Checksum Verification Failed
716 * ____________________
718 * async_commit sync_commit
720 * | GO TO NEXT "Journal Corruption"
723 * {(n+1)th transanction}
725 * _______|______________
727 * Commit block found Commit block not found
729 * "Journal Corruption" |
730 * _____________|_________
732 * nth trans corrupt OR nth trans
733 * and (n+1)th interrupted interrupted
734 * before commit block
735 * could reach the disk.
736 * (Cannot find the difference in above
737 * mentioned conditions. Hence assume
738 * "Interrupted Commit".)
740 commit_time = be64_to_cpu(
741 ((struct commit_header *)bh->b_data)->h_commit_sec);
743 * If need_check_commit_time is set, it means we are in
744 * PASS_SCAN and csum verify failed before. If
745 * commit_time is increasing, it's the same journal,
746 * otherwise it is stale journal block, just end this
749 if (need_check_commit_time) {
750 if (commit_time >= last_trans_commit_time) {
751 pr_err("JBD2: Invalid checksum found in transaction %u\n",
759 * It likely does not belong to same journal,
760 * just end this recovery with success.
762 jbd_debug(1, "JBD2: Invalid checksum ignored in transaction %u, likely stale data\n",
770 * Found an expected commit block: if checksums
771 * are present, verify them in PASS_SCAN; else not
772 * much to do other than move on to the next sequence
775 if (pass == PASS_SCAN &&
776 jbd2_has_feature_checksum(journal)) {
777 struct commit_header *cbh =
778 (struct commit_header *)bh->b_data;
779 unsigned found_chksum =
780 be32_to_cpu(cbh->h_chksum[0]);
782 if (info->end_transaction) {
783 journal->j_failed_commit =
784 info->end_transaction;
789 /* Neither checksum match nor unused? */
790 if (!((crc32_sum == found_chksum &&
791 cbh->h_chksum_type ==
793 cbh->h_chksum_size ==
794 JBD2_CRC32_CHKSUM_SIZE) ||
795 (cbh->h_chksum_type == 0 &&
796 cbh->h_chksum_size == 0 &&
802 if (pass == PASS_SCAN &&
803 !jbd2_commit_block_csum_verify(journal,
806 if (commit_time < last_trans_commit_time)
807 goto ignore_crc_mismatch;
808 info->end_transaction = next_commit_ID;
810 if (!jbd2_has_feature_async_commit(journal)) {
811 journal->j_failed_commit =
817 if (pass == PASS_SCAN)
818 last_trans_commit_time = commit_time;
823 case JBD2_REVOKE_BLOCK:
825 * Check revoke block crc in pass_scan, if csum verify
826 * failed, check commit block time later.
828 if (pass == PASS_SCAN &&
829 !jbd2_descriptor_block_csum_verify(journal,
831 jbd_debug(1, "JBD2: invalid revoke block found in %lu\n",
833 need_check_commit_time = true;
835 /* If we aren't in the REVOKE pass, then we can
836 * just skip over this block. */
837 if (pass != PASS_REVOKE) {
842 err = scan_revoke_records(journal, bh,
843 next_commit_ID, info);
850 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
859 * We broke out of the log scan loop: either we came to the
860 * known end of the log or we found an unexpected block in the
861 * log. If the latter happened, then we know that the "current"
862 * transaction marks the end of the valid log.
865 if (pass == PASS_SCAN) {
866 if (!info->end_transaction)
867 info->end_transaction = next_commit_ID;
869 /* It's really bad news if different passes end up at
870 * different places (but possible due to IO errors). */
871 if (info->end_transaction != next_commit_ID) {
872 printk(KERN_ERR "JBD2: recovery pass %d ended at "
873 "transaction %u, expected %u\n",
874 pass, next_commit_ID, info->end_transaction);
880 if (jbd2_has_feature_fast_commit(journal) && pass != PASS_REVOKE) {
881 err = fc_do_one_pass(journal, info, pass);
886 if (block_error && success == 0)
894 /* Scan a revoke record, marking all blocks mentioned as revoked. */
896 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
897 tid_t sequence, struct recovery_info *info)
899 jbd2_journal_revoke_header_t *header;
905 header = (jbd2_journal_revoke_header_t *) bh->b_data;
906 offset = sizeof(jbd2_journal_revoke_header_t);
907 rcount = be32_to_cpu(header->r_count);
909 if (jbd2_journal_has_csum_v2or3(journal))
910 csum_size = sizeof(struct jbd2_journal_block_tail);
911 if (rcount > journal->j_blocksize - csum_size)
915 if (jbd2_has_feature_64bit(journal))
918 while (offset + record_len <= max) {
919 unsigned long long blocknr;
923 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
925 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
926 offset += record_len;
927 err = jbd2_journal_set_revoke(journal, blocknr, sequence);