2 * linux/fs/jbd2/recovery.c
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
6 * Copyright 1999-2000 Red Hat Software --- All Rights Reserved
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.
12 * Journal recovery routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
19 #include <linux/time.h>
21 #include <linux/jbd2.h>
22 #include <linux/errno.h>
23 #include <linux/crc32.h>
24 #include <linux/blkdev.h>
28 * Maintain information about the progress of the recovery job, so that
29 * the different passes can carry information between them.
33 tid_t start_transaction;
34 tid_t end_transaction;
41 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
42 static int do_one_pass(journal_t *journal,
43 struct recovery_info *info, enum passtype pass);
44 static int scan_revoke_records(journal_t *, struct buffer_head *,
45 tid_t, struct recovery_info *);
49 /* Release readahead buffers after use */
50 static void journal_brelse_array(struct buffer_head *b[], int n)
58 * When reading from the journal, we are going through the block device
59 * layer directly and so there is no readahead being done for us. We
60 * need to implement any readahead ourselves if we want it to happen at
61 * all. Recovery is basically one long sequential read, so make sure we
62 * do the IO in reasonably large chunks.
64 * This is not so critical that we need to be enormously clever about
65 * the readahead size, though. 128K is a purely arbitrary, good-enough
70 static int do_readahead(journal_t *journal, unsigned int start)
73 unsigned int max, nbufs, next;
74 unsigned long long blocknr;
75 struct buffer_head *bh;
77 struct buffer_head * bufs[MAXBUF];
79 /* Do up to 128K of readahead */
80 max = start + (128 * 1024 / journal->j_blocksize);
81 if (max > journal->j_maxlen)
82 max = journal->j_maxlen;
84 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
85 * a time to the block device IO layer. */
89 for (next = start; next < max; next++) {
90 err = jbd2_journal_bmap(journal, next, &blocknr);
93 printk(KERN_ERR "JBD2: bad block at offset %u\n",
98 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
104 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
106 if (nbufs == MAXBUF) {
107 ll_rw_block(READ, nbufs, bufs);
108 journal_brelse_array(bufs, nbufs);
116 ll_rw_block(READ, nbufs, bufs);
121 journal_brelse_array(bufs, nbufs);
125 #endif /* __KERNEL__ */
129 * Read a block from the journal
132 static int jread(struct buffer_head **bhp, journal_t *journal,
136 unsigned long long blocknr;
137 struct buffer_head *bh;
141 if (offset >= journal->j_maxlen) {
142 printk(KERN_ERR "JBD2: corrupted journal superblock\n");
146 err = jbd2_journal_bmap(journal, offset, &blocknr);
149 printk(KERN_ERR "JBD2: bad block at offset %u\n",
154 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
158 if (!buffer_uptodate(bh)) {
159 /* If this is a brand new buffer, start readahead.
160 Otherwise, we assume we are already reading it. */
162 do_readahead(journal, offset);
166 if (!buffer_uptodate(bh)) {
167 printk(KERN_ERR "JBD2: Failed to read block at offset %u\n",
177 static int jbd2_descr_block_csum_verify(journal_t *j,
180 struct jbd2_journal_block_tail *tail;
181 __u32 provided, calculated;
183 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
186 tail = (struct jbd2_journal_block_tail *)(buf + j->j_blocksize -
187 sizeof(struct jbd2_journal_block_tail));
188 provided = tail->t_checksum;
189 tail->t_checksum = 0;
190 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
191 tail->t_checksum = provided;
193 provided = be32_to_cpu(provided);
194 return provided == 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_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
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 tag = (journal_block_tag_t *) tagp;
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); \
237 * jbd2_journal_recover - recovers a on-disk journal
238 * @journal: the journal to recover
240 * The primary function for recovering the log contents when mounting a
243 * Recovery is done in three passes. In the first pass, we look for the
244 * end of the log. In the second, we assemble the list of revoke
245 * blocks. In the third and final pass, we replay any un-revoked blocks
248 int jbd2_journal_recover(journal_t *journal)
251 journal_superblock_t * sb;
253 struct recovery_info info;
255 memset(&info, 0, sizeof(info));
256 sb = journal->j_superblock;
259 * The journal superblock's s_start field (the current log head)
260 * is always zero if, and only if, the journal was cleanly
265 jbd_debug(1, "No recovery required, last transaction %d\n",
266 be32_to_cpu(sb->s_sequence));
267 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
271 err = do_one_pass(journal, &info, PASS_SCAN);
273 err = do_one_pass(journal, &info, PASS_REVOKE);
275 err = do_one_pass(journal, &info, PASS_REPLAY);
277 jbd_debug(1, "JBD2: recovery, exit status %d, "
278 "recovered transactions %u to %u\n",
279 err, info.start_transaction, info.end_transaction);
280 jbd_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
281 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
283 /* Restart the log at the next transaction ID, thus invalidating
284 * any existing commit records in the log. */
285 journal->j_transaction_sequence = ++info.end_transaction;
287 jbd2_journal_clear_revoke(journal);
288 err2 = sync_blockdev(journal->j_fs_dev);
291 /* Make sure all replayed data is on permanent storage */
292 if (journal->j_flags & JBD2_BARRIER) {
293 err2 = blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
301 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
302 * @journal: journal to startup
304 * Locate any valid recovery information from the journal and set up the
305 * journal structures in memory to ignore it (presumably because the
306 * caller has evidence that it is out of date).
307 * This function does'nt appear to be exorted..
309 * We perform one pass over the journal to allow us to tell the user how
310 * much recovery information is being erased, and to let us initialise
311 * the journal transaction sequence numbers to the next unused ID.
313 int jbd2_journal_skip_recovery(journal_t *journal)
317 struct recovery_info info;
319 memset (&info, 0, sizeof(info));
321 err = do_one_pass(journal, &info, PASS_SCAN);
324 printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
325 ++journal->j_transaction_sequence;
327 #ifdef CONFIG_JBD2_DEBUG
328 int dropped = info.end_transaction -
329 be32_to_cpu(journal->j_superblock->s_sequence);
331 "JBD2: ignoring %d transaction%s from the journal.\n",
332 dropped, (dropped == 1) ? "" : "s");
334 journal->j_transaction_sequence = ++info.end_transaction;
341 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
343 unsigned long long block = be32_to_cpu(tag->t_blocknr);
344 if (tag_bytes > JBD2_TAG_SIZE32)
345 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
350 * calc_chksums calculates the checksums for the blocks described in the
353 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
354 unsigned long *next_log_block, __u32 *crc32_sum)
356 int i, num_blks, err;
357 unsigned long io_block;
358 struct buffer_head *obh;
360 num_blks = count_tags(journal, bh);
361 /* Calculate checksum of the descriptor block. */
362 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
364 for (i = 0; i < num_blks; i++) {
365 io_block = (*next_log_block)++;
366 wrap(journal, *next_log_block);
367 err = jread(&obh, journal, io_block);
369 printk(KERN_ERR "JBD2: IO error %d recovering block "
370 "%lu in log\n", err, io_block);
373 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
381 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
383 struct commit_header *h;
384 __u32 provided, calculated;
386 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
390 provided = h->h_chksum[0];
392 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
393 h->h_chksum[0] = provided;
395 provided = be32_to_cpu(provided);
396 return provided == calculated;
399 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
400 void *buf, __u32 sequence)
402 __u32 provided, calculated;
404 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
407 sequence = cpu_to_be32(sequence);
408 calculated = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&sequence,
410 calculated = jbd2_chksum(j, calculated, buf, j->j_blocksize);
411 provided = be32_to_cpu(tag->t_checksum);
413 return provided == cpu_to_be32(calculated);
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;
433 * First thing is to establish what we expect to find in the log
434 * (in terms of transaction IDs), and where (in terms of log
435 * block offsets): query the superblock.
438 sb = journal->j_superblock;
439 next_commit_ID = be32_to_cpu(sb->s_sequence);
440 next_log_block = be32_to_cpu(sb->s_start);
442 first_commit_ID = next_commit_ID;
443 if (pass == PASS_SCAN)
444 info->start_transaction = first_commit_ID;
446 jbd_debug(1, "Starting recovery pass %d\n", pass);
449 * Now we walk through the log, transaction by transaction,
450 * making sure that each transaction has a commit block in the
451 * expected place. Each complete transaction gets replayed back
452 * into the main filesystem.
458 journal_block_tag_t * tag;
459 struct buffer_head * obh;
460 struct buffer_head * nbh;
464 /* If we already know where to stop the log traversal,
465 * check right now that we haven't gone past the end of
468 if (pass != PASS_SCAN)
469 if (tid_geq(next_commit_ID, info->end_transaction))
472 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
473 next_commit_ID, next_log_block, journal->j_last);
475 /* Skip over each chunk of the transaction looking
476 * either the next descriptor block or the final commit
479 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
480 err = jread(&bh, journal, next_log_block);
485 wrap(journal, next_log_block);
487 /* What kind of buffer is it?
489 * If it is a descriptor block, check that it has the
490 * expected sequence number. Otherwise, we're all done
493 tmp = (journal_header_t *)bh->b_data;
495 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
500 blocktype = be32_to_cpu(tmp->h_blocktype);
501 sequence = be32_to_cpu(tmp->h_sequence);
502 jbd_debug(3, "Found magic %d, sequence %d\n",
503 blocktype, sequence);
505 if (sequence != next_commit_ID) {
510 /* OK, we have a valid descriptor block which matches
511 * all of the sequence number checks. What are we going
512 * to do with it? That depends on the pass... */
515 case JBD2_DESCRIPTOR_BLOCK:
516 /* Verify checksum first */
517 if (JBD2_HAS_INCOMPAT_FEATURE(journal,
518 JBD2_FEATURE_INCOMPAT_CSUM_V2))
520 sizeof(struct jbd2_journal_block_tail);
521 if (descr_csum_size > 0 &&
522 !jbd2_descr_block_csum_verify(journal,
528 /* If it is a valid descriptor block, replay it
529 * in pass REPLAY; if journal_checksums enabled, then
530 * calculate checksums in PASS_SCAN, otherwise,
531 * just skip over the blocks it describes. */
532 if (pass != PASS_REPLAY) {
533 if (pass == PASS_SCAN &&
534 JBD2_HAS_COMPAT_FEATURE(journal,
535 JBD2_FEATURE_COMPAT_CHECKSUM) &&
536 !info->end_transaction) {
537 if (calc_chksums(journal, bh,
546 next_log_block += count_tags(journal, bh);
547 wrap(journal, next_log_block);
552 /* A descriptor block: we can now write all of
553 * the data blocks. Yay, useful work is finally
554 * getting done here! */
556 tagp = &bh->b_data[sizeof(journal_header_t)];
557 while ((tagp - bh->b_data + tag_bytes)
558 <= journal->j_blocksize - descr_csum_size) {
559 unsigned long io_block;
561 tag = (journal_block_tag_t *) tagp;
562 flags = be16_to_cpu(tag->t_flags);
564 io_block = next_log_block++;
565 wrap(journal, next_log_block);
566 err = jread(&obh, journal, io_block);
568 /* Recover what we can, but
569 * report failure at the end. */
572 "JBD2: IO error %d recovering "
573 "block %ld in log\n",
576 unsigned long long blocknr;
578 J_ASSERT(obh != NULL);
579 blocknr = read_tag_block(tag_bytes,
582 /* If the block has been
583 * revoked, then we're all done
585 if (jbd2_journal_test_revoke
589 ++info->nr_revoke_hits;
593 /* Look for block corruption */
594 if (!jbd2_block_tag_csum_verify(
595 journal, tag, obh->b_data,
596 be32_to_cpu(tmp->h_sequence))) {
599 printk(KERN_ERR "JBD: Invalid "
600 "checksum recovering "
601 "block %llu in log\n",
607 /* Find a buffer for the new
608 * data being restored */
609 nbh = __getblk(journal->j_fs_dev,
611 journal->j_blocksize);
614 "JBD2: Out of memory "
615 "during recovery.\n");
623 memcpy(nbh->b_data, obh->b_data,
624 journal->j_blocksize);
625 if (flags & JBD2_FLAG_ESCAPE) {
626 *((__be32 *)nbh->b_data) =
627 cpu_to_be32(JBD2_MAGIC_NUMBER);
630 BUFFER_TRACE(nbh, "marking dirty");
631 set_buffer_uptodate(nbh);
632 mark_buffer_dirty(nbh);
633 BUFFER_TRACE(nbh, "marking uptodate");
635 /* ll_rw_block(WRITE, 1, &nbh); */
643 if (!(flags & JBD2_FLAG_SAME_UUID))
646 if (flags & JBD2_FLAG_LAST_TAG)
653 case JBD2_COMMIT_BLOCK:
654 /* How to differentiate between interrupted commit
655 * and journal corruption ?
658 * Checksum Verification Failed
660 * ____________________
662 * async_commit sync_commit
664 * | GO TO NEXT "Journal Corruption"
667 * {(n+1)th transanction}
669 * _______|______________
671 * Commit block found Commit block not found
673 * "Journal Corruption" |
674 * _____________|_________
676 * nth trans corrupt OR nth trans
677 * and (n+1)th interrupted interrupted
678 * before commit block
679 * could reach the disk.
680 * (Cannot find the difference in above
681 * mentioned conditions. Hence assume
682 * "Interrupted Commit".)
685 /* Found an expected commit block: if checksums
686 * are present verify them in PASS_SCAN; else not
687 * much to do other than move on to the next sequence
689 if (pass == PASS_SCAN &&
690 JBD2_HAS_COMPAT_FEATURE(journal,
691 JBD2_FEATURE_COMPAT_CHECKSUM)) {
692 int chksum_err, chksum_seen;
693 struct commit_header *cbh =
694 (struct commit_header *)bh->b_data;
695 unsigned found_chksum =
696 be32_to_cpu(cbh->h_chksum[0]);
698 chksum_err = chksum_seen = 0;
700 if (info->end_transaction) {
701 journal->j_failed_commit =
702 info->end_transaction;
707 if (crc32_sum == found_chksum &&
708 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
709 cbh->h_chksum_size ==
710 JBD2_CRC32_CHKSUM_SIZE)
712 else if (!(cbh->h_chksum_type == 0 &&
713 cbh->h_chksum_size == 0 &&
717 * If fs is mounted using an old kernel and then
718 * kernel with journal_chksum is used then we
719 * get a situation where the journal flag has
720 * checksum flag set but checksums are not
721 * present i.e chksum = 0, in the individual
723 * Hence to avoid checksum failures, in this
724 * situation, this extra check is added.
729 info->end_transaction = next_commit_ID;
731 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
732 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
733 journal->j_failed_commit =
741 if (pass == PASS_SCAN &&
742 !jbd2_commit_block_csum_verify(journal,
744 info->end_transaction = next_commit_ID;
746 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
747 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
748 journal->j_failed_commit =
758 case JBD2_REVOKE_BLOCK:
759 /* If we aren't in the REVOKE pass, then we can
760 * just skip over this block. */
761 if (pass != PASS_REVOKE) {
766 err = scan_revoke_records(journal, bh,
767 next_commit_ID, info);
774 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
783 * We broke out of the log scan loop: either we came to the
784 * known end of the log or we found an unexpected block in the
785 * log. If the latter happened, then we know that the "current"
786 * transaction marks the end of the valid log.
789 if (pass == PASS_SCAN) {
790 if (!info->end_transaction)
791 info->end_transaction = next_commit_ID;
793 /* It's really bad news if different passes end up at
794 * different places (but possible due to IO errors). */
795 if (info->end_transaction != next_commit_ID) {
796 printk(KERN_ERR "JBD2: recovery pass %d ended at "
797 "transaction %u, expected %u\n",
798 pass, next_commit_ID, info->end_transaction);
803 if (block_error && success == 0)
811 static int jbd2_revoke_block_csum_verify(journal_t *j,
814 struct jbd2_journal_revoke_tail *tail;
815 __u32 provided, calculated;
817 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
820 tail = (struct jbd2_journal_revoke_tail *)(buf + j->j_blocksize -
821 sizeof(struct jbd2_journal_revoke_tail));
822 provided = tail->r_checksum;
823 tail->r_checksum = 0;
824 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
825 tail->r_checksum = provided;
827 provided = be32_to_cpu(provided);
828 return provided == calculated;
831 /* Scan a revoke record, marking all blocks mentioned as revoked. */
833 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
834 tid_t sequence, struct recovery_info *info)
836 jbd2_journal_revoke_header_t *header;
840 header = (jbd2_journal_revoke_header_t *) bh->b_data;
841 offset = sizeof(jbd2_journal_revoke_header_t);
842 max = be32_to_cpu(header->r_count);
844 if (!jbd2_revoke_block_csum_verify(journal, header))
847 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
850 while (offset + record_len <= max) {
851 unsigned long long blocknr;
855 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
857 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
858 offset += record_len;
859 err = jbd2_journal_set_revoke(journal, blocknr, sequence);