2 * linux/fs/jbd/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.
20 #include <linux/time.h>
22 #include <linux/jbd.h>
23 #include <linux/errno.h>
24 #include <linux/slab.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 = journal_bmap(journal, next, &blocknr);
93 printk (KERN_ERR "JBD: 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 "JBD: corrupted journal superblock\n");
146 err = journal_bmap(journal, offset, &blocknr);
149 printk (KERN_ERR "JBD: 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 "JBD: Failed to read block at offset %u\n",
177 static int jbd2_descr_block_csum_verify(journal_t *j,
180 struct journal_block_tail *tail;
181 __u32 provided, calculated;
183 if (!JFS_HAS_INCOMPAT_FEATURE(j, JFS_FEATURE_INCOMPAT_CSUM_V2))
186 tail = (struct journal_block_tail *)(buf + j->j_blocksize -
187 sizeof(struct journal_block_tail));
188 provided = tail->t_checksum;
189 tail->t_checksum = 0;
190 calculated = ext2fs_crc32c_le(~0, j->j_superblock->s_uuid,
191 sizeof(j->j_superblock->s_uuid));
192 calculated = ext2fs_crc32c_le(calculated, buf, j->j_blocksize);
193 tail->t_checksum = provided;
195 provided = ext2fs_be32_to_cpu(provided);
196 return provided == calculated;
200 * Count the number of in-use tags in a journal descriptor block.
203 static int count_tags(journal_t *journal, struct buffer_head *bh)
206 journal_block_tag_t * tag;
207 int nr = 0, size = journal->j_blocksize;
208 int tag_bytes = journal_tag_bytes(journal);
210 if (JFS_HAS_INCOMPAT_FEATURE(journal, JFS_FEATURE_INCOMPAT_CSUM_V2))
211 size -= sizeof(struct journal_block_tail);
213 tagp = &bh->b_data[sizeof(journal_header_t)];
215 while ((tagp - bh->b_data + tag_bytes) <= size) {
216 tag = (journal_block_tag_t *) tagp;
220 if (!(tag->t_flags & cpu_to_be16(JFS_FLAG_SAME_UUID)))
223 if (tag->t_flags & cpu_to_be16(JFS_FLAG_LAST_TAG))
231 /* Make sure we wrap around the log correctly! */
232 #define wrap(journal, var) \
234 if (var >= (journal)->j_last) \
235 var -= ((journal)->j_last - (journal)->j_first); \
239 * journal_recover - recovers a on-disk journal
240 * @journal: the journal to recover
242 * The primary function for recovering the log contents when mounting a
245 * Recovery is done in three passes. In the first pass, we look for the
246 * end of the log. In the second, we assemble the list of revoke
247 * blocks. In the third and final pass, we replay any un-revoked blocks
250 int journal_recover(journal_t *journal)
253 journal_superblock_t * sb;
255 struct recovery_info info;
257 memset(&info, 0, sizeof(info));
258 sb = journal->j_superblock;
261 * The journal superblock's s_start field (the current log head)
262 * is always zero if, and only if, the journal was cleanly
267 jbd_debug(1, "No recovery required, last transaction %d\n",
268 be32_to_cpu(sb->s_sequence));
269 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
273 err = do_one_pass(journal, &info, PASS_SCAN);
275 err = do_one_pass(journal, &info, PASS_REVOKE);
277 err = do_one_pass(journal, &info, PASS_REPLAY);
279 jbd_debug(1, "JBD: recovery, exit status %d, "
280 "recovered transactions %u to %u\n",
281 err, info.start_transaction, info.end_transaction);
282 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
283 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
285 /* Restart the log at the next transaction ID, thus invalidating
286 * any existing commit records in the log. */
287 journal->j_transaction_sequence = ++info.end_transaction;
289 journal_clear_revoke(journal);
290 sync_blockdev(journal->j_fs_dev);
295 * journal_skip_recovery - Start journal and wipe exiting records
296 * @journal: journal to startup
298 * Locate any valid recovery information from the journal and set up the
299 * journal structures in memory to ignore it (presumably because the
300 * caller has evidence that it is out of date).
301 * This function does'nt appear to be exorted..
303 * We perform one pass over the journal to allow us to tell the user how
304 * much recovery information is being erased, and to let us initialise
305 * the journal transaction sequence numbers to the next unused ID.
307 int journal_skip_recovery(journal_t *journal)
310 journal_superblock_t * sb;
312 struct recovery_info info;
314 memset (&info, 0, sizeof(info));
315 sb = journal->j_superblock;
317 err = do_one_pass(journal, &info, PASS_SCAN);
320 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
321 ++journal->j_transaction_sequence;
323 #ifdef CONFIG_JBD_DEBUG
324 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
327 "JBD: ignoring %d transaction%s from the journal.\n",
328 dropped, (dropped == 1) ? "" : "s");
329 journal->j_transaction_sequence = ++info.end_transaction;
336 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
338 unsigned long long block = be32_to_cpu(tag->t_blocknr);
339 if (tag_bytes > JBD_TAG_SIZE32)
340 block |= (__u64)be32_to_cpu(tag->t_blocknr_high) << 32;
345 * calc_chksums calculates the checksums for the blocks described in the
348 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
349 unsigned long long *next_log_block, __u32 *crc32_sum)
351 int i, num_blks, err;
352 unsigned long long io_block;
353 struct buffer_head *obh;
355 num_blks = count_tags(journal, bh);
356 /* Calculate checksum of the descriptor block. */
357 *crc32_sum = ext2fs_crc32_be(*crc32_sum, (void *)bh->b_data,
360 for (i = 0; i < num_blks; i++) {
361 io_block = (*next_log_block)++;
362 wrap(journal, *next_log_block);
363 err = jread(&obh, journal, io_block);
365 printk(KERN_ERR "JBD: IO error %d recovering block "
366 "%llu in log\n", err, io_block);
369 *crc32_sum = ext2fs_crc32_be(*crc32_sum,
378 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
380 struct commit_header *h;
381 __u32 provided, calculated;
383 if (!JFS_HAS_INCOMPAT_FEATURE(j, JFS_FEATURE_INCOMPAT_CSUM_V2))
387 provided = h->h_chksum[0];
389 calculated = ext2fs_crc32c_le(~0, j->j_superblock->s_uuid,
390 sizeof(j->j_superblock->s_uuid));
391 calculated = ext2fs_crc32c_le(calculated, buf, j->j_blocksize);
392 h->h_chksum[0] = provided;
394 provided = ext2fs_be32_to_cpu(provided);
395 return provided == calculated;
398 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
399 void *buf, __u32 sequence)
401 __u32 provided, calculated;
403 if (!JFS_HAS_INCOMPAT_FEATURE(j, JFS_FEATURE_INCOMPAT_CSUM_V2))
406 sequence = ext2fs_cpu_to_be32(sequence);
407 calculated = ext2fs_crc32c_le(~0, j->j_superblock->s_uuid,
408 sizeof(j->j_superblock->s_uuid));
409 calculated = ext2fs_crc32c_le(calculated, (__u8 *)&sequence,
411 calculated = ext2fs_crc32c_le(calculated, buf, j->j_blocksize) & 0xffff;
412 provided = ext2fs_be16_to_cpu(tag->t_checksum);
414 return provided == ext2fs_cpu_to_be32(calculated);
417 static int do_one_pass(journal_t *journal,
418 struct recovery_info *info, enum passtype pass)
420 unsigned int first_commit_ID, next_commit_ID;
421 unsigned long long next_log_block;
422 int err, success = 0;
423 journal_superblock_t * sb;
424 journal_header_t * tmp;
425 struct buffer_head * bh;
426 unsigned int sequence;
428 int tag_bytes = journal_tag_bytes(journal);
429 __u32 crc32_sum = ~0; /* Transactional Checksums */
430 int descr_csum_size = 0;
432 /* Precompute the maximum metadata descriptors in a descriptor block */
433 int MAX_BLOCKS_PER_DESC;
434 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
438 * First thing is to establish what we expect to find in the log
439 * (in terms of transaction IDs), and where (in terms of log
440 * block offsets): query the superblock.
443 sb = journal->j_superblock;
444 next_commit_ID = be32_to_cpu(sb->s_sequence);
445 next_log_block = be32_to_cpu(sb->s_start);
447 first_commit_ID = next_commit_ID;
448 if (pass == PASS_SCAN)
449 info->start_transaction = first_commit_ID;
451 jbd_debug(1, "Starting recovery pass %d\n", pass);
454 * Now we walk through the log, transaction by transaction,
455 * making sure that each transaction has a commit block in the
456 * expected place. Each complete transaction gets replayed back
457 * into the main filesystem.
463 journal_block_tag_t * tag;
464 struct buffer_head * obh;
465 struct buffer_head * nbh;
469 /* If we already know where to stop the log traversal,
470 * check right now that we haven't gone past the end of
473 if (pass != PASS_SCAN)
474 if (tid_geq(next_commit_ID, info->end_transaction))
477 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
478 next_commit_ID, next_log_block, journal->j_last);
480 /* Skip over each chunk of the transaction looking
481 * either the next descriptor block or the final commit
484 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
485 err = jread(&bh, journal, next_log_block);
490 wrap(journal, next_log_block);
492 /* What kind of buffer is it?
494 * If it is a descriptor block, check that it has the
495 * expected sequence number. Otherwise, we're all done
498 tmp = (journal_header_t *)bh->b_data;
500 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
505 blocktype = be32_to_cpu(tmp->h_blocktype);
506 sequence = be32_to_cpu(tmp->h_sequence);
507 jbd_debug(3, "Found magic %d, sequence %d\n",
508 blocktype, sequence);
510 if (sequence != next_commit_ID) {
515 /* OK, we have a valid descriptor block which matches
516 * all of the sequence number checks. What are we going
517 * to do with it? That depends on the pass... */
520 case JFS_DESCRIPTOR_BLOCK:
521 /* Verify checksum first */
522 if (JFS_HAS_INCOMPAT_FEATURE(journal,
523 JFS_FEATURE_INCOMPAT_CSUM_V2))
525 sizeof(struct journal_block_tail);
526 if (descr_csum_size > 0 &&
527 !jbd2_descr_block_csum_verify(journal,
533 /* If it is a valid descriptor block, replay it
534 * in pass REPLAY; if journal_checksums enabled, then
535 * calculate checksums in PASS_SCAN, otherwise,
536 * just skip over the blocks it describes. */
537 if (pass != PASS_REPLAY) {
538 if (pass == PASS_SCAN &&
539 JFS_HAS_COMPAT_FEATURE(journal,
540 JFS_FEATURE_COMPAT_CHECKSUM) &&
541 !info->end_transaction) {
542 if (calc_chksums(journal, bh,
551 next_log_block += count_tags(journal, bh);
552 wrap(journal, next_log_block);
557 /* A descriptor block: we can now write all of
558 * the data blocks. Yay, useful work is finally
559 * getting done here! */
561 tagp = &bh->b_data[sizeof(journal_header_t)];
562 while ((tagp - bh->b_data + tag_bytes)
563 <= journal->j_blocksize - descr_csum_size) {
564 unsigned long long io_block;
566 tag = (journal_block_tag_t *) tagp;
567 flags = be16_to_cpu(tag->t_flags);
569 io_block = next_log_block++;
570 wrap(journal, next_log_block);
571 err = jread(&obh, journal, io_block);
573 /* Recover what we can, but
574 * report failure at the end. */
577 "JBD: IO error %d recovering "
578 "block %llu in log\n",
581 unsigned long long blocknr;
583 J_ASSERT(obh != NULL);
584 blocknr = read_tag_block(tag_bytes,
587 /* If the block has been
588 * revoked, then we're all done
590 if (journal_test_revoke
594 ++info->nr_revoke_hits;
598 /* Look for block corruption */
599 if (!jbd2_block_tag_csum_verify(
600 journal, tag, obh->b_data,
601 be32_to_cpu(tmp->h_sequence))) {
604 printk(KERN_ERR "JBD: Invalid "
605 "checksum recovering "
606 "block %ld in log\n",
611 /* Find a buffer for the new
612 * data being restored */
613 nbh = __getblk(journal->j_fs_dev,
615 journal->j_blocksize);
618 "JBD: Out of memory "
619 "during recovery.\n");
627 memcpy(nbh->b_data, obh->b_data,
628 journal->j_blocksize);
629 if (flags & JFS_FLAG_ESCAPE) {
630 journal_header_t *header;
632 header = (journal_header_t *) &nbh->b_data[0];
633 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
636 BUFFER_TRACE(nbh, "marking dirty");
637 set_buffer_uptodate(nbh);
638 mark_buffer_dirty(nbh);
639 BUFFER_TRACE(nbh, "marking uptodate");
641 /* ll_rw_block(WRITE, 1, &nbh); */
649 if (!(flags & JFS_FLAG_SAME_UUID))
652 if (flags & JFS_FLAG_LAST_TAG)
659 case JFS_COMMIT_BLOCK:
660 jbd_debug(3, "Commit block for #%u found\n",
662 /* How to differentiate between interrupted commit
663 * and journal corruption ?
666 * Checksum Verification Failed
668 * ____________________
670 * async_commit sync_commit
672 * | GO TO NEXT "Journal Corruption"
675 * {(n+1)th transanction}
677 * _______|______________
679 * Commit block found Commit block not found
681 * "Journal Corruption" |
682 * _____________|_________
684 * nth trans corrupt OR nth trans
685 * and (n+1)th interrupted interrupted
686 * before commit block
687 * could reach the disk.
688 * (Cannot find the difference in above
689 * mentioned conditions. Hence assume
690 * "Interrupted Commit".)
693 /* Found an expected commit block: if checksums
694 * are present verify them in PASS_SCAN; else not
695 * much to do other than move on to the next sequence
697 if (pass == PASS_SCAN &&
698 JFS_HAS_COMPAT_FEATURE(journal,
699 JFS_FEATURE_COMPAT_CHECKSUM)) {
700 int chksum_err, chksum_seen;
701 struct commit_header *cbh =
702 (struct commit_header *)bh->b_data;
703 unsigned found_chksum =
704 be32_to_cpu(cbh->h_chksum[0]);
706 chksum_err = chksum_seen = 0;
708 jbd_debug(3, "Checksums %x %x\n",
709 crc32_sum, found_chksum);
710 if (info->end_transaction) {
711 journal->j_failed_commit =
712 info->end_transaction;
717 if (crc32_sum == found_chksum &&
718 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
719 cbh->h_chksum_size ==
720 JBD2_CRC32_CHKSUM_SIZE)
722 else if (!(cbh->h_chksum_type == 0 &&
723 cbh->h_chksum_size == 0 &&
727 * If fs is mounted using an old kernel and then
728 * kernel with journal_chksum is used then we
729 * get a situation where the journal flag has
730 * checksum flag set but checksums are not
731 * present i.e chksum = 0, in the individual
733 * Hence to avoid checksum failures, in this
734 * situation, this extra check is added.
739 info->end_transaction = next_commit_ID;
740 jbd_debug(1, "Checksum_err %x %x\n",
741 crc32_sum, found_chksum);
742 if (!JFS_HAS_INCOMPAT_FEATURE(journal,
743 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
744 journal->j_failed_commit =
752 if (pass == PASS_SCAN &&
753 !jbd2_commit_block_csum_verify(journal,
755 info->end_transaction = next_commit_ID;
757 if (!JFS_HAS_INCOMPAT_FEATURE(journal,
758 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
759 journal->j_failed_commit =
769 case JFS_REVOKE_BLOCK:
770 /* If we aren't in the REVOKE pass, then we can
771 * just skip over this block. */
772 if (pass != PASS_REVOKE) {
777 err = scan_revoke_records(journal, bh,
778 next_commit_ID, info);
785 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
794 * We broke out of the log scan loop: either we came to the
795 * known end of the log or we found an unexpected block in the
796 * log. If the latter happened, then we know that the "current"
797 * transaction marks the end of the valid log.
800 if (pass == PASS_SCAN) {
801 if (!info->end_transaction)
802 info->end_transaction = next_commit_ID;
804 /* It's really bad news if different passes end up at
805 * different places (but possible due to IO errors). */
806 if (info->end_transaction != next_commit_ID) {
807 printk (KERN_ERR "JBD: recovery pass %d ended at "
808 "transaction %u, expected %u\n",
809 pass, next_commit_ID, info->end_transaction);
821 static int jbd2_revoke_block_csum_verify(journal_t *j,
824 struct journal_revoke_tail *tail;
825 __u32 provided, calculated;
827 if (!JFS_HAS_INCOMPAT_FEATURE(j, JFS_FEATURE_INCOMPAT_CSUM_V2))
830 tail = (struct journal_revoke_tail *)(buf + j->j_blocksize -
831 sizeof(struct journal_revoke_tail));
832 provided = tail->r_checksum;
833 tail->r_checksum = 0;
834 calculated = ext2fs_crc32c_le(~0, j->j_superblock->s_uuid,
835 sizeof(j->j_superblock->s_uuid));
836 calculated = ext2fs_crc32c_le(calculated, buf, j->j_blocksize);
837 tail->r_checksum = provided;
839 provided = ext2fs_be32_to_cpu(provided);
840 return provided == calculated;
843 /* Scan a revoke record, marking all blocks mentioned as revoked. */
845 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
846 tid_t sequence, struct recovery_info *info)
848 journal_revoke_header_t *header;
852 header = (journal_revoke_header_t *) bh->b_data;
853 offset = sizeof(journal_revoke_header_t);
854 max = be32_to_cpu(header->r_count);
856 if (!jbd2_revoke_block_csum_verify(journal, header))
859 if (JFS_HAS_INCOMPAT_FEATURE(journal, JFS_FEATURE_INCOMPAT_64BIT))
862 while (offset < max) {
863 unsigned long long blocknr;
867 blocknr = ext2fs_be32_to_cpu(*((__be32 *)(bh->b_data +
870 blocknr = ext2fs_be64_to_cpu(*((__be64 *)(bh->b_data +
872 offset += record_len;
873 err = journal_set_revoke(journal, blocknr, sequence);