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 = 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");
143 return -EFSCORRUPTED;
146 err = 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 journal_block_tail *tail;
184 if (!journal_has_csum_v2or3(j))
187 tail = (struct journal_block_tail *)((char *)buf + j->j_blocksize -
188 sizeof(struct 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 == ext2fs_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 (journal_has_csum_v2or3(journal))
209 size -= sizeof(struct 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 & ext2fs_cpu_to_be16(JFS_FLAG_SAME_UUID)))
221 if (tag->t_flags & ext2fs_cpu_to_be16(JFS_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 * 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 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 ext2fs_be32_to_cpu(sb->s_sequence));
267 journal->j_transaction_sequence = ext2fs_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 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 & JFS_BARRIER) {
293 err2 = blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
301 * 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 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_JFS_DEBUG
328 int dropped = info.end_transaction -
329 ext2fs_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 __u32 get_be32(__be32 *p)
343 unsigned char *cp = (unsigned char *) p;
347 ret = (ret << 8) + *cp++;
348 ret = (ret << 8) + *cp++;
349 ret = (ret << 8) + *cp++;
353 static inline unsigned long long read_tag_block(journal_t *journal,
354 journal_block_tag_t *tag)
356 unsigned long long block = get_be32(&tag->t_blocknr);
357 if (jfs_has_feature_64bit(journal))
358 block |= (u64)get_be32(&tag->t_blocknr_high) << 32;
363 * calc_chksums calculates the checksums for the blocks described in the
366 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
367 unsigned long *next_log_block, __u32 *crc32_sum)
369 int i, num_blks, err;
370 unsigned long io_block;
371 struct buffer_head *obh;
373 num_blks = count_tags(journal, bh);
374 /* Calculate checksum of the descriptor block. */
375 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
377 for (i = 0; i < num_blks; i++) {
378 io_block = (*next_log_block)++;
379 wrap(journal, *next_log_block);
380 err = jread(&obh, journal, io_block);
382 printk(KERN_ERR "JBD2: IO error %d recovering block "
383 "%lu in log\n", err, io_block);
386 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
394 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
396 struct commit_header *h;
400 if (!journal_has_csum_v2or3(j))
404 provided = h->h_chksum[0];
406 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
407 h->h_chksum[0] = provided;
409 return provided == ext2fs_cpu_to_be32(calculated);
412 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
413 void *buf, __u32 sequence)
415 journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
419 if (!journal_has_csum_v2or3(j))
422 seq = ext2fs_cpu_to_be32(sequence);
423 csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
424 csum32 = jbd2_chksum(j, csum32, buf, j->j_blocksize);
426 if (jfs_has_feature_csum3(j))
427 return tag3->t_checksum == ext2fs_cpu_to_be32(csum32);
429 return tag->t_checksum == ext2fs_cpu_to_be16(csum32);
432 static int do_one_pass(journal_t *journal,
433 struct recovery_info *info, enum passtype pass)
435 unsigned int first_commit_ID, next_commit_ID;
436 unsigned long next_log_block;
437 int err, success = 0;
438 journal_superblock_t * sb;
439 journal_header_t * tmp;
440 struct buffer_head * bh;
441 unsigned int sequence;
443 int tag_bytes = journal_tag_bytes(journal);
444 __u32 crc32_sum = ~0; /* Transactional Checksums */
445 int descr_csum_size = 0;
449 * First thing is to establish what we expect to find in the log
450 * (in terms of transaction IDs), and where (in terms of log
451 * block offsets): query the superblock.
454 sb = journal->j_superblock;
455 next_commit_ID = ext2fs_be32_to_cpu(sb->s_sequence);
456 next_log_block = ext2fs_be32_to_cpu(sb->s_start);
458 first_commit_ID = next_commit_ID;
459 if (pass == PASS_SCAN)
460 info->start_transaction = first_commit_ID;
462 jbd_debug(1, "Starting recovery pass %d\n", pass);
465 * Now we walk through the log, transaction by transaction,
466 * making sure that each transaction has a commit block in the
467 * expected place. Each complete transaction gets replayed back
468 * into the main filesystem.
474 journal_block_tag_t * tag;
475 struct buffer_head * obh;
476 struct buffer_head * nbh;
480 /* If we already know where to stop the log traversal,
481 * check right now that we haven't gone past the end of
484 if (pass != PASS_SCAN)
485 if (tid_geq(next_commit_ID, info->end_transaction))
488 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
489 next_commit_ID, next_log_block, journal->j_last);
491 /* Skip over each chunk of the transaction looking
492 * either the next descriptor block or the final commit
495 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
496 err = jread(&bh, journal, next_log_block);
501 wrap(journal, next_log_block);
503 /* What kind of buffer is it?
505 * If it is a descriptor block, check that it has the
506 * expected sequence number. Otherwise, we're all done
509 tmp = (journal_header_t *)bh->b_data;
511 if (tmp->h_magic != ext2fs_cpu_to_be32(JFS_MAGIC_NUMBER)) {
516 blocktype = ext2fs_be32_to_cpu(tmp->h_blocktype);
517 sequence = ext2fs_be32_to_cpu(tmp->h_sequence);
518 jbd_debug(3, "Found magic %d, sequence %d\n",
519 blocktype, sequence);
521 if (sequence != next_commit_ID) {
526 /* OK, we have a valid descriptor block which matches
527 * all of the sequence number checks. What are we going
528 * to do with it? That depends on the pass... */
531 case JFS_DESCRIPTOR_BLOCK:
532 /* Verify checksum first */
533 if (journal_has_csum_v2or3(journal))
535 sizeof(struct journal_block_tail);
536 if (descr_csum_size > 0 &&
537 !jbd2_descr_block_csum_verify(journal,
544 /* If it is a valid descriptor block, replay it
545 * in pass REPLAY; if journal_checksums enabled, then
546 * calculate checksums in PASS_SCAN, otherwise,
547 * just skip over the blocks it describes. */
548 if (pass != PASS_REPLAY) {
549 if (pass == PASS_SCAN &&
550 jfs_has_feature_checksum(journal) &&
551 !info->end_transaction) {
552 if (calc_chksums(journal, bh,
561 next_log_block += count_tags(journal, bh);
562 wrap(journal, next_log_block);
567 /* A descriptor block: we can now write all of
568 * the data blocks. Yay, useful work is finally
569 * getting done here! */
571 tagp = &bh->b_data[sizeof(journal_header_t)];
572 while ((tagp - bh->b_data + tag_bytes)
573 <= journal->j_blocksize - descr_csum_size) {
574 unsigned long io_block;
576 tag = (journal_block_tag_t *) tagp;
577 flags = ext2fs_be16_to_cpu(tag->t_flags);
579 io_block = next_log_block++;
580 wrap(journal, next_log_block);
581 err = jread(&obh, journal, io_block);
583 /* Recover what we can, but
584 * report failure at the end. */
587 "JBD2: IO error %d recovering "
588 "block %ld in log\n",
591 unsigned long long blocknr;
593 J_ASSERT(obh != NULL);
594 blocknr = read_tag_block(journal,
597 /* If the block has been
598 * revoked, then we're all done
600 if (journal_test_revoke
604 ++info->nr_revoke_hits;
608 /* Look for block corruption */
609 if (!jbd2_block_tag_csum_verify(
610 journal, tag, obh->b_data,
611 ext2fs_be32_to_cpu(tmp->h_sequence))) {
613 success = -EFSBADCRC;
614 printk(KERN_ERR "JBD2: Invalid "
615 "checksum recovering "
616 "block %llu in log\n",
622 /* Find a buffer for the new
623 * data being restored */
624 nbh = __getblk(journal->j_fs_dev,
626 journal->j_blocksize);
629 "JBD2: Out of memory "
630 "during recovery.\n");
638 memcpy(nbh->b_data, obh->b_data,
639 journal->j_blocksize);
640 if (flags & JFS_FLAG_ESCAPE) {
641 __u32 magic = ext2fs_cpu_to_be32(JFS_MAGIC_NUMBER);
642 memcpy(nbh->b_data, &magic,
646 BUFFER_TRACE(nbh, "marking dirty");
647 set_buffer_uptodate(nbh);
648 mark_buffer_dirty(nbh);
649 BUFFER_TRACE(nbh, "marking uptodate");
651 /* ll_rw_block(WRITE, 1, &nbh); */
659 if (!(flags & JFS_FLAG_SAME_UUID))
662 if (flags & JFS_FLAG_LAST_TAG)
669 case JFS_COMMIT_BLOCK:
670 /* How to differentiate between interrupted commit
671 * and journal corruption ?
674 * Checksum Verification Failed
676 * ____________________
678 * async_commit sync_commit
680 * | GO TO NEXT "Journal Corruption"
683 * {(n+1)th transanction}
685 * _______|______________
687 * Commit block found Commit block not found
689 * "Journal Corruption" |
690 * _____________|_________
692 * nth trans corrupt OR nth trans
693 * and (n+1)th interrupted interrupted
694 * before commit block
695 * could reach the disk.
696 * (Cannot find the difference in above
697 * mentioned conditions. Hence assume
698 * "Interrupted Commit".)
701 /* Found an expected commit block: if checksums
702 * are present verify them in PASS_SCAN; else not
703 * much to do other than move on to the next sequence
705 if (pass == PASS_SCAN &&
706 jfs_has_feature_checksum(journal)) {
707 int chksum_err, chksum_seen;
708 struct commit_header *cbh =
709 (struct commit_header *)bh->b_data;
710 unsigned found_chksum =
711 ext2fs_be32_to_cpu(cbh->h_chksum[0]);
713 chksum_err = chksum_seen = 0;
715 if (info->end_transaction) {
716 journal->j_failed_commit =
717 info->end_transaction;
722 if (crc32_sum == found_chksum &&
723 cbh->h_chksum_type == JFS_CRC32_CHKSUM &&
724 cbh->h_chksum_size ==
725 JFS_CRC32_CHKSUM_SIZE)
727 else if (!(cbh->h_chksum_type == 0 &&
728 cbh->h_chksum_size == 0 &&
732 * If fs is mounted using an old kernel and then
733 * kernel with journal_chksum is used then we
734 * get a situation where the journal flag has
735 * checksum flag set but checksums are not
736 * present i.e chksum = 0, in the individual
738 * Hence to avoid checksum failures, in this
739 * situation, this extra check is added.
744 info->end_transaction = next_commit_ID;
746 if (!jfs_has_feature_async_commit(journal)){
747 journal->j_failed_commit =
755 if (pass == PASS_SCAN &&
756 !jbd2_commit_block_csum_verify(journal,
758 info->end_transaction = next_commit_ID;
760 if (!jfs_has_feature_async_commit(journal)) {
761 journal->j_failed_commit =
771 case JFS_REVOKE_BLOCK:
772 /* If we aren't in the REVOKE pass, then we can
773 * just skip over this block. */
774 if (pass != PASS_REVOKE) {
779 err = scan_revoke_records(journal, bh,
780 next_commit_ID, info);
787 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
796 * We broke out of the log scan loop: either we came to the
797 * known end of the log or we found an unexpected block in the
798 * log. If the latter happened, then we know that the "current"
799 * transaction marks the end of the valid log.
802 if (pass == PASS_SCAN) {
803 if (!info->end_transaction)
804 info->end_transaction = next_commit_ID;
806 /* It's really bad news if different passes end up at
807 * different places (but possible due to IO errors). */
808 if (info->end_transaction != next_commit_ID) {
809 printk(KERN_ERR "JBD2: recovery pass %d ended at "
810 "transaction %u, expected %u\n",
811 pass, next_commit_ID, info->end_transaction);
816 if (block_error && success == 0)
824 static int jbd2_revoke_block_csum_verify(journal_t *j,
827 struct journal_revoke_tail *tail;
831 if (!journal_has_csum_v2or3(j))
834 tail = (struct journal_revoke_tail *)((char *)buf + j->j_blocksize -
835 sizeof(struct journal_revoke_tail));
836 provided = tail->r_checksum;
837 tail->r_checksum = 0;
838 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
839 tail->r_checksum = provided;
841 return provided == ext2fs_cpu_to_be32(calculated);
844 /* Scan a revoke record, marking all blocks mentioned as revoked. */
846 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
847 tid_t sequence, struct recovery_info *info)
849 journal_revoke_header_t *header;
851 unsigned csum_size = 0;
855 header = (journal_revoke_header_t *) bh->b_data;
856 offset = sizeof(journal_revoke_header_t);
857 rcount = ext2fs_be32_to_cpu(header->r_count);
859 if (!jbd2_revoke_block_csum_verify(journal, header))
862 if (journal_has_csum_v2or3(journal))
863 csum_size = sizeof(struct journal_revoke_tail);
864 if (rcount > journal->j_blocksize - csum_size)
868 if (jfs_has_feature_64bit(journal))
871 while (offset + record_len <= max) {
872 unsigned long long blocknr;
876 blocknr = ext2fs_be32_to_cpu(* ((__u32 *) (bh->b_data+offset)));
878 blocknr = ext2fs_be64_to_cpu(* ((__u64 *) (bh->b_data+offset)));
879 offset += record_len;
880 err = journal_set_revoke(journal, blocknr, sequence);