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 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
298 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
299 * @journal: journal to startup
301 * Locate any valid recovery information from the journal and set up the
302 * journal structures in memory to ignore it (presumably because the
303 * caller has evidence that it is out of date).
304 * This function does'nt appear to be exorted..
306 * We perform one pass over the journal to allow us to tell the user how
307 * much recovery information is being erased, and to let us initialise
308 * the journal transaction sequence numbers to the next unused ID.
310 int jbd2_journal_skip_recovery(journal_t *journal)
314 struct recovery_info info;
316 memset (&info, 0, sizeof(info));
318 err = do_one_pass(journal, &info, PASS_SCAN);
321 printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
322 ++journal->j_transaction_sequence;
324 #ifdef CONFIG_JBD2_DEBUG
325 int dropped = info.end_transaction -
326 be32_to_cpu(journal->j_superblock->s_sequence);
328 "JBD2: ignoring %d transaction%s from the journal.\n",
329 dropped, (dropped == 1) ? "" : "s");
331 journal->j_transaction_sequence = ++info.end_transaction;
338 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
340 unsigned long long block = be32_to_cpu(tag->t_blocknr);
341 if (tag_bytes > JBD2_TAG_SIZE32)
342 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
347 * calc_chksums calculates the checksums for the blocks described in the
350 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
351 unsigned long *next_log_block, __u32 *crc32_sum)
353 int i, num_blks, err;
354 unsigned long io_block;
355 struct buffer_head *obh;
357 num_blks = count_tags(journal, bh);
358 /* Calculate checksum of the descriptor block. */
359 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
361 for (i = 0; i < num_blks; i++) {
362 io_block = (*next_log_block)++;
363 wrap(journal, *next_log_block);
364 err = jread(&obh, journal, io_block);
366 printk(KERN_ERR "JBD2: IO error %d recovering block "
367 "%lu in log\n", err, io_block);
370 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
378 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
380 struct commit_header *h;
381 __u32 provided, calculated;
383 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
387 provided = h->h_chksum[0];
389 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
390 h->h_chksum[0] = provided;
392 provided = be32_to_cpu(provided);
393 return provided == calculated;
396 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
397 void *buf, __u32 sequence)
399 __u32 provided, calculated;
401 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
404 sequence = cpu_to_be32(sequence);
405 calculated = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&sequence,
407 calculated = jbd2_chksum(j, calculated, buf, j->j_blocksize);
408 provided = be32_to_cpu(tag->t_checksum);
410 return provided == cpu_to_be32(calculated);
413 static int do_one_pass(journal_t *journal,
414 struct recovery_info *info, enum passtype pass)
416 unsigned int first_commit_ID, next_commit_ID;
417 unsigned long next_log_block;
418 int err, success = 0;
419 journal_superblock_t * sb;
420 journal_header_t * tmp;
421 struct buffer_head * bh;
422 unsigned int sequence;
424 int tag_bytes = journal_tag_bytes(journal);
425 __u32 crc32_sum = ~0; /* Transactional Checksums */
426 int descr_csum_size = 0;
429 * First thing is to establish what we expect to find in the log
430 * (in terms of transaction IDs), and where (in terms of log
431 * block offsets): query the superblock.
434 sb = journal->j_superblock;
435 next_commit_ID = be32_to_cpu(sb->s_sequence);
436 next_log_block = be32_to_cpu(sb->s_start);
438 first_commit_ID = next_commit_ID;
439 if (pass == PASS_SCAN)
440 info->start_transaction = first_commit_ID;
442 jbd_debug(1, "Starting recovery pass %d\n", pass);
445 * Now we walk through the log, transaction by transaction,
446 * making sure that each transaction has a commit block in the
447 * expected place. Each complete transaction gets replayed back
448 * into the main filesystem.
454 journal_block_tag_t * tag;
455 struct buffer_head * obh;
456 struct buffer_head * nbh;
460 /* If we already know where to stop the log traversal,
461 * check right now that we haven't gone past the end of
464 if (pass != PASS_SCAN)
465 if (tid_geq(next_commit_ID, info->end_transaction))
468 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
469 next_commit_ID, next_log_block, journal->j_last);
471 /* Skip over each chunk of the transaction looking
472 * either the next descriptor block or the final commit
475 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
476 err = jread(&bh, journal, next_log_block);
481 wrap(journal, next_log_block);
483 /* What kind of buffer is it?
485 * If it is a descriptor block, check that it has the
486 * expected sequence number. Otherwise, we're all done
489 tmp = (journal_header_t *)bh->b_data;
491 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
496 blocktype = be32_to_cpu(tmp->h_blocktype);
497 sequence = be32_to_cpu(tmp->h_sequence);
498 jbd_debug(3, "Found magic %d, sequence %d\n",
499 blocktype, sequence);
501 if (sequence != next_commit_ID) {
506 /* OK, we have a valid descriptor block which matches
507 * all of the sequence number checks. What are we going
508 * to do with it? That depends on the pass... */
511 case JBD2_DESCRIPTOR_BLOCK:
512 /* Verify checksum first */
513 if (JBD2_HAS_INCOMPAT_FEATURE(journal,
514 JBD2_FEATURE_INCOMPAT_CSUM_V2))
516 sizeof(struct jbd2_journal_block_tail);
517 if (descr_csum_size > 0 &&
518 !jbd2_descr_block_csum_verify(journal,
524 /* If it is a valid descriptor block, replay it
525 * in pass REPLAY; if journal_checksums enabled, then
526 * calculate checksums in PASS_SCAN, otherwise,
527 * just skip over the blocks it describes. */
528 if (pass != PASS_REPLAY) {
529 if (pass == PASS_SCAN &&
530 JBD2_HAS_COMPAT_FEATURE(journal,
531 JBD2_FEATURE_COMPAT_CHECKSUM) &&
532 !info->end_transaction) {
533 if (calc_chksums(journal, bh,
542 next_log_block += count_tags(journal, bh);
543 wrap(journal, next_log_block);
548 /* A descriptor block: we can now write all of
549 * the data blocks. Yay, useful work is finally
550 * getting done here! */
552 tagp = &bh->b_data[sizeof(journal_header_t)];
553 while ((tagp - bh->b_data + tag_bytes)
554 <= journal->j_blocksize - descr_csum_size) {
555 unsigned long io_block;
557 tag = (journal_block_tag_t *) tagp;
558 flags = be16_to_cpu(tag->t_flags);
560 io_block = next_log_block++;
561 wrap(journal, next_log_block);
562 err = jread(&obh, journal, io_block);
564 /* Recover what we can, but
565 * report failure at the end. */
568 "JBD2: IO error %d recovering "
569 "block %ld in log\n",
572 unsigned long long blocknr;
574 J_ASSERT(obh != NULL);
575 blocknr = read_tag_block(tag_bytes,
578 /* If the block has been
579 * revoked, then we're all done
581 if (jbd2_journal_test_revoke
585 ++info->nr_revoke_hits;
589 /* Look for block corruption */
590 if (!jbd2_block_tag_csum_verify(
591 journal, tag, obh->b_data,
592 be32_to_cpu(tmp->h_sequence))) {
595 printk(KERN_ERR "JBD: Invalid "
596 "checksum recovering "
597 "block %llu in log\n",
602 /* Find a buffer for the new
603 * data being restored */
604 nbh = __getblk(journal->j_fs_dev,
606 journal->j_blocksize);
609 "JBD2: Out of memory "
610 "during recovery.\n");
618 memcpy(nbh->b_data, obh->b_data,
619 journal->j_blocksize);
620 if (flags & JBD2_FLAG_ESCAPE) {
621 *((__be32 *)nbh->b_data) =
622 cpu_to_be32(JBD2_MAGIC_NUMBER);
625 BUFFER_TRACE(nbh, "marking dirty");
626 set_buffer_uptodate(nbh);
627 mark_buffer_dirty(nbh);
628 BUFFER_TRACE(nbh, "marking uptodate");
630 /* ll_rw_block(WRITE, 1, &nbh); */
638 if (!(flags & JBD2_FLAG_SAME_UUID))
641 if (flags & JBD2_FLAG_LAST_TAG)
648 case JBD2_COMMIT_BLOCK:
649 /* How to differentiate between interrupted commit
650 * and journal corruption ?
653 * Checksum Verification Failed
655 * ____________________
657 * async_commit sync_commit
659 * | GO TO NEXT "Journal Corruption"
662 * {(n+1)th transanction}
664 * _______|______________
666 * Commit block found Commit block not found
668 * "Journal Corruption" |
669 * _____________|_________
671 * nth trans corrupt OR nth trans
672 * and (n+1)th interrupted interrupted
673 * before commit block
674 * could reach the disk.
675 * (Cannot find the difference in above
676 * mentioned conditions. Hence assume
677 * "Interrupted Commit".)
680 /* Found an expected commit block: if checksums
681 * are present verify them in PASS_SCAN; else not
682 * much to do other than move on to the next sequence
684 if (pass == PASS_SCAN &&
685 JBD2_HAS_COMPAT_FEATURE(journal,
686 JBD2_FEATURE_COMPAT_CHECKSUM)) {
687 int chksum_err, chksum_seen;
688 struct commit_header *cbh =
689 (struct commit_header *)bh->b_data;
690 unsigned found_chksum =
691 be32_to_cpu(cbh->h_chksum[0]);
693 chksum_err = chksum_seen = 0;
695 if (info->end_transaction) {
696 journal->j_failed_commit =
697 info->end_transaction;
702 if (crc32_sum == found_chksum &&
703 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
704 cbh->h_chksum_size ==
705 JBD2_CRC32_CHKSUM_SIZE)
707 else if (!(cbh->h_chksum_type == 0 &&
708 cbh->h_chksum_size == 0 &&
712 * If fs is mounted using an old kernel and then
713 * kernel with journal_chksum is used then we
714 * get a situation where the journal flag has
715 * checksum flag set but checksums are not
716 * present i.e chksum = 0, in the individual
718 * Hence to avoid checksum failures, in this
719 * situation, this extra check is added.
724 info->end_transaction = next_commit_ID;
726 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
727 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
728 journal->j_failed_commit =
736 if (pass == PASS_SCAN &&
737 !jbd2_commit_block_csum_verify(journal,
739 info->end_transaction = next_commit_ID;
741 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
742 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
743 journal->j_failed_commit =
753 case JBD2_REVOKE_BLOCK:
754 /* If we aren't in the REVOKE pass, then we can
755 * just skip over this block. */
756 if (pass != PASS_REVOKE) {
761 err = scan_revoke_records(journal, bh,
762 next_commit_ID, info);
769 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
778 * We broke out of the log scan loop: either we came to the
779 * known end of the log or we found an unexpected block in the
780 * log. If the latter happened, then we know that the "current"
781 * transaction marks the end of the valid log.
784 if (pass == PASS_SCAN) {
785 if (!info->end_transaction)
786 info->end_transaction = next_commit_ID;
788 /* It's really bad news if different passes end up at
789 * different places (but possible due to IO errors). */
790 if (info->end_transaction != next_commit_ID) {
791 printk(KERN_ERR "JBD2: recovery pass %d ended at "
792 "transaction %u, expected %u\n",
793 pass, next_commit_ID, info->end_transaction);
805 static int jbd2_revoke_block_csum_verify(journal_t *j,
808 struct jbd2_journal_revoke_tail *tail;
809 __u32 provided, calculated;
811 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
814 tail = (struct jbd2_journal_revoke_tail *)(buf + j->j_blocksize -
815 sizeof(struct jbd2_journal_revoke_tail));
816 provided = tail->r_checksum;
817 tail->r_checksum = 0;
818 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
819 tail->r_checksum = provided;
821 provided = be32_to_cpu(provided);
822 return provided == calculated;
825 /* Scan a revoke record, marking all blocks mentioned as revoked. */
827 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
828 tid_t sequence, struct recovery_info *info)
830 jbd2_journal_revoke_header_t *header;
834 header = (jbd2_journal_revoke_header_t *) bh->b_data;
835 offset = sizeof(jbd2_journal_revoke_header_t);
836 max = be32_to_cpu(header->r_count);
838 if (!jbd2_revoke_block_csum_verify(journal, header))
841 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
844 while (offset + record_len <= max) {
845 unsigned long long blocknr;
849 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
851 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
852 offset += record_len;
853 err = jbd2_journal_set_revoke(journal, blocknr, sequence);