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>
27 * Maintain information about the progress of the recovery job, so that
28 * the different passes can carry information between them.
32 tid_t start_transaction;
33 tid_t end_transaction;
40 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
41 static int do_one_pass(journal_t *journal,
42 struct recovery_info *info, enum passtype pass);
43 static int scan_revoke_records(journal_t *, struct buffer_head *,
44 tid_t, struct recovery_info *);
48 /* Release readahead buffers after use */
49 static void journal_brelse_array(struct buffer_head *b[], int n)
57 * When reading from the journal, we are going through the block device
58 * layer directly and so there is no readahead being done for us. We
59 * need to implement any readahead ourselves if we want it to happen at
60 * all. Recovery is basically one long sequential read, so make sure we
61 * do the IO in reasonably large chunks.
63 * This is not so critical that we need to be enormously clever about
64 * the readahead size, though. 128K is a purely arbitrary, good-enough
69 static int do_readahead(journal_t *journal, unsigned int start)
72 unsigned int max, nbufs, next;
73 unsigned long long blocknr;
74 struct buffer_head *bh;
76 struct buffer_head * bufs[MAXBUF];
78 /* Do up to 128K of readahead */
79 max = start + (128 * 1024 / journal->j_blocksize);
80 if (max > journal->j_maxlen)
81 max = journal->j_maxlen;
83 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
84 * a time to the block device IO layer. */
88 for (next = start; next < max; next++) {
89 err = jbd2_journal_bmap(journal, next, &blocknr);
92 printk (KERN_ERR "JBD: bad block at offset %u\n",
97 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
103 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
105 if (nbufs == MAXBUF) {
106 ll_rw_block(READ, nbufs, bufs);
107 journal_brelse_array(bufs, nbufs);
115 ll_rw_block(READ, nbufs, bufs);
120 journal_brelse_array(bufs, nbufs);
124 #endif /* __KERNEL__ */
128 * Read a block from the journal
131 static int jread(struct buffer_head **bhp, journal_t *journal,
135 unsigned long long blocknr;
136 struct buffer_head *bh;
140 if (offset >= journal->j_maxlen) {
141 printk(KERN_ERR "JBD: corrupted journal superblock\n");
145 err = jbd2_journal_bmap(journal, offset, &blocknr);
148 printk (KERN_ERR "JBD: bad block at offset %u\n",
153 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
157 if (!buffer_uptodate(bh)) {
158 /* If this is a brand new buffer, start readahead.
159 Otherwise, we assume we are already reading it. */
161 do_readahead(journal, offset);
165 if (!buffer_uptodate(bh)) {
166 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
178 * Count the number of in-use tags in a journal descriptor block.
181 static int count_tags(journal_t *journal, struct buffer_head *bh)
184 journal_block_tag_t * tag;
185 int nr = 0, size = journal->j_blocksize;
186 int tag_bytes = journal_tag_bytes(journal);
188 tagp = &bh->b_data[sizeof(journal_header_t)];
190 while ((tagp - bh->b_data + tag_bytes) <= size) {
191 tag = (journal_block_tag_t *) tagp;
195 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID)))
198 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG))
206 /* Make sure we wrap around the log correctly! */
207 #define wrap(journal, var) \
209 if (var >= (journal)->j_last) \
210 var -= ((journal)->j_last - (journal)->j_first); \
214 * jbd2_journal_recover - recovers a on-disk journal
215 * @journal: the journal to recover
217 * The primary function for recovering the log contents when mounting a
220 * Recovery is done in three passes. In the first pass, we look for the
221 * end of the log. In the second, we assemble the list of revoke
222 * blocks. In the third and final pass, we replay any un-revoked blocks
225 int jbd2_journal_recover(journal_t *journal)
228 journal_superblock_t * sb;
230 struct recovery_info info;
232 memset(&info, 0, sizeof(info));
233 sb = journal->j_superblock;
236 * The journal superblock's s_start field (the current log head)
237 * is always zero if, and only if, the journal was cleanly
242 jbd_debug(1, "No recovery required, last transaction %d\n",
243 be32_to_cpu(sb->s_sequence));
244 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
248 err = do_one_pass(journal, &info, PASS_SCAN);
250 err = do_one_pass(journal, &info, PASS_REVOKE);
252 err = do_one_pass(journal, &info, PASS_REPLAY);
254 jbd_debug(1, "JBD: recovery, exit status %d, "
255 "recovered transactions %u to %u\n",
256 err, info.start_transaction, info.end_transaction);
257 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
258 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
260 /* Restart the log at the next transaction ID, thus invalidating
261 * any existing commit records in the log. */
262 journal->j_transaction_sequence = ++info.end_transaction;
264 jbd2_journal_clear_revoke(journal);
265 err2 = sync_blockdev(journal->j_fs_dev);
273 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
274 * @journal: journal to startup
276 * Locate any valid recovery information from the journal and set up the
277 * journal structures in memory to ignore it (presumably because the
278 * caller has evidence that it is out of date).
279 * This function does'nt appear to be exorted..
281 * We perform one pass over the journal to allow us to tell the user how
282 * much recovery information is being erased, and to let us initialise
283 * the journal transaction sequence numbers to the next unused ID.
285 int jbd2_journal_skip_recovery(journal_t *journal)
288 journal_superblock_t * sb;
290 struct recovery_info info;
292 memset (&info, 0, sizeof(info));
293 sb = journal->j_superblock;
295 err = do_one_pass(journal, &info, PASS_SCAN);
298 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
299 ++journal->j_transaction_sequence;
301 #ifdef CONFIG_JBD2_DEBUG
302 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
305 "JBD: ignoring %d transaction%s from the journal.\n",
306 dropped, (dropped == 1) ? "" : "s");
307 journal->j_transaction_sequence = ++info.end_transaction;
314 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
316 unsigned long long block = be32_to_cpu(tag->t_blocknr);
317 if (tag_bytes > JBD2_TAG_SIZE32)
318 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
323 * calc_chksums calculates the checksums for the blocks described in the
326 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
327 unsigned long *next_log_block, __u32 *crc32_sum)
329 int i, num_blks, err;
330 unsigned long io_block;
331 struct buffer_head *obh;
333 num_blks = count_tags(journal, bh);
334 /* Calculate checksum of the descriptor block. */
335 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
337 for (i = 0; i < num_blks; i++) {
338 io_block = (*next_log_block)++;
339 wrap(journal, *next_log_block);
340 err = jread(&obh, journal, io_block);
342 printk(KERN_ERR "JBD: IO error %d recovering block "
343 "%lu in log\n", err, io_block);
346 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
354 static int do_one_pass(journal_t *journal,
355 struct recovery_info *info, enum passtype pass)
357 unsigned int first_commit_ID, next_commit_ID;
358 unsigned long next_log_block;
359 int err, success = 0;
360 journal_superblock_t * sb;
361 journal_header_t * tmp;
362 struct buffer_head * bh;
363 unsigned int sequence;
365 int tag_bytes = journal_tag_bytes(journal);
366 __u32 crc32_sum = ~0; /* Transactional Checksums */
368 /* Precompute the maximum metadata descriptors in a descriptor block */
369 int MAX_BLOCKS_PER_DESC;
370 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
374 * First thing is to establish what we expect to find in the log
375 * (in terms of transaction IDs), and where (in terms of log
376 * block offsets): query the superblock.
379 sb = journal->j_superblock;
380 next_commit_ID = be32_to_cpu(sb->s_sequence);
381 next_log_block = be32_to_cpu(sb->s_start);
383 first_commit_ID = next_commit_ID;
384 if (pass == PASS_SCAN)
385 info->start_transaction = first_commit_ID;
387 jbd_debug(1, "Starting recovery pass %d\n", pass);
390 * Now we walk through the log, transaction by transaction,
391 * making sure that each transaction has a commit block in the
392 * expected place. Each complete transaction gets replayed back
393 * into the main filesystem.
399 journal_block_tag_t * tag;
400 struct buffer_head * obh;
401 struct buffer_head * nbh;
405 /* If we already know where to stop the log traversal,
406 * check right now that we haven't gone past the end of
409 if (pass != PASS_SCAN)
410 if (tid_geq(next_commit_ID, info->end_transaction))
413 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
414 next_commit_ID, next_log_block, journal->j_last);
416 /* Skip over each chunk of the transaction looking
417 * either the next descriptor block or the final commit
420 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
421 err = jread(&bh, journal, next_log_block);
426 wrap(journal, next_log_block);
428 /* What kind of buffer is it?
430 * If it is a descriptor block, check that it has the
431 * expected sequence number. Otherwise, we're all done
434 tmp = (journal_header_t *)bh->b_data;
436 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
441 blocktype = be32_to_cpu(tmp->h_blocktype);
442 sequence = be32_to_cpu(tmp->h_sequence);
443 jbd_debug(3, "Found magic %d, sequence %d\n",
444 blocktype, sequence);
446 if (sequence != next_commit_ID) {
451 /* OK, we have a valid descriptor block which matches
452 * all of the sequence number checks. What are we going
453 * to do with it? That depends on the pass... */
456 case JBD2_DESCRIPTOR_BLOCK:
457 /* If it is a valid descriptor block, replay it
458 * in pass REPLAY; if journal_checksums enabled, then
459 * calculate checksums in PASS_SCAN, otherwise,
460 * just skip over the blocks it describes. */
461 if (pass != PASS_REPLAY) {
462 if (pass == PASS_SCAN &&
463 JBD2_HAS_COMPAT_FEATURE(journal,
464 JBD2_FEATURE_COMPAT_CHECKSUM) &&
465 !info->end_transaction) {
466 if (calc_chksums(journal, bh,
475 next_log_block += count_tags(journal, bh);
476 wrap(journal, next_log_block);
481 /* A descriptor block: we can now write all of
482 * the data blocks. Yay, useful work is finally
483 * getting done here! */
485 tagp = &bh->b_data[sizeof(journal_header_t)];
486 while ((tagp - bh->b_data + tag_bytes)
487 <= journal->j_blocksize) {
488 unsigned long io_block;
490 tag = (journal_block_tag_t *) tagp;
491 flags = be32_to_cpu(tag->t_flags);
493 io_block = next_log_block++;
494 wrap(journal, next_log_block);
495 err = jread(&obh, journal, io_block);
497 /* Recover what we can, but
498 * report failure at the end. */
501 "JBD: IO error %d recovering "
502 "block %ld in log\n",
505 unsigned long long blocknr;
507 J_ASSERT(obh != NULL);
508 blocknr = read_tag_block(tag_bytes,
511 /* If the block has been
512 * revoked, then we're all done
514 if (jbd2_journal_test_revoke
518 ++info->nr_revoke_hits;
522 /* Find a buffer for the new
523 * data being restored */
524 nbh = __getblk(journal->j_fs_dev,
526 journal->j_blocksize);
529 "JBD: Out of memory "
530 "during recovery.\n");
538 memcpy(nbh->b_data, obh->b_data,
539 journal->j_blocksize);
540 if (flags & JBD2_FLAG_ESCAPE) {
541 *((__be32 *)nbh->b_data) =
542 cpu_to_be32(JBD2_MAGIC_NUMBER);
545 BUFFER_TRACE(nbh, "marking dirty");
546 set_buffer_uptodate(nbh);
547 mark_buffer_dirty(nbh);
548 BUFFER_TRACE(nbh, "marking uptodate");
550 /* ll_rw_block(WRITE, 1, &nbh); */
558 if (!(flags & JBD2_FLAG_SAME_UUID))
561 if (flags & JBD2_FLAG_LAST_TAG)
568 case JBD2_COMMIT_BLOCK:
569 /* How to differentiate between interrupted commit
570 * and journal corruption ?
573 * Checksum Verification Failed
575 * ____________________
577 * async_commit sync_commit
579 * | GO TO NEXT "Journal Corruption"
582 * {(n+1)th transanction}
584 * _______|______________
586 * Commit block found Commit block not found
588 * "Journal Corruption" |
589 * _____________|_________
591 * nth trans corrupt OR nth trans
592 * and (n+1)th interrupted interrupted
593 * before commit block
594 * could reach the disk.
595 * (Cannot find the difference in above
596 * mentioned conditions. Hence assume
597 * "Interrupted Commit".)
600 /* Found an expected commit block: if checksums
601 * are present verify them in PASS_SCAN; else not
602 * much to do other than move on to the next sequence
604 if (pass == PASS_SCAN &&
605 JBD2_HAS_COMPAT_FEATURE(journal,
606 JBD2_FEATURE_COMPAT_CHECKSUM)) {
607 int chksum_err, chksum_seen;
608 struct commit_header *cbh =
609 (struct commit_header *)bh->b_data;
610 unsigned found_chksum =
611 be32_to_cpu(cbh->h_chksum[0]);
613 chksum_err = chksum_seen = 0;
615 if (info->end_transaction) {
616 journal->j_failed_commit =
617 info->end_transaction;
622 if (crc32_sum == found_chksum &&
623 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
624 cbh->h_chksum_size ==
625 JBD2_CRC32_CHKSUM_SIZE)
627 else if (!(cbh->h_chksum_type == 0 &&
628 cbh->h_chksum_size == 0 &&
632 * If fs is mounted using an old kernel and then
633 * kernel with journal_chksum is used then we
634 * get a situation where the journal flag has
635 * checksum flag set but checksums are not
636 * present i.e chksum = 0, in the individual
638 * Hence to avoid checksum failures, in this
639 * situation, this extra check is added.
644 info->end_transaction = next_commit_ID;
646 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
647 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
648 journal->j_failed_commit =
660 case JBD2_REVOKE_BLOCK:
661 /* If we aren't in the REVOKE pass, then we can
662 * just skip over this block. */
663 if (pass != PASS_REVOKE) {
668 err = scan_revoke_records(journal, bh,
669 next_commit_ID, info);
676 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
685 * We broke out of the log scan loop: either we came to the
686 * known end of the log or we found an unexpected block in the
687 * log. If the latter happened, then we know that the "current"
688 * transaction marks the end of the valid log.
691 if (pass == PASS_SCAN) {
692 if (!info->end_transaction)
693 info->end_transaction = next_commit_ID;
695 /* It's really bad news if different passes end up at
696 * different places (but possible due to IO errors). */
697 if (info->end_transaction != next_commit_ID) {
698 printk (KERN_ERR "JBD: recovery pass %d ended at "
699 "transaction %u, expected %u\n",
700 pass, next_commit_ID, info->end_transaction);
713 /* Scan a revoke record, marking all blocks mentioned as revoked. */
715 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
716 tid_t sequence, struct recovery_info *info)
718 jbd2_journal_revoke_header_t *header;
722 header = (jbd2_journal_revoke_header_t *) bh->b_data;
723 offset = sizeof(jbd2_journal_revoke_header_t);
724 max = be32_to_cpu(header->r_count);
726 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
729 while (offset + record_len <= max) {
730 unsigned long long blocknr;
734 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
736 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
737 offset += record_len;
738 err = jbd2_journal_set_revoke(journal, blocknr, sequence);