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.
19 #include <linux/time.h>
21 #include <linux/jbd.h>
22 #include <linux/errno.h>
23 #include <linux/slab.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 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 = 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 blocknr;
136 struct buffer_head *bh;
140 if (offset >= journal->j_maxlen) {
141 printk(KERN_ERR "JBD: corrupted journal superblock\n");
145 err = 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(JFS_FLAG_SAME_UUID)))
198 if (tag->t_flags & cpu_to_be32(JFS_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 * 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 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 journal_clear_revoke(journal);
265 sync_blockdev(journal->j_fs_dev);
270 * journal_skip_recovery - Start journal and wipe exiting records
271 * @journal: journal to startup
273 * Locate any valid recovery information from the journal and set up the
274 * journal structures in memory to ignore it (presumably because the
275 * caller has evidence that it is out of date).
276 * This function does'nt appear to be exorted..
278 * We perform one pass over the journal to allow us to tell the user how
279 * much recovery information is being erased, and to let us initialise
280 * the journal transaction sequence numbers to the next unused ID.
282 int journal_skip_recovery(journal_t *journal)
285 journal_superblock_t * sb;
287 struct recovery_info info;
289 memset (&info, 0, sizeof(info));
290 sb = journal->j_superblock;
292 err = do_one_pass(journal, &info, PASS_SCAN);
295 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
296 ++journal->j_transaction_sequence;
298 #ifdef CONFIG_JBD_DEBUG
299 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
302 "JBD: ignoring %d transaction%s from the journal.\n",
303 dropped, (dropped == 1) ? "" : "s");
304 journal->j_transaction_sequence = ++info.end_transaction;
312 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
314 unsigned long long block = be32_to_cpu(tag->t_blocknr);
315 if (tag_bytes > JBD_TAG_SIZE32)
316 block |= (__u64)be32_to_cpu(tag->t_blocknr_high) << 32;
322 * calc_chksums calculates the checksums for the blocks described in the
325 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
326 unsigned long *next_log_block, __u32 *crc32_sum)
328 int i, num_blks, err;
329 unsigned long io_block;
330 struct buffer_head *obh;
332 num_blks = count_tags(journal, bh);
333 /* Calculate checksum of the descriptor block. */
334 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
336 for (i = 0; i < num_blks; i++) {
337 io_block = (*next_log_block)++;
338 wrap(journal, *next_log_block);
339 err = jread(&obh, journal, io_block);
341 printk(KERN_ERR "JBD: IO error %d recovering block "
342 "%lu in log\n", err, io_block);
345 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
353 static int do_one_pass(journal_t *journal,
354 struct recovery_info *info, enum passtype pass)
356 unsigned int first_commit_ID, next_commit_ID;
357 unsigned long next_log_block;
358 int err, success = 0;
359 journal_superblock_t * sb;
360 journal_header_t * tmp;
361 struct buffer_head * bh;
362 unsigned int sequence;
364 int tag_bytes = journal_tag_bytes(journal);
365 __u32 crc32_sum = ~0; /* Transactional Checksums */
367 /* Precompute the maximum metadata descriptors in a descriptor block */
368 int MAX_BLOCKS_PER_DESC;
369 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
373 * First thing is to establish what we expect to find in the log
374 * (in terms of transaction IDs), and where (in terms of log
375 * block offsets): query the superblock.
378 sb = journal->j_superblock;
379 next_commit_ID = be32_to_cpu(sb->s_sequence);
380 next_log_block = be32_to_cpu(sb->s_start);
382 first_commit_ID = next_commit_ID;
383 if (pass == PASS_SCAN)
384 info->start_transaction = first_commit_ID;
386 jbd_debug(1, "Starting recovery pass %d\n", pass);
389 * Now we walk through the log, transaction by transaction,
390 * making sure that each transaction has a commit block in the
391 * expected place. Each complete transaction gets replayed back
392 * into the main filesystem.
398 journal_block_tag_t * tag;
399 struct buffer_head * obh;
400 struct buffer_head * nbh;
404 /* If we already know where to stop the log traversal,
405 * check right now that we haven't gone past the end of
408 if (pass != PASS_SCAN)
409 if (tid_geq(next_commit_ID, info->end_transaction))
412 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
413 next_commit_ID, next_log_block, journal->j_last);
415 /* Skip over each chunk of the transaction looking
416 * either the next descriptor block or the final commit
419 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
420 err = jread(&bh, journal, next_log_block);
425 wrap(journal, next_log_block);
427 /* What kind of buffer is it?
429 * If it is a descriptor block, check that it has the
430 * expected sequence number. Otherwise, we're all done
433 tmp = (journal_header_t *)bh->b_data;
435 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
440 blocktype = be32_to_cpu(tmp->h_blocktype);
441 sequence = be32_to_cpu(tmp->h_sequence);
442 jbd_debug(3, "Found magic %d, sequence %d\n",
443 blocktype, sequence);
445 if (sequence != next_commit_ID) {
450 /* OK, we have a valid descriptor block which matches
451 * all of the sequence number checks. What are we going
452 * to do with it? That depends on the pass... */
455 case JFS_DESCRIPTOR_BLOCK:
456 /* If it is a valid descriptor block, replay it
457 * in pass REPLAY; if journal_checksums enabled, then
458 * calculate checksums in PASS_SCAN, otherwise,
459 * just skip over the blocks it describes. */
460 if (pass != PASS_REPLAY) {
461 if (pass == PASS_SCAN &&
462 JFS_HAS_COMPAT_FEATURE(journal,
463 JFS_FEATURE_COMPAT_CHECKSUM) &&
464 !info->end_transaction) {
465 if (calc_chksums(journal, bh,
474 next_log_block += count_tags(journal, bh);
475 wrap(journal, next_log_block);
480 /* A descriptor block: we can now write all of
481 * the data blocks. Yay, useful work is finally
482 * getting done here! */
484 tagp = &bh->b_data[sizeof(journal_header_t)];
485 while ((tagp - bh->b_data + tag_bytes)
486 <= journal->j_blocksize) {
487 unsigned long io_block;
489 tag = (journal_block_tag_t *) tagp;
490 flags = be32_to_cpu(tag->t_flags);
492 io_block = next_log_block++;
493 wrap(journal, next_log_block);
494 err = jread(&obh, journal, io_block);
496 /* Recover what we can, but
497 * report failure at the end. */
500 "JBD: IO error %d recovering "
501 "block %ld in log\n",
504 unsigned long blocknr;
506 J_ASSERT(obh != NULL);
507 blocknr = be32_to_cpu(tag->t_blocknr);
509 /* If the block has been
510 * revoked, then we're all done
512 if (journal_test_revoke
516 ++info->nr_revoke_hits;
520 /* Find a buffer for the new
521 * data being restored */
522 nbh = __getblk(journal->j_fs_dev,
524 journal->j_blocksize);
527 "JBD: Out of memory "
528 "during recovery.\n");
536 memcpy(nbh->b_data, obh->b_data,
537 journal->j_blocksize);
538 if (flags & JFS_FLAG_ESCAPE) {
539 *((__be32 *)nbh->b_data) =
540 cpu_to_be32(JFS_MAGIC_NUMBER);
543 BUFFER_TRACE(nbh, "marking dirty");
544 set_buffer_uptodate(nbh);
545 mark_buffer_dirty(nbh);
546 BUFFER_TRACE(nbh, "marking uptodate");
548 /* ll_rw_block(WRITE, 1, &nbh); */
556 if (!(flags & JFS_FLAG_SAME_UUID))
559 if (flags & JFS_FLAG_LAST_TAG)
566 case JFS_COMMIT_BLOCK:
567 jbd_debug(3, "Commit block for #%u found\n",
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 JFS_HAS_COMPAT_FEATURE(journal,
606 JFS_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 jbd_debug(3, "Checksums %x %x\n",
616 crc32_sum, found_chksum);
617 if (info->end_transaction) {
618 journal->j_failed_commit =
619 info->end_transaction;
624 if (crc32_sum == found_chksum &&
625 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
626 cbh->h_chksum_size ==
627 JBD2_CRC32_CHKSUM_SIZE)
629 else if (!(cbh->h_chksum_type == 0 &&
630 cbh->h_chksum_size == 0 &&
634 * If fs is mounted using an old kernel and then
635 * kernel with journal_chksum is used then we
636 * get a situation where the journal flag has
637 * checksum flag set but checksums are not
638 * present i.e chksum = 0, in the individual
640 * Hence to avoid checksum failures, in this
641 * situation, this extra check is added.
646 info->end_transaction = next_commit_ID;
647 jbd_debug(1, "Checksum_err %x %x\n",
648 crc32_sum, found_chksum);
649 if (!JFS_HAS_INCOMPAT_FEATURE(journal,
650 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
651 journal->j_failed_commit =
663 case JFS_REVOKE_BLOCK:
664 /* If we aren't in the REVOKE pass, then we can
665 * just skip over this block. */
666 if (pass != PASS_REVOKE) {
671 err = scan_revoke_records(journal, bh,
672 next_commit_ID, info);
679 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
688 * We broke out of the log scan loop: either we came to the
689 * known end of the log or we found an unexpected block in the
690 * log. If the latter happened, then we know that the "current"
691 * transaction marks the end of the valid log.
694 if (pass == PASS_SCAN) {
695 if (!info->end_transaction)
696 info->end_transaction = next_commit_ID;
698 /* It's really bad news if different passes end up at
699 * different places (but possible due to IO errors). */
700 if (info->end_transaction != next_commit_ID) {
701 printk (KERN_ERR "JBD: recovery pass %d ended at "
702 "transaction %u, expected %u\n",
703 pass, next_commit_ID, info->end_transaction);
716 /* Scan a revoke record, marking all blocks mentioned as revoked. */
718 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
719 tid_t sequence, struct recovery_info *info)
721 journal_revoke_header_t *header;
724 header = (journal_revoke_header_t *) bh->b_data;
725 offset = sizeof(journal_revoke_header_t);
726 max = be32_to_cpu(header->r_count);
728 while (offset < max) {
729 unsigned long blocknr;
732 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
734 err = journal_set_revoke(journal, blocknr, sequence);