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",
179 * Count the number of in-use tags in a journal descriptor block.
182 static int count_tags(journal_t *journal, struct buffer_head *bh)
185 journal_block_tag_t * tag;
186 int nr = 0, size = journal->j_blocksize;
187 int tag_bytes = journal_tag_bytes(journal);
189 tagp = &bh->b_data[sizeof(journal_header_t)];
191 while ((tagp - bh->b_data + tag_bytes) <= size) {
192 tag = (journal_block_tag_t *) tagp;
196 if (!(tag->t_flags & cpu_to_be32(JFS_FLAG_SAME_UUID)))
199 if (tag->t_flags & cpu_to_be32(JFS_FLAG_LAST_TAG))
207 /* Make sure we wrap around the log correctly! */
208 #define wrap(journal, var) \
210 if (var >= (journal)->j_last) \
211 var -= ((journal)->j_last - (journal)->j_first); \
215 * journal_recover - recovers a on-disk journal
216 * @journal: the journal to recover
218 * The primary function for recovering the log contents when mounting a
221 * Recovery is done in three passes. In the first pass, we look for the
222 * end of the log. In the second, we assemble the list of revoke
223 * blocks. In the third and final pass, we replay any un-revoked blocks
226 int journal_recover(journal_t *journal)
229 journal_superblock_t * sb;
231 struct recovery_info info;
233 memset(&info, 0, sizeof(info));
234 sb = journal->j_superblock;
237 * The journal superblock's s_start field (the current log head)
238 * is always zero if, and only if, the journal was cleanly
243 jbd_debug(1, "No recovery required, last transaction %d\n",
244 be32_to_cpu(sb->s_sequence));
245 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
249 err = do_one_pass(journal, &info, PASS_SCAN);
251 err = do_one_pass(journal, &info, PASS_REVOKE);
253 err = do_one_pass(journal, &info, PASS_REPLAY);
255 jbd_debug(1, "JBD: recovery, exit status %d, "
256 "recovered transactions %u to %u\n",
257 err, info.start_transaction, info.end_transaction);
258 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
259 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
261 /* Restart the log at the next transaction ID, thus invalidating
262 * any existing commit records in the log. */
263 journal->j_transaction_sequence = ++info.end_transaction;
265 journal_clear_revoke(journal);
266 sync_blockdev(journal->j_fs_dev);
271 * journal_skip_recovery - Start journal and wipe exiting records
272 * @journal: journal to startup
274 * Locate any valid recovery information from the journal and set up the
275 * journal structures in memory to ignore it (presumably because the
276 * caller has evidence that it is out of date).
277 * This function does'nt appear to be exorted..
279 * We perform one pass over the journal to allow us to tell the user how
280 * much recovery information is being erased, and to let us initialise
281 * the journal transaction sequence numbers to the next unused ID.
283 int journal_skip_recovery(journal_t *journal)
286 struct recovery_info info;
288 memset (&info, 0, sizeof(info));
290 err = do_one_pass(journal, &info, PASS_SCAN);
293 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
294 ++journal->j_transaction_sequence;
296 #ifdef CONFIG_JBD_DEBUG
297 journal_superblock_t *sb = journal->j_superblock;
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;
311 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
313 unsigned long long block = be32_to_cpu(tag->t_blocknr);
314 if (tag_bytes > JBD_TAG_SIZE32)
315 block |= (__u64)be32_to_cpu(tag->t_blocknr_high) << 32;
320 * calc_chksums calculates the checksums for the blocks described in the
323 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
324 unsigned long long *next_log_block, __u32 *crc32_sum)
326 int i, num_blks, err;
327 unsigned long long io_block;
328 struct buffer_head *obh;
330 num_blks = count_tags(journal, bh);
331 /* Calculate checksum of the descriptor block. */
332 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
334 for (i = 0; i < num_blks; i++) {
335 io_block = (*next_log_block)++;
336 wrap(journal, *next_log_block);
337 err = jread(&obh, journal, io_block);
339 printk(KERN_ERR "JBD: IO error %d recovering block "
340 "%llu in log\n", err, io_block);
343 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
351 static int do_one_pass(journal_t *journal,
352 struct recovery_info *info, enum passtype pass)
354 unsigned int first_commit_ID, next_commit_ID;
355 unsigned long long next_log_block;
356 int err, success = 0;
357 journal_superblock_t * sb;
358 journal_header_t * tmp;
359 struct buffer_head * bh;
360 unsigned int sequence;
362 int tag_bytes = journal_tag_bytes(journal);
363 __u32 crc32_sum = ~0; /* Transactional Checksums */
366 * First thing is to establish what we expect to find in the log
367 * (in terms of transaction IDs), and where (in terms of log
368 * block offsets): query the superblock.
371 sb = journal->j_superblock;
372 next_commit_ID = be32_to_cpu(sb->s_sequence);
373 next_log_block = be32_to_cpu(sb->s_start);
375 first_commit_ID = next_commit_ID;
376 if (pass == PASS_SCAN)
377 info->start_transaction = first_commit_ID;
379 jbd_debug(1, "Starting recovery pass %d\n", pass);
382 * Now we walk through the log, transaction by transaction,
383 * making sure that each transaction has a commit block in the
384 * expected place. Each complete transaction gets replayed back
385 * into the main filesystem.
391 journal_block_tag_t * tag;
392 struct buffer_head * obh;
393 struct buffer_head * nbh;
397 /* If we already know where to stop the log traversal,
398 * check right now that we haven't gone past the end of
401 if (pass != PASS_SCAN)
402 if (tid_geq(next_commit_ID, info->end_transaction))
405 jbd_debug(2, "Scanning for sequence ID %u at %llu/%lu\n",
406 next_commit_ID, next_log_block, journal->j_last);
408 /* Skip over each chunk of the transaction looking
409 * either the next descriptor block or the final commit
412 jbd_debug(3, "JBD: checking block %llu\n", next_log_block);
413 err = jread(&bh, journal, next_log_block);
418 wrap(journal, next_log_block);
420 /* What kind of buffer is it?
422 * If it is a descriptor block, check that it has the
423 * expected sequence number. Otherwise, we're all done
426 tmp = (journal_header_t *)bh->b_data;
428 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
433 blocktype = be32_to_cpu(tmp->h_blocktype);
434 sequence = be32_to_cpu(tmp->h_sequence);
435 jbd_debug(3, "Found magic %d, sequence %d\n",
436 blocktype, sequence);
438 if (sequence != next_commit_ID) {
443 /* OK, we have a valid descriptor block which matches
444 * all of the sequence number checks. What are we going
445 * to do with it? That depends on the pass... */
448 case JFS_DESCRIPTOR_BLOCK:
449 /* If it is a valid descriptor block, replay it
450 * in pass REPLAY; if journal_checksums enabled, then
451 * calculate checksums in PASS_SCAN, otherwise,
452 * just skip over the blocks it describes. */
453 if (pass != PASS_REPLAY) {
454 if (pass == PASS_SCAN &&
455 JFS_HAS_COMPAT_FEATURE(journal,
456 JFS_FEATURE_COMPAT_CHECKSUM) &&
457 !info->end_transaction) {
458 if (calc_chksums(journal, bh,
467 next_log_block += count_tags(journal, bh);
468 wrap(journal, next_log_block);
473 /* A descriptor block: we can now write all of
474 * the data blocks. Yay, useful work is finally
475 * getting done here! */
477 tagp = &bh->b_data[sizeof(journal_header_t)];
478 while ((tagp - bh->b_data + tag_bytes)
479 <= journal->j_blocksize) {
480 unsigned long long io_block;
482 tag = (journal_block_tag_t *) tagp;
483 flags = be32_to_cpu(tag->t_flags);
485 io_block = next_log_block++;
486 wrap(journal, next_log_block);
487 err = jread(&obh, journal, io_block);
489 /* Recover what we can, but
490 * report failure at the end. */
493 "JBD: IO error %d recovering "
494 "block %llu in log\n",
497 unsigned long long blocknr;
499 J_ASSERT(obh != NULL);
500 blocknr = read_tag_block(tag_bytes,
503 /* If the block has been
504 * revoked, then we're all done
506 if (journal_test_revoke
510 ++info->nr_revoke_hits;
514 /* Find a buffer for the new
515 * data being restored */
516 nbh = __getblk(journal->j_fs_dev,
518 journal->j_blocksize);
521 "JBD: Out of memory "
522 "during recovery.\n");
530 memcpy(nbh->b_data, obh->b_data,
531 journal->j_blocksize);
532 if (flags & JFS_FLAG_ESCAPE) {
533 journal_header_t *header;
535 header = (journal_header_t *) &nbh->b_data[0];
536 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
539 BUFFER_TRACE(nbh, "marking dirty");
540 set_buffer_uptodate(nbh);
541 mark_buffer_dirty(nbh);
542 BUFFER_TRACE(nbh, "marking uptodate");
544 /* ll_rw_block(WRITE, 1, &nbh); */
552 if (!(flags & JFS_FLAG_SAME_UUID))
555 if (flags & JFS_FLAG_LAST_TAG)
562 case JFS_COMMIT_BLOCK:
563 jbd_debug(3, "Commit block for #%u found\n",
565 /* How to differentiate between interrupted commit
566 * and journal corruption ?
569 * Checksum Verification Failed
571 * ____________________
573 * async_commit sync_commit
575 * | GO TO NEXT "Journal Corruption"
578 * {(n+1)th transanction}
580 * _______|______________
582 * Commit block found Commit block not found
584 * "Journal Corruption" |
585 * _____________|_________
587 * nth trans corrupt OR nth trans
588 * and (n+1)th interrupted interrupted
589 * before commit block
590 * could reach the disk.
591 * (Cannot find the difference in above
592 * mentioned conditions. Hence assume
593 * "Interrupted Commit".)
596 /* Found an expected commit block: if checksums
597 * are present verify them in PASS_SCAN; else not
598 * much to do other than move on to the next sequence
600 if (pass == PASS_SCAN &&
601 JFS_HAS_COMPAT_FEATURE(journal,
602 JFS_FEATURE_COMPAT_CHECKSUM)) {
603 int chksum_err, chksum_seen;
604 struct commit_header *cbh =
605 (struct commit_header *)bh->b_data;
606 unsigned found_chksum =
607 be32_to_cpu(cbh->h_chksum[0]);
609 chksum_err = chksum_seen = 0;
611 jbd_debug(3, "Checksums %x %x\n",
612 crc32_sum, found_chksum);
613 if (info->end_transaction) {
614 journal->j_failed_commit =
615 info->end_transaction;
620 if (crc32_sum == found_chksum &&
621 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
622 cbh->h_chksum_size ==
623 JBD2_CRC32_CHKSUM_SIZE)
625 else if (!(cbh->h_chksum_type == 0 &&
626 cbh->h_chksum_size == 0 &&
630 * If fs is mounted using an old kernel and then
631 * kernel with journal_chksum is used then we
632 * get a situation where the journal flag has
633 * checksum flag set but checksums are not
634 * present i.e chksum = 0, in the individual
636 * Hence to avoid checksum failures, in this
637 * situation, this extra check is added.
642 info->end_transaction = next_commit_ID;
643 jbd_debug(1, "Checksum_err %x %x\n",
644 crc32_sum, found_chksum);
645 if (!JFS_HAS_INCOMPAT_FEATURE(journal,
646 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
647 journal->j_failed_commit =
659 case JFS_REVOKE_BLOCK:
660 /* If we aren't in the REVOKE pass, then we can
661 * just skip over this block. */
662 if (pass != PASS_REVOKE) {
667 err = scan_revoke_records(journal, bh,
668 next_commit_ID, info);
675 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
684 * We broke out of the log scan loop: either we came to the
685 * known end of the log or we found an unexpected block in the
686 * log. If the latter happened, then we know that the "current"
687 * transaction marks the end of the valid log.
690 if (pass == PASS_SCAN) {
691 if (!info->end_transaction)
692 info->end_transaction = next_commit_ID;
694 /* It's really bad news if different passes end up at
695 * different places (but possible due to IO errors). */
696 if (info->end_transaction != next_commit_ID) {
697 printk (KERN_ERR "JBD: recovery pass %d ended at "
698 "transaction %u, expected %u\n",
699 pass, next_commit_ID, info->end_transaction);
712 /* Scan a revoke record, marking all blocks mentioned as revoked. */
714 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
715 tid_t sequence, struct recovery_info *info)
717 journal_revoke_header_t *header;
721 header = (journal_revoke_header_t *) bh->b_data;
722 offset = sizeof(journal_revoke_header_t);
723 max = be32_to_cpu(header->r_count);
725 if (JFS_HAS_INCOMPAT_FEATURE(journal, JFS_FEATURE_INCOMPAT_64BIT))
728 while (offset < max) {
729 unsigned long long blocknr;
732 if (record_len == 4) {
734 memcpy(&b, bh->b_data + offset, sizeof(__be32));
735 blocknr = ext2fs_be32_to_cpu(b);
738 memcpy(&b, bh->b_data + offset, sizeof(__be64));
739 blocknr = ext2fs_be64_to_cpu(b);
742 offset += record_len;
743 err = journal_set_revoke(journal, blocknr, sequence);