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 do_one_pass(journal_t *journal,
379 struct recovery_info *info, enum passtype pass)
381 unsigned int first_commit_ID, next_commit_ID;
382 unsigned long next_log_block;
383 int err, success = 0;
384 journal_superblock_t * sb;
385 journal_header_t * tmp;
386 struct buffer_head * bh;
387 unsigned int sequence;
389 int tag_bytes = journal_tag_bytes(journal);
390 __u32 crc32_sum = ~0; /* Transactional Checksums */
391 int descr_csum_size = 0;
394 * First thing is to establish what we expect to find in the log
395 * (in terms of transaction IDs), and where (in terms of log
396 * block offsets): query the superblock.
399 sb = journal->j_superblock;
400 next_commit_ID = be32_to_cpu(sb->s_sequence);
401 next_log_block = be32_to_cpu(sb->s_start);
403 first_commit_ID = next_commit_ID;
404 if (pass == PASS_SCAN)
405 info->start_transaction = first_commit_ID;
407 jbd_debug(1, "Starting recovery pass %d\n", pass);
410 * Now we walk through the log, transaction by transaction,
411 * making sure that each transaction has a commit block in the
412 * expected place. Each complete transaction gets replayed back
413 * into the main filesystem.
419 journal_block_tag_t * tag;
420 struct buffer_head * obh;
421 struct buffer_head * nbh;
425 /* If we already know where to stop the log traversal,
426 * check right now that we haven't gone past the end of
429 if (pass != PASS_SCAN)
430 if (tid_geq(next_commit_ID, info->end_transaction))
433 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
434 next_commit_ID, next_log_block, journal->j_last);
436 /* Skip over each chunk of the transaction looking
437 * either the next descriptor block or the final commit
440 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
441 err = jread(&bh, journal, next_log_block);
446 wrap(journal, next_log_block);
448 /* What kind of buffer is it?
450 * If it is a descriptor block, check that it has the
451 * expected sequence number. Otherwise, we're all done
454 tmp = (journal_header_t *)bh->b_data;
456 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
461 blocktype = be32_to_cpu(tmp->h_blocktype);
462 sequence = be32_to_cpu(tmp->h_sequence);
463 jbd_debug(3, "Found magic %d, sequence %d\n",
464 blocktype, sequence);
466 if (sequence != next_commit_ID) {
471 /* OK, we have a valid descriptor block which matches
472 * all of the sequence number checks. What are we going
473 * to do with it? That depends on the pass... */
476 case JBD2_DESCRIPTOR_BLOCK:
477 /* Verify checksum first */
478 if (JBD2_HAS_INCOMPAT_FEATURE(journal,
479 JBD2_FEATURE_INCOMPAT_CSUM_V2))
481 sizeof(struct jbd2_journal_block_tail);
482 if (descr_csum_size > 0 &&
483 !jbd2_descr_block_csum_verify(journal,
489 /* If it is a valid descriptor block, replay it
490 * in pass REPLAY; if journal_checksums enabled, then
491 * calculate checksums in PASS_SCAN, otherwise,
492 * just skip over the blocks it describes. */
493 if (pass != PASS_REPLAY) {
494 if (pass == PASS_SCAN &&
495 JBD2_HAS_COMPAT_FEATURE(journal,
496 JBD2_FEATURE_COMPAT_CHECKSUM) &&
497 !info->end_transaction) {
498 if (calc_chksums(journal, bh,
507 next_log_block += count_tags(journal, bh);
508 wrap(journal, next_log_block);
513 /* A descriptor block: we can now write all of
514 * the data blocks. Yay, useful work is finally
515 * getting done here! */
517 tagp = &bh->b_data[sizeof(journal_header_t)];
518 while ((tagp - bh->b_data + tag_bytes)
519 <= journal->j_blocksize - descr_csum_size) {
520 unsigned long io_block;
522 tag = (journal_block_tag_t *) tagp;
523 flags = be16_to_cpu(tag->t_flags);
525 io_block = next_log_block++;
526 wrap(journal, next_log_block);
527 err = jread(&obh, journal, io_block);
529 /* Recover what we can, but
530 * report failure at the end. */
533 "JBD2: IO error %d recovering "
534 "block %ld in log\n",
537 unsigned long long blocknr;
539 J_ASSERT(obh != NULL);
540 blocknr = read_tag_block(tag_bytes,
543 /* If the block has been
544 * revoked, then we're all done
546 if (jbd2_journal_test_revoke
550 ++info->nr_revoke_hits;
554 /* Find a buffer for the new
555 * data being restored */
556 nbh = __getblk(journal->j_fs_dev,
558 journal->j_blocksize);
561 "JBD2: Out of memory "
562 "during recovery.\n");
570 memcpy(nbh->b_data, obh->b_data,
571 journal->j_blocksize);
572 if (flags & JBD2_FLAG_ESCAPE) {
573 *((__be32 *)nbh->b_data) =
574 cpu_to_be32(JBD2_MAGIC_NUMBER);
577 BUFFER_TRACE(nbh, "marking dirty");
578 set_buffer_uptodate(nbh);
579 mark_buffer_dirty(nbh);
580 BUFFER_TRACE(nbh, "marking uptodate");
582 /* ll_rw_block(WRITE, 1, &nbh); */
590 if (!(flags & JBD2_FLAG_SAME_UUID))
593 if (flags & JBD2_FLAG_LAST_TAG)
600 case JBD2_COMMIT_BLOCK:
601 /* How to differentiate between interrupted commit
602 * and journal corruption ?
605 * Checksum Verification Failed
607 * ____________________
609 * async_commit sync_commit
611 * | GO TO NEXT "Journal Corruption"
614 * {(n+1)th transanction}
616 * _______|______________
618 * Commit block found Commit block not found
620 * "Journal Corruption" |
621 * _____________|_________
623 * nth trans corrupt OR nth trans
624 * and (n+1)th interrupted interrupted
625 * before commit block
626 * could reach the disk.
627 * (Cannot find the difference in above
628 * mentioned conditions. Hence assume
629 * "Interrupted Commit".)
632 /* Found an expected commit block: if checksums
633 * are present verify them in PASS_SCAN; else not
634 * much to do other than move on to the next sequence
636 if (pass == PASS_SCAN &&
637 JBD2_HAS_COMPAT_FEATURE(journal,
638 JBD2_FEATURE_COMPAT_CHECKSUM)) {
639 int chksum_err, chksum_seen;
640 struct commit_header *cbh =
641 (struct commit_header *)bh->b_data;
642 unsigned found_chksum =
643 be32_to_cpu(cbh->h_chksum[0]);
645 chksum_err = chksum_seen = 0;
647 if (info->end_transaction) {
648 journal->j_failed_commit =
649 info->end_transaction;
654 if (crc32_sum == found_chksum &&
655 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
656 cbh->h_chksum_size ==
657 JBD2_CRC32_CHKSUM_SIZE)
659 else if (!(cbh->h_chksum_type == 0 &&
660 cbh->h_chksum_size == 0 &&
664 * If fs is mounted using an old kernel and then
665 * kernel with journal_chksum is used then we
666 * get a situation where the journal flag has
667 * checksum flag set but checksums are not
668 * present i.e chksum = 0, in the individual
670 * Hence to avoid checksum failures, in this
671 * situation, this extra check is added.
676 info->end_transaction = next_commit_ID;
678 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
679 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
680 journal->j_failed_commit =
692 case JBD2_REVOKE_BLOCK:
693 /* If we aren't in the REVOKE pass, then we can
694 * just skip over this block. */
695 if (pass != PASS_REVOKE) {
700 err = scan_revoke_records(journal, bh,
701 next_commit_ID, info);
708 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
717 * We broke out of the log scan loop: either we came to the
718 * known end of the log or we found an unexpected block in the
719 * log. If the latter happened, then we know that the "current"
720 * transaction marks the end of the valid log.
723 if (pass == PASS_SCAN) {
724 if (!info->end_transaction)
725 info->end_transaction = next_commit_ID;
727 /* It's really bad news if different passes end up at
728 * different places (but possible due to IO errors). */
729 if (info->end_transaction != next_commit_ID) {
730 printk(KERN_ERR "JBD2: recovery pass %d ended at "
731 "transaction %u, expected %u\n",
732 pass, next_commit_ID, info->end_transaction);
744 static int jbd2_revoke_block_csum_verify(journal_t *j,
747 struct jbd2_journal_revoke_tail *tail;
748 __u32 provided, calculated;
750 if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
753 tail = (struct jbd2_journal_revoke_tail *)(buf + j->j_blocksize -
754 sizeof(struct jbd2_journal_revoke_tail));
755 provided = tail->r_checksum;
756 tail->r_checksum = 0;
757 calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
758 tail->r_checksum = provided;
760 provided = be32_to_cpu(provided);
761 return provided == calculated;
764 /* Scan a revoke record, marking all blocks mentioned as revoked. */
766 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
767 tid_t sequence, struct recovery_info *info)
769 jbd2_journal_revoke_header_t *header;
773 header = (jbd2_journal_revoke_header_t *) bh->b_data;
774 offset = sizeof(jbd2_journal_revoke_header_t);
775 max = be32_to_cpu(header->r_count);
777 if (!jbd2_revoke_block_csum_verify(journal, header))
780 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
783 while (offset + record_len <= max) {
784 unsigned long long blocknr;
788 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
790 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
791 offset += record_len;
792 err = jbd2_journal_set_revoke(journal, blocknr, sequence);