Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq
[linux-2.6/cjktty.git] / fs / jbd2 / recovery.c
blob2bc4d5f116f190048782825738cb19485376227c
1 /*
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.
16 #ifndef __KERNEL__
17 #include "jfs_user.h"
18 #else
19 #include <linux/time.h>
20 #include <linux/fs.h>
21 #include <linux/jbd2.h>
22 #include <linux/errno.h>
23 #include <linux/crc32.h>
24 #endif
27 * Maintain information about the progress of the recovery job, so that
28 * the different passes can carry information between them.
30 struct recovery_info
32 tid_t start_transaction;
33 tid_t end_transaction;
35 int nr_replays;
36 int nr_revokes;
37 int nr_revoke_hits;
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 *);
46 #ifdef __KERNEL__
48 /* Release readahead buffers after use */
49 static void journal_brelse_array(struct buffer_head *b[], int n)
51 while (--n >= 0)
52 brelse (b[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
65 * fixed value.
68 #define MAXBUF 8
69 static int do_readahead(journal_t *journal, unsigned int start)
71 int err;
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. */
86 nbufs = 0;
88 for (next = start; next < max; next++) {
89 err = jbd2_journal_bmap(journal, next, &blocknr);
91 if (err) {
92 printk (KERN_ERR "JBD: bad block at offset %u\n",
93 next);
94 goto failed;
97 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
98 if (!bh) {
99 err = -ENOMEM;
100 goto failed;
103 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
104 bufs[nbufs++] = bh;
105 if (nbufs == MAXBUF) {
106 ll_rw_block(READ, nbufs, bufs);
107 journal_brelse_array(bufs, nbufs);
108 nbufs = 0;
110 } else
111 brelse(bh);
114 if (nbufs)
115 ll_rw_block(READ, nbufs, bufs);
116 err = 0;
118 failed:
119 if (nbufs)
120 journal_brelse_array(bufs, nbufs);
121 return err;
124 #endif /* __KERNEL__ */
128 * Read a block from the journal
131 static int jread(struct buffer_head **bhp, journal_t *journal,
132 unsigned int offset)
134 int err;
135 unsigned long long blocknr;
136 struct buffer_head *bh;
138 *bhp = NULL;
140 if (offset >= journal->j_maxlen) {
141 printk(KERN_ERR "JBD: corrupted journal superblock\n");
142 return -EIO;
145 err = jbd2_journal_bmap(journal, offset, &blocknr);
147 if (err) {
148 printk (KERN_ERR "JBD: bad block at offset %u\n",
149 offset);
150 return err;
153 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
154 if (!bh)
155 return -ENOMEM;
157 if (!buffer_uptodate(bh)) {
158 /* If this is a brand new buffer, start readahead.
159 Otherwise, we assume we are already reading it. */
160 if (!buffer_req(bh))
161 do_readahead(journal, offset);
162 wait_on_buffer(bh);
165 if (!buffer_uptodate(bh)) {
166 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
167 offset);
168 brelse(bh);
169 return -EIO;
172 *bhp = bh;
173 return 0;
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)
183 char * tagp;
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;
193 nr++;
194 tagp += tag_bytes;
195 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID)))
196 tagp += 16;
198 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG))
199 break;
202 return nr;
206 /* Make sure we wrap around the log correctly! */
207 #define wrap(journal, var) \
208 do { \
209 if (var >= (journal)->j_last) \
210 var -= ((journal)->j_last - (journal)->j_first); \
211 } while (0)
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
218 * journaled device.
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
223 * in the log.
225 int jbd2_journal_recover(journal_t *journal)
227 int err, err2;
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
238 * unmounted.
241 if (!sb->s_start) {
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;
245 return 0;
248 err = do_one_pass(journal, &info, PASS_SCAN);
249 if (!err)
250 err = do_one_pass(journal, &info, PASS_REVOKE);
251 if (!err)
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);
266 if (!err)
267 err = err2;
269 return err;
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)
287 int err;
289 struct recovery_info info;
291 memset (&info, 0, sizeof(info));
293 err = do_one_pass(journal, &info, PASS_SCAN);
295 if (err) {
296 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
297 ++journal->j_transaction_sequence;
298 } else {
299 #ifdef CONFIG_JBD2_DEBUG
300 int dropped = info.end_transaction -
301 be32_to_cpu(journal->j_superblock->s_sequence);
302 #endif
303 jbd_debug(1,
304 "JBD: ignoring %d transaction%s from the journal.\n",
305 dropped, (dropped == 1) ? "" : "s");
306 journal->j_transaction_sequence = ++info.end_transaction;
309 journal->j_tail = 0;
310 return err;
313 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
315 unsigned long long block = be32_to_cpu(tag->t_blocknr);
316 if (tag_bytes > JBD2_TAG_SIZE32)
317 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
318 return block;
322 * calc_chksums calculates the checksums for the blocks described in the
323 * descriptor block.
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);
340 if (err) {
341 printk(KERN_ERR "JBD: IO error %d recovering block "
342 "%lu in log\n", err, io_block);
343 return 1;
344 } else {
345 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
346 obh->b_size);
348 put_bh(obh);
350 return 0;
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;
363 int blocktype;
364 int tag_bytes = journal_tag_bytes(journal);
365 __u32 crc32_sum = ~0; /* Transactional Checksums */
368 * First thing is to establish what we expect to find in the log
369 * (in terms of transaction IDs), and where (in terms of log
370 * block offsets): query the superblock.
373 sb = journal->j_superblock;
374 next_commit_ID = be32_to_cpu(sb->s_sequence);
375 next_log_block = be32_to_cpu(sb->s_start);
377 first_commit_ID = next_commit_ID;
378 if (pass == PASS_SCAN)
379 info->start_transaction = first_commit_ID;
381 jbd_debug(1, "Starting recovery pass %d\n", pass);
384 * Now we walk through the log, transaction by transaction,
385 * making sure that each transaction has a commit block in the
386 * expected place. Each complete transaction gets replayed back
387 * into the main filesystem.
390 while (1) {
391 int flags;
392 char * tagp;
393 journal_block_tag_t * tag;
394 struct buffer_head * obh;
395 struct buffer_head * nbh;
397 cond_resched();
399 /* If we already know where to stop the log traversal,
400 * check right now that we haven't gone past the end of
401 * the log. */
403 if (pass != PASS_SCAN)
404 if (tid_geq(next_commit_ID, info->end_transaction))
405 break;
407 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
408 next_commit_ID, next_log_block, journal->j_last);
410 /* Skip over each chunk of the transaction looking
411 * either the next descriptor block or the final commit
412 * record. */
414 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
415 err = jread(&bh, journal, next_log_block);
416 if (err)
417 goto failed;
419 next_log_block++;
420 wrap(journal, next_log_block);
422 /* What kind of buffer is it?
424 * If it is a descriptor block, check that it has the
425 * expected sequence number. Otherwise, we're all done
426 * here. */
428 tmp = (journal_header_t *)bh->b_data;
430 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
431 brelse(bh);
432 break;
435 blocktype = be32_to_cpu(tmp->h_blocktype);
436 sequence = be32_to_cpu(tmp->h_sequence);
437 jbd_debug(3, "Found magic %d, sequence %d\n",
438 blocktype, sequence);
440 if (sequence != next_commit_ID) {
441 brelse(bh);
442 break;
445 /* OK, we have a valid descriptor block which matches
446 * all of the sequence number checks. What are we going
447 * to do with it? That depends on the pass... */
449 switch(blocktype) {
450 case JBD2_DESCRIPTOR_BLOCK:
451 /* If it is a valid descriptor block, replay it
452 * in pass REPLAY; if journal_checksums enabled, then
453 * calculate checksums in PASS_SCAN, otherwise,
454 * just skip over the blocks it describes. */
455 if (pass != PASS_REPLAY) {
456 if (pass == PASS_SCAN &&
457 JBD2_HAS_COMPAT_FEATURE(journal,
458 JBD2_FEATURE_COMPAT_CHECKSUM) &&
459 !info->end_transaction) {
460 if (calc_chksums(journal, bh,
461 &next_log_block,
462 &crc32_sum)) {
463 put_bh(bh);
464 break;
466 put_bh(bh);
467 continue;
469 next_log_block += count_tags(journal, bh);
470 wrap(journal, next_log_block);
471 put_bh(bh);
472 continue;
475 /* A descriptor block: we can now write all of
476 * the data blocks. Yay, useful work is finally
477 * getting done here! */
479 tagp = &bh->b_data[sizeof(journal_header_t)];
480 while ((tagp - bh->b_data + tag_bytes)
481 <= journal->j_blocksize) {
482 unsigned long io_block;
484 tag = (journal_block_tag_t *) tagp;
485 flags = be32_to_cpu(tag->t_flags);
487 io_block = next_log_block++;
488 wrap(journal, next_log_block);
489 err = jread(&obh, journal, io_block);
490 if (err) {
491 /* Recover what we can, but
492 * report failure at the end. */
493 success = err;
494 printk (KERN_ERR
495 "JBD: IO error %d recovering "
496 "block %ld in log\n",
497 err, io_block);
498 } else {
499 unsigned long long blocknr;
501 J_ASSERT(obh != NULL);
502 blocknr = read_tag_block(tag_bytes,
503 tag);
505 /* If the block has been
506 * revoked, then we're all done
507 * here. */
508 if (jbd2_journal_test_revoke
509 (journal, blocknr,
510 next_commit_ID)) {
511 brelse(obh);
512 ++info->nr_revoke_hits;
513 goto skip_write;
516 /* Find a buffer for the new
517 * data being restored */
518 nbh = __getblk(journal->j_fs_dev,
519 blocknr,
520 journal->j_blocksize);
521 if (nbh == NULL) {
522 printk(KERN_ERR
523 "JBD: Out of memory "
524 "during recovery.\n");
525 err = -ENOMEM;
526 brelse(bh);
527 brelse(obh);
528 goto failed;
531 lock_buffer(nbh);
532 memcpy(nbh->b_data, obh->b_data,
533 journal->j_blocksize);
534 if (flags & JBD2_FLAG_ESCAPE) {
535 *((__be32 *)nbh->b_data) =
536 cpu_to_be32(JBD2_MAGIC_NUMBER);
539 BUFFER_TRACE(nbh, "marking dirty");
540 set_buffer_uptodate(nbh);
541 mark_buffer_dirty(nbh);
542 BUFFER_TRACE(nbh, "marking uptodate");
543 ++info->nr_replays;
544 /* ll_rw_block(WRITE, 1, &nbh); */
545 unlock_buffer(nbh);
546 brelse(obh);
547 brelse(nbh);
550 skip_write:
551 tagp += tag_bytes;
552 if (!(flags & JBD2_FLAG_SAME_UUID))
553 tagp += 16;
555 if (flags & JBD2_FLAG_LAST_TAG)
556 break;
559 brelse(bh);
560 continue;
562 case JBD2_COMMIT_BLOCK:
563 /* How to differentiate between interrupted commit
564 * and journal corruption ?
566 * {nth transaction}
567 * Checksum Verification Failed
569 * ____________________
570 * | |
571 * async_commit sync_commit
572 * | |
573 * | GO TO NEXT "Journal Corruption"
574 * | TRANSACTION
576 * {(n+1)th transanction}
578 * _______|______________
579 * | |
580 * Commit block found Commit block not found
581 * | |
582 * "Journal Corruption" |
583 * _____________|_________
584 * | |
585 * nth trans corrupt OR nth trans
586 * and (n+1)th interrupted interrupted
587 * before commit block
588 * could reach the disk.
589 * (Cannot find the difference in above
590 * mentioned conditions. Hence assume
591 * "Interrupted Commit".)
594 /* Found an expected commit block: if checksums
595 * are present verify them in PASS_SCAN; else not
596 * much to do other than move on to the next sequence
597 * number. */
598 if (pass == PASS_SCAN &&
599 JBD2_HAS_COMPAT_FEATURE(journal,
600 JBD2_FEATURE_COMPAT_CHECKSUM)) {
601 int chksum_err, chksum_seen;
602 struct commit_header *cbh =
603 (struct commit_header *)bh->b_data;
604 unsigned found_chksum =
605 be32_to_cpu(cbh->h_chksum[0]);
607 chksum_err = chksum_seen = 0;
609 if (info->end_transaction) {
610 journal->j_failed_commit =
611 info->end_transaction;
612 brelse(bh);
613 break;
616 if (crc32_sum == found_chksum &&
617 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
618 cbh->h_chksum_size ==
619 JBD2_CRC32_CHKSUM_SIZE)
620 chksum_seen = 1;
621 else if (!(cbh->h_chksum_type == 0 &&
622 cbh->h_chksum_size == 0 &&
623 found_chksum == 0 &&
624 !chksum_seen))
626 * If fs is mounted using an old kernel and then
627 * kernel with journal_chksum is used then we
628 * get a situation where the journal flag has
629 * checksum flag set but checksums are not
630 * present i.e chksum = 0, in the individual
631 * commit blocks.
632 * Hence to avoid checksum failures, in this
633 * situation, this extra check is added.
635 chksum_err = 1;
637 if (chksum_err) {
638 info->end_transaction = next_commit_ID;
640 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
641 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
642 journal->j_failed_commit =
643 next_commit_ID;
644 brelse(bh);
645 break;
648 crc32_sum = ~0;
650 brelse(bh);
651 next_commit_ID++;
652 continue;
654 case JBD2_REVOKE_BLOCK:
655 /* If we aren't in the REVOKE pass, then we can
656 * just skip over this block. */
657 if (pass != PASS_REVOKE) {
658 brelse(bh);
659 continue;
662 err = scan_revoke_records(journal, bh,
663 next_commit_ID, info);
664 brelse(bh);
665 if (err)
666 goto failed;
667 continue;
669 default:
670 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
671 blocktype);
672 brelse(bh);
673 goto done;
677 done:
679 * We broke out of the log scan loop: either we came to the
680 * known end of the log or we found an unexpected block in the
681 * log. If the latter happened, then we know that the "current"
682 * transaction marks the end of the valid log.
685 if (pass == PASS_SCAN) {
686 if (!info->end_transaction)
687 info->end_transaction = next_commit_ID;
688 } else {
689 /* It's really bad news if different passes end up at
690 * different places (but possible due to IO errors). */
691 if (info->end_transaction != next_commit_ID) {
692 printk (KERN_ERR "JBD: recovery pass %d ended at "
693 "transaction %u, expected %u\n",
694 pass, next_commit_ID, info->end_transaction);
695 if (!success)
696 success = -EIO;
700 return success;
702 failed:
703 return err;
707 /* Scan a revoke record, marking all blocks mentioned as revoked. */
709 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
710 tid_t sequence, struct recovery_info *info)
712 jbd2_journal_revoke_header_t *header;
713 int offset, max;
714 int record_len = 4;
716 header = (jbd2_journal_revoke_header_t *) bh->b_data;
717 offset = sizeof(jbd2_journal_revoke_header_t);
718 max = be32_to_cpu(header->r_count);
720 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
721 record_len = 8;
723 while (offset + record_len <= max) {
724 unsigned long long blocknr;
725 int err;
727 if (record_len == 4)
728 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
729 else
730 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
731 offset += record_len;
732 err = jbd2_journal_set_revoke(journal, blocknr, sequence);
733 if (err)
734 return err;
735 ++info->nr_revokes;
737 return 0;