?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
block: unify flags for struct bio and struct request
[linux-2.6/next.git] / fs / jbd2 / recovery.c
blob049281b7cb8966ab236f8085c65f5230f891aaae
1 /*
2 * linux/fs/jbd2/recovery.c
3 *
4 * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
5 *
6 * Copyright 1999-2000 Red Hat Software --- All Rights Reserved
7 *
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.
11 *
12 * Journal recovery routines for the generic filesystem journaling code;
13 * part of the ext2fs journaling system.
14 */
15
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
25
26 /*
27 * Maintain information about the progress of the recovery job, so that
28 * the different passes can carry information between them.
29 */
30 struct recovery_info
31 {
32 tid_t start_transaction;
33 tid_t end_transaction;
34
35 int nr_replays;
36 int nr_revokes;
37 int nr_revoke_hits;
38 };
39
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 *);
45
46 #ifdef __KERNEL__
47
48 /* Release readahead buffers after use */
49 static void journal_brelse_array(struct buffer_head *b[], int n)
50 {
51 while (--n >= 0)
52 brelse (b[n]);
53 }
54
55
56 /*
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.
62 *
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.
66 */
67
68 #define MAXBUF 8
69 static int do_readahead(journal_t *journal, unsigned int start)
70 {
71 int err;
72 unsigned int max, nbufs, next;
73 unsigned long long blocknr;
74 struct buffer_head *bh;
75
76 struct buffer_head * bufs[MAXBUF];
77
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;
82
83 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
84 * a time to the block device IO layer. */
85
86 nbufs = 0;
87
88 for (next = start; next < max; next++) {
89 err = jbd2_journal_bmap(journal, next, &blocknr);
90
91 if (err) {
92 printk (KERN_ERR "JBD: bad block at offset %u\n",
93 next);
94 goto failed;
95 }
96
97 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
98 if (!bh) {
99 err = -ENOMEM;
100 goto failed;
101 }
102
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;
109 }
110 } else
111 brelse(bh);
112 }
113
114 if (nbufs)
115 ll_rw_block(READ, nbufs, bufs);
116 err = 0;
117
118 failed:
119 if (nbufs)
120 journal_brelse_array(bufs, nbufs);
121 return err;
122 }
123
124 #endif /* __KERNEL__ */
125
126
127 /*
128 * Read a block from the journal
129 */
130
131 static int jread(struct buffer_head **bhp, journal_t *journal,
132 unsigned int offset)
133 {
134 int err;
135 unsigned long long blocknr;
136 struct buffer_head *bh;
137
138 *bhp = NULL;
139
140 if (offset >= journal->j_maxlen) {
141 printk(KERN_ERR "JBD: corrupted journal superblock\n");
142 return -EIO;
143 }
144
145 err = jbd2_journal_bmap(journal, offset, &blocknr);
146
147 if (err) {
148 printk (KERN_ERR "JBD: bad block at offset %u\n",
149 offset);
150 return err;
151 }
152
153 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
154 if (!bh)
155 return -ENOMEM;
156
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);
163 }
164
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;
170 }
171
172 *bhp = bh;
173 return 0;
174 }
175
176
177 /*
178 * Count the number of in-use tags in a journal descriptor block.
179 */
180
181 static int count_tags(journal_t *journal, struct buffer_head *bh)
182 {
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);
187
188 tagp = &bh->b_data[sizeof(journal_header_t)];
189
190 while ((tagp - bh->b_data + tag_bytes) <= size) {
191 tag = (journal_block_tag_t *) tagp;
192
193 nr++;
194 tagp += tag_bytes;
195 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID)))
196 tagp += 16;
197
198 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG))
199 break;
200 }
201
202 return nr;
203 }
204
205
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)
212
213 /**
214 * jbd2_journal_recover - recovers a on-disk journal
215 * @journal: the journal to recover
216 *
217 * The primary function for recovering the log contents when mounting a
218 * journaled device.
219 *
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.
224 */
225 int jbd2_journal_recover(journal_t *journal)
226 {
227 int err, err2;
228 journal_superblock_t * sb;
229
230 struct recovery_info info;
231
232 memset(&info, 0, sizeof(info));
233 sb = journal->j_superblock;
234
235 /*
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.
239 */
240
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;
246 }
247
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);
253
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);
259
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;
263
264 jbd2_journal_clear_revoke(journal);
265 err2 = sync_blockdev(journal->j_fs_dev);
266 if (!err)
267 err = err2;
268
269 return err;
270 }
271
272 /**
273 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
274 * @journal: journal to startup
275 *
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..
280 *
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.
284 */
285 int jbd2_journal_skip_recovery(journal_t *journal)
286 {
287 int err;
288 journal_superblock_t * sb;
289
290 struct recovery_info info;
291
292 memset (&info, 0, sizeof(info));
293 sb = journal->j_superblock;
294
295 err = do_one_pass(journal, &info, PASS_SCAN);
296
297 if (err) {
298 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
299 ++journal->j_transaction_sequence;
300 } else {
301 #ifdef CONFIG_JBD2_DEBUG
302 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
303 #endif
304 jbd_debug(1,
305 "JBD: ignoring %d transaction%s from the journal.\n",
306 dropped, (dropped == 1) ? "" : "s");
307 journal->j_transaction_sequence = ++info.end_transaction;
308 }
309
310 journal->j_tail = 0;
311 return err;
312 }
313
314 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
315 {
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;
319 return block;
320 }
321
322 /*
323 * calc_chksums calculates the checksums for the blocks described in the
324 * descriptor block.
325 */
326 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
327 unsigned long *next_log_block, __u32 *crc32_sum)
328 {
329 int i, num_blks, err;
330 unsigned long io_block;
331 struct buffer_head *obh;
332
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);
336
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);
341 if (err) {
342 printk(KERN_ERR "JBD: IO error %d recovering block "
343 "%lu in log\n", err, io_block);
344 return 1;
345 } else {
346 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
347 obh->b_size);
348 }
349 put_bh(obh);
350 }
351 return 0;
352 }
353
354 static int do_one_pass(journal_t *journal,
355 struct recovery_info *info, enum passtype pass)
356 {
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;
364 int blocktype;
365 int tag_bytes = journal_tag_bytes(journal);
366 __u32 crc32_sum = ~0; /* Transactional Checksums */
367
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))
371 / tag_bytes);
372
373 /*
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.
377 */
378
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);
382
383 first_commit_ID = next_commit_ID;
384 if (pass == PASS_SCAN)
385 info->start_transaction = first_commit_ID;
386
387 jbd_debug(1, "Starting recovery pass %d\n", pass);
388
389 /*
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.
394 */
395
396 while (1) {
397 int flags;
398 char * tagp;
399 journal_block_tag_t * tag;
400 struct buffer_head * obh;
401 struct buffer_head * nbh;
402
403 cond_resched();
404
405 /* If we already know where to stop the log traversal,
406 * check right now that we haven't gone past the end of
407 * the log. */
408
409 if (pass != PASS_SCAN)
410 if (tid_geq(next_commit_ID, info->end_transaction))
411 break;
412
413 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
414 next_commit_ID, next_log_block, journal->j_last);
415
416 /* Skip over each chunk of the transaction looking
417 * either the next descriptor block or the final commit
418 * record. */
419
420 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
421 err = jread(&bh, journal, next_log_block);
422 if (err)
423 goto failed;
424
425 next_log_block++;
426 wrap(journal, next_log_block);
427
428 /* What kind of buffer is it?
429 *
430 * If it is a descriptor block, check that it has the
431 * expected sequence number. Otherwise, we're all done
432 * here. */
433
434 tmp = (journal_header_t *)bh->b_data;
435
436 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
437 brelse(bh);
438 break;
439 }
440
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);
445
446 if (sequence != next_commit_ID) {
447 brelse(bh);
448 break;
449 }
450
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... */
454
455 switch(blocktype) {
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,
467 &next_log_block,
468 &crc32_sum)) {
469 put_bh(bh);
470 break;
471 }
472 put_bh(bh);
473 continue;
474 }
475 next_log_block += count_tags(journal, bh);
476 wrap(journal, next_log_block);
477 put_bh(bh);
478 continue;
479 }
480
481 /* A descriptor block: we can now write all of
482 * the data blocks. Yay, useful work is finally
483 * getting done here! */
484
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;
489
490 tag = (journal_block_tag_t *) tagp;
491 flags = be32_to_cpu(tag->t_flags);
492
493 io_block = next_log_block++;
494 wrap(journal, next_log_block);
495 err = jread(&obh, journal, io_block);
496 if (err) {
497 /* Recover what we can, but
498 * report failure at the end. */
499 success = err;
500 printk (KERN_ERR
501 "JBD: IO error %d recovering "
502 "block %ld in log\n",
503 err, io_block);
504 } else {
505 unsigned long long blocknr;
506
507 J_ASSERT(obh != NULL);
508 blocknr = read_tag_block(tag_bytes,
509 tag);
510
511 /* If the block has been
512 * revoked, then we're all done
513 * here. */
514 if (jbd2_journal_test_revoke
515 (journal, blocknr,
516 next_commit_ID)) {
517 brelse(obh);
518 ++info->nr_revoke_hits;
519 goto skip_write;
520 }
521
522 /* Find a buffer for the new
523 * data being restored */
524 nbh = __getblk(journal->j_fs_dev,
525 blocknr,
526 journal->j_blocksize);
527 if (nbh == NULL) {
528 printk(KERN_ERR
529 "JBD: Out of memory "
530 "during recovery.\n");
531 err = -ENOMEM;
532 brelse(bh);
533 brelse(obh);
534 goto failed;
535 }
536
537 lock_buffer(nbh);
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);
543 }
544
545 BUFFER_TRACE(nbh, "marking dirty");
546 set_buffer_uptodate(nbh);
547 mark_buffer_dirty(nbh);
548 BUFFER_TRACE(nbh, "marking uptodate");
549 ++info->nr_replays;
550 /* ll_rw_block(WRITE, 1, &nbh); */
551 unlock_buffer(nbh);
552 brelse(obh);
553 brelse(nbh);
554 }
555
556 skip_write:
557 tagp += tag_bytes;
558 if (!(flags & JBD2_FLAG_SAME_UUID))
559 tagp += 16;
560
561 if (flags & JBD2_FLAG_LAST_TAG)
562 break;
563 }
564
565 brelse(bh);
566 continue;
567
568 case JBD2_COMMIT_BLOCK:
569 /* How to differentiate between interrupted commit
570 * and journal corruption ?
571 *
572 * {nth transaction}
573 * Checksum Verification Failed
574 * |
575 * ____________________
576 * | |
577 * async_commit sync_commit
578 * | |
579 * | GO TO NEXT "Journal Corruption"
580 * | TRANSACTION
581 * |
582 * {(n+1)th transanction}
583 * |
584 * _______|______________
585 * | |
586 * Commit block found Commit block not found
587 * | |
588 * "Journal Corruption" |
589 * _____________|_________
590 * | |
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".)
598 */
599
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
603 * number. */
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]);
612
613 chksum_err = chksum_seen = 0;
614
615 if (info->end_transaction) {
616 journal->j_failed_commit =
617 info->end_transaction;
618 brelse(bh);
619 break;
620 }
621
622 if (crc32_sum == found_chksum &&
623 cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
624 cbh->h_chksum_size ==
625 JBD2_CRC32_CHKSUM_SIZE)
626 chksum_seen = 1;
627 else if (!(cbh->h_chksum_type == 0 &&
628 cbh->h_chksum_size == 0 &&
629 found_chksum == 0 &&
630 !chksum_seen))
631 /*
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
637 * commit blocks.
638 * Hence to avoid checksum failures, in this
639 * situation, this extra check is added.
640 */
641 chksum_err = 1;
642
643 if (chksum_err) {
644 info->end_transaction = next_commit_ID;
645
646 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
647 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
648 journal->j_failed_commit =
649 next_commit_ID;
650 brelse(bh);
651 break;
652 }
653 }
654 crc32_sum = ~0;
655 }
656 brelse(bh);
657 next_commit_ID++;
658 continue;
659
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) {
664 brelse(bh);
665 continue;
666 }
667
668 err = scan_revoke_records(journal, bh,
669 next_commit_ID, info);
670 brelse(bh);
671 if (err)
672 goto failed;
673 continue;
674
675 default:
676 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
677 blocktype);
678 brelse(bh);
679 goto done;
680 }
681 }
682
683 done:
684 /*
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.
689 */
690
691 if (pass == PASS_SCAN) {
692 if (!info->end_transaction)
693 info->end_transaction = next_commit_ID;
694 } else {
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);
701 if (!success)
702 success = -EIO;
703 }
704 }
705
706 return success;
707
708 failed:
709 return err;
710 }
711
712
713 /* Scan a revoke record, marking all blocks mentioned as revoked. */
714
715 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
716 tid_t sequence, struct recovery_info *info)
717 {
718 jbd2_journal_revoke_header_t *header;
719 int offset, max;
720 int record_len = 4;
721
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);
725
726 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
727 record_len = 8;
728
729 while (offset + record_len <= max) {
730 unsigned long long blocknr;
731 int err;
732
733 if (record_len == 4)
734 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
735 else
736 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
737 offset += record_len;
738 err = jbd2_journal_set_revoke(journal, blocknr, sequence);
739 if (err)
740 return err;
741 ++info->nr_revokes;
742 }
743 return 0;
744 }
linux-next