blob: 54c9bc9e1b17e01487cca2eb81d5bfdf7a649585 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Kleine-König58862692007-05-09 07:51:49 +02002 * linux/fs/jbd/recovery.c
Mingming Caoae6ddcc2006-09-27 01:49:27 -07003 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07004 * 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;
Mingming Caoae6ddcc2006-09-27 01:49:27 -070013 * part of the ext2fs journaling system.
Linus Torvalds1da177e2005-04-16 15:20:36 -070014 */
15
16#ifndef __KERNEL__
17#include "jfs_user.h"
18#else
19#include <linux/time.h>
20#include <linux/fs.h>
21#include <linux/jbd.h>
22#include <linux/errno.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#endif
24
25/*
26 * Maintain information about the progress of the recovery job, so that
Mingming Caoae6ddcc2006-09-27 01:49:27 -070027 * the different passes can carry information between them.
Linus Torvalds1da177e2005-04-16 15:20:36 -070028 */
Mingming Caoae6ddcc2006-09-27 01:49:27 -070029struct recovery_info
Linus Torvalds1da177e2005-04-16 15:20:36 -070030{
31 tid_t start_transaction;
32 tid_t end_transaction;
33
34 int nr_replays;
35 int nr_revokes;
36 int nr_revoke_hits;
37};
38
39enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
40static int do_one_pass(journal_t *journal,
41 struct recovery_info *info, enum passtype pass);
42static int scan_revoke_records(journal_t *, struct buffer_head *,
43 tid_t, struct recovery_info *);
44
45#ifdef __KERNEL__
46
47/* Release readahead buffers after use */
Dave Kleikampf71b2f12006-09-29 01:58:39 -070048static void journal_brelse_array(struct buffer_head *b[], int n)
Linus Torvalds1da177e2005-04-16 15:20:36 -070049{
50 while (--n >= 0)
51 brelse (b[n]);
52}
53
54
55/*
56 * When reading from the journal, we are going through the block device
57 * layer directly and so there is no readahead being done for us. We
58 * need to implement any readahead ourselves if we want it to happen at
59 * all. Recovery is basically one long sequential read, so make sure we
60 * do the IO in reasonably large chunks.
61 *
62 * This is not so critical that we need to be enormously clever about
63 * the readahead size, though. 128K is a purely arbitrary, good-enough
64 * fixed value.
65 */
66
67#define MAXBUF 8
68static int do_readahead(journal_t *journal, unsigned int start)
69{
70 int err;
71 unsigned int max, nbufs, next;
Jan Kara9c28cbc2009-08-03 19:21:00 +020072 unsigned int blocknr;
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 struct buffer_head *bh;
74
75 struct buffer_head * bufs[MAXBUF];
76
77 /* Do up to 128K of readahead */
78 max = start + (128 * 1024 / journal->j_blocksize);
79 if (max > journal->j_maxlen)
80 max = journal->j_maxlen;
81
82 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
83 * a time to the block device IO layer. */
84
85 nbufs = 0;
86
87 for (next = start; next < max; next++) {
88 err = journal_bmap(journal, next, &blocknr);
89
90 if (err) {
91 printk (KERN_ERR "JBD: bad block at offset %u\n",
92 next);
93 goto failed;
94 }
95
96 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
97 if (!bh) {
98 err = -ENOMEM;
99 goto failed;
100 }
101
102 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
103 bufs[nbufs++] = bh;
104 if (nbufs == MAXBUF) {
105 ll_rw_block(READ, nbufs, bufs);
106 journal_brelse_array(bufs, nbufs);
107 nbufs = 0;
108 }
109 } else
110 brelse(bh);
111 }
112
113 if (nbufs)
114 ll_rw_block(READ, nbufs, bufs);
115 err = 0;
116
117failed:
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700118 if (nbufs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 journal_brelse_array(bufs, nbufs);
120 return err;
121}
122
123#endif /* __KERNEL__ */
124
125
126/*
127 * Read a block from the journal
128 */
129
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700130static int jread(struct buffer_head **bhp, journal_t *journal,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 unsigned int offset)
132{
133 int err;
Jan Kara9c28cbc2009-08-03 19:21:00 +0200134 unsigned int blocknr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 struct buffer_head *bh;
136
137 *bhp = NULL;
138
139 if (offset >= journal->j_maxlen) {
140 printk(KERN_ERR "JBD: corrupted journal superblock\n");
141 return -EIO;
142 }
143
144 err = journal_bmap(journal, offset, &blocknr);
145
146 if (err) {
147 printk (KERN_ERR "JBD: bad block at offset %u\n",
148 offset);
149 return err;
150 }
151
152 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
153 if (!bh)
154 return -ENOMEM;
155
156 if (!buffer_uptodate(bh)) {
157 /* If this is a brand new buffer, start readahead.
158 Otherwise, we assume we are already reading it. */
159 if (!buffer_req(bh))
160 do_readahead(journal, offset);
161 wait_on_buffer(bh);
162 }
163
164 if (!buffer_uptodate(bh)) {
165 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
166 offset);
167 brelse(bh);
168 return -EIO;
169 }
170
171 *bhp = bh;
172 return 0;
173}
174
175
176/*
177 * Count the number of in-use tags in a journal descriptor block.
178 */
179
180static int count_tags(struct buffer_head *bh, int size)
181{
182 char * tagp;
183 journal_block_tag_t * tag;
184 int nr = 0;
185
186 tagp = &bh->b_data[sizeof(journal_header_t)];
187
188 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
189 tag = (journal_block_tag_t *) tagp;
190
191 nr++;
192 tagp += sizeof(journal_block_tag_t);
193 if (!(tag->t_flags & cpu_to_be32(JFS_FLAG_SAME_UUID)))
194 tagp += 16;
195
196 if (tag->t_flags & cpu_to_be32(JFS_FLAG_LAST_TAG))
197 break;
198 }
199
200 return nr;
201}
202
203
204/* Make sure we wrap around the log correctly! */
205#define wrap(journal, var) \
206do { \
207 if (var >= (journal)->j_last) \
208 var -= ((journal)->j_last - (journal)->j_first); \
209} while (0)
210
211/**
Randy Dunlap6c8bec62005-11-07 01:01:04 -0800212 * journal_recover - recovers a on-disk journal
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 * @journal: the journal to recover
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700214 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215 * The primary function for recovering the log contents when mounting a
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700216 * journaled device.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 *
218 * Recovery is done in three passes. In the first pass, we look for the
219 * end of the log. In the second, we assemble the list of revoke
220 * blocks. In the third and final pass, we replay any un-revoked blocks
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700221 * in the log.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 */
223int journal_recover(journal_t *journal)
224{
Hidehiro Kawai4afe9782008-10-22 14:15:00 -0700225 int err, err2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 journal_superblock_t * sb;
227
228 struct recovery_info info;
229
230 memset(&info, 0, sizeof(info));
231 sb = journal->j_superblock;
232
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700233 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 * The journal superblock's s_start field (the current log head)
235 * is always zero if, and only if, the journal was cleanly
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700236 * unmounted.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 */
238
239 if (!sb->s_start) {
240 jbd_debug(1, "No recovery required, last transaction %d\n",
241 be32_to_cpu(sb->s_sequence));
242 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
243 return 0;
244 }
245
246 err = do_one_pass(journal, &info, PASS_SCAN);
247 if (!err)
248 err = do_one_pass(journal, &info, PASS_REVOKE);
249 if (!err)
250 err = do_one_pass(journal, &info, PASS_REPLAY);
251
Jose R. Santos9ad163a2007-10-18 23:39:23 -0700252 jbd_debug(1, "JBD: recovery, exit status %d, "
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 "recovered transactions %u to %u\n",
254 err, info.start_transaction, info.end_transaction);
Jose R. Santos9ad163a2007-10-18 23:39:23 -0700255 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
257
258 /* Restart the log at the next transaction ID, thus invalidating
259 * any existing commit records in the log. */
260 journal->j_transaction_sequence = ++info.end_transaction;
261
262 journal_clear_revoke(journal);
Hidehiro Kawai4afe9782008-10-22 14:15:00 -0700263 err2 = sync_blockdev(journal->j_fs_dev);
264 if (!err)
265 err = err2;
266
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 return err;
268}
269
270/**
Randy Dunlap6c8bec62005-11-07 01:01:04 -0800271 * journal_skip_recovery - Start journal and wipe exiting records
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 * @journal: journal to startup
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700273 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 * Locate any valid recovery information from the journal and set up the
275 * journal structures in memory to ignore it (presumably because the
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700276 * caller has evidence that it is out of date).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 * This function does'nt appear to be exorted..
278 *
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
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700281 * the journal transaction sequence numbers to the next unused ID.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 */
283int journal_skip_recovery(journal_t *journal)
284{
285 int err;
286 journal_superblock_t * sb;
287
288 struct recovery_info info;
289
290 memset (&info, 0, sizeof(info));
291 sb = journal->j_superblock;
292
293 err = do_one_pass(journal, &info, PASS_SCAN);
294
295 if (err) {
296 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
297 ++journal->j_transaction_sequence;
298 } else {
299#ifdef CONFIG_JBD_DEBUG
300 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
301#endif
Jose R. Santos9ad163a2007-10-18 23:39:23 -0700302 jbd_debug(1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 "JBD: ignoring %d transaction%s from the journal.\n",
304 dropped, (dropped == 1) ? "" : "s");
305 journal->j_transaction_sequence = ++info.end_transaction;
306 }
307
308 journal->j_tail = 0;
309 return err;
310}
311
312static int do_one_pass(journal_t *journal,
313 struct recovery_info *info, enum passtype pass)
314{
315 unsigned int first_commit_ID, next_commit_ID;
Jan Kara9c28cbc2009-08-03 19:21:00 +0200316 unsigned int next_log_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317 int err, success = 0;
318 journal_superblock_t * sb;
Dave Kleikampe9ad5622006-09-27 01:49:35 -0700319 journal_header_t * tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320 struct buffer_head * bh;
321 unsigned int sequence;
322 int blocktype;
323
324 /* Precompute the maximum metadata descriptors in a descriptor block */
325 int MAX_BLOCKS_PER_DESC;
326 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
327 / sizeof(journal_block_tag_t));
328
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700329 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700330 * First thing is to establish what we expect to find in the log
331 * (in terms of transaction IDs), and where (in terms of log
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700332 * block offsets): query the superblock.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 */
334
335 sb = journal->j_superblock;
336 next_commit_ID = be32_to_cpu(sb->s_sequence);
337 next_log_block = be32_to_cpu(sb->s_start);
338
339 first_commit_ID = next_commit_ID;
340 if (pass == PASS_SCAN)
341 info->start_transaction = first_commit_ID;
342
343 jbd_debug(1, "Starting recovery pass %d\n", pass);
344
345 /*
346 * Now we walk through the log, transaction by transaction,
347 * making sure that each transaction has a commit block in the
348 * expected place. Each complete transaction gets replayed back
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700349 * into the main filesystem.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 */
351
352 while (1) {
353 int flags;
354 char * tagp;
355 journal_block_tag_t * tag;
356 struct buffer_head * obh;
357 struct buffer_head * nbh;
358
Andi Kleene86e1432008-02-06 01:40:11 -0800359 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360
361 /* If we already know where to stop the log traversal,
362 * check right now that we haven't gone past the end of
363 * the log. */
364
365 if (pass != PASS_SCAN)
366 if (tid_geq(next_commit_ID, info->end_transaction))
367 break;
368
Jan Kara9c28cbc2009-08-03 19:21:00 +0200369 jbd_debug(2, "Scanning for sequence ID %u at %u/%u\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 next_commit_ID, next_log_block, journal->j_last);
371
372 /* Skip over each chunk of the transaction looking
373 * either the next descriptor block or the final commit
374 * record. */
375
Jan Kara9c28cbc2009-08-03 19:21:00 +0200376 jbd_debug(3, "JBD: checking block %u\n", next_log_block);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 err = jread(&bh, journal, next_log_block);
378 if (err)
379 goto failed;
380
381 next_log_block++;
382 wrap(journal, next_log_block);
383
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700384 /* What kind of buffer is it?
385 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 * If it is a descriptor block, check that it has the
387 * expected sequence number. Otherwise, we're all done
388 * here. */
389
390 tmp = (journal_header_t *)bh->b_data;
391
392 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
393 brelse(bh);
394 break;
395 }
396
397 blocktype = be32_to_cpu(tmp->h_blocktype);
398 sequence = be32_to_cpu(tmp->h_sequence);
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700399 jbd_debug(3, "Found magic %d, sequence %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 blocktype, sequence);
401
402 if (sequence != next_commit_ID) {
403 brelse(bh);
404 break;
405 }
406
407 /* OK, we have a valid descriptor block which matches
408 * all of the sequence number checks. What are we going
409 * to do with it? That depends on the pass... */
410
411 switch(blocktype) {
412 case JFS_DESCRIPTOR_BLOCK:
413 /* If it is a valid descriptor block, replay it
414 * in pass REPLAY; otherwise, just skip over the
415 * blocks it describes. */
416 if (pass != PASS_REPLAY) {
417 next_log_block +=
418 count_tags(bh, journal->j_blocksize);
419 wrap(journal, next_log_block);
420 brelse(bh);
421 continue;
422 }
423
424 /* A descriptor block: we can now write all of
425 * the data blocks. Yay, useful work is finally
426 * getting done here! */
427
428 tagp = &bh->b_data[sizeof(journal_header_t)];
429 while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
430 <= journal->j_blocksize) {
Jan Kara9c28cbc2009-08-03 19:21:00 +0200431 unsigned int io_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432
433 tag = (journal_block_tag_t *) tagp;
434 flags = be32_to_cpu(tag->t_flags);
435
436 io_block = next_log_block++;
437 wrap(journal, next_log_block);
438 err = jread(&obh, journal, io_block);
439 if (err) {
440 /* Recover what we can, but
441 * report failure at the end. */
442 success = err;
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700443 printk (KERN_ERR
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 "JBD: IO error %d recovering "
Jan Kara9c28cbc2009-08-03 19:21:00 +0200445 "block %u in log\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 err, io_block);
447 } else {
Jan Kara9c28cbc2009-08-03 19:21:00 +0200448 unsigned int blocknr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449
450 J_ASSERT(obh != NULL);
451 blocknr = be32_to_cpu(tag->t_blocknr);
452
453 /* If the block has been
454 * revoked, then we're all done
455 * here. */
456 if (journal_test_revoke
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700457 (journal, blocknr,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 next_commit_ID)) {
459 brelse(obh);
460 ++info->nr_revoke_hits;
461 goto skip_write;
462 }
463
464 /* Find a buffer for the new
465 * data being restored */
466 nbh = __getblk(journal->j_fs_dev,
467 blocknr,
468 journal->j_blocksize);
469 if (nbh == NULL) {
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700470 printk(KERN_ERR
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 "JBD: Out of memory "
472 "during recovery.\n");
473 err = -ENOMEM;
474 brelse(bh);
475 brelse(obh);
476 goto failed;
477 }
478
479 lock_buffer(nbh);
480 memcpy(nbh->b_data, obh->b_data,
481 journal->j_blocksize);
482 if (flags & JFS_FLAG_ESCAPE) {
Duane Griffin439aeec2008-03-19 17:00:53 -0700483 *((__be32 *)nbh->b_data) =
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484 cpu_to_be32(JFS_MAGIC_NUMBER);
485 }
486
487 BUFFER_TRACE(nbh, "marking dirty");
488 set_buffer_uptodate(nbh);
489 mark_buffer_dirty(nbh);
490 BUFFER_TRACE(nbh, "marking uptodate");
491 ++info->nr_replays;
492 /* ll_rw_block(WRITE, 1, &nbh); */
493 unlock_buffer(nbh);
494 brelse(obh);
495 brelse(nbh);
496 }
497
498 skip_write:
499 tagp += sizeof(journal_block_tag_t);
500 if (!(flags & JFS_FLAG_SAME_UUID))
501 tagp += 16;
502
503 if (flags & JFS_FLAG_LAST_TAG)
504 break;
505 }
506
507 brelse(bh);
508 continue;
509
510 case JFS_COMMIT_BLOCK:
511 /* Found an expected commit block: not much to
512 * do other than move on to the next sequence
513 * number. */
514 brelse(bh);
515 next_commit_ID++;
516 continue;
517
518 case JFS_REVOKE_BLOCK:
519 /* If we aren't in the REVOKE pass, then we can
520 * just skip over this block. */
521 if (pass != PASS_REVOKE) {
522 brelse(bh);
523 continue;
524 }
525
526 err = scan_revoke_records(journal, bh,
527 next_commit_ID, info);
528 brelse(bh);
529 if (err)
530 goto failed;
531 continue;
532
533 default:
534 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
535 blocktype);
Theodore Ts'oe8f1c622006-06-25 05:47:51 -0700536 brelse(bh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537 goto done;
538 }
539 }
540
541 done:
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700542 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 * We broke out of the log scan loop: either we came to the
544 * known end of the log or we found an unexpected block in the
545 * log. If the latter happened, then we know that the "current"
546 * transaction marks the end of the valid log.
547 */
548
549 if (pass == PASS_SCAN)
550 info->end_transaction = next_commit_ID;
551 else {
552 /* It's really bad news if different passes end up at
553 * different places (but possible due to IO errors). */
554 if (info->end_transaction != next_commit_ID) {
555 printk (KERN_ERR "JBD: recovery pass %d ended at "
556 "transaction %u, expected %u\n",
557 pass, next_commit_ID, info->end_transaction);
558 if (!success)
559 success = -EIO;
560 }
561 }
562
563 return success;
564
565 failed:
566 return err;
567}
568
569
570/* Scan a revoke record, marking all blocks mentioned as revoked. */
571
Mingming Caoae6ddcc2006-09-27 01:49:27 -0700572static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 tid_t sequence, struct recovery_info *info)
574{
575 journal_revoke_header_t *header;
576 int offset, max;
577
578 header = (journal_revoke_header_t *) bh->b_data;
579 offset = sizeof(journal_revoke_header_t);
580 max = be32_to_cpu(header->r_count);
581
582 while (offset < max) {
Jan Kara9c28cbc2009-08-03 19:21:00 +0200583 unsigned int blocknr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 int err;
585
586 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
587 offset += 4;
588 err = journal_set_revoke(journal, blocknr, sequence);
589 if (err)
590 return err;
591 ++info->nr_revokes;
592 }
593 return 0;
594}