1 /*
2 * linux/fs/jbd/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/jbd.h>
22 #include <linux/errno.h>
23 #include <linux/blkdev.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 */
journal_brelse_array(struct buffer_head * b[],int n)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
do_readahead(journal_t * journal,unsigned int start)69 static int do_readahead(journal_t *journal, unsigned int start)
70 {
71 int err;
72 unsigned int max, nbufs, next;
73 unsigned int 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 = 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
jread(struct buffer_head ** bhp,journal_t * journal,unsigned int offset)131 static int jread(struct buffer_head **bhp, journal_t *journal,
132 unsigned int offset)
133 {
134 int err;
135 unsigned int 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 = 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
count_tags(struct buffer_head * bh,int size)181 static int count_tags(struct buffer_head *bh, int size)
182 {
183 char * tagp;
184 journal_block_tag_t * tag;
185 int nr = 0;
186
187 tagp = &bh->b_data[sizeof(journal_header_t)];
188
189 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
190 tag = (journal_block_tag_t *) tagp;
191
192 nr++;
193 tagp += sizeof(journal_block_tag_t);
194 if (!(tag->t_flags & cpu_to_be32(JFS_FLAG_SAME_UUID)))
195 tagp += 16;
196
197 if (tag->t_flags & cpu_to_be32(JFS_FLAG_LAST_TAG))
198 break;
199 }
200
201 return nr;
202 }
203
204
205 /* Make sure we wrap around the log correctly! */
206 #define wrap(journal, var) \
207 do { \
208 if (var >= (journal)->j_last) \
209 var -= ((journal)->j_last - (journal)->j_first); \
210 } while (0)
211
212 /**
213 * journal_recover - recovers a on-disk journal
214 * @journal: the journal to recover
215 *
216 * The primary function for recovering the log contents when mounting a
217 * journaled device.
218 *
219 * Recovery is done in three passes. In the first pass, we look for the
220 * end of the log. In the second, we assemble the list of revoke
221 * blocks. In the third and final pass, we replay any un-revoked blocks
222 * in the log.
223 */
journal_recover(journal_t * journal)224 int journal_recover(journal_t *journal)
225 {
226 int err, err2;
227 journal_superblock_t * sb;
228
229 struct recovery_info info;
230
231 memset(&info, 0, sizeof(info));
232 sb = journal->j_superblock;
233
234 /*
235 * The journal superblock's s_start field (the current log head)
236 * is always zero if, and only if, the journal was cleanly
237 * unmounted.
238 */
239
240 if (!sb->s_start) {
241 jbd_debug(1, "No recovery required, last transaction %d\n",
242 be32_to_cpu(sb->s_sequence));
243 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
244 return 0;
245 }
246
247 err = do_one_pass(journal, &info, PASS_SCAN);
248 if (!err)
249 err = do_one_pass(journal, &info, PASS_REVOKE);
250 if (!err)
251 err = do_one_pass(journal, &info, PASS_REPLAY);
252
253 jbd_debug(1, "JBD: recovery, exit status %d, "
254 "recovered transactions %u to %u\n",
255 err, info.start_transaction, info.end_transaction);
256 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
257 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
258
259 /* Restart the log at the next transaction ID, thus invalidating
260 * any existing commit records in the log. */
261 journal->j_transaction_sequence = ++info.end_transaction;
262
263 journal_clear_revoke(journal);
264 err2 = sync_blockdev(journal->j_fs_dev);
265 if (!err)
266 err = err2;
267 /* Flush disk caches to get replayed data on the permanent storage */
268 if (journal->j_flags & JFS_BARRIER)
269 blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
270
271 return err;
272 }
273
274 /**
275 * journal_skip_recovery - Start journal and wipe exiting records
276 * @journal: journal to startup
277 *
278 * Locate any valid recovery information from the journal and set up the
279 * journal structures in memory to ignore it (presumably because the
280 * caller has evidence that it is out of date).
281 * This function does'nt appear to be exorted..
282 *
283 * We perform one pass over the journal to allow us to tell the user how
284 * much recovery information is being erased, and to let us initialise
285 * the journal transaction sequence numbers to the next unused ID.
286 */
journal_skip_recovery(journal_t * journal)287 int journal_skip_recovery(journal_t *journal)
288 {
289 int err;
290 struct recovery_info info;
291
292 memset (&info, 0, sizeof(info));
293
294 err = do_one_pass(journal, &info, PASS_SCAN);
295
296 if (err) {
297 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
298 ++journal->j_transaction_sequence;
299 } else {
300 #ifdef CONFIG_JBD_DEBUG
301 int dropped = info.end_transaction -
302 be32_to_cpu(journal->j_superblock->s_sequence);
303 jbd_debug(1,
304 "JBD: ignoring %d transaction%s from the journal.\n",
305 dropped, (dropped == 1) ? "" : "s");
306 #endif
307 journal->j_transaction_sequence = ++info.end_transaction;
308 }
309
310 journal->j_tail = 0;
311 return err;
312 }
313
do_one_pass(journal_t * journal,struct recovery_info * info,enum passtype pass)314 static int do_one_pass(journal_t *journal,
315 struct recovery_info *info, enum passtype pass)
316 {
317 unsigned int first_commit_ID, next_commit_ID;
318 unsigned int next_log_block;
319 int err, success = 0;
320 journal_superblock_t * sb;
321 journal_header_t * tmp;
322 struct buffer_head * bh;
323 unsigned int sequence;
324 int blocktype;
325
326 /*
327 * First thing is to establish what we expect to find in the log
328 * (in terms of transaction IDs), and where (in terms of log
329 * block offsets): query the superblock.
330 */
331
332 sb = journal->j_superblock;
333 next_commit_ID = be32_to_cpu(sb->s_sequence);
334 next_log_block = be32_to_cpu(sb->s_start);
335
336 first_commit_ID = next_commit_ID;
337 if (pass == PASS_SCAN)
338 info->start_transaction = first_commit_ID;
339
340 jbd_debug(1, "Starting recovery pass %d\n", pass);
341
342 /*
343 * Now we walk through the log, transaction by transaction,
344 * making sure that each transaction has a commit block in the
345 * expected place. Each complete transaction gets replayed back
346 * into the main filesystem.
347 */
348
349 while (1) {
350 int flags;
351 char * tagp;
352 journal_block_tag_t * tag;
353 struct buffer_head * obh;
354 struct buffer_head * nbh;
355
356 cond_resched();
357
358 /* If we already know where to stop the log traversal,
359 * check right now that we haven't gone past the end of
360 * the log. */
361
362 if (pass != PASS_SCAN)
363 if (tid_geq(next_commit_ID, info->end_transaction))
364 break;
365
366 jbd_debug(2, "Scanning for sequence ID %u at %u/%u\n",
367 next_commit_ID, next_log_block, journal->j_last);
368
369 /* Skip over each chunk of the transaction looking
370 * either the next descriptor block or the final commit
371 * record. */
372
373 jbd_debug(3, "JBD: checking block %u\n", next_log_block);
374 err = jread(&bh, journal, next_log_block);
375 if (err)
376 goto failed;
377
378 next_log_block++;
379 wrap(journal, next_log_block);
380
381 /* What kind of buffer is it?
382 *
383 * If it is a descriptor block, check that it has the
384 * expected sequence number. Otherwise, we're all done
385 * here. */
386
387 tmp = (journal_header_t *)bh->b_data;
388
389 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
390 brelse(bh);
391 break;
392 }
393
394 blocktype = be32_to_cpu(tmp->h_blocktype);
395 sequence = be32_to_cpu(tmp->h_sequence);
396 jbd_debug(3, "Found magic %d, sequence %d\n",
397 blocktype, sequence);
398
399 if (sequence != next_commit_ID) {
400 brelse(bh);
401 break;
402 }
403
404 /* OK, we have a valid descriptor block which matches
405 * all of the sequence number checks. What are we going
406 * to do with it? That depends on the pass... */
407
408 switch(blocktype) {
409 case JFS_DESCRIPTOR_BLOCK:
410 /* If it is a valid descriptor block, replay it
411 * in pass REPLAY; otherwise, just skip over the
412 * blocks it describes. */
413 if (pass != PASS_REPLAY) {
414 next_log_block +=
415 count_tags(bh, journal->j_blocksize);
416 wrap(journal, next_log_block);
417 brelse(bh);
418 continue;
419 }
420
421 /* A descriptor block: we can now write all of
422 * the data blocks. Yay, useful work is finally
423 * getting done here! */
424
425 tagp = &bh->b_data[sizeof(journal_header_t)];
426 while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
427 <= journal->j_blocksize) {
428 unsigned int io_block;
429
430 tag = (journal_block_tag_t *) tagp;
431 flags = be32_to_cpu(tag->t_flags);
432
433 io_block = next_log_block++;
434 wrap(journal, next_log_block);
435 err = jread(&obh, journal, io_block);
436 if (err) {
437 /* Recover what we can, but
438 * report failure at the end. */
439 success = err;
440 printk (KERN_ERR
441 "JBD: IO error %d recovering "
442 "block %u in log\n",
443 err, io_block);
444 } else {
445 unsigned int blocknr;
446
447 J_ASSERT(obh != NULL);
448 blocknr = be32_to_cpu(tag->t_blocknr);
449
450 /* If the block has been
451 * revoked, then we're all done
452 * here. */
453 if (journal_test_revoke
454 (journal, blocknr,
455 next_commit_ID)) {
456 brelse(obh);
457 ++info->nr_revoke_hits;
458 goto skip_write;
459 }
460
461 /* Find a buffer for the new
462 * data being restored */
463 nbh = __getblk(journal->j_fs_dev,
464 blocknr,
465 journal->j_blocksize);
466 if (nbh == NULL) {
467 printk(KERN_ERR
468 "JBD: Out of memory "
469 "during recovery.\n");
470 err = -ENOMEM;
471 brelse(bh);
472 brelse(obh);
473 goto failed;
474 }
475
476 lock_buffer(nbh);
477 memcpy(nbh->b_data, obh->b_data,
478 journal->j_blocksize);
479 if (flags & JFS_FLAG_ESCAPE) {
480 *((__be32 *)nbh->b_data) =
481 cpu_to_be32(JFS_MAGIC_NUMBER);
482 }
483
484 BUFFER_TRACE(nbh, "marking dirty");
485 set_buffer_uptodate(nbh);
486 mark_buffer_dirty(nbh);
487 BUFFER_TRACE(nbh, "marking uptodate");
488 ++info->nr_replays;
489 /* ll_rw_block(WRITE, 1, &nbh); */
490 unlock_buffer(nbh);
491 brelse(obh);
492 brelse(nbh);
493 }
494
495 skip_write:
496 tagp += sizeof(journal_block_tag_t);
497 if (!(flags & JFS_FLAG_SAME_UUID))
498 tagp += 16;
499
500 if (flags & JFS_FLAG_LAST_TAG)
501 break;
502 }
503
504 brelse(bh);
505 continue;
506
507 case JFS_COMMIT_BLOCK:
508 /* Found an expected commit block: not much to
509 * do other than move on to the next sequence
510 * number. */
511 brelse(bh);
512 next_commit_ID++;
513 continue;
514
515 case JFS_REVOKE_BLOCK:
516 /* If we aren't in the REVOKE pass, then we can
517 * just skip over this block. */
518 if (pass != PASS_REVOKE) {
519 brelse(bh);
520 continue;
521 }
522
523 err = scan_revoke_records(journal, bh,
524 next_commit_ID, info);
525 brelse(bh);
526 if (err)
527 goto failed;
528 continue;
529
530 default:
531 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
532 blocktype);
533 brelse(bh);
534 goto done;
535 }
536 }
537
538 done:
539 /*
540 * We broke out of the log scan loop: either we came to the
541 * known end of the log or we found an unexpected block in the
542 * log. If the latter happened, then we know that the "current"
543 * transaction marks the end of the valid log.
544 */
545
546 if (pass == PASS_SCAN)
547 info->end_transaction = next_commit_ID;
548 else {
549 /* It's really bad news if different passes end up at
550 * different places (but possible due to IO errors). */
551 if (info->end_transaction != next_commit_ID) {
552 printk (KERN_ERR "JBD: recovery pass %d ended at "
553 "transaction %u, expected %u\n",
554 pass, next_commit_ID, info->end_transaction);
555 if (!success)
556 success = -EIO;
557 }
558 }
559
560 return success;
561
562 failed:
563 return err;
564 }
565
566
567 /* Scan a revoke record, marking all blocks mentioned as revoked. */
568
scan_revoke_records(journal_t * journal,struct buffer_head * bh,tid_t sequence,struct recovery_info * info)569 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
570 tid_t sequence, struct recovery_info *info)
571 {
572 journal_revoke_header_t *header;
573 int offset, max;
574
575 header = (journal_revoke_header_t *) bh->b_data;
576 offset = sizeof(journal_revoke_header_t);
577 max = be32_to_cpu(header->r_count);
578
579 while (offset < max) {
580 unsigned int blocknr;
581 int err;
582
583 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
584 offset += 4;
585 err = journal_set_revoke(journal, blocknr, sequence);
586 if (err)
587 return err;
588 ++info->nr_revokes;
589 }
590 return 0;
591 }
592