1 /*
2 * linux/fs/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/slab.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 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 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 = 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 long blocknr;
136 struct buffer_head *bh;
137
138 *bhp = NULL;
139
140 J_ASSERT (offset < journal->j_maxlen);
141
142 err = journal_bmap(journal, offset, &blocknr);
143
144 if (err) {
145 printk (KERN_ERR "JBD: bad block at offset %u\n",
146 offset);
147 return err;
148 }
149
150 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
151 if (!bh)
152 return -ENOMEM;
153
154 if (!buffer_uptodate(bh)) {
155 /* If this is a brand new buffer, start readahead.
156 Otherwise, we assume we are already reading it. */
157 if (!buffer_req(bh))
158 do_readahead(journal, offset);
159 wait_on_buffer(bh);
160 }
161
162 if (!buffer_uptodate(bh)) {
163 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
164 offset);
165 brelse(bh);
166 return -EIO;
167 }
168
169 *bhp = bh;
170 return 0;
171 }
172
173
174 /*
175 * Count the number of in-use tags in a journal descriptor block.
176 */
177
count_tags(struct buffer_head * bh,int size)178 static int count_tags(struct buffer_head *bh, int size)
179 {
180 char * tagp;
181 journal_block_tag_t * tag;
182 int nr = 0;
183
184 tagp = &bh->b_data[sizeof(journal_header_t)];
185
186 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
187 tag = (journal_block_tag_t *) tagp;
188
189 nr++;
190 tagp += sizeof(journal_block_tag_t);
191 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
192 tagp += 16;
193
194 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
195 break;
196 }
197
198 return nr;
199 }
200
201
202 /* Make sure we wrap around the log correctly! */
203 #define wrap(journal, var) \
204 do { \
205 if (var >= (journal)->j_last) \
206 var -= ((journal)->j_last - (journal)->j_first); \
207 } while (0)
208
209 /**
210 * int journal_recover(journal_t *journal) - recovers a on-disk journal
211 * @journal: the journal to recover
212 *
213 * The primary function for recovering the log contents when mounting a
214 * journaled device.
215 *
216 * Recovery is done in three passes. In the first pass, we look for the
217 * end of the log. In the second, we assemble the list of revoke
218 * blocks. In the third and final pass, we replay any un-revoked blocks
219 * in the log.
220 */
journal_recover(journal_t * journal)221 int journal_recover(journal_t *journal)
222 {
223 int err;
224 journal_superblock_t * sb;
225
226 struct recovery_info info;
227
228 memset(&info, 0, sizeof(info));
229 sb = journal->j_superblock;
230
231 /*
232 * The journal superblock's s_start field (the current log head)
233 * is always zero if, and only if, the journal was cleanly
234 * unmounted.
235 */
236
237 if (!sb->s_start) {
238 jbd_debug(1, "No recovery required, last transaction %d\n",
239 (int)ntohl(sb->s_sequence));
240 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
241 return 0;
242 }
243
244 err = do_one_pass(journal, &info, PASS_SCAN);
245 if (!err)
246 err = do_one_pass(journal, &info, PASS_REVOKE);
247 if (!err)
248 err = do_one_pass(journal, &info, PASS_REPLAY);
249
250 jbd_debug(0, "JBD: recovery, exit status %d, "
251 "recovered transactions %u to %u\n",
252 err, info.start_transaction, info.end_transaction);
253 jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
254 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
255
256 /* Restart the log at the next transaction ID, thus invalidating
257 * any existing commit records in the log. */
258 journal->j_transaction_sequence = ++info.end_transaction;
259
260 journal_clear_revoke(journal);
261 sync_blockdev(journal->j_fs_dev);
262 return err;
263 }
264
265 /**
266 * int journal_skip_recovery() - Start journal and wipe exiting records
267 * @journal: journal to startup
268 *
269 * Locate any valid recovery information from the journal and set up the
270 * journal structures in memory to ignore it (presumably because the
271 * caller has evidence that it is out of date).
272 * This function does'nt appear to be exorted..
273 *
274 * We perform one pass over the journal to allow us to tell the user how
275 * much recovery information is being erased, and to let us initialise
276 * the journal transaction sequence numbers to the next unused ID.
277 */
journal_skip_recovery(journal_t * journal)278 int journal_skip_recovery(journal_t *journal)
279 {
280 int err;
281 journal_superblock_t * sb;
282
283 struct recovery_info info;
284
285 memset (&info, 0, sizeof(info));
286 sb = journal->j_superblock;
287
288 err = do_one_pass(journal, &info, PASS_SCAN);
289
290 if (err) {
291 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
292 ++journal->j_transaction_sequence;
293 } else {
294 #ifdef CONFIG_JBD_DEBUG
295 int dropped = info.end_transaction - ntohl(sb->s_sequence);
296 #endif
297 jbd_debug(0,
298 "JBD: ignoring %d transaction%s from the journal.\n",
299 dropped, (dropped == 1) ? "" : "s");
300 journal->j_transaction_sequence = ++info.end_transaction;
301 }
302
303 journal->j_tail = 0;
304 return err;
305 }
306
do_one_pass(journal_t * journal,struct recovery_info * info,enum passtype pass)307 static int do_one_pass(journal_t *journal,
308 struct recovery_info *info, enum passtype pass)
309 {
310 unsigned int first_commit_ID, next_commit_ID;
311 unsigned long next_log_block;
312 int err, success = 0;
313 journal_superblock_t * sb;
314 journal_header_t * tmp;
315 struct buffer_head * bh;
316 unsigned int sequence;
317 int blocktype;
318
319 /* Precompute the maximum metadata descriptors in a descriptor block */
320 int MAX_BLOCKS_PER_DESC;
321 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
322 / sizeof(journal_block_tag_t));
323
324 /*
325 * First thing is to establish what we expect to find in the log
326 * (in terms of transaction IDs), and where (in terms of log
327 * block offsets): query the superblock.
328 */
329
330 sb = journal->j_superblock;
331 next_commit_ID = ntohl(sb->s_sequence);
332 next_log_block = ntohl(sb->s_start);
333
334 first_commit_ID = next_commit_ID;
335 if (pass == PASS_SCAN)
336 info->start_transaction = first_commit_ID;
337
338 jbd_debug(1, "Starting recovery pass %d\n", pass);
339
340 /*
341 * Now we walk through the log, transaction by transaction,
342 * making sure that each transaction has a commit block in the
343 * expected place. Each complete transaction gets replayed back
344 * into the main filesystem.
345 */
346
347 while (1) {
348 int flags;
349 char * tagp;
350 journal_block_tag_t * tag;
351 struct buffer_head * obh;
352 struct buffer_head * nbh;
353
354 /* If we already know where to stop the log traversal,
355 * check right now that we haven't gone past the end of
356 * the log. */
357
358 if (pass != PASS_SCAN)
359 if (tid_geq(next_commit_ID, info->end_transaction))
360 break;
361
362 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
363 next_commit_ID, next_log_block, journal->j_last);
364
365 /* Skip over each chunk of the transaction looking
366 * either the next descriptor block or the final commit
367 * record. */
368
369 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
370 err = jread(&bh, journal, next_log_block);
371 if (err)
372 goto failed;
373
374 next_log_block++;
375 wrap(journal, next_log_block);
376
377 /* What kind of buffer is it?
378 *
379 * If it is a descriptor block, check that it has the
380 * expected sequence number. Otherwise, we're all done
381 * here. */
382
383 tmp = (journal_header_t *)bh->b_data;
384
385 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
386 brelse(bh);
387 break;
388 }
389
390 blocktype = ntohl(tmp->h_blocktype);
391 sequence = ntohl(tmp->h_sequence);
392 jbd_debug(3, "Found magic %d, sequence %d\n",
393 blocktype, sequence);
394
395 if (sequence != next_commit_ID) {
396 brelse(bh);
397 break;
398 }
399
400 /* OK, we have a valid descriptor block which matches
401 * all of the sequence number checks. What are we going
402 * to do with it? That depends on the pass... */
403
404 switch(blocktype) {
405 case JFS_DESCRIPTOR_BLOCK:
406 /* If it is a valid descriptor block, replay it
407 * in pass REPLAY; otherwise, just skip over the
408 * blocks it describes. */
409 if (pass != PASS_REPLAY) {
410 next_log_block +=
411 count_tags(bh, journal->j_blocksize);
412 wrap(journal, next_log_block);
413 brelse(bh);
414 continue;
415 }
416
417 /* A descriptor block: we can now write all of
418 * the data blocks. Yay, useful work is finally
419 * getting done here! */
420
421 tagp = &bh->b_data[sizeof(journal_header_t)];
422 while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
423 <= journal->j_blocksize) {
424 unsigned long io_block;
425
426 tag = (journal_block_tag_t *) tagp;
427 flags = ntohl(tag->t_flags);
428
429 io_block = next_log_block++;
430 wrap(journal, next_log_block);
431 err = jread(&obh, journal, io_block);
432 if (err) {
433 /* Recover what we can, but
434 * report failure at the end. */
435 success = err;
436 printk (KERN_ERR
437 "JBD: IO error %d recovering "
438 "block %lu in log\n",
439 err, io_block);
440 } else {
441 unsigned long blocknr;
442
443 J_ASSERT(obh != NULL);
444 blocknr = ntohl(tag->t_blocknr);
445
446 /* If the block has been
447 * revoked, then we're all done
448 * here. */
449 if (journal_test_revoke
450 (journal, blocknr,
451 next_commit_ID)) {
452 brelse(obh);
453 ++info->nr_revoke_hits;
454 goto skip_write;
455 }
456
457 /* Find a buffer for the new
458 * data being restored */
459 nbh = __getblk(journal->j_fs_dev,
460 blocknr,
461 journal->j_blocksize);
462 if (nbh == NULL) {
463 printk(KERN_ERR
464 "JBD: Out of memory "
465 "during recovery.\n");
466 err = -ENOMEM;
467 brelse(bh);
468 brelse(obh);
469 goto failed;
470 }
471
472 lock_buffer(nbh);
473 memcpy(nbh->b_data, obh->b_data,
474 journal->j_blocksize);
475 if (flags & JFS_FLAG_ESCAPE) {
476 *((unsigned int *)bh->b_data) =
477 htonl(JFS_MAGIC_NUMBER);
478 }
479
480 BUFFER_TRACE(nbh, "marking dirty");
481 set_buffer_uptodate(nbh);
482 mark_buffer_dirty(nbh);
483 BUFFER_TRACE(nbh, "marking uptodate");
484 ++info->nr_replays;
485 /* ll_rw_block(WRITE, 1, &nbh); */
486 unlock_buffer(nbh);
487 brelse(obh);
488 brelse(nbh);
489 }
490
491 skip_write:
492 tagp += sizeof(journal_block_tag_t);
493 if (!(flags & JFS_FLAG_SAME_UUID))
494 tagp += 16;
495
496 if (flags & JFS_FLAG_LAST_TAG)
497 break;
498 }
499
500 brelse(bh);
501 continue;
502
503 case JFS_COMMIT_BLOCK:
504 /* Found an expected commit block: not much to
505 * do other than move on to the next sequence
506 * number. */
507 brelse(bh);
508 next_commit_ID++;
509 continue;
510
511 case JFS_REVOKE_BLOCK:
512 /* If we aren't in the REVOKE pass, then we can
513 * just skip over this block. */
514 if (pass != PASS_REVOKE) {
515 brelse(bh);
516 continue;
517 }
518
519 err = scan_revoke_records(journal, bh,
520 next_commit_ID, info);
521 brelse(bh);
522 if (err)
523 goto failed;
524 continue;
525
526 default:
527 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
528 blocktype);
529 brelse(bh);
530 goto done;
531 }
532 }
533
534 done:
535 /*
536 * We broke out of the log scan loop: either we came to the
537 * known end of the log or we found an unexpected block in the
538 * log. If the latter happened, then we know that the "current"
539 * transaction marks the end of the valid log.
540 */
541
542 if (pass == PASS_SCAN)
543 info->end_transaction = next_commit_ID;
544 else {
545 /* It's really bad news if different passes end up at
546 * different places (but possible due to IO errors). */
547 if (info->end_transaction != next_commit_ID) {
548 printk (KERN_ERR "JBD: recovery pass %d ended at "
549 "transaction %u, expected %u\n",
550 pass, next_commit_ID, info->end_transaction);
551 if (!success)
552 success = -EIO;
553 }
554 }
555
556 return success;
557
558 failed:
559 return err;
560 }
561
562
563 /* Scan a revoke record, marking all blocks mentioned as revoked. */
564
scan_revoke_records(journal_t * journal,struct buffer_head * bh,tid_t sequence,struct recovery_info * info)565 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
566 tid_t sequence, struct recovery_info *info)
567 {
568 journal_revoke_header_t *header;
569 int offset, max;
570
571 header = (journal_revoke_header_t *) bh->b_data;
572 offset = sizeof(journal_revoke_header_t);
573 max = ntohl(header->r_count);
574
575 while (offset < max) {
576 unsigned long blocknr;
577 int err;
578
579 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
580 offset += 4;
581 err = journal_set_revoke(journal, blocknr, sequence);
582 if (err)
583 return err;
584 ++info->nr_revokes;
585 }
586 return 0;
587 }
588