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.
19 #include <linux/sched.h>
21 #include <linux/jbd.h>
22 #include <linux/errno.h>
23 #include <linux/slab.h>
24 #include <linux/locks.h>
28 * Maintain information about the progress of the recovery job, so that
29 * the different passes can carry information between them.
33 tid_t start_transaction;
34 tid_t end_transaction;
41 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
42 static int do_one_pass(journal_t *journal,
43 struct recovery_info *info, enum passtype pass);
44 static int scan_revoke_records(journal_t *, struct buffer_head *,
45 tid_t, struct recovery_info *);
49 /* Release readahead buffers after use */
50 void journal_brelse_array(struct buffer_head *b[], int n)
58 * When reading from the journal, we are going through the block device
59 * layer directly and so there is no readahead being done for us. We
60 * need to implement any readahead ourselves if we want it to happen at
61 * all. Recovery is basically one long sequential read, so make sure we
62 * do the IO in reasonably large chunks.
64 * This is not so critical that we need to be enormously clever about
65 * the readahead size, though. 128K is a purely arbitrary, good-enough
70 static int do_readahead(journal_t *journal, unsigned int start)
73 unsigned int max, nbufs, next;
74 unsigned long blocknr;
75 struct buffer_head *bh;
77 struct buffer_head * bufs[MAXBUF];
79 /* Do up to 128K of readahead */
80 max = start + (128 * 1024 / journal->j_blocksize);
81 if (max > journal->j_maxlen)
82 max = journal->j_maxlen;
84 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
85 * a time to the block device IO layer. */
89 for (next = start; next < max; next++) {
90 err = journal_bmap(journal, next, &blocknr);
93 printk (KERN_ERR "JBD: bad block at offset %u\n",
98 bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
104 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
106 if (nbufs == MAXBUF) {
107 ll_rw_block(READ, nbufs, bufs);
108 journal_brelse_array(bufs, nbufs);
116 ll_rw_block(READ, nbufs, bufs);
121 journal_brelse_array(bufs, nbufs);
125 #endif /* __KERNEL__ */
129 * Read a block from the journal
132 static int jread(struct buffer_head **bhp, journal_t *journal,
136 unsigned long blocknr;
137 struct buffer_head *bh;
141 J_ASSERT (offset < journal->j_maxlen);
143 err = journal_bmap(journal, offset, &blocknr);
146 printk (KERN_ERR "JBD: bad block at offset %u\n",
151 bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
155 if (!buffer_uptodate(bh)) {
156 /* If this is a brand new buffer, start readahead.
157 Otherwise, we assume we are already reading it. */
159 do_readahead(journal, offset);
163 if (!buffer_uptodate(bh)) {
164 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
176 * Count the number of in-use tags in a journal descriptor block.
179 static int count_tags(struct buffer_head *bh, int size)
182 journal_block_tag_t * tag;
185 tagp = &bh->b_data[sizeof(journal_header_t)];
187 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
188 tag = (journal_block_tag_t *) tagp;
191 tagp += sizeof(journal_block_tag_t);
192 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
195 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
203 /* Make sure we wrap around the log correctly! */
204 #define wrap(journal, var) \
206 if (var >= (journal)->j_last) \
207 var -= ((journal)->j_last - (journal)->j_first); \
213 * The primary function for recovering the log contents when mounting a
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
222 int journal_recover(journal_t *journal)
225 journal_superblock_t * sb;
227 struct recovery_info info;
229 memset(&info, 0, sizeof(info));
230 sb = journal->j_superblock;
233 * The journal superblock's s_start field (the current log head)
234 * is always zero if, and only if, the journal was cleanly
239 jbd_debug(1, "No recovery required, last transaction %d\n",
240 ntohl(sb->s_sequence));
241 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
246 err = do_one_pass(journal, &info, PASS_SCAN);
248 err = do_one_pass(journal, &info, PASS_REVOKE);
250 err = do_one_pass(journal, &info, PASS_REPLAY);
252 jbd_debug(0, "JBD: recovery, exit status %d, "
253 "recovered transactions %u to %u\n",
254 err, info.start_transaction, info.end_transaction);
255 jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
256 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
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;
262 journal_clear_revoke(journal);
263 fsync_no_super(journal->j_fs_dev);
268 * journal_skip_recovery
270 * Locate any valid recovery information from the journal and set up the
271 * journal structures in memory to ignore it (presumably because the
272 * caller has evidence that it is out of date).
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.
279 int journal_skip_recovery(journal_t *journal)
282 journal_superblock_t * sb;
284 struct recovery_info info;
286 memset (&info, 0, sizeof(info));
287 sb = journal->j_superblock;
289 err = do_one_pass(journal, &info, PASS_SCAN);
292 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
293 ++journal->j_transaction_sequence;
295 #ifdef CONFIG_JBD_DEBUG
296 int dropped = info.end_transaction - ntohl(sb->s_sequence);
300 "JBD: ignoring %d transaction%s from the journal.\n",
301 dropped, (dropped == 1) ? "" : "s");
302 journal->j_transaction_sequence = ++info.end_transaction;
310 static int do_one_pass(journal_t *journal,
311 struct recovery_info *info, enum passtype pass)
314 unsigned int first_commit_ID, next_commit_ID;
315 unsigned long next_log_block;
316 int err, success = 0;
317 journal_superblock_t * sb;
318 journal_header_t * tmp;
319 struct buffer_head * bh;
320 unsigned int sequence;
323 /* Precompute the maximum metadata descriptors in a descriptor block */
324 int MAX_BLOCKS_PER_DESC;
325 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
326 / sizeof(journal_block_tag_t));
329 * First thing is to establish what we expect to find in the log
330 * (in terms of transaction IDs), and where (in terms of log
331 * block offsets): query the superblock.
334 sb = journal->j_superblock;
335 next_commit_ID = ntohl(sb->s_sequence);
336 next_log_block = ntohl(sb->s_start);
338 first_commit_ID = next_commit_ID;
339 if (pass == PASS_SCAN)
340 info->start_transaction = first_commit_ID;
342 jbd_debug(1, "Starting recovery pass %d\n", pass);
345 * Now we walk through the log, transaction by transaction,
346 * making sure that each transaction has a commit block in the
347 * expected place. Each complete transaction gets replayed back
348 * into the main filesystem.
354 journal_block_tag_t * tag;
355 struct buffer_head * obh;
356 struct buffer_head * nbh;
358 /* If we already know where to stop the log traversal,
359 * check right now that we haven't gone past the end of
362 if (pass != PASS_SCAN)
363 if (tid_geq(next_commit_ID, info->end_transaction))
366 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
367 next_commit_ID, next_log_block, journal->j_last);
369 /* Skip over each chunk of the transaction looking
370 * either the next descriptor block or the final commit
373 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
374 err = jread(&bh, journal, next_log_block);
379 wrap(journal, next_log_block);
381 /* What kind of buffer is it?
383 * If it is a descriptor block, check that it has the
384 * expected sequence number. Otherwise, we're all done
387 tmp = (journal_header_t *)bh->b_data;
389 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
394 blocktype = ntohl(tmp->h_blocktype);
395 sequence = ntohl(tmp->h_sequence);
396 jbd_debug(3, "Found magic %d, sequence %d\n",
397 blocktype, sequence);
399 if (sequence != next_commit_ID) {
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... */
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) {
415 count_tags(bh, journal->j_blocksize);
416 wrap(journal, next_log_block);
421 /* A descriptor block: we can now write all of
422 * the data blocks. Yay, useful work is finally
423 * getting done here! */
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 long io_block;
430 tag = (journal_block_tag_t *) tagp;
431 flags = ntohl(tag->t_flags);
433 io_block = next_log_block++;
434 wrap(journal, next_log_block);
435 err = jread(&obh, journal, io_block);
437 /* Recover what we can, but
438 * report failure at the end. */
441 "JBD: IO error %d recovering "
442 "block %ld in log\n",
445 unsigned long blocknr;
447 J_ASSERT(obh != NULL);
448 blocknr = ntohl(tag->t_blocknr);
450 /* If the block has been
451 * revoked, then we're all done
453 if (journal_test_revoke
457 ++info->nr_revoke_hits;
461 /* Find a buffer for the new
462 * data being restored */
463 nbh = getblk(journal->j_fs_dev, blocknr,
464 journal->j_blocksize);
467 "JBD: Out of memory "
468 "during recovery.\n");
475 memcpy(nbh->b_data, obh->b_data,
476 journal->j_blocksize);
477 if (flags & JFS_FLAG_ESCAPE) {
478 *((unsigned int *)bh->b_data) =
479 htonl(JFS_MAGIC_NUMBER);
482 BUFFER_TRACE(nbh, "marking dirty");
483 mark_buffer_dirty(nbh);
484 BUFFER_TRACE(nbh, "marking uptodate");
485 mark_buffer_uptodate(nbh, 1);
487 /* ll_rw_block(WRITE, 1, &nbh); */
493 tagp += sizeof(journal_block_tag_t);
494 if (!(flags & JFS_FLAG_SAME_UUID))
497 if (flags & JFS_FLAG_LAST_TAG)
504 case JFS_COMMIT_BLOCK:
505 /* Found an expected commit block: not much to
506 * do other than move on to the next sequence
512 case JFS_REVOKE_BLOCK:
513 /* If we aren't in the REVOKE pass, then we can
514 * just skip over this block. */
515 if (pass != PASS_REVOKE) {
520 err = scan_revoke_records(journal, bh,
521 next_commit_ID, info);
528 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
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.
542 if (pass == PASS_SCAN)
543 info->end_transaction = next_commit_ID;
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);
563 /* Scan a revoke record, marking all blocks mentioned as revoked. */
565 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
566 tid_t sequence, struct recovery_info *info)
568 journal_revoke_header_t *header;
571 header = (journal_revoke_header_t *) bh->b_data;
572 offset = sizeof(journal_revoke_header_t);
573 max = ntohl(header->r_count);
575 while (offset < max) {
576 unsigned long blocknr;
579 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
581 err = journal_set_revoke(journal, blocknr, sequence);