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/time.h>
21 #include <linux/jbd.h>
22 #include <linux/errno.h>
23 #include <linux/slab.h>
27 * Maintain information about the progress of the recovery job, so that
28 * the different passes can carry information between them.
32 tid_t start_transaction;
33 tid_t end_transaction;
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 *);
48 /* Release readahead buffers after use */
49 void journal_brelse_array(struct buffer_head *b[], int n)
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.
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
69 static int do_readahead(journal_t *journal, unsigned int start)
72 unsigned int max, nbufs, next;
73 unsigned long blocknr;
74 struct buffer_head *bh;
76 struct buffer_head * bufs[MAXBUF];
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;
83 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at
84 * a time to the block device IO layer. */
88 for (next = start; next < max; next++) {
89 err = journal_bmap(journal, next, &blocknr);
92 printk (KERN_ERR "JBD: bad block at offset %u\n",
97 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
103 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
105 if (nbufs == MAXBUF) {
106 ll_rw_block(READ, nbufs, bufs);
107 journal_brelse_array(bufs, nbufs);
115 ll_rw_block(READ, nbufs, bufs);
120 journal_brelse_array(bufs, nbufs);
124 #endif /* __KERNEL__ */
128 * Read a block from the journal
131 static int jread(struct buffer_head **bhp, journal_t *journal,
135 unsigned long blocknr;
136 struct buffer_head *bh;
140 J_ASSERT (offset < journal->j_maxlen);
142 err = journal_bmap(journal, offset, &blocknr);
145 printk (KERN_ERR "JBD: bad block at offset %u\n",
150 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
154 if (!buffer_uptodate(bh)) {
155 /* If this is a brand new buffer, start readahead.
156 Otherwise, we assume we are already reading it. */
158 do_readahead(journal, offset);
162 if (!buffer_uptodate(bh)) {
163 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
175 * Count the number of in-use tags in a journal descriptor block.
178 static int count_tags(struct buffer_head *bh, int size)
181 journal_block_tag_t * tag;
184 tagp = &bh->b_data[sizeof(journal_header_t)];
186 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
187 tag = (journal_block_tag_t *) tagp;
190 tagp += sizeof(journal_block_tag_t);
191 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
194 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
202 /* Make sure we wrap around the log correctly! */
203 #define wrap(journal, var) \
205 if (var >= (journal)->j_last) \
206 var -= ((journal)->j_last - (journal)->j_first); \
210 * int journal_recover(journal_t *journal) - recovers a on-disk journal
211 * @journal: the journal to recover
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
221 int journal_recover(journal_t *journal)
224 journal_superblock_t * sb;
226 struct recovery_info info;
228 memset(&info, 0, sizeof(info));
229 sb = journal->j_superblock;
232 * The journal superblock's s_start field (the current log head)
233 * is always zero if, and only if, the journal was cleanly
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;
244 err = do_one_pass(journal, &info, PASS_SCAN);
246 err = do_one_pass(journal, &info, PASS_REVOKE);
248 err = do_one_pass(journal, &info, PASS_REPLAY);
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);
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;
260 journal_clear_revoke(journal);
261 sync_blockdev(journal->j_fs_dev);
266 * int journal_skip_recovery() - Start journal and wipe exiting records
267 * @journal: journal to startup
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..
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.
278 int journal_skip_recovery(journal_t *journal)
281 journal_superblock_t * sb;
283 struct recovery_info info;
285 memset (&info, 0, sizeof(info));
286 sb = journal->j_superblock;
288 err = do_one_pass(journal, &info, PASS_SCAN);
291 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
292 ++journal->j_transaction_sequence;
294 #ifdef CONFIG_JBD_DEBUG
295 int dropped = info.end_transaction - ntohl(sb->s_sequence);
298 "JBD: ignoring %d transaction%s from the journal.\n",
299 dropped, (dropped == 1) ? "" : "s");
300 journal->j_transaction_sequence = ++info.end_transaction;
307 static int do_one_pass(journal_t *journal,
308 struct recovery_info *info, enum passtype pass)
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;
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));
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.
330 sb = journal->j_superblock;
331 next_commit_ID = ntohl(sb->s_sequence);
332 next_log_block = ntohl(sb->s_start);
334 first_commit_ID = next_commit_ID;
335 if (pass == PASS_SCAN)
336 info->start_transaction = first_commit_ID;
338 jbd_debug(1, "Starting recovery pass %d\n", pass);
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.
350 journal_block_tag_t * tag;
351 struct buffer_head * obh;
352 struct buffer_head * nbh;
354 /* If we already know where to stop the log traversal,
355 * check right now that we haven't gone past the end of
358 if (pass != PASS_SCAN)
359 if (tid_geq(next_commit_ID, info->end_transaction))
362 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
363 next_commit_ID, next_log_block, journal->j_last);
365 /* Skip over each chunk of the transaction looking
366 * either the next descriptor block or the final commit
369 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
370 err = jread(&bh, journal, next_log_block);
375 wrap(journal, next_log_block);
377 /* What kind of buffer is it?
379 * If it is a descriptor block, check that it has the
380 * expected sequence number. Otherwise, we're all done
383 tmp = (journal_header_t *)bh->b_data;
385 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
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);
395 if (sequence != next_commit_ID) {
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... */
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) {
411 count_tags(bh, journal->j_blocksize);
412 wrap(journal, next_log_block);
417 /* A descriptor block: we can now write all of
418 * the data blocks. Yay, useful work is finally
419 * getting done here! */
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;
426 tag = (journal_block_tag_t *) tagp;
427 flags = ntohl(tag->t_flags);
429 io_block = next_log_block++;
430 wrap(journal, next_log_block);
431 err = jread(&obh, journal, io_block);
433 /* Recover what we can, but
434 * report failure at the end. */
437 "JBD: IO error %d recovering "
438 "block %lu in log\n",
441 unsigned long blocknr;
443 J_ASSERT(obh != NULL);
444 blocknr = ntohl(tag->t_blocknr);
446 /* If the block has been
447 * revoked, then we're all done
449 if (journal_test_revoke
453 ++info->nr_revoke_hits;
457 /* Find a buffer for the new
458 * data being restored */
459 nbh = __getblk(journal->j_fs_dev,
461 journal->j_blocksize);
464 "JBD: Out of memory "
465 "during recovery.\n");
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);
480 BUFFER_TRACE(nbh, "marking dirty");
481 set_buffer_uptodate(nbh);
482 mark_buffer_dirty(nbh);
483 BUFFER_TRACE(nbh, "marking uptodate");
485 /* ll_rw_block(WRITE, 1, &nbh); */
492 tagp += sizeof(journal_block_tag_t);
493 if (!(flags & JFS_FLAG_SAME_UUID))
496 if (flags & JFS_FLAG_LAST_TAG)
503 case JFS_COMMIT_BLOCK:
504 /* Found an expected commit block: not much to
505 * do other than move on to the next sequence
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) {
519 err = scan_revoke_records(journal, bh,
520 next_commit_ID, info);
527 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);