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 if (offset >= journal->j_maxlen) {
141 printk(KERN_ERR "JBD: corrupted journal superblock\n");
145 err = journal_bmap(journal, offset, &blocknr);
148 printk (KERN_ERR "JBD: bad block at offset %u\n",
153 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
157 if (!buffer_uptodate(bh)) {
158 /* If this is a brand new buffer, start readahead.
159 Otherwise, we assume we are already reading it. */
161 do_readahead(journal, offset);
165 if (!buffer_uptodate(bh)) {
166 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
178 * Count the number of in-use tags in a journal descriptor block.
181 static int count_tags(struct buffer_head *bh, int size)
184 journal_block_tag_t * tag;
187 tagp = &bh->b_data[sizeof(journal_header_t)];
189 while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
190 tag = (journal_block_tag_t *) tagp;
193 tagp += sizeof(journal_block_tag_t);
194 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
197 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
205 /* Make sure we wrap around the log correctly! */
206 #define wrap(journal, var) \
208 if (var >= (journal)->j_last) \
209 var -= ((journal)->j_last - (journal)->j_first); \
213 * int journal_recover(journal_t *journal) - recovers a on-disk journal
214 * @journal: the journal to recover
216 * The primary function for recovering the log contents when mounting a
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
224 int journal_recover(journal_t *journal)
227 journal_superblock_t * sb;
229 struct recovery_info info;
231 memset(&info, 0, sizeof(info));
232 sb = journal->j_superblock;
235 * The journal superblock's s_start field (the current log head)
236 * is always zero if, and only if, the journal was cleanly
241 jbd_debug(1, "No recovery required, last transaction %d\n",
242 ntohl(sb->s_sequence));
243 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
247 err = do_one_pass(journal, &info, PASS_SCAN);
249 err = do_one_pass(journal, &info, PASS_REVOKE);
251 err = do_one_pass(journal, &info, PASS_REPLAY);
253 jbd_debug(0, "JBD: recovery, exit status %d, "
254 "recovered transactions %u to %u\n",
255 err, info.start_transaction, info.end_transaction);
256 jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
257 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
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;
263 journal_clear_revoke(journal);
264 sync_blockdev(journal->j_fs_dev);
269 * int journal_skip_recovery() - Start journal and wipe exiting records
270 * @journal: journal to startup
272 * Locate any valid recovery information from the journal and set up the
273 * journal structures in memory to ignore it (presumably because the
274 * caller has evidence that it is out of date).
275 * This function does'nt appear to be exorted..
277 * We perform one pass over the journal to allow us to tell the user how
278 * much recovery information is being erased, and to let us initialise
279 * the journal transaction sequence numbers to the next unused ID.
281 int journal_skip_recovery(journal_t *journal)
284 journal_superblock_t * sb;
286 struct recovery_info info;
288 memset (&info, 0, sizeof(info));
289 sb = journal->j_superblock;
291 err = do_one_pass(journal, &info, PASS_SCAN);
294 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
295 ++journal->j_transaction_sequence;
297 #ifdef CONFIG_JBD_DEBUG
298 int dropped = info.end_transaction - ntohl(sb->s_sequence);
301 "JBD: ignoring %d transaction%s from the journal.\n",
302 dropped, (dropped == 1) ? "" : "s");
303 journal->j_transaction_sequence = ++info.end_transaction;
310 static int do_one_pass(journal_t *journal,
311 struct recovery_info *info, enum passtype pass)
313 unsigned int first_commit_ID, next_commit_ID;
314 unsigned long next_log_block;
315 int err, success = 0;
316 journal_superblock_t * sb;
317 journal_header_t * tmp;
318 struct buffer_head * bh;
319 unsigned int sequence;
322 /* Precompute the maximum metadata descriptors in a descriptor block */
323 int MAX_BLOCKS_PER_DESC;
324 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
325 / sizeof(journal_block_tag_t));
328 * First thing is to establish what we expect to find in the log
329 * (in terms of transaction IDs), and where (in terms of log
330 * block offsets): query the superblock.
333 sb = journal->j_superblock;
334 next_commit_ID = ntohl(sb->s_sequence);
335 next_log_block = ntohl(sb->s_start);
337 first_commit_ID = next_commit_ID;
338 if (pass == PASS_SCAN)
339 info->start_transaction = first_commit_ID;
341 jbd_debug(1, "Starting recovery pass %d\n", pass);
344 * Now we walk through the log, transaction by transaction,
345 * making sure that each transaction has a commit block in the
346 * expected place. Each complete transaction gets replayed back
347 * into the main filesystem.
353 journal_block_tag_t * tag;
354 struct buffer_head * obh;
355 struct buffer_head * nbh;
357 /* If we already know where to stop the log traversal,
358 * check right now that we haven't gone past the end of
361 if (pass != PASS_SCAN)
362 if (tid_geq(next_commit_ID, info->end_transaction))
365 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
366 next_commit_ID, next_log_block, journal->j_last);
368 /* Skip over each chunk of the transaction looking
369 * either the next descriptor block or the final commit
372 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
373 err = jread(&bh, journal, next_log_block);
378 wrap(journal, next_log_block);
380 /* What kind of buffer is it?
382 * If it is a descriptor block, check that it has the
383 * expected sequence number. Otherwise, we're all done
386 tmp = (journal_header_t *)bh->b_data;
388 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
393 blocktype = ntohl(tmp->h_blocktype);
394 sequence = ntohl(tmp->h_sequence);
395 jbd_debug(3, "Found magic %d, sequence %d\n",
396 blocktype, sequence);
398 if (sequence != next_commit_ID) {
403 /* OK, we have a valid descriptor block which matches
404 * all of the sequence number checks. What are we going
405 * to do with it? That depends on the pass... */
408 case JFS_DESCRIPTOR_BLOCK:
409 /* If it is a valid descriptor block, replay it
410 * in pass REPLAY; otherwise, just skip over the
411 * blocks it describes. */
412 if (pass != PASS_REPLAY) {
414 count_tags(bh, journal->j_blocksize);
415 wrap(journal, next_log_block);
420 /* A descriptor block: we can now write all of
421 * the data blocks. Yay, useful work is finally
422 * getting done here! */
424 tagp = &bh->b_data[sizeof(journal_header_t)];
425 while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
426 <= journal->j_blocksize) {
427 unsigned long io_block;
429 tag = (journal_block_tag_t *) tagp;
430 flags = ntohl(tag->t_flags);
432 io_block = next_log_block++;
433 wrap(journal, next_log_block);
434 err = jread(&obh, journal, io_block);
436 /* Recover what we can, but
437 * report failure at the end. */
440 "JBD: IO error %d recovering "
441 "block %ld in log\n",
444 unsigned long blocknr;
446 J_ASSERT(obh != NULL);
447 blocknr = ntohl(tag->t_blocknr);
449 /* If the block has been
450 * revoked, then we're all done
452 if (journal_test_revoke
456 ++info->nr_revoke_hits;
460 /* Find a buffer for the new
461 * data being restored */
462 nbh = __getblk(journal->j_fs_dev,
464 journal->j_blocksize);
467 "JBD: Out of memory "
468 "during recovery.\n");
476 memcpy(nbh->b_data, obh->b_data,
477 journal->j_blocksize);
478 if (flags & JFS_FLAG_ESCAPE) {
479 *((unsigned int *)bh->b_data) =
480 htonl(JFS_MAGIC_NUMBER);
483 BUFFER_TRACE(nbh, "marking dirty");
484 set_buffer_uptodate(nbh);
485 mark_buffer_dirty(nbh);
486 BUFFER_TRACE(nbh, "marking uptodate");
488 /* ll_rw_block(WRITE, 1, &nbh); */
495 tagp += sizeof(journal_block_tag_t);
496 if (!(flags & JFS_FLAG_SAME_UUID))
499 if (flags & JFS_FLAG_LAST_TAG)
506 case JFS_COMMIT_BLOCK:
507 /* Found an expected commit block: not much to
508 * do other than move on to the next sequence
514 case JFS_REVOKE_BLOCK:
515 /* If we aren't in the REVOKE pass, then we can
516 * just skip over this block. */
517 if (pass != PASS_REVOKE) {
522 err = scan_revoke_records(journal, bh,
523 next_commit_ID, info);
530 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
538 * We broke out of the log scan loop: either we came to the
539 * known end of the log or we found an unexpected block in the
540 * log. If the latter happened, then we know that the "current"
541 * transaction marks the end of the valid log.
544 if (pass == PASS_SCAN)
545 info->end_transaction = next_commit_ID;
547 /* It's really bad news if different passes end up at
548 * different places (but possible due to IO errors). */
549 if (info->end_transaction != next_commit_ID) {
550 printk (KERN_ERR "JBD: recovery pass %d ended at "
551 "transaction %u, expected %u\n",
552 pass, next_commit_ID, info->end_transaction);
565 /* Scan a revoke record, marking all blocks mentioned as revoked. */
567 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
568 tid_t sequence, struct recovery_info *info)
570 journal_revoke_header_t *header;
573 header = (journal_revoke_header_t *) bh->b_data;
574 offset = sizeof(journal_revoke_header_t);
575 max = ntohl(header->r_count);
577 while (offset < max) {
578 unsigned long blocknr;
581 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
583 err = journal_set_revoke(journal, blocknr, sequence);