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); \
211 * int journal_recover(journal_t *journal) - recovers a on-disk journal
212 * @journal: the journal to recover
214 * The primary function for recovering the log contents when mounting a
217 int journal_recover(journal_t *journal)
220 * Recovery is done in three passes. In the first pass, we look for the
221 * end of the log. In the second, we assemble the list of revoke
222 * blocks. In the third and final pass, we replay any un-revoked blocks
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;
248 err = do_one_pass(journal, &info, PASS_SCAN);
250 err = do_one_pass(journal, &info, PASS_REVOKE);
252 err = do_one_pass(journal, &info, PASS_REPLAY);
254 jbd_debug(0, "JBD: recovery, exit status %d, "
255 "recovered transactions %u to %u\n",
256 err, info.start_transaction, info.end_transaction);
257 jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
258 info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
260 /* Restart the log at the next transaction ID, thus invalidating
261 * any existing commit records in the log. */
262 journal->j_transaction_sequence = ++info.end_transaction;
264 journal_clear_revoke(journal);
265 fsync_no_super(journal->j_fs_dev);
270 * int journal_skip_recovery() - Start journal and wipe exiting records
271 * @journal: journal to startup
273 * Locate any valid recovery information from the journal and set up the
274 * journal structures in memory to ignore it (presumably because the
275 * caller has evidence that it is out of date).
276 * This function does'nt appear to be exorted..
278 int journal_skip_recovery(journal_t *journal)
281 * We perform one pass over the journal to allow us to tell the user how
282 * much recovery information is being erased, and to let us initialise
283 * the journal transaction sequence numbers to the next unused ID.
287 journal_superblock_t * sb;
289 struct recovery_info info;
291 memset (&info, 0, sizeof(info));
292 sb = journal->j_superblock;
294 err = do_one_pass(journal, &info, PASS_SCAN);
297 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
298 ++journal->j_transaction_sequence;
300 #ifdef CONFIG_JBD_DEBUG
301 int dropped = info.end_transaction - ntohl(sb->s_sequence);
305 "JBD: ignoring %d transaction%s from the journal.\n",
306 dropped, (dropped == 1) ? "" : "s");
307 journal->j_transaction_sequence = ++info.end_transaction;
315 static int do_one_pass(journal_t *journal,
316 struct recovery_info *info, enum passtype pass)
319 unsigned int first_commit_ID, next_commit_ID;
320 unsigned long next_log_block;
321 int err, success = 0;
322 journal_superblock_t * sb;
323 journal_header_t * tmp;
324 struct buffer_head * bh;
325 unsigned int sequence;
328 /* Precompute the maximum metadata descriptors in a descriptor block */
329 int MAX_BLOCKS_PER_DESC;
330 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
331 / sizeof(journal_block_tag_t));
334 * First thing is to establish what we expect to find in the log
335 * (in terms of transaction IDs), and where (in terms of log
336 * block offsets): query the superblock.
339 sb = journal->j_superblock;
340 next_commit_ID = ntohl(sb->s_sequence);
341 next_log_block = ntohl(sb->s_start);
343 first_commit_ID = next_commit_ID;
344 if (pass == PASS_SCAN)
345 info->start_transaction = first_commit_ID;
347 jbd_debug(1, "Starting recovery pass %d\n", pass);
350 * Now we walk through the log, transaction by transaction,
351 * making sure that each transaction has a commit block in the
352 * expected place. Each complete transaction gets replayed back
353 * into the main filesystem.
359 journal_block_tag_t * tag;
360 struct buffer_head * obh;
361 struct buffer_head * nbh;
363 /* If we already know where to stop the log traversal,
364 * check right now that we haven't gone past the end of
367 if (pass != PASS_SCAN)
368 if (tid_geq(next_commit_ID, info->end_transaction))
371 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
372 next_commit_ID, next_log_block, journal->j_last);
374 /* Skip over each chunk of the transaction looking
375 * either the next descriptor block or the final commit
378 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
379 err = jread(&bh, journal, next_log_block);
384 wrap(journal, next_log_block);
386 /* What kind of buffer is it?
388 * If it is a descriptor block, check that it has the
389 * expected sequence number. Otherwise, we're all done
392 tmp = (journal_header_t *)bh->b_data;
394 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
399 blocktype = ntohl(tmp->h_blocktype);
400 sequence = ntohl(tmp->h_sequence);
401 jbd_debug(3, "Found magic %d, sequence %d\n",
402 blocktype, sequence);
404 if (sequence != next_commit_ID) {
409 /* OK, we have a valid descriptor block which matches
410 * all of the sequence number checks. What are we going
411 * to do with it? That depends on the pass... */
414 case JFS_DESCRIPTOR_BLOCK:
415 /* If it is a valid descriptor block, replay it
416 * in pass REPLAY; otherwise, just skip over the
417 * blocks it describes. */
418 if (pass != PASS_REPLAY) {
420 count_tags(bh, journal->j_blocksize);
421 wrap(journal, next_log_block);
426 /* A descriptor block: we can now write all of
427 * the data blocks. Yay, useful work is finally
428 * getting done here! */
430 tagp = &bh->b_data[sizeof(journal_header_t)];
431 while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
432 <= journal->j_blocksize) {
433 unsigned long io_block;
435 tag = (journal_block_tag_t *) tagp;
436 flags = ntohl(tag->t_flags);
438 io_block = next_log_block++;
439 wrap(journal, next_log_block);
440 err = jread(&obh, journal, io_block);
442 /* Recover what we can, but
443 * report failure at the end. */
446 "JBD: IO error %d recovering "
447 "block %ld in log\n",
450 unsigned long blocknr;
452 J_ASSERT(obh != NULL);
453 blocknr = ntohl(tag->t_blocknr);
455 /* If the block has been
456 * revoked, then we're all done
458 if (journal_test_revoke
462 ++info->nr_revoke_hits;
466 /* Find a buffer for the new
467 * data being restored */
468 nbh = getblk(journal->j_fs_dev, blocknr,
469 journal->j_blocksize);
472 "JBD: Out of memory "
473 "during recovery.\n");
481 memcpy(nbh->b_data, obh->b_data,
482 journal->j_blocksize);
483 if (flags & JFS_FLAG_ESCAPE) {
484 *((unsigned int *)bh->b_data) =
485 htonl(JFS_MAGIC_NUMBER);
488 BUFFER_TRACE(nbh, "marking dirty");
489 mark_buffer_dirty(nbh);
490 BUFFER_TRACE(nbh, "marking uptodate");
491 mark_buffer_uptodate(nbh, 1);
494 /* ll_rw_block(WRITE, 1, &nbh); */
500 tagp += sizeof(journal_block_tag_t);
501 if (!(flags & JFS_FLAG_SAME_UUID))
504 if (flags & JFS_FLAG_LAST_TAG)
511 case JFS_COMMIT_BLOCK:
512 /* Found an expected commit block: not much to
513 * do other than move on to the next sequence
519 case JFS_REVOKE_BLOCK:
520 /* If we aren't in the REVOKE pass, then we can
521 * just skip over this block. */
522 if (pass != PASS_REVOKE) {
527 err = scan_revoke_records(journal, bh,
528 next_commit_ID, info);
535 jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
543 * We broke out of the log scan loop: either we came to the
544 * known end of the log or we found an unexpected block in the
545 * log. If the latter happened, then we know that the "current"
546 * transaction marks the end of the valid log.
549 if (pass == PASS_SCAN)
550 info->end_transaction = next_commit_ID;
552 /* It's really bad news if different passes end up at
553 * different places (but possible due to IO errors). */
554 if (info->end_transaction != next_commit_ID) {
555 printk (KERN_ERR "JBD: recovery pass %d ended at "
556 "transaction %u, expected %u\n",
557 pass, next_commit_ID, info->end_transaction);
570 /* Scan a revoke record, marking all blocks mentioned as revoked. */
572 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
573 tid_t sequence, struct recovery_info *info)
575 journal_revoke_header_t *header;
578 header = (journal_revoke_header_t *) bh->b_data;
579 offset = sizeof(journal_revoke_header_t);
580 max = ntohl(header->r_count);
582 while (offset < max) {
583 unsigned long blocknr;
586 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
588 err = journal_set_revoke(journal, blocknr, sequence);