1/*
2 * linux/fs/jbd2/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/jbd2.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 */
30struct 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
40enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
41static int do_one_pass(journal_t *journal,
42				struct recovery_info *info, enum passtype pass);
43static 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 */
49static 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
69static int do_readahead(journal_t *journal, unsigned int start)
70{
71	int err;
72	unsigned int max, nbufs, next;
73	unsigned long 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 = jbd2_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
118failed:
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
131static int jread(struct buffer_head **bhp, journal_t *journal,
132		 unsigned int offset)
133{
134	int err;
135	unsigned long long blocknr;
136	struct buffer_head *bh;
137
138	*bhp = NULL;
139
140	if (offset >= journal->j_maxlen) {
141		printk(KERN_ERR "JBD: corrupted journal superblock\n");
142		return -EIO;
143	}
144
145	err = jbd2_journal_bmap(journal, offset, &blocknr);
146
147	if (err) {
148		printk (KERN_ERR "JBD: bad block at offset %u\n",
149			offset);
150		return err;
151	}
152
153	bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize);
154	if (!bh)
155		return -ENOMEM;
156
157	if (!buffer_uptodate(bh)) {
158		/* If this is a brand new buffer, start readahead.
159                   Otherwise, we assume we are already reading it.  */
160		if (!buffer_req(bh))
161			do_readahead(journal, offset);
162		wait_on_buffer(bh);
163	}
164
165	if (!buffer_uptodate(bh)) {
166		printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
167			offset);
168		brelse(bh);
169		return -EIO;
170	}
171
172	*bhp = bh;
173	return 0;
174}
175
176
177/*
178 * Count the number of in-use tags in a journal descriptor block.
179 */
180
181static int count_tags(journal_t *journal, struct buffer_head *bh)
182{
183	char *			tagp;
184	journal_block_tag_t *	tag;
185	int			nr = 0, size = journal->j_blocksize;
186	int			tag_bytes = journal_tag_bytes(journal);
187
188	tagp = &bh->b_data[sizeof(journal_header_t)];
189
190	while ((tagp - bh->b_data + tag_bytes) <= size) {
191		tag = (journal_block_tag_t *) tagp;
192
193		nr++;
194		tagp += tag_bytes;
195		if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID)))
196			tagp += 16;
197
198		if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG))
199			break;
200	}
201
202	return nr;
203}
204
205
206/* Make sure we wrap around the log correctly! */
207#define wrap(journal, var)						\
208do {									\
209	if (var >= (journal)->j_last)					\
210		var -= ((journal)->j_last - (journal)->j_first);	\
211} while (0)
212
213/**
214 * jbd2_journal_recover - recovers a on-disk journal
215 * @journal: the journal to recover
216 *
217 * The primary function for recovering the log contents when mounting a
218 * journaled device.
219 *
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
223 * in the log.
224 */
225int jbd2_journal_recover(journal_t *journal)
226{
227	int			err;
228	journal_superblock_t *	sb;
229
230	struct recovery_info	info;
231
232	memset(&info, 0, sizeof(info));
233	sb = journal->j_superblock;
234
235	/*
236	 * The journal superblock's s_start field (the current log head)
237	 * is always zero if, and only if, the journal was cleanly
238	 * unmounted.
239	 */
240
241	if (!sb->s_start) {
242		jbd_debug(1, "No recovery required, last transaction %d\n",
243			  be32_to_cpu(sb->s_sequence));
244		journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
245		return 0;
246	}
247
248	err = do_one_pass(journal, &info, PASS_SCAN);
249	if (!err)
250		err = do_one_pass(journal, &info, PASS_REVOKE);
251	if (!err)
252		err = do_one_pass(journal, &info, PASS_REPLAY);
253
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);
259
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;
263
264	jbd2_journal_clear_revoke(journal);
265	sync_blockdev(journal->j_fs_dev);
266	return err;
267}
268
269/**
270 * jbd2_journal_skip_recovery - Start journal and wipe exiting records
271 * @journal: journal to startup
272 *
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..
277 *
278 * We perform one pass over the journal to allow us to tell the user how
279 * much recovery information is being erased, and to let us initialise
280 * the journal transaction sequence numbers to the next unused ID.
281 */
282int jbd2_journal_skip_recovery(journal_t *journal)
283{
284	int			err;
285	journal_superblock_t *	sb;
286
287	struct recovery_info	info;
288
289	memset (&info, 0, sizeof(info));
290	sb = journal->j_superblock;
291
292	err = do_one_pass(journal, &info, PASS_SCAN);
293
294	if (err) {
295		printk(KERN_ERR "JBD: error %d scanning journal\n", err);
296		++journal->j_transaction_sequence;
297	} else {
298#ifdef CONFIG_JBD_DEBUG
299		int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
300#endif
301		jbd_debug(0,
302			  "JBD: ignoring %d transaction%s from the journal.\n",
303			  dropped, (dropped == 1) ? "" : "s");
304		journal->j_transaction_sequence = ++info.end_transaction;
305	}
306
307	journal->j_tail = 0;
308	return err;
309}
310
311static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
312{
313	unsigned long long block = be32_to_cpu(tag->t_blocknr);
314	if (tag_bytes > JBD_TAG_SIZE32)
315		block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
316	return block;
317}
318
319static int do_one_pass(journal_t *journal,
320			struct recovery_info *info, enum passtype pass)
321{
322	unsigned int		first_commit_ID, next_commit_ID;
323	unsigned long		next_log_block;
324	int			err, success = 0;
325	journal_superblock_t *	sb;
326	journal_header_t *	tmp;
327	struct buffer_head *	bh;
328	unsigned int		sequence;
329	int			blocktype;
330	int			tag_bytes = journal_tag_bytes(journal);
331
332	/* Precompute the maximum metadata descriptors in a descriptor block */
333	int			MAX_BLOCKS_PER_DESC;
334	MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
335			       / tag_bytes);
336
337	/*
338	 * First thing is to establish what we expect to find in the log
339	 * (in terms of transaction IDs), and where (in terms of log
340	 * block offsets): query the superblock.
341	 */
342
343	sb = journal->j_superblock;
344	next_commit_ID = be32_to_cpu(sb->s_sequence);
345	next_log_block = be32_to_cpu(sb->s_start);
346
347	first_commit_ID = next_commit_ID;
348	if (pass == PASS_SCAN)
349		info->start_transaction = first_commit_ID;
350
351	jbd_debug(1, "Starting recovery pass %d\n", pass);
352
353	/*
354	 * Now we walk through the log, transaction by transaction,
355	 * making sure that each transaction has a commit block in the
356	 * expected place.  Each complete transaction gets replayed back
357	 * into the main filesystem.
358	 */
359
360	while (1) {
361		int			flags;
362		char *			tagp;
363		journal_block_tag_t *	tag;
364		struct buffer_head *	obh;
365		struct buffer_head *	nbh;
366
367		cond_resched();		/* We're under lock_kernel() */
368
369		/* If we already know where to stop the log traversal,
370		 * check right now that we haven't gone past the end of
371		 * the log. */
372
373		if (pass != PASS_SCAN)
374			if (tid_geq(next_commit_ID, info->end_transaction))
375				break;
376
377		jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
378			  next_commit_ID, next_log_block, journal->j_last);
379
380		/* Skip over each chunk of the transaction looking
381		 * either the next descriptor block or the final commit
382		 * record. */
383
384		jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
385		err = jread(&bh, journal, next_log_block);
386		if (err)
387			goto failed;
388
389		next_log_block++;
390		wrap(journal, next_log_block);
391
392		/* What kind of buffer is it?
393		 *
394		 * If it is a descriptor block, check that it has the
395		 * expected sequence number.  Otherwise, we're all done
396		 * here. */
397
398		tmp = (journal_header_t *)bh->b_data;
399
400		if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
401			brelse(bh);
402			break;
403		}
404
405		blocktype = be32_to_cpu(tmp->h_blocktype);
406		sequence = be32_to_cpu(tmp->h_sequence);
407		jbd_debug(3, "Found magic %d, sequence %d\n",
408			  blocktype, sequence);
409
410		if (sequence != next_commit_ID) {
411			brelse(bh);
412			break;
413		}
414
415		/* OK, we have a valid descriptor block which matches
416		 * all of the sequence number checks.  What are we going
417		 * to do with it?  That depends on the pass... */
418
419		switch(blocktype) {
420		case JBD2_DESCRIPTOR_BLOCK:
421			/* If it is a valid descriptor block, replay it
422			 * in pass REPLAY; otherwise, just skip over the
423			 * blocks it describes. */
424			if (pass != PASS_REPLAY) {
425				next_log_block += count_tags(journal, bh);
426				wrap(journal, next_log_block);
427				brelse(bh);
428				continue;
429			}
430
431			/* A descriptor block: we can now write all of
432			 * the data blocks.  Yay, useful work is finally
433			 * getting done here! */
434
435			tagp = &bh->b_data[sizeof(journal_header_t)];
436			while ((tagp - bh->b_data + tag_bytes)
437			       <= journal->j_blocksize) {
438				unsigned long io_block;
439
440				tag = (journal_block_tag_t *) tagp;
441				flags = be32_to_cpu(tag->t_flags);
442
443				io_block = next_log_block++;
444				wrap(journal, next_log_block);
445				err = jread(&obh, journal, io_block);
446				if (err) {
447					/* Recover what we can, but
448					 * report failure at the end. */
449					success = err;
450					printk (KERN_ERR
451						"JBD: IO error %d recovering "
452						"block %ld in log\n",
453						err, io_block);
454				} else {
455					unsigned long long blocknr;
456
457					J_ASSERT(obh != NULL);
458					blocknr = read_tag_block(tag_bytes,
459								 tag);
460
461					/* If the block has been
462					 * revoked, then we're all done
463					 * here. */
464					if (jbd2_journal_test_revoke
465					    (journal, blocknr,
466					     next_commit_ID)) {
467						brelse(obh);
468						++info->nr_revoke_hits;
469						goto skip_write;
470					}
471
472					/* Find a buffer for the new
473					 * data being restored */
474					nbh = __getblk(journal->j_fs_dev,
475							blocknr,
476							journal->j_blocksize);
477					if (nbh == NULL) {
478						printk(KERN_ERR
479						       "JBD: Out of memory "
480						       "during recovery.\n");
481						err = -ENOMEM;
482						brelse(bh);
483						brelse(obh);
484						goto failed;
485					}
486
487					lock_buffer(nbh);
488					memcpy(nbh->b_data, obh->b_data,
489							journal->j_blocksize);
490					if (flags & JBD2_FLAG_ESCAPE) {
491						*((__be32 *)bh->b_data) =
492						cpu_to_be32(JBD2_MAGIC_NUMBER);
493					}
494
495					BUFFER_TRACE(nbh, "marking dirty");
496					set_buffer_uptodate(nbh);
497					mark_buffer_dirty(nbh);
498					BUFFER_TRACE(nbh, "marking uptodate");
499					++info->nr_replays;
500					/* ll_rw_block(WRITE, 1, &nbh); */
501					unlock_buffer(nbh);
502					brelse(obh);
503					brelse(nbh);
504				}
505
506			skip_write:
507				tagp += tag_bytes;
508				if (!(flags & JBD2_FLAG_SAME_UUID))
509					tagp += 16;
510
511				if (flags & JBD2_FLAG_LAST_TAG)
512					break;
513			}
514
515			brelse(bh);
516			continue;
517
518		case JBD2_COMMIT_BLOCK:
519			/* Found an expected commit block: not much to
520			 * do other than move on to the next sequence
521			 * number. */
522			brelse(bh);
523			next_commit_ID++;
524			continue;
525
526		case JBD2_REVOKE_BLOCK:
527			/* If we aren't in the REVOKE pass, then we can
528			 * just skip over this block. */
529			if (pass != PASS_REVOKE) {
530				brelse(bh);
531				continue;
532			}
533
534			err = scan_revoke_records(journal, bh,
535						  next_commit_ID, info);
536			brelse(bh);
537			if (err)
538				goto failed;
539			continue;
540
541		default:
542			jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
543				  blocktype);
544			brelse(bh);
545			goto done;
546		}
547	}
548
549 done:
550	/*
551	 * We broke out of the log scan loop: either we came to the
552	 * known end of the log or we found an unexpected block in the
553	 * log.  If the latter happened, then we know that the "current"
554	 * transaction marks the end of the valid log.
555	 */
556
557	if (pass == PASS_SCAN)
558		info->end_transaction = next_commit_ID;
559	else {
560		/* It's really bad news if different passes end up at
561		 * different places (but possible due to IO errors). */
562		if (info->end_transaction != next_commit_ID) {
563			printk (KERN_ERR "JBD: recovery pass %d ended at "
564				"transaction %u, expected %u\n",
565				pass, next_commit_ID, info->end_transaction);
566			if (!success)
567				success = -EIO;
568		}
569	}
570
571	return success;
572
573 failed:
574	return err;
575}
576
577
578/* Scan a revoke record, marking all blocks mentioned as revoked. */
579
580static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
581			       tid_t sequence, struct recovery_info *info)
582{
583	jbd2_journal_revoke_header_t *header;
584	int offset, max;
585	int record_len = 4;
586
587	header = (jbd2_journal_revoke_header_t *) bh->b_data;
588	offset = sizeof(jbd2_journal_revoke_header_t);
589	max = be32_to_cpu(header->r_count);
590
591	if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
592		record_len = 8;
593
594	while (offset + record_len <= max) {
595		unsigned long long blocknr;
596		int err;
597
598		if (record_len == 4)
599			blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
600		else
601			blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
602		offset += record_len;
603		err = jbd2_journal_set_revoke(journal, blocknr, sequence);
604		if (err)
605			return err;
606		++info->nr_revokes;
607	}
608	return 0;
609}
610