1/*
2 * linux/fs/jbd/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/jbd.h>
22#include <linux/errno.h>
23#include <linux/blkdev.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 int 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 = 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 int 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 = 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(struct buffer_head *bh, int size)
182{
183	char *			tagp;
184	journal_block_tag_t *	tag;
185	int			nr = 0;
186
187	tagp = &bh->b_data[sizeof(journal_header_t)];
188
189	while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
190		tag = (journal_block_tag_t *) tagp;
191
192		nr++;
193		tagp += sizeof(journal_block_tag_t);
194		if (!(tag->t_flags & cpu_to_be32(JFS_FLAG_SAME_UUID)))
195			tagp += 16;
196
197		if (tag->t_flags & cpu_to_be32(JFS_FLAG_LAST_TAG))
198			break;
199	}
200
201	return nr;
202}
203
204
205/* Make sure we wrap around the log correctly! */
206#define wrap(journal, var)						\
207do {									\
208	if (var >= (journal)->j_last)					\
209		var -= ((journal)->j_last - (journal)->j_first);	\
210} while (0)
211
212/**
213 * journal_recover - recovers a on-disk journal
214 * @journal: the journal to recover
215 *
216 * The primary function for recovering the log contents when mounting a
217 * journaled device.
218 *
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
222 * in the log.
223 */
224int journal_recover(journal_t *journal)
225{
226	int			err, err2;
227	journal_superblock_t *	sb;
228
229	struct recovery_info	info;
230
231	memset(&info, 0, sizeof(info));
232	sb = journal->j_superblock;
233
234	/*
235	 * The journal superblock's s_start field (the current log head)
236	 * is always zero if, and only if, the journal was cleanly
237	 * unmounted.
238	 */
239
240	if (!sb->s_start) {
241		jbd_debug(1, "No recovery required, last transaction %d\n",
242			  be32_to_cpu(sb->s_sequence));
243		journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
244		return 0;
245	}
246
247	err = do_one_pass(journal, &info, PASS_SCAN);
248	if (!err)
249		err = do_one_pass(journal, &info, PASS_REVOKE);
250	if (!err)
251		err = do_one_pass(journal, &info, PASS_REPLAY);
252
253	jbd_debug(1, "JBD: recovery, exit status %d, "
254		  "recovered transactions %u to %u\n",
255		  err, info.start_transaction, info.end_transaction);
256	jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n",
257		  info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
258
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;
262
263	journal_clear_revoke(journal);
264	err2 = sync_blockdev(journal->j_fs_dev);
265	if (!err)
266		err = err2;
267	/* Flush disk caches to get replayed data on the permanent storage */
268	if (journal->j_flags & JFS_BARRIER) {
269		err2 = blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
270		if (!err)
271			err = err2;
272	}
273
274	return err;
275}
276
277/**
278 * journal_skip_recovery - Start journal and wipe exiting records
279 * @journal: journal to startup
280 *
281 * Locate any valid recovery information from the journal and set up the
282 * journal structures in memory to ignore it (presumably because the
283 * caller has evidence that it is out of date).
284 * This function does'nt appear to be exorted..
285 *
286 * We perform one pass over the journal to allow us to tell the user how
287 * much recovery information is being erased, and to let us initialise
288 * the journal transaction sequence numbers to the next unused ID.
289 */
290int journal_skip_recovery(journal_t *journal)
291{
292	int			err;
293	struct recovery_info	info;
294
295	memset (&info, 0, sizeof(info));
296
297	err = do_one_pass(journal, &info, PASS_SCAN);
298
299	if (err) {
300		printk(KERN_ERR "JBD: error %d scanning journal\n", err);
301		++journal->j_transaction_sequence;
302	} else {
303#ifdef CONFIG_JBD_DEBUG
304		int dropped = info.end_transaction -
305			      be32_to_cpu(journal->j_superblock->s_sequence);
306		jbd_debug(1,
307			  "JBD: ignoring %d transaction%s from the journal.\n",
308			  dropped, (dropped == 1) ? "" : "s");
309#endif
310		journal->j_transaction_sequence = ++info.end_transaction;
311	}
312
313	journal->j_tail = 0;
314	return err;
315}
316
317static int do_one_pass(journal_t *journal,
318			struct recovery_info *info, enum passtype pass)
319{
320	unsigned int		first_commit_ID, next_commit_ID;
321	unsigned int		next_log_block;
322	int			err, success = 0;
323	journal_superblock_t *	sb;
324	journal_header_t *	tmp;
325	struct buffer_head *	bh;
326	unsigned int		sequence;
327	int			blocktype;
328
329	/*
330	 * First thing is to establish what we expect to find in the log
331	 * (in terms of transaction IDs), and where (in terms of log
332	 * block offsets): query the superblock.
333	 */
334
335	sb = journal->j_superblock;
336	next_commit_ID = be32_to_cpu(sb->s_sequence);
337	next_log_block = be32_to_cpu(sb->s_start);
338
339	first_commit_ID = next_commit_ID;
340	if (pass == PASS_SCAN)
341		info->start_transaction = first_commit_ID;
342
343	jbd_debug(1, "Starting recovery pass %d\n", pass);
344
345	/*
346	 * Now we walk through the log, transaction by transaction,
347	 * making sure that each transaction has a commit block in the
348	 * expected place.  Each complete transaction gets replayed back
349	 * into the main filesystem.
350	 */
351
352	while (1) {
353		int			flags;
354		char *			tagp;
355		journal_block_tag_t *	tag;
356		struct buffer_head *	obh;
357		struct buffer_head *	nbh;
358
359		cond_resched();
360
361		/* If we already know where to stop the log traversal,
362		 * check right now that we haven't gone past the end of
363		 * the log. */
364
365		if (pass != PASS_SCAN)
366			if (tid_geq(next_commit_ID, info->end_transaction))
367				break;
368
369		jbd_debug(2, "Scanning for sequence ID %u at %u/%u\n",
370			  next_commit_ID, next_log_block, journal->j_last);
371
372		/* Skip over each chunk of the transaction looking
373		 * either the next descriptor block or the final commit
374		 * record. */
375
376		jbd_debug(3, "JBD: checking block %u\n", next_log_block);
377		err = jread(&bh, journal, next_log_block);
378		if (err)
379			goto failed;
380
381		next_log_block++;
382		wrap(journal, next_log_block);
383
384		/* What kind of buffer is it?
385		 *
386		 * If it is a descriptor block, check that it has the
387		 * expected sequence number.  Otherwise, we're all done
388		 * here. */
389
390		tmp = (journal_header_t *)bh->b_data;
391
392		if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
393			brelse(bh);
394			break;
395		}
396
397		blocktype = be32_to_cpu(tmp->h_blocktype);
398		sequence = be32_to_cpu(tmp->h_sequence);
399		jbd_debug(3, "Found magic %d, sequence %d\n",
400			  blocktype, sequence);
401
402		if (sequence != next_commit_ID) {
403			brelse(bh);
404			break;
405		}
406
407		/* OK, we have a valid descriptor block which matches
408		 * all of the sequence number checks.  What are we going
409		 * to do with it?  That depends on the pass... */
410
411		switch(blocktype) {
412		case JFS_DESCRIPTOR_BLOCK:
413			/* If it is a valid descriptor block, replay it
414			 * in pass REPLAY; otherwise, just skip over the
415			 * blocks it describes. */
416			if (pass != PASS_REPLAY) {
417				next_log_block +=
418					count_tags(bh, journal->j_blocksize);
419				wrap(journal, next_log_block);
420				brelse(bh);
421				continue;
422			}
423
424			/* A descriptor block: we can now write all of
425			 * the data blocks.  Yay, useful work is finally
426			 * getting done here! */
427
428			tagp = &bh->b_data[sizeof(journal_header_t)];
429			while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
430			       <= journal->j_blocksize) {
431				unsigned int io_block;
432
433				tag = (journal_block_tag_t *) tagp;
434				flags = be32_to_cpu(tag->t_flags);
435
436				io_block = next_log_block++;
437				wrap(journal, next_log_block);
438				err = jread(&obh, journal, io_block);
439				if (err) {
440					/* Recover what we can, but
441					 * report failure at the end. */
442					success = err;
443					printk (KERN_ERR
444						"JBD: IO error %d recovering "
445						"block %u in log\n",
446						err, io_block);
447				} else {
448					unsigned int blocknr;
449
450					J_ASSERT(obh != NULL);
451					blocknr = be32_to_cpu(tag->t_blocknr);
452
453					/* If the block has been
454					 * revoked, then we're all done
455					 * here. */
456					if (journal_test_revoke
457					    (journal, blocknr,
458					     next_commit_ID)) {
459						brelse(obh);
460						++info->nr_revoke_hits;
461						goto skip_write;
462					}
463
464					/* Find a buffer for the new
465					 * data being restored */
466					nbh = __getblk(journal->j_fs_dev,
467							blocknr,
468							journal->j_blocksize);
469					if (nbh == NULL) {
470						printk(KERN_ERR
471						       "JBD: Out of memory "
472						       "during recovery.\n");
473						err = -ENOMEM;
474						brelse(bh);
475						brelse(obh);
476						goto failed;
477					}
478
479					lock_buffer(nbh);
480					memcpy(nbh->b_data, obh->b_data,
481							journal->j_blocksize);
482					if (flags & JFS_FLAG_ESCAPE) {
483						*((__be32 *)nbh->b_data) =
484						cpu_to_be32(JFS_MAGIC_NUMBER);
485					}
486
487					BUFFER_TRACE(nbh, "marking dirty");
488					set_buffer_uptodate(nbh);
489					mark_buffer_dirty(nbh);
490					BUFFER_TRACE(nbh, "marking uptodate");
491					++info->nr_replays;
492					/* ll_rw_block(WRITE, 1, &nbh); */
493					unlock_buffer(nbh);
494					brelse(obh);
495					brelse(nbh);
496				}
497
498			skip_write:
499				tagp += sizeof(journal_block_tag_t);
500				if (!(flags & JFS_FLAG_SAME_UUID))
501					tagp += 16;
502
503				if (flags & JFS_FLAG_LAST_TAG)
504					break;
505			}
506
507			brelse(bh);
508			continue;
509
510		case JFS_COMMIT_BLOCK:
511			/* Found an expected commit block: not much to
512			 * do other than move on to the next sequence
513			 * number. */
514			brelse(bh);
515			next_commit_ID++;
516			continue;
517
518		case JFS_REVOKE_BLOCK:
519			/* If we aren't in the REVOKE pass, then we can
520			 * just skip over this block. */
521			if (pass != PASS_REVOKE) {
522				brelse(bh);
523				continue;
524			}
525
526			err = scan_revoke_records(journal, bh,
527						  next_commit_ID, info);
528			brelse(bh);
529			if (err)
530				goto failed;
531			continue;
532
533		default:
534			jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
535				  blocktype);
536			brelse(bh);
537			goto done;
538		}
539	}
540
541 done:
542	/*
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.
547	 */
548
549	if (pass == PASS_SCAN)
550		info->end_transaction = next_commit_ID;
551	else {
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);
558			if (!success)
559				success = -EIO;
560		}
561	}
562
563	return success;
564
565 failed:
566	return err;
567}
568
569
570/* Scan a revoke record, marking all blocks mentioned as revoked. */
571
572static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
573			       tid_t sequence, struct recovery_info *info)
574{
575	journal_revoke_header_t *header;
576	int offset, max;
577
578	header = (journal_revoke_header_t *) bh->b_data;
579	offset = sizeof(journal_revoke_header_t);
580	max = be32_to_cpu(header->r_count);
581
582	while (offset < max) {
583		unsigned int blocknr;
584		int err;
585
586		blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
587		offset += 4;
588		err = journal_set_revoke(journal, blocknr, sequence);
589		if (err)
590			return err;
591		++info->nr_revokes;
592	}
593	return 0;
594}
595