[PATCH] VFS: Permit filesystem to perform statfs with a known root dentry
[linux-2.6.git] / fs / jbd / recovery.c
1 /*
2  * linux/fs/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/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  */
30 struct 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
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 *);
45
46 #ifdef __KERNEL__
47
48 /* Release readahead buffers after use */
49 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
69 static int do_readahead(journal_t *journal, unsigned int start)
70 {
71         int err;
72         unsigned int max, nbufs, next;
73         unsigned 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 = 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
118 failed:
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
131 static int jread(struct buffer_head **bhp, journal_t *journal, 
132                  unsigned int offset)
133 {
134         int err;
135         unsigned 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 = 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
181 static 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)                                              \
207 do {                                                                    \
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  */
224 int journal_recover(journal_t *journal)
225 {
226         int                     err;
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(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);
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         sync_blockdev(journal->j_fs_dev);
265         return err;
266 }
267
268 /**
269  * journal_skip_recovery - Start journal and wipe exiting records
270  * @journal: journal to startup
271  * 
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..
276  *
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. 
280  */
281 int journal_skip_recovery(journal_t *journal)
282 {
283         int                     err;
284         journal_superblock_t *  sb;
285
286         struct recovery_info    info;
287
288         memset (&info, 0, sizeof(info));
289         sb = journal->j_superblock;
290
291         err = do_one_pass(journal, &info, PASS_SCAN);
292
293         if (err) {
294                 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
295                 ++journal->j_transaction_sequence;
296         } else {
297 #ifdef CONFIG_JBD_DEBUG
298                 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence);
299 #endif
300                 jbd_debug(0, 
301                           "JBD: ignoring %d transaction%s from the journal.\n",
302                           dropped, (dropped == 1) ? "" : "s");
303                 journal->j_transaction_sequence = ++info.end_transaction;
304         }
305
306         journal->j_tail = 0;
307         return err;
308 }
309
310 static int do_one_pass(journal_t *journal,
311                         struct recovery_info *info, enum passtype pass)
312 {
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;
320         int                     blocktype;
321
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));
326
327         /* 
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.  
331          */
332
333         sb = journal->j_superblock;
334         next_commit_ID = be32_to_cpu(sb->s_sequence);
335         next_log_block = be32_to_cpu(sb->s_start);
336
337         first_commit_ID = next_commit_ID;
338         if (pass == PASS_SCAN)
339                 info->start_transaction = first_commit_ID;
340
341         jbd_debug(1, "Starting recovery pass %d\n", pass);
342
343         /*
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. 
348          */
349
350         while (1) {
351                 int                     flags;
352                 char *                  tagp;
353                 journal_block_tag_t *   tag;
354                 struct buffer_head *    obh;
355                 struct buffer_head *    nbh;
356
357                 cond_resched();         /* We're under lock_kernel() */
358
359                 /* If we already know where to stop the log traversal,
360                  * check right now that we haven't gone past the end of
361                  * the log. */
362
363                 if (pass != PASS_SCAN)
364                         if (tid_geq(next_commit_ID, info->end_transaction))
365                                 break;
366
367                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
368                           next_commit_ID, next_log_block, journal->j_last);
369
370                 /* Skip over each chunk of the transaction looking
371                  * either the next descriptor block or the final commit
372                  * record. */
373
374                 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
375                 err = jread(&bh, journal, next_log_block);
376                 if (err)
377                         goto failed;
378
379                 next_log_block++;
380                 wrap(journal, next_log_block);
381
382                 /* What kind of buffer is it? 
383                  * 
384                  * If it is a descriptor block, check that it has the
385                  * expected sequence number.  Otherwise, we're all done
386                  * here. */
387
388                 tmp = (journal_header_t *)bh->b_data;
389
390                 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) {
391                         brelse(bh);
392                         break;
393                 }
394
395                 blocktype = be32_to_cpu(tmp->h_blocktype);
396                 sequence = be32_to_cpu(tmp->h_sequence);
397                 jbd_debug(3, "Found magic %d, sequence %d\n", 
398                           blocktype, sequence);
399
400                 if (sequence != next_commit_ID) {
401                         brelse(bh);
402                         break;
403                 }
404
405                 /* OK, we have a valid descriptor block which matches
406                  * all of the sequence number checks.  What are we going
407                  * to do with it?  That depends on the pass... */
408
409                 switch(blocktype) {
410                 case JFS_DESCRIPTOR_BLOCK:
411                         /* If it is a valid descriptor block, replay it
412                          * in pass REPLAY; otherwise, just skip over the
413                          * blocks it describes. */
414                         if (pass != PASS_REPLAY) {
415                                 next_log_block +=
416                                         count_tags(bh, journal->j_blocksize);
417                                 wrap(journal, next_log_block);
418                                 brelse(bh);
419                                 continue;
420                         }
421
422                         /* A descriptor block: we can now write all of
423                          * the data blocks.  Yay, useful work is finally
424                          * getting done here! */
425
426                         tagp = &bh->b_data[sizeof(journal_header_t)];
427                         while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
428                                <= journal->j_blocksize) {
429                                 unsigned long io_block;
430
431                                 tag = (journal_block_tag_t *) tagp;
432                                 flags = be32_to_cpu(tag->t_flags);
433
434                                 io_block = next_log_block++;
435                                 wrap(journal, next_log_block);
436                                 err = jread(&obh, journal, io_block);
437                                 if (err) {
438                                         /* Recover what we can, but
439                                          * report failure at the end. */
440                                         success = err;
441                                         printk (KERN_ERR 
442                                                 "JBD: IO error %d recovering "
443                                                 "block %ld in log\n",
444                                                 err, io_block);
445                                 } else {
446                                         unsigned long blocknr;
447
448                                         J_ASSERT(obh != NULL);
449                                         blocknr = be32_to_cpu(tag->t_blocknr);
450
451                                         /* If the block has been
452                                          * revoked, then we're all done
453                                          * here. */
454                                         if (journal_test_revoke
455                                             (journal, blocknr, 
456                                              next_commit_ID)) {
457                                                 brelse(obh);
458                                                 ++info->nr_revoke_hits;
459                                                 goto skip_write;
460                                         }
461
462                                         /* Find a buffer for the new
463                                          * data being restored */
464                                         nbh = __getblk(journal->j_fs_dev,
465                                                         blocknr,
466                                                         journal->j_blocksize);
467                                         if (nbh == NULL) {
468                                                 printk(KERN_ERR 
469                                                        "JBD: Out of memory "
470                                                        "during recovery.\n");
471                                                 err = -ENOMEM;
472                                                 brelse(bh);
473                                                 brelse(obh);
474                                                 goto failed;
475                                         }
476
477                                         lock_buffer(nbh);
478                                         memcpy(nbh->b_data, obh->b_data,
479                                                         journal->j_blocksize);
480                                         if (flags & JFS_FLAG_ESCAPE) {
481                                                 *((__be32 *)bh->b_data) =
482                                                 cpu_to_be32(JFS_MAGIC_NUMBER);
483                                         }
484
485                                         BUFFER_TRACE(nbh, "marking dirty");
486                                         set_buffer_uptodate(nbh);
487                                         mark_buffer_dirty(nbh);
488                                         BUFFER_TRACE(nbh, "marking uptodate");
489                                         ++info->nr_replays;
490                                         /* ll_rw_block(WRITE, 1, &nbh); */
491                                         unlock_buffer(nbh);
492                                         brelse(obh);
493                                         brelse(nbh);
494                                 }
495
496                         skip_write:
497                                 tagp += sizeof(journal_block_tag_t);
498                                 if (!(flags & JFS_FLAG_SAME_UUID))
499                                         tagp += 16;
500
501                                 if (flags & JFS_FLAG_LAST_TAG)
502                                         break;
503                         }
504
505                         brelse(bh);
506                         continue;
507
508                 case JFS_COMMIT_BLOCK:
509                         /* Found an expected commit block: not much to
510                          * do other than move on to the next sequence
511                          * number. */
512                         brelse(bh);
513                         next_commit_ID++;
514                         continue;
515
516                 case JFS_REVOKE_BLOCK:
517                         /* If we aren't in the REVOKE pass, then we can
518                          * just skip over this block. */
519                         if (pass != PASS_REVOKE) {
520                                 brelse(bh);
521                                 continue;
522                         }
523
524                         err = scan_revoke_records(journal, bh,
525                                                   next_commit_ID, info);
526                         brelse(bh);
527                         if (err)
528                                 goto failed;
529                         continue;
530
531                 default:
532                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
533                                   blocktype);
534                         goto done;
535                 }
536         }
537
538  done:
539         /* 
540          * We broke out of the log scan loop: either we came to the
541          * known end of the log or we found an unexpected block in the
542          * log.  If the latter happened, then we know that the "current"
543          * transaction marks the end of the valid log.
544          */
545
546         if (pass == PASS_SCAN)
547                 info->end_transaction = next_commit_ID;
548         else {
549                 /* It's really bad news if different passes end up at
550                  * different places (but possible due to IO errors). */
551                 if (info->end_transaction != next_commit_ID) {
552                         printk (KERN_ERR "JBD: recovery pass %d ended at "
553                                 "transaction %u, expected %u\n",
554                                 pass, next_commit_ID, info->end_transaction);
555                         if (!success)
556                                 success = -EIO;
557                 }
558         }
559
560         return success;
561
562  failed:
563         return err;
564 }
565
566
567 /* Scan a revoke record, marking all blocks mentioned as revoked. */
568
569 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 
570                                tid_t sequence, struct recovery_info *info)
571 {
572         journal_revoke_header_t *header;
573         int offset, max;
574
575         header = (journal_revoke_header_t *) bh->b_data;
576         offset = sizeof(journal_revoke_header_t);
577         max = be32_to_cpu(header->r_count);
578
579         while (offset < max) {
580                 unsigned long blocknr;
581                 int err;
582
583                 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
584                 offset += 4;
585                 err = journal_set_revoke(journal, blocknr, sequence);
586                 if (err)
587                         return err;
588                 ++info->nr_revokes;
589         }
590         return 0;
591 }