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/sched.h>
  20 #include <linux/fs.h>
  21 #include <linux/jbd.h>
  22 #include <linux/errno.h>
  23 #include <linux/slab.h>
  24 #include <linux/locks.h>
  25 #endif
  26
  27 /*
  28  * Maintain information about the progress of the recovery job, so that
  29  * the different passes can carry information between them.
  30  */
  31 struct recovery_info
  32 {
  33         tid_t           start_transaction;
  34         tid_t           end_transaction;
  35
  36         int             nr_replays;
  37         int             nr_revokes;
  38         int             nr_revoke_hits;
  39 };
  40
  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 *);
  46
  47 #ifdef __KERNEL__
  48
  49 /* Release readahead buffers after use */
  50 void journal_brelse_array(struct buffer_head *b[], int n)
  51 {
  52         while (--n >= 0)
  53                 brelse (b[n]);
  54 }
  55
  56
  57 /*
  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.
  63  *
  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
  66  * fixed value.
  67  */
  68
  69 #define MAXBUF 8
  70 static int do_readahead(journal_t *journal, unsigned int start)
  71 {
  72         int err;
  73         unsigned int max, nbufs, next;
  74         unsigned long blocknr;
  75         struct buffer_head *bh;
  76
  77         struct buffer_head * bufs[MAXBUF];
  78
  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;
  83
  84         /* Do the readahead itself.  We'll submit MAXBUF buffer_heads at
  85          * a time to the block device IO layer. */
  86
  87         nbufs = 0;
  88
  89         for (next = start; next < max; next++) {
  90                 err = journal_bmap(journal, next, &blocknr);
  91
  92                 if (err) {
  93                         printk (KERN_ERR "JBD: bad block at offset %u\n",
  94                                 next);
  95                         goto failed;
  96                 }
  97
  98                 bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
  99                 if (!bh) {
 100                         err = -ENOMEM;
 101                         goto failed;
 102                 }
 103
 104                 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
 105                         bufs[nbufs++] = bh;
 106                         if (nbufs == MAXBUF) {
 107                                 ll_rw_block(READ, nbufs, bufs);
 108                                 journal_brelse_array(bufs, nbufs);
 109                                 nbufs = 0;
 110                         }
 111                 } else
 112                         brelse(bh);
 113         }
 114
 115         if (nbufs)
 116                 ll_rw_block(READ, nbufs, bufs);
 117         err = 0;
 118
 119 failed:
 120         if (nbufs)
 121                 journal_brelse_array(bufs, nbufs);
 122         return err;
 123 }
 124
 125 #endif /* __KERNEL__ */
 126
 127
 128 /*
 129  * Read a block from the journal
 130  */
 131
 132 static int jread(struct buffer_head **bhp, journal_t *journal,
 133                  unsigned int offset)
 134 {
 135         int err;
 136         unsigned long blocknr;
 137         struct buffer_head *bh;
 138
 139         *bhp = NULL;
 140
 141         J_ASSERT (offset < journal->j_maxlen);
 142
 143         err = journal_bmap(journal, offset, &blocknr);
 144
 145         if (err) {
 146                 printk (KERN_ERR "JBD: bad block at offset %u\n",
 147                         offset);
 148                 return err;
 149         }
 150
 151         bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
 152         if (!bh)
 153                 return -ENOMEM;
 154
 155         if (!buffer_uptodate(bh)) {
 156                 /* If this is a brand new buffer, start readahead.
 157                    Otherwise, we assume we are already reading it.  */
 158                 if (!buffer_req(bh))
 159                         do_readahead(journal, offset);
 160                 wait_on_buffer(bh);
 161         }
 162
 163         if (!buffer_uptodate(bh)) {
 164                 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
 165                         offset);
 166                 brelse(bh);
 167                 return -EIO;
 168         }
 169
 170         *bhp = bh;
 171         return 0;
 172 }
 173
 174
 175 /*
 176  * Count the number of in-use tags in a journal descriptor block.
 177  */
 178
 179 static int count_tags(struct buffer_head *bh, int size)
 180 {
 181         char *                  tagp;
 182         journal_block_tag_t *   tag;
 183         int                     nr = 0;
 184
 185         tagp = &bh->b_data[sizeof(journal_header_t)];
 186
 187         while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
 188                 tag = (journal_block_tag_t *) tagp;
 189
 190                 nr++;
 191                 tagp += sizeof(journal_block_tag_t);
 192                 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
 193                         tagp += 16;
 194
 195                 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
 196                         break;
 197         }
 198
 199         return nr;
 200 }
 201
 202
 203 /* Make sure we wrap around the log correctly! */
 204 #define wrap(journal, var)                                              \
 205 do {                                                                    \
 206         if (var >= (journal)->j_last)                                   \
 207                 var -= ((journal)->j_last - (journal)->j_first);        \
 208 } while (0)
 209
 210 /**
 211  * int journal_recover(journal_t *journal) - recovers a on-disk journal
 212  * @journal: the journal to recover
 213  *
 214  * The primary function for recovering the log contents when mounting a
 215  * journaled device.
 216  */
 217 int journal_recover(journal_t *journal)
 218 {
 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  */
 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                           ntohl(sb->s_sequence));
 243                 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
 244                 return 0;
 245         }
 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         journal_clear_revoke(journal);
 265         fsync_no_super(journal->j_fs_dev);
 266         return err;
 267 }
 268
 269 /**
 270  * int 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 int journal_skip_recovery(journal_t *journal)
 279 {
 280 /*
 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.
 284  */
 285
 286         int                     err;
 287         journal_superblock_t *  sb;
 288
 289         struct recovery_info    info;
 290
 291         memset (&info, 0, sizeof(info));
 292         sb = journal->j_superblock;
 293
 294         err = do_one_pass(journal, &info, PASS_SCAN);
 295
 296         if (err) {
 297                 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
 298                 ++journal->j_transaction_sequence;
 299         } else {
 300 #ifdef CONFIG_JBD_DEBUG
 301                 int dropped = info.end_transaction - ntohl(sb->s_sequence);
 302 #endif
 303
 304                 jbd_debug(0,
 305                           "JBD: ignoring %d transaction%s from the journal.\n",
 306                           dropped, (dropped == 1) ? "" : "s");
 307                 journal->j_transaction_sequence = ++info.end_transaction;
 308         }
 309
 310         journal->j_tail = 0;
 311
 312         return err;
 313 }
 314
 315 static int do_one_pass(journal_t *journal,
 316                         struct recovery_info *info, enum passtype pass)
 317 {
 318
 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;
 326         int                     blocktype;
 327
 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));
 332
 333         /*
 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.
 337          */
 338
 339         sb = journal->j_superblock;
 340         next_commit_ID = ntohl(sb->s_sequence);
 341         next_log_block = ntohl(sb->s_start);
 342
 343         first_commit_ID = next_commit_ID;
 344         if (pass == PASS_SCAN)
 345                 info->start_transaction = first_commit_ID;
 346
 347         jbd_debug(1, "Starting recovery pass %d\n", pass);
 348
 349         /*
 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.
 354          */
 355
 356         while (1) {
 357                 int                     flags;
 358                 char *                  tagp;
 359                 journal_block_tag_t *   tag;
 360                 struct buffer_head *    obh;
 361                 struct buffer_head *    nbh;
 362
 363                 /* If we already know where to stop the log traversal,
 364                  * check right now that we haven't gone past the end of
 365                  * the log. */
 366
 367                 if (pass != PASS_SCAN)
 368                         if (tid_geq(next_commit_ID, info->end_transaction))
 369                                 break;
 370
 371                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 372                           next_commit_ID, next_log_block, journal->j_last);
 373
 374                 /* Skip over each chunk of the transaction looking
 375                  * either the next descriptor block or the final commit
 376                  * record. */
 377
 378                 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
 379                 err = jread(&bh, journal, next_log_block);
 380                 if (err)
 381                         goto failed;
 382
 383                 next_log_block++;
 384                 wrap(journal, next_log_block);
 385
 386                 /* What kind of buffer is it?
 387                  *
 388                  * If it is a descriptor block, check that it has the
 389                  * expected sequence number.  Otherwise, we're all done
 390                  * here. */
 391
 392                 tmp = (journal_header_t *)bh->b_data;
 393
 394                 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
 395                         brelse(bh);
 396                         break;
 397                 }
 398
 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);
 403
 404                 if (sequence != next_commit_ID) {
 405                         brelse(bh);
 406                         break;
 407                 }
 408
 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... */
 412
 413                 switch(blocktype) {
 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) {
 419                                 next_log_block +=
 420                                         count_tags(bh, journal->j_blocksize);
 421                                 wrap(journal, next_log_block);
 422                                 brelse(bh);
 423                                 continue;
 424                         }
 425
 426                         /* A descriptor block: we can now write all of
 427                          * the data blocks.  Yay, useful work is finally
 428                          * getting done here! */
 429
 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;
 434
 435                                 tag = (journal_block_tag_t *) tagp;
 436                                 flags = ntohl(tag->t_flags);
 437
 438                                 io_block = next_log_block++;
 439                                 wrap(journal, next_log_block);
 440                                 err = jread(&obh, journal, io_block);
 441                                 if (err) {
 442                                         /* Recover what we can, but
 443                                          * report failure at the end. */
 444                                         success = err;
 445                                         printk (KERN_ERR
 446                                                 "JBD: IO error %d recovering "
 447                                                 "block %ld in log\n",
 448                                                 err, io_block);
 449                                 } else {
 450                                         unsigned long blocknr;
 451
 452                                         J_ASSERT(obh != NULL);
 453                                         blocknr = ntohl(tag->t_blocknr);
 454
 455                                         /* If the block has been
 456                                          * revoked, then we're all done
 457                                          * here. */
 458                                         if (journal_test_revoke
 459                                             (journal, blocknr,
 460                                              next_commit_ID)) {
 461                                                 brelse(obh);
 462                                                 ++info->nr_revoke_hits;
 463                                                 goto skip_write;
 464                                         }
 465
 466                                         /* Find a buffer for the new
 467                                          * data being restored */
 468                                         nbh = getblk(journal->j_fs_dev, blocknr,
 469                                                      journal->j_blocksize);
 470                                         if (nbh == NULL) {
 471                                                 printk(KERN_ERR
 472                                                        "JBD: Out of memory "
 473                                                        "during recovery.\n");
 474                                                 err = -ENOMEM;
 475                                                 brelse(bh);
 476                                                 brelse(obh);
 477                                                 goto failed;
 478                                         }
 479
 480                                         lock_buffer(nbh);
 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);
 486                                         }
 487
 488                                         BUFFER_TRACE(nbh, "marking dirty");
 489                                         mark_buffer_dirty(nbh);
 490                                         BUFFER_TRACE(nbh, "marking uptodate");
 491                                         mark_buffer_uptodate(nbh, 1);
 492                                         unlock_buffer(nbh);
 493                                         ++info->nr_replays;
 494                                         /* ll_rw_block(WRITE, 1, &nbh); */
 495                                         brelse(obh);
 496                                         brelse(nbh);
 497                                 }
 498
 499                         skip_write:
 500                                 tagp += sizeof(journal_block_tag_t);
 501                                 if (!(flags & JFS_FLAG_SAME_UUID))
 502                                         tagp += 16;
 503
 504                                 if (flags & JFS_FLAG_LAST_TAG)
 505                                         break;
 506                         }
 507
 508                         brelse(bh);
 509                         continue;
 510
 511                 case JFS_COMMIT_BLOCK:
 512                         /* Found an expected commit block: not much to
 513                          * do other than move on to the next sequence
 514                          * number. */
 515                         brelse(bh);
 516                         next_commit_ID++;
 517                         continue;
 518
 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) {
 523                                 brelse(bh);
 524                                 continue;
 525                         }
 526
 527                         err = scan_revoke_records(journal, bh,
 528                                                   next_commit_ID, info);
 529                         brelse(bh);
 530                         if (err)
 531                                 goto failed;
 532                         continue;
 533
 534                 default:
 535                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
 536                                   blocktype);
 537                         brelse(bh);
 538                         goto done;
 539                 }
 540         }
 541
 542  done:
 543         /*
 544          * We broke out of the log scan loop: either we came to the
 545          * known end of the log or we found an unexpected block in the
 546          * log.  If the latter happened, then we know that the "current"
 547          * transaction marks the end of the valid log.
 548          */
 549
 550         if (pass == PASS_SCAN)
 551                 info->end_transaction = next_commit_ID;
 552         else {
 553                 /* It's really bad news if different passes end up at
 554                  * different places (but possible due to IO errors). */
 555                 if (info->end_transaction != next_commit_ID) {
 556                         printk (KERN_ERR "JBD: recovery pass %d ended at "
 557                                 "transaction %u, expected %u\n",
 558                                 pass, next_commit_ID, info->end_transaction);
 559                         if (!success)
 560                                 success = -EIO;
 561                 }
 562         }
 563
 564         return success;
 565
 566  failed:
 567         return err;
 568 }
 569
 570
 571 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 572
 573 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 574                                tid_t sequence, struct recovery_info *info)
 575 {
 576         journal_revoke_header_t *header;
 577         int offset, max;
 578
 579         header = (journal_revoke_header_t *) bh->b_data;
 580         offset = sizeof(journal_revoke_header_t);
 581         max = ntohl(header->r_count);
 582
 583         while (offset < max) {
 584                 unsigned long blocknr;
 585                 int err;
 586
 587                 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
 588                 offset += 4;
 589                 err = journal_set_revoke(journal, blocknr, sequence);
 590                 if (err)
 591                         return err;
 592                 ++info->nr_revokes;
 593         }
 594         return 0;
 595 }