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         if (offset >= journal->j_maxlen) {
 142                 printk(KERN_ERR "JBD: corrupted journal superblock\n");
 143                 return -EIO;
 144         }
 145
 146         err = journal_bmap(journal, offset, &blocknr);
 147
 148         if (err) {
 149                 printk (KERN_ERR "JBD: bad block at offset %u\n",
 150                         offset);
 151                 return err;
 152         }
 153
 154         bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
 155         if (!bh)
 156                 return -ENOMEM;
 157
 158         if (!buffer_uptodate(bh)) {
 159                 /* If this is a brand new buffer, start readahead.
 160                    Otherwise, we assume we are already reading it.  */
 161                 if (!buffer_req(bh))
 162                         do_readahead(journal, offset);
 163                 wait_on_buffer(bh);
 164         }
 165
 166         if (!buffer_uptodate(bh)) {
 167                 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
 168                         offset);
 169                 brelse(bh);
 170                 return -EIO;
 171         }
 172
 173         *bhp = bh;
 174         return 0;
 175 }
 176
 177
 178 /*
 179  * Count the number of in-use tags in a journal descriptor block.
 180  */
 181
 182 static int count_tags(struct buffer_head *bh, int size)
 183 {
 184         char *                  tagp;
 185         journal_block_tag_t *   tag;
 186         int                     nr = 0;
 187
 188         tagp = &bh->b_data[sizeof(journal_header_t)];
 189
 190         while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
 191                 tag = (journal_block_tag_t *) tagp;
 192
 193                 nr++;
 194                 tagp += sizeof(journal_block_tag_t);
 195                 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
 196                         tagp += 16;
 197
 198                 if (tag->t_flags & htonl(JFS_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)                                              \
 208 do {                                                                    \
 209         if (var >= (journal)->j_last)                                   \
 210                 var -= ((journal)->j_last - (journal)->j_first);        \
 211 } while (0)
 212
 213 /**
 214  * int journal_recover(journal_t *journal) - 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 int journal_recover(journal_t *journal)
 221 {
 222 /*
 223  * Recovery is done in three passes.  In the first pass, we look for the
 224  * end of the log.  In the second, we assemble the list of revoke
 225  * blocks.  In the third and final pass, we replay any un-revoked blocks
 226  * in the log.
 227  */
 228
 229         int                     err;
 230         journal_superblock_t *  sb;
 231
 232         struct recovery_info    info;
 233
 234         memset(&info, 0, sizeof(info));
 235         sb = journal->j_superblock;
 236
 237         /*
 238          * The journal superblock's s_start field (the current log head)
 239          * is always zero if, and only if, the journal was cleanly
 240          * unmounted.
 241          */
 242
 243         if (!sb->s_start) {
 244                 jbd_debug(1, "No recovery required, last transaction %d\n",
 245                           ntohl(sb->s_sequence));
 246                 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
 247                 return 0;
 248         }
 249
 250
 251         err = do_one_pass(journal, &info, PASS_SCAN);
 252         if (!err)
 253                 err = do_one_pass(journal, &info, PASS_REVOKE);
 254         if (!err)
 255                 err = do_one_pass(journal, &info, PASS_REPLAY);
 256
 257         jbd_debug(0, "JBD: recovery, exit status %d, "
 258                   "recovered transactions %u to %u\n",
 259                   err, info.start_transaction, info.end_transaction);
 260         jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
 261                   info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
 262
 263         /* Restart the log at the next transaction ID, thus invalidating
 264          * any existing commit records in the log. */
 265         journal->j_transaction_sequence = ++info.end_transaction;
 266
 267         journal_clear_revoke(journal);
 268         fsync_no_super(journal->j_fs_dev);
 269         return err;
 270 }
 271
 272 /**
 273  * int journal_skip_recovery() - Start journal and wipe exiting records
 274  * @journal: journal to startup
 275  *
 276  * Locate any valid recovery information from the journal and set up the
 277  * journal structures in memory to ignore it (presumably because the
 278  * caller has evidence that it is out of date).
 279  * This function does'nt appear to be exorted..
 280  */
 281 int journal_skip_recovery(journal_t *journal)
 282 {
 283 /*
 284  * We perform one pass over the journal to allow us to tell the user how
 285  * much recovery information is being erased, and to let us initialise
 286  * the journal transaction sequence numbers to the next unused ID.
 287  */
 288
 289         int                     err;
 290         journal_superblock_t *  sb;
 291
 292         struct recovery_info    info;
 293
 294         memset (&info, 0, sizeof(info));
 295         sb = journal->j_superblock;
 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 - ntohl(sb->s_sequence);
 305 #endif
 306
 307                 jbd_debug(0,
 308                           "JBD: ignoring %d transaction%s from the journal.\n",
 309                           dropped, (dropped == 1) ? "" : "s");
 310                 journal->j_transaction_sequence = ++info.end_transaction;
 311         }
 312
 313         journal->j_tail = 0;
 314
 315         return err;
 316 }
 317
 318 static int do_one_pass(journal_t *journal,
 319                         struct recovery_info *info, enum passtype pass)
 320 {
 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
 331         /* Precompute the maximum metadata descriptors in a descriptor block */
 332         int                     MAX_BLOCKS_PER_DESC;
 333         MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
 334                                / sizeof(journal_block_tag_t));
 335
 336         /*
 337          * First thing is to establish what we expect to find in the log
 338          * (in terms of transaction IDs), and where (in terms of log
 339          * block offsets): query the superblock.
 340          */
 341
 342         sb = journal->j_superblock;
 343         next_commit_ID = ntohl(sb->s_sequence);
 344         next_log_block = ntohl(sb->s_start);
 345
 346         first_commit_ID = next_commit_ID;
 347         if (pass == PASS_SCAN)
 348                 info->start_transaction = first_commit_ID;
 349
 350         jbd_debug(1, "Starting recovery pass %d\n", pass);
 351
 352         /*
 353          * Now we walk through the log, transaction by transaction,
 354          * making sure that each transaction has a commit block in the
 355          * expected place.  Each complete transaction gets replayed back
 356          * into the main filesystem.
 357          */
 358
 359         while (1) {
 360                 int                     flags;
 361                 char *                  tagp;
 362                 journal_block_tag_t *   tag;
 363                 struct buffer_head *    obh;
 364                 struct buffer_head *    nbh;
 365
 366                 /* If we already know where to stop the log traversal,
 367                  * check right now that we haven't gone past the end of
 368                  * the log. */
 369
 370                 if (pass != PASS_SCAN)
 371                         if (tid_geq(next_commit_ID, info->end_transaction))
 372                                 break;
 373
 374                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 375                           next_commit_ID, next_log_block, journal->j_last);
 376
 377                 /* Skip over each chunk of the transaction looking
 378                  * either the next descriptor block or the final commit
 379                  * record. */
 380
 381                 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
 382                 err = jread(&bh, journal, next_log_block);
 383                 if (err)
 384                         goto failed;
 385
 386                 next_log_block++;
 387                 wrap(journal, next_log_block);
 388
 389                 /* What kind of buffer is it?
 390                  *
 391                  * If it is a descriptor block, check that it has the
 392                  * expected sequence number.  Otherwise, we're all done
 393                  * here. */
 394
 395                 tmp = (journal_header_t *)bh->b_data;
 396
 397                 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
 398                         brelse(bh);
 399                         break;
 400                 }
 401
 402                 blocktype = ntohl(tmp->h_blocktype);
 403                 sequence = ntohl(tmp->h_sequence);
 404                 jbd_debug(3, "Found magic %d, sequence %d\n",
 405                           blocktype, sequence);
 406
 407                 if (sequence != next_commit_ID) {
 408                         brelse(bh);
 409                         break;
 410                 }
 411
 412                 /* OK, we have a valid descriptor block which matches
 413                  * all of the sequence number checks.  What are we going
 414                  * to do with it?  That depends on the pass... */
 415
 416                 switch(blocktype) {
 417                 case JFS_DESCRIPTOR_BLOCK:
 418                         /* If it is a valid descriptor block, replay it
 419                          * in pass REPLAY; otherwise, just skip over the
 420                          * blocks it describes. */
 421                         if (pass != PASS_REPLAY) {
 422                                 next_log_block +=
 423                                         count_tags(bh, journal->j_blocksize);
 424                                 wrap(journal, next_log_block);
 425                                 brelse(bh);
 426                                 continue;
 427                         }
 428
 429                         /* A descriptor block: we can now write all of
 430                          * the data blocks.  Yay, useful work is finally
 431                          * getting done here! */
 432
 433                         tagp = &bh->b_data[sizeof(journal_header_t)];
 434                         while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
 435                                <= journal->j_blocksize) {
 436                                 unsigned long io_block;
 437
 438                                 tag = (journal_block_tag_t *) tagp;
 439                                 flags = ntohl(tag->t_flags);
 440
 441                                 io_block = next_log_block++;
 442                                 wrap(journal, next_log_block);
 443                                 err = jread(&obh, journal, io_block);
 444                                 if (err) {
 445                                         /* Recover what we can, but
 446                                          * report failure at the end. */
 447                                         success = err;
 448                                         printk (KERN_ERR
 449                                                 "JBD: IO error %d recovering "
 450                                                 "block %ld in log\n",
 451                                                 err, io_block);
 452                                 } else {
 453                                         unsigned long blocknr;
 454
 455                                         J_ASSERT(obh != NULL);
 456                                         blocknr = ntohl(tag->t_blocknr);
 457
 458                                         /* If the block has been
 459                                          * revoked, then we're all done
 460                                          * here. */
 461                                         if (journal_test_revoke
 462                                             (journal, blocknr,
 463                                              next_commit_ID)) {
 464                                                 brelse(obh);
 465                                                 ++info->nr_revoke_hits;
 466                                                 goto skip_write;
 467                                         }
 468
 469                                         /* Find a buffer for the new
 470                                          * data being restored */
 471                                         nbh = getblk(journal->j_fs_dev, blocknr,
 472                                                      journal->j_blocksize);
 473                                         if (nbh == NULL) {
 474                                                 printk(KERN_ERR
 475                                                        "JBD: Out of memory "
 476                                                        "during recovery.\n");
 477                                                 err = -ENOMEM;
 478                                                 brelse(bh);
 479                                                 brelse(obh);
 480                                                 goto failed;
 481                                         }
 482
 483                                         lock_buffer(nbh);
 484                                         memcpy(nbh->b_data, obh->b_data,
 485                                                         journal->j_blocksize);
 486                                         if (flags & JFS_FLAG_ESCAPE) {
 487                                                 *((unsigned int *)bh->b_data) =
 488                                                         htonl(JFS_MAGIC_NUMBER);
 489                                         }
 490
 491                                         BUFFER_TRACE(nbh, "marking dirty");
 492                                         mark_buffer_dirty(nbh);
 493                                         BUFFER_TRACE(nbh, "marking uptodate");
 494                                         mark_buffer_uptodate(nbh, 1);
 495                                         unlock_buffer(nbh);
 496                                         ++info->nr_replays;
 497                                         /* ll_rw_block(WRITE, 1, &nbh); */
 498                                         brelse(obh);
 499                                         brelse(nbh);
 500                                 }
 501
 502                         skip_write:
 503                                 tagp += sizeof(journal_block_tag_t);
 504                                 if (!(flags & JFS_FLAG_SAME_UUID))
 505                                         tagp += 16;
 506
 507                                 if (flags & JFS_FLAG_LAST_TAG)
 508                                         break;
 509                         }
 510
 511                         brelse(bh);
 512                         continue;
 513
 514                 case JFS_COMMIT_BLOCK:
 515                         /* Found an expected commit block: not much to
 516                          * do other than move on to the next sequence
 517                          * number. */
 518                         brelse(bh);
 519                         next_commit_ID++;
 520                         continue;
 521
 522                 case JFS_REVOKE_BLOCK:
 523                         /* If we aren't in the REVOKE pass, then we can
 524                          * just skip over this block. */
 525                         if (pass != PASS_REVOKE) {
 526                                 brelse(bh);
 527                                 continue;
 528                         }
 529
 530                         err = scan_revoke_records(journal, bh,
 531                                                   next_commit_ID, info);
 532                         brelse(bh);
 533                         if (err)
 534                                 goto failed;
 535                         continue;
 536
 537                 default:
 538                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
 539                                   blocktype);
 540                         brelse(bh);
 541                         goto done;
 542                 }
 543         }
 544
 545  done:
 546         /*
 547          * We broke out of the log scan loop: either we came to the
 548          * known end of the log or we found an unexpected block in the
 549          * log.  If the latter happened, then we know that the "current"
 550          * transaction marks the end of the valid log.
 551          */
 552
 553         if (pass == PASS_SCAN)
 554                 info->end_transaction = next_commit_ID;
 555         else {
 556                 /* It's really bad news if different passes end up at
 557                  * different places (but possible due to IO errors). */
 558                 if (info->end_transaction != next_commit_ID) {
 559                         printk (KERN_ERR "JBD: recovery pass %d ended at "
 560                                 "transaction %u, expected %u\n",
 561                                 pass, next_commit_ID, info->end_transaction);
 562                         if (!success)
 563                                 success = -EIO;
 564                 }
 565         }
 566
 567         return success;
 568
 569  failed:
 570         return err;
 571 }
 572
 573
 574 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 575
 576 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 577                                tid_t sequence, struct recovery_info *info)
 578 {
 579         journal_revoke_header_t *header;
 580         int offset, max;
 581
 582         header = (journal_revoke_header_t *) bh->b_data;
 583         offset = sizeof(journal_revoke_header_t);
 584         max = ntohl(header->r_count);
 585
 586         while (offset < max) {
 587                 unsigned long blocknr;
 588                 int err;
 589
 590                 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
 591                 offset += 4;
 592                 err = journal_set_revoke(journal, blocknr, sequence);
 593                 if (err)
 594                         return err;
 595                 ++info->nr_revokes;
 596         }
 597         return 0;
 598 }