fs/jbd2/recovery.c

   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/crc32.h>
  24 #include <linux/blkdev.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 static 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 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 = jbd2_journal_bmap(journal, next, &blocknr);
  91
  92                 if (err) {
  93                         printk(KERN_ERR "JBD2: 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 long blocknr;
 137         struct buffer_head *bh;
 138
 139         *bhp = NULL;
 140
 141         if (offset >= journal->j_maxlen) {
 142                 printk(KERN_ERR "JBD2: corrupted journal superblock\n");
 143                 return -EIO;
 144         }
 145
 146         err = jbd2_journal_bmap(journal, offset, &blocknr);
 147
 148         if (err) {
 149                 printk(KERN_ERR "JBD2: 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 "JBD2: 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 static int jbd2_descr_block_csum_verify(journal_t *j,
 178                                         void *buf)
 179 {
 180         struct jbd2_journal_block_tail *tail;
 181         __u32 provided, calculated;
 182
 183         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 184                 return 1;
 185
 186         tail = (struct jbd2_journal_block_tail *)(buf + j->j_blocksize -
 187                         sizeof(struct jbd2_journal_block_tail));
 188         provided = tail->t_checksum;
 189         tail->t_checksum = 0;
 190         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 191         tail->t_checksum = provided;
 192
 193         provided = be32_to_cpu(provided);
 194         return provided == calculated;
 195 }
 196
 197 /*
 198  * Count the number of in-use tags in a journal descriptor block.
 199  */
 200
 201 static int count_tags(journal_t *journal, struct buffer_head *bh)
 202 {
 203         char *                  tagp;
 204         journal_block_tag_t *   tag;
 205         int                     nr = 0, size = journal->j_blocksize;
 206         int                     tag_bytes = journal_tag_bytes(journal);
 207
 208         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 209                 size -= sizeof(struct jbd2_journal_block_tail);
 210
 211         tagp = &bh->b_data[sizeof(journal_header_t)];
 212
 213         while ((tagp - bh->b_data + tag_bytes) <= size) {
 214                 tag = (journal_block_tag_t *) tagp;
 215
 216                 nr++;
 217                 tagp += tag_bytes;
 218                 if (!(tag->t_flags & cpu_to_be16(JBD2_FLAG_SAME_UUID)))
 219                         tagp += 16;
 220
 221                 if (tag->t_flags & cpu_to_be16(JBD2_FLAG_LAST_TAG))
 222                         break;
 223         }
 224
 225         return nr;
 226 }
 227
 228
 229 /* Make sure we wrap around the log correctly! */
 230 #define wrap(journal, var)                                              \
 231 do {                                                                    \
 232         if (var >= (journal)->j_last)                                   \
 233                 var -= ((journal)->j_last - (journal)->j_first);        \
 234 } while (0)
 235
 236 /**
 237  * jbd2_journal_recover - recovers a on-disk journal
 238  * @journal: the journal to recover
 239  *
 240  * The primary function for recovering the log contents when mounting a
 241  * journaled device.
 242  *
 243  * Recovery is done in three passes.  In the first pass, we look for the
 244  * end of the log.  In the second, we assemble the list of revoke
 245  * blocks.  In the third and final pass, we replay any un-revoked blocks
 246  * in the log.
 247  */
 248 int jbd2_journal_recover(journal_t *journal)
 249 {
 250         int                     err, err2;
 251         journal_superblock_t *  sb;
 252
 253         struct recovery_info    info;
 254
 255         memset(&info, 0, sizeof(info));
 256         sb = journal->j_superblock;
 257
 258         /*
 259          * The journal superblock's s_start field (the current log head)
 260          * is always zero if, and only if, the journal was cleanly
 261          * unmounted.
 262          */
 263
 264         if (!sb->s_start) {
 265                 jbd_debug(1, "No recovery required, last transaction %d\n",
 266                           be32_to_cpu(sb->s_sequence));
 267                 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1;
 268                 return 0;
 269         }
 270
 271         err = do_one_pass(journal, &info, PASS_SCAN);
 272         if (!err)
 273                 err = do_one_pass(journal, &info, PASS_REVOKE);
 274         if (!err)
 275                 err = do_one_pass(journal, &info, PASS_REPLAY);
 276
 277         jbd_debug(1, "JBD2: recovery, exit status %d, "
 278                   "recovered transactions %u to %u\n",
 279                   err, info.start_transaction, info.end_transaction);
 280         jbd_debug(1, "JBD2: Replayed %d and revoked %d/%d blocks\n",
 281                   info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
 282
 283         /* Restart the log at the next transaction ID, thus invalidating
 284          * any existing commit records in the log. */
 285         journal->j_transaction_sequence = ++info.end_transaction;
 286
 287         jbd2_journal_clear_revoke(journal);
 288         err2 = sync_blockdev(journal->j_fs_dev);
 289         if (!err)
 290                 err = err2;
 291         /* Make sure all replayed data is on permanent storage */
 292         if (journal->j_flags & JBD2_BARRIER) {
 293                 err2 = blkdev_issue_flush(journal->j_fs_dev, GFP_KERNEL, NULL);
 294                 if (!err)
 295                         err = err2;
 296         }
 297         return err;
 298 }
 299
 300 /**
 301  * jbd2_journal_skip_recovery - Start journal and wipe exiting records
 302  * @journal: journal to startup
 303  *
 304  * Locate any valid recovery information from the journal and set up the
 305  * journal structures in memory to ignore it (presumably because the
 306  * caller has evidence that it is out of date).
 307  * This function does'nt appear to be exorted..
 308  *
 309  * We perform one pass over the journal to allow us to tell the user how
 310  * much recovery information is being erased, and to let us initialise
 311  * the journal transaction sequence numbers to the next unused ID.
 312  */
 313 int jbd2_journal_skip_recovery(journal_t *journal)
 314 {
 315         int                     err;
 316
 317         struct recovery_info    info;
 318
 319         memset (&info, 0, sizeof(info));
 320
 321         err = do_one_pass(journal, &info, PASS_SCAN);
 322
 323         if (err) {
 324                 printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
 325                 ++journal->j_transaction_sequence;
 326         } else {
 327 #ifdef CONFIG_JBD2_DEBUG
 328                 int dropped = info.end_transaction -
 329                         be32_to_cpu(journal->j_superblock->s_sequence);
 330                 jbd_debug(1,
 331                           "JBD2: ignoring %d transaction%s from the journal.\n",
 332                           dropped, (dropped == 1) ? "" : "s");
 333 #endif
 334                 journal->j_transaction_sequence = ++info.end_transaction;
 335         }
 336
 337         journal->j_tail = 0;
 338         return err;
 339 }
 340
 341 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag)
 342 {
 343         unsigned long long block = be32_to_cpu(tag->t_blocknr);
 344         if (tag_bytes > JBD2_TAG_SIZE32)
 345                 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
 346         return block;
 347 }
 348
 349 /*
 350  * calc_chksums calculates the checksums for the blocks described in the
 351  * descriptor block.
 352  */
 353 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
 354                         unsigned long *next_log_block, __u32 *crc32_sum)
 355 {
 356         int i, num_blks, err;
 357         unsigned long io_block;
 358         struct buffer_head *obh;
 359
 360         num_blks = count_tags(journal, bh);
 361         /* Calculate checksum of the descriptor block. */
 362         *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
 363
 364         for (i = 0; i < num_blks; i++) {
 365                 io_block = (*next_log_block)++;
 366                 wrap(journal, *next_log_block);
 367                 err = jread(&obh, journal, io_block);
 368                 if (err) {
 369                         printk(KERN_ERR "JBD2: IO error %d recovering block "
 370                                 "%lu in log\n", err, io_block);
 371                         return 1;
 372                 } else {
 373                         *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
 374                                      obh->b_size);
 375                 }
 376                 put_bh(obh);
 377         }
 378         return 0;
 379 }
 380
 381 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
 382 {
 383         struct commit_header *h;
 384         __u32 provided, calculated;
 385
 386         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 387                 return 1;
 388
 389         h = buf;
 390         provided = h->h_chksum[0];
 391         h->h_chksum[0] = 0;
 392         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 393         h->h_chksum[0] = provided;
 394
 395         provided = be32_to_cpu(provided);
 396         return provided == calculated;
 397 }
 398
 399 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
 400                                       void *buf, __u32 sequence)
 401 {
 402         __u32 provided, calculated;
 403
 404         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 405                 return 1;
 406
 407         sequence = cpu_to_be32(sequence);
 408         calculated = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&sequence,
 409                                  sizeof(sequence));
 410         calculated = jbd2_chksum(j, calculated, buf, j->j_blocksize);
 411         provided = be32_to_cpu(tag->t_checksum);
 412
 413         return provided == cpu_to_be32(calculated);
 414 }
 415
 416 static int do_one_pass(journal_t *journal,
 417                         struct recovery_info *info, enum passtype pass)
 418 {
 419         unsigned int            first_commit_ID, next_commit_ID;
 420         unsigned long           next_log_block;
 421         int                     err, success = 0;
 422         journal_superblock_t *  sb;
 423         journal_header_t *      tmp;
 424         struct buffer_head *    bh;
 425         unsigned int            sequence;
 426         int                     blocktype;
 427         int                     tag_bytes = journal_tag_bytes(journal);
 428         __u32                   crc32_sum = ~0; /* Transactional Checksums */
 429         int                     descr_csum_size = 0;
 430         int                     block_error = 0;
 431
 432         /*
 433          * First thing is to establish what we expect to find in the log
 434          * (in terms of transaction IDs), and where (in terms of log
 435          * block offsets): query the superblock.
 436          */
 437
 438         sb = journal->j_superblock;
 439         next_commit_ID = be32_to_cpu(sb->s_sequence);
 440         next_log_block = be32_to_cpu(sb->s_start);
 441
 442         first_commit_ID = next_commit_ID;
 443         if (pass == PASS_SCAN)
 444                 info->start_transaction = first_commit_ID;
 445
 446         jbd_debug(1, "Starting recovery pass %d\n", pass);
 447
 448         /*
 449          * Now we walk through the log, transaction by transaction,
 450          * making sure that each transaction has a commit block in the
 451          * expected place.  Each complete transaction gets replayed back
 452          * into the main filesystem.
 453          */
 454
 455         while (1) {
 456                 int                     flags;
 457                 char *                  tagp;
 458                 journal_block_tag_t *   tag;
 459                 struct buffer_head *    obh;
 460                 struct buffer_head *    nbh;
 461
 462                 cond_resched();
 463
 464                 /* If we already know where to stop the log traversal,
 465                  * check right now that we haven't gone past the end of
 466                  * the log. */
 467
 468                 if (pass != PASS_SCAN)
 469                         if (tid_geq(next_commit_ID, info->end_transaction))
 470                                 break;
 471
 472                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 473                           next_commit_ID, next_log_block, journal->j_last);
 474
 475                 /* Skip over each chunk of the transaction looking
 476                  * either the next descriptor block or the final commit
 477                  * record. */
 478
 479                 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
 480                 err = jread(&bh, journal, next_log_block);
 481                 if (err)
 482                         goto failed;
 483
 484                 next_log_block++;
 485                 wrap(journal, next_log_block);
 486
 487                 /* What kind of buffer is it?
 488                  *
 489                  * If it is a descriptor block, check that it has the
 490                  * expected sequence number.  Otherwise, we're all done
 491                  * here. */
 492
 493                 tmp = (journal_header_t *)bh->b_data;
 494
 495                 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
 496                         brelse(bh);
 497                         break;
 498                 }
 499
 500                 blocktype = be32_to_cpu(tmp->h_blocktype);
 501                 sequence = be32_to_cpu(tmp->h_sequence);
 502                 jbd_debug(3, "Found magic %d, sequence %d\n",
 503                           blocktype, sequence);
 504
 505                 if (sequence != next_commit_ID) {
 506                         brelse(bh);
 507                         break;
 508                 }
 509
 510                 /* OK, we have a valid descriptor block which matches
 511                  * all of the sequence number checks.  What are we going
 512                  * to do with it?  That depends on the pass... */
 513
 514                 switch(blocktype) {
 515                 case JBD2_DESCRIPTOR_BLOCK:
 516                         /* Verify checksum first */
 517                         if (JBD2_HAS_INCOMPAT_FEATURE(journal,
 518                                         JBD2_FEATURE_INCOMPAT_CSUM_V2))
 519                                 descr_csum_size =
 520                                         sizeof(struct jbd2_journal_block_tail);
 521                         if (descr_csum_size > 0 &&
 522                             !jbd2_descr_block_csum_verify(journal,
 523                                                           bh->b_data)) {
 524                                 err = -EIO;
 525                                 brelse(bh);
 526                                 goto failed;
 527                         }
 528
 529                         /* If it is a valid descriptor block, replay it
 530                          * in pass REPLAY; if journal_checksums enabled, then
 531                          * calculate checksums in PASS_SCAN, otherwise,
 532                          * just skip over the blocks it describes. */
 533                         if (pass != PASS_REPLAY) {
 534                                 if (pass == PASS_SCAN &&
 535                                     JBD2_HAS_COMPAT_FEATURE(journal,
 536                                             JBD2_FEATURE_COMPAT_CHECKSUM) &&
 537                                     !info->end_transaction) {
 538                                         if (calc_chksums(journal, bh,
 539                                                         &next_log_block,
 540                                                         &crc32_sum)) {
 541                                                 put_bh(bh);
 542                                                 break;
 543                                         }
 544                                         put_bh(bh);
 545                                         continue;
 546                                 }
 547                                 next_log_block += count_tags(journal, bh);
 548                                 wrap(journal, next_log_block);
 549                                 put_bh(bh);
 550                                 continue;
 551                         }
 552
 553                         /* A descriptor block: we can now write all of
 554                          * the data blocks.  Yay, useful work is finally
 555                          * getting done here! */
 556
 557                         tagp = &bh->b_data[sizeof(journal_header_t)];
 558                         while ((tagp - bh->b_data + tag_bytes)
 559                                <= journal->j_blocksize - descr_csum_size) {
 560                                 unsigned long io_block;
 561
 562                                 tag = (journal_block_tag_t *) tagp;
 563                                 flags = be16_to_cpu(tag->t_flags);
 564
 565                                 io_block = next_log_block++;
 566                                 wrap(journal, next_log_block);
 567                                 err = jread(&obh, journal, io_block);
 568                                 if (err) {
 569                                         /* Recover what we can, but
 570                                          * report failure at the end. */
 571                                         success = err;
 572                                         printk(KERN_ERR
 573                                                 "JBD2: IO error %d recovering "
 574                                                 "block %ld in log\n",
 575                                                 err, io_block);
 576                                 } else {
 577                                         unsigned long long blocknr;
 578
 579                                         J_ASSERT(obh != NULL);
 580                                         blocknr = read_tag_block(tag_bytes,
 581                                                                  tag);
 582
 583                                         /* If the block has been
 584                                          * revoked, then we're all done
 585                                          * here. */
 586                                         if (jbd2_journal_test_revoke
 587                                             (journal, blocknr,
 588                                              next_commit_ID)) {
 589                                                 brelse(obh);
 590                                                 ++info->nr_revoke_hits;
 591                                                 goto skip_write;
 592                                         }
 593
 594                                         /* Look for block corruption */
 595                                         if (!jbd2_block_tag_csum_verify(
 596                                                 journal, tag, obh->b_data,
 597                                                 be32_to_cpu(tmp->h_sequence))) {
 598                                                 brelse(obh);
 599                                                 success = -EIO;
 600                                                 printk(KERN_ERR "JBD: Invalid "
 601                                                        "checksum recovering "
 602                                                        "block %llu in log\n",
 603                                                        blocknr);
 604                                                 block_error = 1;
 605                                                 goto skip_write;
 606                                         }
 607
 608                                         /* Find a buffer for the new
 609                                          * data being restored */
 610                                         nbh = __getblk(journal->j_fs_dev,
 611                                                         blocknr,
 612                                                         journal->j_blocksize);
 613                                         if (nbh == NULL) {
 614                                                 printk(KERN_ERR
 615                                                        "JBD2: Out of memory "
 616                                                        "during recovery.\n");
 617                                                 err = -ENOMEM;
 618                                                 brelse(bh);
 619                                                 brelse(obh);
 620                                                 goto failed;
 621                                         }
 622
 623                                         lock_buffer(nbh);
 624                                         memcpy(nbh->b_data, obh->b_data,
 625                                                         journal->j_blocksize);
 626                                         if (flags & JBD2_FLAG_ESCAPE) {
 627                                                 *((__be32 *)nbh->b_data) =
 628                                                 cpu_to_be32(JBD2_MAGIC_NUMBER);
 629                                         }
 630
 631                                         BUFFER_TRACE(nbh, "marking dirty");
 632                                         set_buffer_uptodate(nbh);
 633                                         mark_buffer_dirty(nbh);
 634                                         BUFFER_TRACE(nbh, "marking uptodate");
 635                                         ++info->nr_replays;
 636                                         /* ll_rw_block(WRITE, 1, &nbh); */
 637                                         unlock_buffer(nbh);
 638                                         brelse(obh);
 639                                         brelse(nbh);
 640                                 }
 641
 642                         skip_write:
 643                                 tagp += tag_bytes;
 644                                 if (!(flags & JBD2_FLAG_SAME_UUID))
 645                                         tagp += 16;
 646
 647                                 if (flags & JBD2_FLAG_LAST_TAG)
 648                                         break;
 649                         }
 650
 651                         brelse(bh);
 652                         continue;
 653
 654                 case JBD2_COMMIT_BLOCK:
 655                         /*     How to differentiate between interrupted commit
 656                          *               and journal corruption ?
 657                          *
 658                          * {nth transaction}
 659                          *        Checksum Verification Failed
 660                          *                       |
 661                          *               ____________________
 662                          *              |                    |
 663                          *      async_commit             sync_commit
 664                          *              |                    |
 665                          *              | GO TO NEXT    "Journal Corruption"
 666                          *              | TRANSACTION
 667                          *              |
 668                          * {(n+1)th transanction}
 669                          *              |
 670                          *       _______|______________
 671                          *      |                     |
 672                          * Commit block found   Commit block not found
 673                          *      |                     |
 674                          * "Journal Corruption"       |
 675                          *               _____________|_________
 676                          *              |                       |
 677                          *      nth trans corrupt       OR   nth trans
 678                          *      and (n+1)th interrupted     interrupted
 679                          *      before commit block
 680                          *      could reach the disk.
 681                          *      (Cannot find the difference in above
 682                          *       mentioned conditions. Hence assume
 683                          *       "Interrupted Commit".)
 684                          */
 685
 686                         /* Found an expected commit block: if checksums
 687                          * are present verify them in PASS_SCAN; else not
 688                          * much to do other than move on to the next sequence
 689                          * number. */
 690                         if (pass == PASS_SCAN &&
 691                             JBD2_HAS_COMPAT_FEATURE(journal,
 692                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
 693                                 int chksum_err, chksum_seen;
 694                                 struct commit_header *cbh =
 695                                         (struct commit_header *)bh->b_data;
 696                                 unsigned found_chksum =
 697                                         be32_to_cpu(cbh->h_chksum[0]);
 698
 699                                 chksum_err = chksum_seen = 0;
 700
 701                                 if (info->end_transaction) {
 702                                         journal->j_failed_commit =
 703                                                 info->end_transaction;
 704                                         brelse(bh);
 705                                         break;
 706                                 }
 707
 708                                 if (crc32_sum == found_chksum &&
 709                                     cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
 710                                     cbh->h_chksum_size ==
 711                                                 JBD2_CRC32_CHKSUM_SIZE)
 712                                        chksum_seen = 1;
 713                                 else if (!(cbh->h_chksum_type == 0 &&
 714                                              cbh->h_chksum_size == 0 &&
 715                                              found_chksum == 0 &&
 716                                              !chksum_seen))
 717                                 /*
 718                                  * If fs is mounted using an old kernel and then
 719                                  * kernel with journal_chksum is used then we
 720                                  * get a situation where the journal flag has
 721                                  * checksum flag set but checksums are not
 722                                  * present i.e chksum = 0, in the individual
 723                                  * commit blocks.
 724                                  * Hence to avoid checksum failures, in this
 725                                  * situation, this extra check is added.
 726                                  */
 727                                                 chksum_err = 1;
 728
 729                                 if (chksum_err) {
 730                                         info->end_transaction = next_commit_ID;
 731
 732                                         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 733                                            JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
 734                                                 journal->j_failed_commit =
 735                                                         next_commit_ID;
 736                                                 brelse(bh);
 737                                                 break;
 738                                         }
 739                                 }
 740                                 crc32_sum = ~0;
 741                         }
 742                         if (pass == PASS_SCAN &&
 743                             !jbd2_commit_block_csum_verify(journal,
 744                                                            bh->b_data)) {
 745                                 info->end_transaction = next_commit_ID;
 746
 747                                 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 748                                      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 749                                         journal->j_failed_commit =
 750                                                 next_commit_ID;
 751                                         brelse(bh);
 752                                         break;
 753                                 }
 754                         }
 755                         brelse(bh);
 756                         next_commit_ID++;
 757                         continue;
 758
 759                 case JBD2_REVOKE_BLOCK:
 760                         /* If we aren't in the REVOKE pass, then we can
 761                          * just skip over this block. */
 762                         if (pass != PASS_REVOKE) {
 763                                 brelse(bh);
 764                                 continue;
 765                         }
 766
 767                         err = scan_revoke_records(journal, bh,
 768                                                   next_commit_ID, info);
 769                         brelse(bh);
 770                         if (err)
 771                                 goto failed;
 772                         continue;
 773
 774                 default:
 775                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
 776                                   blocktype);
 777                         brelse(bh);
 778                         goto done;
 779                 }
 780         }
 781
 782  done:
 783         /*
 784          * We broke out of the log scan loop: either we came to the
 785          * known end of the log or we found an unexpected block in the
 786          * log.  If the latter happened, then we know that the "current"
 787          * transaction marks the end of the valid log.
 788          */
 789
 790         if (pass == PASS_SCAN) {
 791                 if (!info->end_transaction)
 792                         info->end_transaction = next_commit_ID;
 793         } else {
 794                 /* It's really bad news if different passes end up at
 795                  * different places (but possible due to IO errors). */
 796                 if (info->end_transaction != next_commit_ID) {
 797                         printk(KERN_ERR "JBD2: recovery pass %d ended at "
 798                                 "transaction %u, expected %u\n",
 799                                 pass, next_commit_ID, info->end_transaction);
 800                         if (!success)
 801                                 success = -EIO;
 802                 }
 803         }
 804         if (block_error && success == 0)
 805                 success = -EIO;
 806         return success;
 807
 808  failed:
 809         return err;
 810 }
 811
 812 static int jbd2_revoke_block_csum_verify(journal_t *j,
 813                                          void *buf)
 814 {
 815         struct jbd2_journal_revoke_tail *tail;
 816         __u32 provided, calculated;
 817
 818         if (!JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V2))
 819                 return 1;
 820
 821         tail = (struct jbd2_journal_revoke_tail *)(buf + j->j_blocksize -
 822                         sizeof(struct jbd2_journal_revoke_tail));
 823         provided = tail->r_checksum;
 824         tail->r_checksum = 0;
 825         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 826         tail->r_checksum = provided;
 827
 828         provided = be32_to_cpu(provided);
 829         return provided == calculated;
 830 }
 831
 832 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 833
 834 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 835                                tid_t sequence, struct recovery_info *info)
 836 {
 837         jbd2_journal_revoke_header_t *header;
 838         int offset, max;
 839         int record_len = 4;
 840
 841         header = (jbd2_journal_revoke_header_t *) bh->b_data;
 842         offset = sizeof(jbd2_journal_revoke_header_t);
 843         max = be32_to_cpu(header->r_count);
 844
 845         if (!jbd2_revoke_block_csum_verify(journal, header))
 846                 return -EINVAL;
 847
 848         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
 849                 record_len = 8;
 850
 851         while (offset + record_len <= max) {
 852                 unsigned long long blocknr;
 853                 int err;
 854
 855                 if (record_len == 4)
 856                         blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
 857                 else
 858                         blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
 859                 offset += record_len;
 860                 err = jbd2_journal_set_revoke(journal, blocknr, sequence);
 861                 if (err)
 862                         return err;
 863                 ++info->nr_revokes;
 864         }
 865         return 0;
 866 }