kernel/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         __be32 provided;
 182         __u32 calculated;
 183
 184         if (!jbd2_journal_has_csum_v2or3(j))
 185                 return 1;
 186
 187         tail = (struct jbd2_journal_block_tail *)(buf + j->j_blocksize -
 188                         sizeof(struct jbd2_journal_block_tail));
 189         provided = tail->t_checksum;
 190         tail->t_checksum = 0;
 191         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 192         tail->t_checksum = provided;
 193
 194         return provided == cpu_to_be32(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_journal_has_csum_v2or3(journal))
 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(journal_t *journal,
 342                                                 journal_block_tag_t *tag)
 343 {
 344         unsigned long long block = be32_to_cpu(tag->t_blocknr);
 345         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
 346                 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32;
 347         return block;
 348 }
 349
 350 /*
 351  * calc_chksums calculates the checksums for the blocks described in the
 352  * descriptor block.
 353  */
 354 static int calc_chksums(journal_t *journal, struct buffer_head *bh,
 355                         unsigned long *next_log_block, __u32 *crc32_sum)
 356 {
 357         int i, num_blks, err;
 358         unsigned long io_block;
 359         struct buffer_head *obh;
 360
 361         num_blks = count_tags(journal, bh);
 362         /* Calculate checksum of the descriptor block. */
 363         *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
 364
 365         for (i = 0; i < num_blks; i++) {
 366                 io_block = (*next_log_block)++;
 367                 wrap(journal, *next_log_block);
 368                 err = jread(&obh, journal, io_block);
 369                 if (err) {
 370                         printk(KERN_ERR "JBD2: IO error %d recovering block "
 371                                 "%lu in log\n", err, io_block);
 372                         return 1;
 373                 } else {
 374                         *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
 375                                      obh->b_size);
 376                 }
 377                 put_bh(obh);
 378         }
 379         return 0;
 380 }
 381
 382 static int jbd2_commit_block_csum_verify(journal_t *j, void *buf)
 383 {
 384         struct commit_header *h;
 385         __be32 provided;
 386         __u32 calculated;
 387
 388         if (!jbd2_journal_has_csum_v2or3(j))
 389                 return 1;
 390
 391         h = buf;
 392         provided = h->h_chksum[0];
 393         h->h_chksum[0] = 0;
 394         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 395         h->h_chksum[0] = provided;
 396
 397         return provided == cpu_to_be32(calculated);
 398 }
 399
 400 static int jbd2_block_tag_csum_verify(journal_t *j, journal_block_tag_t *tag,
 401                                       void *buf, __u32 sequence)
 402 {
 403         journal_block_tag3_t *tag3 = (journal_block_tag3_t *)tag;
 404         __u32 csum32;
 405         __be32 seq;
 406
 407         if (!jbd2_journal_has_csum_v2or3(j))
 408                 return 1;
 409
 410         seq = cpu_to_be32(sequence);
 411         csum32 = jbd2_chksum(j, j->j_csum_seed, (__u8 *)&seq, sizeof(seq));
 412         csum32 = jbd2_chksum(j, csum32, buf, j->j_blocksize);
 413
 414         if (JBD2_HAS_INCOMPAT_FEATURE(j, JBD2_FEATURE_INCOMPAT_CSUM_V3))
 415                 return tag3->t_checksum == cpu_to_be32(csum32);
 416         else
 417                 return tag->t_checksum == cpu_to_be16(csum32);
 418 }
 419
 420 static int do_one_pass(journal_t *journal,
 421                         struct recovery_info *info, enum passtype pass)
 422 {
 423         unsigned int            first_commit_ID, next_commit_ID;
 424         unsigned long           next_log_block;
 425         int                     err, success = 0;
 426         journal_superblock_t *  sb;
 427         journal_header_t *      tmp;
 428         struct buffer_head *    bh;
 429         unsigned int            sequence;
 430         int                     blocktype;
 431         int                     tag_bytes = journal_tag_bytes(journal);
 432         __u32                   crc32_sum = ~0; /* Transactional Checksums */
 433         int                     descr_csum_size = 0;
 434         int                     block_error = 0;
 435
 436         /*
 437          * First thing is to establish what we expect to find in the log
 438          * (in terms of transaction IDs), and where (in terms of log
 439          * block offsets): query the superblock.
 440          */
 441
 442         sb = journal->j_superblock;
 443         next_commit_ID = be32_to_cpu(sb->s_sequence);
 444         next_log_block = be32_to_cpu(sb->s_start);
 445
 446         first_commit_ID = next_commit_ID;
 447         if (pass == PASS_SCAN)
 448                 info->start_transaction = first_commit_ID;
 449
 450         jbd_debug(1, "Starting recovery pass %d\n", pass);
 451
 452         /*
 453          * Now we walk through the log, transaction by transaction,
 454          * making sure that each transaction has a commit block in the
 455          * expected place.  Each complete transaction gets replayed back
 456          * into the main filesystem.
 457          */
 458
 459         while (1) {
 460                 int                     flags;
 461                 char *                  tagp;
 462                 journal_block_tag_t *   tag;
 463                 struct buffer_head *    obh;
 464                 struct buffer_head *    nbh;
 465
 466                 cond_resched();
 467
 468                 /* If we already know where to stop the log traversal,
 469                  * check right now that we haven't gone past the end of
 470                  * the log. */
 471
 472                 if (pass != PASS_SCAN)
 473                         if (tid_geq(next_commit_ID, info->end_transaction))
 474                                 break;
 475
 476                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 477                           next_commit_ID, next_log_block, journal->j_last);
 478
 479                 /* Skip over each chunk of the transaction looking
 480                  * either the next descriptor block or the final commit
 481                  * record. */
 482
 483                 jbd_debug(3, "JBD2: checking block %ld\n", next_log_block);
 484                 err = jread(&bh, journal, next_log_block);
 485                 if (err)
 486                         goto failed;
 487
 488                 next_log_block++;
 489                 wrap(journal, next_log_block);
 490
 491                 /* What kind of buffer is it?
 492                  *
 493                  * If it is a descriptor block, check that it has the
 494                  * expected sequence number.  Otherwise, we're all done
 495                  * here. */
 496
 497                 tmp = (journal_header_t *)bh->b_data;
 498
 499                 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) {
 500                         brelse(bh);
 501                         break;
 502                 }
 503
 504                 blocktype = be32_to_cpu(tmp->h_blocktype);
 505                 sequence = be32_to_cpu(tmp->h_sequence);
 506                 jbd_debug(3, "Found magic %d, sequence %d\n",
 507                           blocktype, sequence);
 508
 509                 if (sequence != next_commit_ID) {
 510                         brelse(bh);
 511                         break;
 512                 }
 513
 514                 /* OK, we have a valid descriptor block which matches
 515                  * all of the sequence number checks.  What are we going
 516                  * to do with it?  That depends on the pass... */
 517
 518                 switch(blocktype) {
 519                 case JBD2_DESCRIPTOR_BLOCK:
 520                         /* Verify checksum first */
 521                         if (jbd2_journal_has_csum_v2or3(journal))
 522                                 descr_csum_size =
 523                                         sizeof(struct jbd2_journal_block_tail);
 524                         if (descr_csum_size > 0 &&
 525                             !jbd2_descr_block_csum_verify(journal,
 526                                                           bh->b_data)) {
 527                                 printk(KERN_ERR "JBD2: Invalid checksum "
 528                                        "recovering block %lu in log\n",
 529                                        next_log_block);
 530                                 err = -EIO;
 531                                 brelse(bh);
 532                                 goto failed;
 533                         }
 534
 535                         /* If it is a valid descriptor block, replay it
 536                          * in pass REPLAY; if journal_checksums enabled, then
 537                          * calculate checksums in PASS_SCAN, otherwise,
 538                          * just skip over the blocks it describes. */
 539                         if (pass != PASS_REPLAY) {
 540                                 if (pass == PASS_SCAN &&
 541                                     JBD2_HAS_COMPAT_FEATURE(journal,
 542                                             JBD2_FEATURE_COMPAT_CHECKSUM) &&
 543                                     !info->end_transaction) {
 544                                         if (calc_chksums(journal, bh,
 545                                                         &next_log_block,
 546                                                         &crc32_sum)) {
 547                                                 put_bh(bh);
 548                                                 break;
 549                                         }
 550                                         put_bh(bh);
 551                                         continue;
 552                                 }
 553                                 next_log_block += count_tags(journal, bh);
 554                                 wrap(journal, next_log_block);
 555                                 put_bh(bh);
 556                                 continue;
 557                         }
 558
 559                         /* A descriptor block: we can now write all of
 560                          * the data blocks.  Yay, useful work is finally
 561                          * getting done here! */
 562
 563                         tagp = &bh->b_data[sizeof(journal_header_t)];
 564                         while ((tagp - bh->b_data + tag_bytes)
 565                                <= journal->j_blocksize - descr_csum_size) {
 566                                 unsigned long io_block;
 567
 568                                 tag = (journal_block_tag_t *) tagp;
 569                                 flags = be16_to_cpu(tag->t_flags);
 570
 571                                 io_block = next_log_block++;
 572                                 wrap(journal, next_log_block);
 573                                 err = jread(&obh, journal, io_block);
 574                                 if (err) {
 575                                         /* Recover what we can, but
 576                                          * report failure at the end. */
 577                                         success = err;
 578                                         printk(KERN_ERR
 579                                                 "JBD2: IO error %d recovering "
 580                                                 "block %ld in log\n",
 581                                                 err, io_block);
 582                                 } else {
 583                                         unsigned long long blocknr;
 584
 585                                         J_ASSERT(obh != NULL);
 586                                         blocknr = read_tag_block(journal,
 587                                                                  tag);
 588
 589                                         /* If the block has been
 590                                          * revoked, then we're all done
 591                                          * here. */
 592                                         if (jbd2_journal_test_revoke
 593                                             (journal, blocknr,
 594                                              next_commit_ID)) {
 595                                                 brelse(obh);
 596                                                 ++info->nr_revoke_hits;
 597                                                 goto skip_write;
 598                                         }
 599
 600                                         /* Look for block corruption */
 601                                         if (!jbd2_block_tag_csum_verify(
 602                                                 journal, tag, obh->b_data,
 603                                                 be32_to_cpu(tmp->h_sequence))) {
 604                                                 brelse(obh);
 605                                                 success = -EIO;
 606                                                 printk(KERN_ERR "JBD2: Invalid "
 607                                                        "checksum recovering "
 608                                                        "block %llu in log\n",
 609                                                        blocknr);
 610                                                 block_error = 1;
 611                                                 goto skip_write;
 612                                         }
 613
 614                                         /* Find a buffer for the new
 615                                          * data being restored */
 616                                         nbh = __getblk(journal->j_fs_dev,
 617                                                         blocknr,
 618                                                         journal->j_blocksize);
 619                                         if (nbh == NULL) {
 620                                                 printk(KERN_ERR
 621                                                        "JBD2: Out of memory "
 622                                                        "during recovery.\n");
 623                                                 err = -ENOMEM;
 624                                                 brelse(bh);
 625                                                 brelse(obh);
 626                                                 goto failed;
 627                                         }
 628
 629                                         lock_buffer(nbh);
 630                                         memcpy(nbh->b_data, obh->b_data,
 631                                                         journal->j_blocksize);
 632                                         if (flags & JBD2_FLAG_ESCAPE) {
 633                                                 *((__be32 *)nbh->b_data) =
 634                                                 cpu_to_be32(JBD2_MAGIC_NUMBER);
 635                                         }
 636
 637                                         BUFFER_TRACE(nbh, "marking dirty");
 638                                         set_buffer_uptodate(nbh);
 639                                         mark_buffer_dirty(nbh);
 640                                         BUFFER_TRACE(nbh, "marking uptodate");
 641                                         ++info->nr_replays;
 642                                         /* ll_rw_block(WRITE, 1, &nbh); */
 643                                         unlock_buffer(nbh);
 644                                         brelse(obh);
 645                                         brelse(nbh);
 646                                 }
 647
 648                         skip_write:
 649                                 tagp += tag_bytes;
 650                                 if (!(flags & JBD2_FLAG_SAME_UUID))
 651                                         tagp += 16;
 652
 653                                 if (flags & JBD2_FLAG_LAST_TAG)
 654                                         break;
 655                         }
 656
 657                         brelse(bh);
 658                         continue;
 659
 660                 case JBD2_COMMIT_BLOCK:
 661                         /*     How to differentiate between interrupted commit
 662                          *               and journal corruption ?
 663                          *
 664                          * {nth transaction}
 665                          *        Checksum Verification Failed
 666                          *                       |
 667                          *               ____________________
 668                          *              |                    |
 669                          *      async_commit             sync_commit
 670                          *              |                    |
 671                          *              | GO TO NEXT    "Journal Corruption"
 672                          *              | TRANSACTION
 673                          *              |
 674                          * {(n+1)th transanction}
 675                          *              |
 676                          *       _______|______________
 677                          *      |                     |
 678                          * Commit block found   Commit block not found
 679                          *      |                     |
 680                          * "Journal Corruption"       |
 681                          *               _____________|_________
 682                          *              |                       |
 683                          *      nth trans corrupt       OR   nth trans
 684                          *      and (n+1)th interrupted     interrupted
 685                          *      before commit block
 686                          *      could reach the disk.
 687                          *      (Cannot find the difference in above
 688                          *       mentioned conditions. Hence assume
 689                          *       "Interrupted Commit".)
 690                          */
 691
 692                         /* Found an expected commit block: if checksums
 693                          * are present verify them in PASS_SCAN; else not
 694                          * much to do other than move on to the next sequence
 695                          * number. */
 696                         if (pass == PASS_SCAN &&
 697                             JBD2_HAS_COMPAT_FEATURE(journal,
 698                                     JBD2_FEATURE_COMPAT_CHECKSUM)) {
 699                                 int chksum_err, chksum_seen;
 700                                 struct commit_header *cbh =
 701                                         (struct commit_header *)bh->b_data;
 702                                 unsigned found_chksum =
 703                                         be32_to_cpu(cbh->h_chksum[0]);
 704
 705                                 chksum_err = chksum_seen = 0;
 706
 707                                 if (info->end_transaction) {
 708                                         journal->j_failed_commit =
 709                                                 info->end_transaction;
 710                                         brelse(bh);
 711                                         break;
 712                                 }
 713
 714                                 if (crc32_sum == found_chksum &&
 715                                     cbh->h_chksum_type == JBD2_CRC32_CHKSUM &&
 716                                     cbh->h_chksum_size ==
 717                                                 JBD2_CRC32_CHKSUM_SIZE)
 718                                        chksum_seen = 1;
 719                                 else if (!(cbh->h_chksum_type == 0 &&
 720                                              cbh->h_chksum_size == 0 &&
 721                                              found_chksum == 0 &&
 722                                              !chksum_seen))
 723                                 /*
 724                                  * If fs is mounted using an old kernel and then
 725                                  * kernel with journal_chksum is used then we
 726                                  * get a situation where the journal flag has
 727                                  * checksum flag set but checksums are not
 728                                  * present i.e chksum = 0, in the individual
 729                                  * commit blocks.
 730                                  * Hence to avoid checksum failures, in this
 731                                  * situation, this extra check is added.
 732                                  */
 733                                                 chksum_err = 1;
 734
 735                                 if (chksum_err) {
 736                                         info->end_transaction = next_commit_ID;
 737
 738                                         if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 739                                            JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){
 740                                                 journal->j_failed_commit =
 741                                                         next_commit_ID;
 742                                                 brelse(bh);
 743                                                 break;
 744                                         }
 745                                 }
 746                                 crc32_sum = ~0;
 747                         }
 748                         if (pass == PASS_SCAN &&
 749                             !jbd2_commit_block_csum_verify(journal,
 750                                                            bh->b_data)) {
 751                                 info->end_transaction = next_commit_ID;
 752
 753                                 if (!JBD2_HAS_INCOMPAT_FEATURE(journal,
 754                                      JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
 755                                         journal->j_failed_commit =
 756                                                 next_commit_ID;
 757                                         brelse(bh);
 758                                         break;
 759                                 }
 760                         }
 761                         brelse(bh);
 762                         next_commit_ID++;
 763                         continue;
 764
 765                 case JBD2_REVOKE_BLOCK:
 766                         /* If we aren't in the REVOKE pass, then we can
 767                          * just skip over this block. */
 768                         if (pass != PASS_REVOKE) {
 769                                 brelse(bh);
 770                                 continue;
 771                         }
 772
 773                         err = scan_revoke_records(journal, bh,
 774                                                   next_commit_ID, info);
 775                         brelse(bh);
 776                         if (err)
 777                                 goto failed;
 778                         continue;
 779
 780                 default:
 781                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
 782                                   blocktype);
 783                         brelse(bh);
 784                         goto done;
 785                 }
 786         }
 787
 788  done:
 789         /*
 790          * We broke out of the log scan loop: either we came to the
 791          * known end of the log or we found an unexpected block in the
 792          * log.  If the latter happened, then we know that the "current"
 793          * transaction marks the end of the valid log.
 794          */
 795
 796         if (pass == PASS_SCAN) {
 797                 if (!info->end_transaction)
 798                         info->end_transaction = next_commit_ID;
 799         } else {
 800                 /* It's really bad news if different passes end up at
 801                  * different places (but possible due to IO errors). */
 802                 if (info->end_transaction != next_commit_ID) {
 803                         printk(KERN_ERR "JBD2: recovery pass %d ended at "
 804                                 "transaction %u, expected %u\n",
 805                                 pass, next_commit_ID, info->end_transaction);
 806                         if (!success)
 807                                 success = -EIO;
 808                 }
 809         }
 810         if (block_error && success == 0)
 811                 success = -EIO;
 812         return success;
 813
 814  failed:
 815         return err;
 816 }
 817
 818 static int jbd2_revoke_block_csum_verify(journal_t *j,
 819                                          void *buf)
 820 {
 821         struct jbd2_journal_revoke_tail *tail;
 822         __be32 provided;
 823         __u32 calculated;
 824
 825         if (!jbd2_journal_has_csum_v2or3(j))
 826                 return 1;
 827
 828         tail = (struct jbd2_journal_revoke_tail *)(buf + j->j_blocksize -
 829                         sizeof(struct jbd2_journal_revoke_tail));
 830         provided = tail->r_checksum;
 831         tail->r_checksum = 0;
 832         calculated = jbd2_chksum(j, j->j_csum_seed, buf, j->j_blocksize);
 833         tail->r_checksum = provided;
 834
 835         return provided == cpu_to_be32(calculated);
 836 }
 837
 838 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 839
 840 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 841                                tid_t sequence, struct recovery_info *info)
 842 {
 843         jbd2_journal_revoke_header_t *header;
 844         int offset, max;
 845         int csum_size = 0;
 846         __u32 rcount;
 847         int record_len = 4;
 848
 849         header = (jbd2_journal_revoke_header_t *) bh->b_data;
 850         offset = sizeof(jbd2_journal_revoke_header_t);
 851         rcount = be32_to_cpu(header->r_count);
 852
 853         if (!jbd2_revoke_block_csum_verify(journal, header))
 854                 return -EINVAL;
 855
 856         if (jbd2_journal_has_csum_v2or3(journal))
 857                 csum_size = sizeof(struct jbd2_journal_revoke_tail);
 858         if (rcount > journal->j_blocksize - csum_size)
 859                 return -EINVAL;
 860         max = rcount;
 861
 862         if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT))
 863                 record_len = 8;
 864
 865         while (offset + record_len <= max) {
 866                 unsigned long long blocknr;
 867                 int err;
 868
 869                 if (record_len == 4)
 870                         blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset)));
 871                 else
 872                         blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset)));
 873                 offset += record_len;
 874                 err = jbd2_journal_set_revoke(journal, blocknr, sequence);
 875                 if (err)
 876                         return err;
 877                 ++info->nr_revokes;
 878         }
 879         return 0;
 880 }