e2fsck/recovery.c

   1 /*
   2  * linux/fs/recovery.c
   3  *
   4  * Written by Stephen C. Tweedie <sct@redhat.com>, 1999
   5  *
   6  * Copyright 1999-2000 Red Hat Software --- All Rights Reserved
   7  *
   8  * This file is part of the Linux kernel and is made available under
   9  * the terms of the GNU General Public License, version 2, or at your
  10  * option, any later version, incorporated herein by reference.
  11  *
  12  * Journal recovery routines for the generic filesystem journaling code;
  13  * part of the ext2fs journaling system.
  14  */
  15
  16 #ifndef __KERNEL__
  17 #include "jfs_user.h"
  18 #else
  19 #include <linux/sched.h>
  20 #include <linux/fs.h>
  21 #include <linux/jbd.h>
  22 #include <linux/errno.h>
  23 #include <linux/slab.h>
  24 #include <linux/locks.h>
  25 #endif
  26
  27 /*
  28  * Maintain information about the progress of the recovery job, so that
  29  * the different passes can carry information between them.
  30  */
  31 struct recovery_info
  32 {
  33         tid_t           start_transaction;
  34         tid_t           end_transaction;
  35
  36         int             nr_replays;
  37         int             nr_revokes;
  38         int             nr_revoke_hits;
  39 };
  40
  41 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
  42 static int do_one_pass(journal_t *journal,
  43                                 struct recovery_info *info, enum passtype pass);
  44 static int scan_revoke_records(journal_t *, struct buffer_head *,
  45                                 tid_t, struct recovery_info *);
  46
  47 #ifdef __KERNEL__
  48
  49 /* Release readahead buffers after use */
  50 void journal_brelse_array(struct buffer_head *b[], int n)
  51 {
  52         while (--n >= 0)
  53                 brelse (b[n]);
  54 }
  55
  56
  57 /*
  58  * When reading from the journal, we are going through the block device
  59  * layer directly and so there is no readahead being done for us.  We
  60  * need to implement any readahead ourselves if we want it to happen at
  61  * all.  Recovery is basically one long sequential read, so make sure we
  62  * do the IO in reasonably large chunks.
  63  *
  64  * This is not so critical that we need to be enormously clever about
  65  * the readahead size, though.  128K is a purely arbitrary, good-enough
  66  * fixed value.
  67  */
  68
  69 #define MAXBUF 8
  70 static int do_readahead(journal_t *journal, unsigned int start)
  71 {
  72         int err;
  73         unsigned int max, nbufs, next;
  74         unsigned long blocknr;
  75         struct buffer_head *bh;
  76
  77         struct buffer_head * bufs[MAXBUF];
  78
  79         /* Do up to 128K of readahead */
  80         max = start + (128 * 1024 / journal->j_blocksize);
  81         if (max > journal->j_maxlen)
  82                 max = journal->j_maxlen;
  83
  84         /* Do the readahead itself.  We'll submit MAXBUF buffer_heads at
  85          * a time to the block device IO layer. */
  86
  87         nbufs = 0;
  88
  89         for (next = start; next < max; next++) {
  90                 err = journal_bmap(journal, next, &blocknr);
  91
  92                 if (err) {
  93                         printk (KERN_ERR "JBD: bad block at offset %u\n",
  94                                 next);
  95                         goto failed;
  96                 }
  97
  98                 bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
  99                 if (!bh) {
 100                         err = -ENOMEM;
 101                         goto failed;
 102                 }
 103
 104                 if (!buffer_uptodate(bh) && !buffer_locked(bh)) {
 105                         bufs[nbufs++] = bh;
 106                         if (nbufs == MAXBUF) {
 107                                 ll_rw_block(READ, nbufs, bufs);
 108                                 journal_brelse_array(bufs, nbufs);
 109                                 nbufs = 0;
 110                         }
 111                 } else
 112                         brelse(bh);
 113         }
 114
 115         if (nbufs)
 116                 ll_rw_block(READ, nbufs, bufs);
 117         err = 0;
 118
 119 failed:
 120         if (nbufs)
 121                 journal_brelse_array(bufs, nbufs);
 122         return err;
 123 }
 124
 125 #endif /* __KERNEL__ */
 126
 127
 128 /*
 129  * Read a block from the journal
 130  */
 131
 132 static int jread(struct buffer_head **bhp, journal_t *journal,
 133                  unsigned int offset)
 134 {
 135         int err;
 136         unsigned long blocknr;
 137         struct buffer_head *bh;
 138
 139         *bhp = NULL;
 140
 141         J_ASSERT (offset < journal->j_maxlen);
 142
 143         err = journal_bmap(journal, offset, &blocknr);
 144
 145         if (err) {
 146                 printk (KERN_ERR "JBD: bad block at offset %u\n",
 147                         offset);
 148                 return err;
 149         }
 150
 151         bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
 152         if (!bh)
 153                 return -ENOMEM;
 154
 155         if (!buffer_uptodate(bh)) {
 156                 /* If this is a brand new buffer, start readahead.
 157                    Otherwise, we assume we are already reading it.  */
 158                 if (!buffer_req(bh))
 159                         do_readahead(journal, offset);
 160                 wait_on_buffer(bh);
 161         }
 162
 163         if (!buffer_uptodate(bh)) {
 164                 printk (KERN_ERR "JBD: Failed to read block at offset %u\n",
 165                         offset);
 166                 brelse(bh);
 167                 return -EIO;
 168         }
 169
 170         *bhp = bh;
 171         return 0;
 172 }
 173
 174
 175 /*
 176  * Count the number of in-use tags in a journal descriptor block.
 177  */
 178
 179 static int count_tags(struct buffer_head *bh, int size)
 180 {
 181         char *                  tagp;
 182         journal_block_tag_t *   tag;
 183         int                     nr = 0;
 184
 185         tagp = &bh->b_data[sizeof(journal_header_t)];
 186
 187         while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
 188                 tag = (journal_block_tag_t *) tagp;
 189
 190                 nr++;
 191                 tagp += sizeof(journal_block_tag_t);
 192                 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
 193                         tagp += 16;
 194
 195                 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
 196                         break;
 197         }
 198
 199         return nr;
 200 }
 201
 202
 203 /* Make sure we wrap around the log correctly! */
 204 #define wrap(journal, var)                                              \
 205 do {                                                                    \
 206         if (var >= (journal)->j_last)                                   \
 207                 var -= ((journal)->j_last - (journal)->j_first);        \
 208 } while (0)
 209
 210 /*
 211  * journal_recover
 212  *
 213  * The primary function for recovering the log contents when mounting a
 214  * journaled device.
 215  *
 216  * Recovery is done in three passes.  In the first pass, we look for the
 217  * end of the log.  In the second, we assemble the list of revoke
 218  * blocks.  In the third and final pass, we replay any un-revoked blocks
 219  * in the log.
 220  */
 221
 222 int journal_recover(journal_t *journal)
 223 {
 224         int                     err;
 225         journal_superblock_t *  sb;
 226
 227         struct recovery_info    info;
 228
 229         memset(&info, 0, sizeof(info));
 230         sb = journal->j_superblock;
 231
 232         /*
 233          * The journal superblock's s_start field (the current log head)
 234          * is always zero if, and only if, the journal was cleanly
 235          * unmounted.
 236          */
 237
 238         if (!sb->s_start) {
 239                 jbd_debug(1, "No recovery required, last transaction %d\n",
 240                           ntohl(sb->s_sequence));
 241                 journal->j_transaction_sequence = ntohl(sb->s_sequence) + 1;
 242                 return 0;
 243         }
 244
 245
 246         err = do_one_pass(journal, &info, PASS_SCAN);
 247         if (!err)
 248                 err = do_one_pass(journal, &info, PASS_REVOKE);
 249         if (!err)
 250                 err = do_one_pass(journal, &info, PASS_REPLAY);
 251
 252         jbd_debug(0, "JBD: recovery, exit status %d, "
 253                   "recovered transactions %u to %u\n",
 254                   err, info.start_transaction, info.end_transaction);
 255         jbd_debug(0, "JBD: Replayed %d and revoked %d/%d blocks\n",
 256                   info.nr_replays, info.nr_revoke_hits, info.nr_revokes);
 257
 258         /* Restart the log at the next transaction ID, thus invalidating
 259          * any existing commit records in the log. */
 260         journal->j_transaction_sequence = ++info.end_transaction;
 261
 262         journal_clear_revoke(journal);
 263         fsync_no_super(journal->j_fs_dev);
 264         return err;
 265 }
 266
 267 /*
 268  * journal_skip_recovery
 269  *
 270  * Locate any valid recovery information from the journal and set up the
 271  * journal structures in memory to ignore it (presumably because the
 272  * caller has evidence that it is out of date).
 273  *
 274  * We perform one pass over the journal to allow us to tell the user how
 275  * much recovery information is being erased, and to let us initialise
 276  * the journal transaction sequence numbers to the next unused ID.
 277  */
 278
 279 int journal_skip_recovery(journal_t *journal)
 280 {
 281         int                     err;
 282         journal_superblock_t *  sb;
 283
 284         struct recovery_info    info;
 285
 286         memset (&info, 0, sizeof(info));
 287         sb = journal->j_superblock;
 288
 289         err = do_one_pass(journal, &info, PASS_SCAN);
 290
 291         if (err) {
 292                 printk(KERN_ERR "JBD: error %d scanning journal\n", err);
 293                 ++journal->j_transaction_sequence;
 294         } else {
 295 #ifdef CONFIG_JBD_DEBUG
 296                 int dropped = info.end_transaction - ntohl(sb->s_sequence);
 297 #endif
 298
 299                 jbd_debug(0,
 300                           "JBD: ignoring %d transaction%s from the journal.\n",
 301                           dropped, (dropped == 1) ? "" : "s");
 302                 journal->j_transaction_sequence = ++info.end_transaction;
 303         }
 304
 305         journal->j_tail = 0;
 306
 307         return err;
 308 }
 309
 310 static int do_one_pass(journal_t *journal,
 311                         struct recovery_info *info, enum passtype pass)
 312 {
 313
 314         unsigned int            first_commit_ID, next_commit_ID;
 315         unsigned long           next_log_block;
 316         int                     err, success = 0;
 317         journal_superblock_t *  sb;
 318         journal_header_t *      tmp;
 319         struct buffer_head *    bh;
 320         unsigned int            sequence;
 321         int                     blocktype;
 322
 323         /* Precompute the maximum metadata descriptors in a descriptor block */
 324         int                     MAX_BLOCKS_PER_DESC;
 325         MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
 326                                / sizeof(journal_block_tag_t));
 327
 328         /*
 329          * First thing is to establish what we expect to find in the log
 330          * (in terms of transaction IDs), and where (in terms of log
 331          * block offsets): query the superblock.
 332          */
 333
 334         sb = journal->j_superblock;
 335         next_commit_ID = ntohl(sb->s_sequence);
 336         next_log_block = ntohl(sb->s_start);
 337
 338         first_commit_ID = next_commit_ID;
 339         if (pass == PASS_SCAN)
 340                 info->start_transaction = first_commit_ID;
 341
 342         jbd_debug(1, "Starting recovery pass %d\n", pass);
 343
 344         /*
 345          * Now we walk through the log, transaction by transaction,
 346          * making sure that each transaction has a commit block in the
 347          * expected place.  Each complete transaction gets replayed back
 348          * into the main filesystem.
 349          */
 350
 351         while (1) {
 352                 int                     flags;
 353                 char *                  tagp;
 354                 journal_block_tag_t *   tag;
 355                 struct buffer_head *    obh;
 356                 struct buffer_head *    nbh;
 357
 358                 /* If we already know where to stop the log traversal,
 359                  * check right now that we haven't gone past the end of
 360                  * the log. */
 361
 362                 if (pass != PASS_SCAN)
 363                         if (tid_geq(next_commit_ID, info->end_transaction))
 364                                 break;
 365
 366                 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 367                           next_commit_ID, next_log_block, journal->j_last);
 368
 369                 /* Skip over each chunk of the transaction looking
 370                  * either the next descriptor block or the final commit
 371                  * record. */
 372
 373                 jbd_debug(3, "JBD: checking block %ld\n", next_log_block);
 374                 err = jread(&bh, journal, next_log_block);
 375                 if (err)
 376                         goto failed;
 377
 378                 next_log_block++;
 379                 wrap(journal, next_log_block);
 380
 381                 /* What kind of buffer is it?
 382                  *
 383                  * If it is a descriptor block, check that it has the
 384                  * expected sequence number.  Otherwise, we're all done
 385                  * here. */
 386
 387                 tmp = (journal_header_t *)bh->b_data;
 388
 389                 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
 390                         brelse(bh);
 391                         break;
 392                 }
 393
 394                 blocktype = ntohl(tmp->h_blocktype);
 395                 sequence = ntohl(tmp->h_sequence);
 396                 jbd_debug(3, "Found magic %d, sequence %d\n",
 397                           blocktype, sequence);
 398
 399                 if (sequence != next_commit_ID) {
 400                         brelse(bh);
 401                         break;
 402                 }
 403
 404                 /* OK, we have a valid descriptor block which matches
 405                  * all of the sequence number checks.  What are we going
 406                  * to do with it?  That depends on the pass... */
 407
 408                 switch(blocktype) {
 409                 case JFS_DESCRIPTOR_BLOCK:
 410                         /* If it is a valid descriptor block, replay it
 411                          * in pass REPLAY; otherwise, just skip over the
 412                          * blocks it describes. */
 413                         if (pass != PASS_REPLAY) {
 414                                 next_log_block +=
 415                                         count_tags(bh, journal->j_blocksize);
 416                                 wrap(journal, next_log_block);
 417                                 brelse(bh);
 418                                 continue;
 419                         }
 420
 421                         /* A descriptor block: we can now write all of
 422                          * the data blocks.  Yay, useful work is finally
 423                          * getting done here! */
 424
 425                         tagp = &bh->b_data[sizeof(journal_header_t)];
 426                         while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
 427                                <= journal->j_blocksize) {
 428                                 unsigned long io_block;
 429
 430                                 tag = (journal_block_tag_t *) tagp;
 431                                 flags = ntohl(tag->t_flags);
 432
 433                                 io_block = next_log_block++;
 434                                 wrap(journal, next_log_block);
 435                                 err = jread(&obh, journal, io_block);
 436                                 if (err) {
 437                                         /* Recover what we can, but
 438                                          * report failure at the end. */
 439                                         success = err;
 440                                         printk (KERN_ERR
 441                                                 "JBD: IO error %d recovering "
 442                                                 "block %ld in log\n",
 443                                                 err, io_block);
 444                                 } else {
 445                                         unsigned long blocknr;
 446
 447                                         J_ASSERT(obh != NULL);
 448                                         blocknr = ntohl(tag->t_blocknr);
 449
 450                                         /* If the block has been
 451                                          * revoked, then we're all done
 452                                          * here. */
 453                                         if (journal_test_revoke
 454                                             (journal, blocknr,
 455                                              next_commit_ID)) {
 456                                                 brelse(obh);
 457                                                 ++info->nr_revoke_hits;
 458                                                 goto skip_write;
 459                                         }
 460
 461                                         /* Find a buffer for the new
 462                                          * data being restored */
 463                                         nbh = getblk(journal->j_fs_dev, blocknr,
 464                                                      journal->j_blocksize);
 465                                         if (nbh == NULL) {
 466                                                 printk(KERN_ERR
 467                                                        "JBD: Out of memory "
 468                                                        "during recovery.\n");
 469                                                 err = -ENOMEM;
 470                                                 brelse(bh);
 471                                                 brelse(obh);
 472                                                 goto failed;
 473                                         }
 474
 475                                         memcpy(nbh->b_data, obh->b_data,
 476                                                         journal->j_blocksize);
 477                                         if (flags & JFS_FLAG_ESCAPE) {
 478                                                 *((unsigned int *)bh->b_data) =
 479                                                         htonl(JFS_MAGIC_NUMBER);
 480                                         }
 481
 482                                         BUFFER_TRACE(nbh, "marking dirty");
 483                                         mark_buffer_dirty(nbh);
 484                                         BUFFER_TRACE(nbh, "marking uptodate");
 485                                         mark_buffer_uptodate(nbh, 1);
 486                                         ++info->nr_replays;
 487                                         /* ll_rw_block(WRITE, 1, &nbh); */
 488                                         brelse(obh);
 489                                         brelse(nbh);
 490                                 }
 491
 492                         skip_write:
 493                                 tagp += sizeof(journal_block_tag_t);
 494                                 if (!(flags & JFS_FLAG_SAME_UUID))
 495                                         tagp += 16;
 496
 497                                 if (flags & JFS_FLAG_LAST_TAG)
 498                                         break;
 499                         }
 500
 501                         brelse(bh);
 502                         continue;
 503
 504                 case JFS_COMMIT_BLOCK:
 505                         /* Found an expected commit block: not much to
 506                          * do other than move on to the next sequence
 507                          * number. */
 508                         brelse(bh);
 509                         next_commit_ID++;
 510                         continue;
 511
 512                 case JFS_REVOKE_BLOCK:
 513                         /* If we aren't in the REVOKE pass, then we can
 514                          * just skip over this block. */
 515                         if (pass != PASS_REVOKE) {
 516                                 brelse(bh);
 517                                 continue;
 518                         }
 519
 520                         err = scan_revoke_records(journal, bh,
 521                                                   next_commit_ID, info);
 522                         brelse(bh);
 523                         if (err)
 524                                 goto failed;
 525                         continue;
 526
 527                 default:
 528                         jbd_debug(3, "Unrecognised magic %d, end of scan.\n",
 529                                   blocktype);
 530                         goto done;
 531                 }
 532         }
 533
 534  done:
 535         /*
 536          * We broke out of the log scan loop: either we came to the
 537          * known end of the log or we found an unexpected block in the
 538          * log.  If the latter happened, then we know that the "current"
 539          * transaction marks the end of the valid log.
 540          */
 541
 542         if (pass == PASS_SCAN)
 543                 info->end_transaction = next_commit_ID;
 544         else {
 545                 /* It's really bad news if different passes end up at
 546                  * different places (but possible due to IO errors). */
 547                 if (info->end_transaction != next_commit_ID) {
 548                         printk (KERN_ERR "JBD: recovery pass %d ended at "
 549                                 "transaction %u, expected %u\n",
 550                                 pass, next_commit_ID, info->end_transaction);
 551                         if (!success)
 552                                 success = -EIO;
 553                 }
 554         }
 555
 556         return success;
 557
 558  failed:
 559         return err;
 560 }
 561
 562
 563 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 564
 565 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 566                                tid_t sequence, struct recovery_info *info)
 567 {
 568         journal_revoke_header_t *header;
 569         int offset, max;
 570
 571         header = (journal_revoke_header_t *) bh->b_data;
 572         offset = sizeof(journal_revoke_header_t);
 573         max = ntohl(header->r_count);
 574
 575         while (offset < max) {
 576                 unsigned long blocknr;
 577                 int err;
 578
 579                 blocknr = ntohl(* ((unsigned int *) (bh->b_data+offset)));
 580                 offset += 4;
 581                 err = journal_set_revoke(journal, blocknr, sequence);
 582                 if (err)
 583                         return err;
 584                 ++info->nr_revokes;
 585         }
 586         return 0;
 587 }