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/jfs.h>
  22 #include <linux/errno.h>
  23 #include <linux/malloc.h>
  24 #include <linux/locks.h>
  25 #include <linux/buffer.h>
  26 #endif
  27
  28 /*
  29  * Maintain information about the progress of the recovery job, so that
  30  * the different passes can carry information between them.
  31  */
  32 struct recovery_info
  33 {
  34         tid_t           start_transaction;
  35         tid_t           end_transaction;
  36
  37         int             nr_replays;
  38         int             nr_revokes;
  39         int             nr_revoke_hits;
  40 };
  41
  42 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
  43 static int do_one_pass(journal_t *, struct recovery_info *, enum passtype);
  44 static int scan_revoke_records(journal_t *, struct buffer_head *, tid_t, struct recovery_info *);
  45
  46 #ifdef __KERNEL__
  47
  48 /* Release readahead buffers after use */
  49 static void brelse_array(struct buffer_head *b[], int n)
  50 {
  51         while (--n >= 0)
  52                 brelse (b[n]);
  53 }
  54
  55
  56 /*
  57  * When reading from the journal, we are going through the block device
  58  * layer directly and so there is no readahead being done for us.  We
  59  * need to implement any readahead ourselves if we want it to happen at
  60  * all.  Recovery is basically one long sequential read, so make sure we
  61  * do the IO in reasonably large chunks.
  62  *
  63  * This is not so critical that we need to be enormously clever about
  64  * the readahead size, though.  128K is a purely arbitrary, good-enough
  65  * fixed value.
  66  */
  67
  68 static int do_readahead(journal_t *journal, unsigned int start)
  69 {
  70         int err;
  71         unsigned int max, nbufs, next, blocknr;
  72         struct buffer_head *bh;
  73
  74         #define MAXBUF 8
  75         struct buffer_head * bufs[MAXBUF];
  76
  77         /* Do up to 128K of readahead */
  78         max = start + (128 * 1024 / journal->j_blocksize);
  79         if (max > journal->j_maxlen)
  80                 max = journal->j_maxlen;
  81
  82         /* Do the readahead itself.  We'll submit MAXBUF buffer_heads at
  83          * a time to the block device IO layer. */
  84
  85         nbufs = 0;
  86
  87         for (next = start; next < max; next++) {
  88                 blocknr = next;
  89                 if (journal->j_inode)
  90                         blocknr = bmap(journal->j_inode, next);
  91                 if (!blocknr) {
  92                         printk (KERN_ERR "JFS: bad block at offset %u\n",
  93                                 next);
  94                         err = -EIO;
  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                                 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                 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         unsigned int blocknr;
 136         struct buffer_head *bh;
 137
 138         *bhp = NULL;
 139
 140         J_ASSERT (offset < journal->j_maxlen);
 141
 142         blocknr = offset;
 143         if (journal->j_inode)
 144                 blocknr = bmap(journal->j_inode, offset);
 145
 146         if (!blocknr) {
 147                 printk (KERN_ERR "JFS: bad block at offset %u\n",
 148                         offset);
 149                 return -EIO;
 150         }
 151
 152         bh = getblk(journal->j_dev, blocknr, journal->j_blocksize);
 153         if (!bh)
 154                 return -ENOMEM;
 155
 156         if (!buffer_uptodate(bh)) {
 157                 /* If this is a brand new buffer, start readahead.
 158                    Otherwise, we assume we are already reading it.  */
 159                 if (!buffer_req(bh))
 160                         do_readahead(journal, offset);
 161                 wait_on_buffer(bh);
 162         }
 163
 164         if (!buffer_uptodate(bh)) {
 165                 printk (KERN_ERR "JFS: Failed to read block at offset %u\n",
 166                         offset);
 167                 brelse(bh);
 168                 return -EIO;
 169         }
 170
 171         *bhp = bh;
 172         return 0;
 173 }
 174
 175
 176 /*
 177  * Count the number of in-use tags in a journal descriptor block.
 178  */
 179
 180 int count_tags(struct buffer_head *bh, int size)
 181 {
 182         char *                  tagp;
 183         journal_block_tag_t *   tag;
 184         int                     nr = 0;
 185
 186         tagp = &bh->b_data[sizeof(journal_header_t)];
 187
 188         while ((tagp - bh->b_data + sizeof(journal_block_tag_t)) <= size) {
 189                 tag = (journal_block_tag_t *) tagp;
 190
 191                 nr++;
 192                 tagp += sizeof(journal_block_tag_t);
 193                 if (!(tag->t_flags & htonl(JFS_FLAG_SAME_UUID)))
 194                         tagp += 16;
 195
 196                 if (tag->t_flags & htonl(JFS_FLAG_LAST_TAG))
 197                         break;
 198         }
 199
 200         return nr;
 201 }
 202
 203
 204 /* Make sure we wrap around the log correctly! */
 205 #define wrap(journal, var)                                              \
 206 do {                                                                    \
 207         if (var >= (journal)->j_last)                                   \
 208                 var -= ((journal)->j_last - (journal)->j_first);        \
 209 } while (0)
 210
 211 /*
 212  * journal_recover
 213  *
 214  * The primary function for recovering the log contents when mounting a
 215  * journaled device.
 216  *
 217  * Recovery is done in three passes.  In the first pass, we look for the
 218  * end of the log.  In the second, we assemble the list of revoke
 219  * blocks.  In the third and final pass, we replay any un-revoked blocks
 220  * in the log.
 221  */
 222
 223 int journal_recover(journal_t *journal)
 224 {
 225         int                     err;
 226         journal_superblock_t *  sb;
 227
 228         struct recovery_info    info = {};
 229
 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                 jfs_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         jfs_debug(0, "JFS: recovery, exit status %d, "
 253                   "recovered transactions %u to %u\n",
 254                   err, info.start_transaction, info.end_transaction);
 255         jfs_debug(0, "JFS: 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_dev(journal->j_dev);
 264         return err;
 265 }
 266
 267 static int do_one_pass(journal_t *journal, struct recovery_info *info,
 268                        enum passtype pass)
 269 {
 270
 271         unsigned int            first_commit_ID, next_commit_ID;
 272         unsigned long           next_log_block;
 273         int                     err, success = 0;
 274         journal_superblock_t *  sb;
 275         journal_header_t *      tmp;
 276         struct buffer_head *    bh;
 277         unsigned int            sequence;
 278         int                     blocktype;
 279
 280         /* Precompute the maximum metadata descriptors in a descriptor block */
 281         int                     MAX_BLOCKS_PER_DESC;
 282         MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t))
 283                                / sizeof(journal_block_tag_t));
 284
 285         /*
 286          * First thing is to establish what we expect to find in the log
 287          * (in terms of transaction IDs), and where (in terms of log
 288          * block offsets): query the superblock.
 289          */
 290
 291         sb = journal->j_superblock;
 292         next_commit_ID = ntohl(sb->s_sequence);
 293         next_log_block = ntohl(sb->s_start);
 294
 295         first_commit_ID = next_commit_ID;
 296         if (pass == PASS_SCAN)
 297                 info->start_transaction = first_commit_ID;
 298
 299         jfs_debug(1, "Starting recovery pass %d\n", pass);
 300
 301         /*
 302          * Now we walk through the log, transaction by transaction,
 303          * making sure that each transaction has a commit block in the
 304          * expected place.  Each complete transaction gets replayed back
 305          * into the main filesystem.
 306          */
 307
 308         while (1) {
 309                 int                     flags;
 310                 char *                  tagp;
 311                 journal_block_tag_t *   tag;
 312                 struct buffer_head *    obh;
 313                 struct buffer_head *    nbh;
 314
 315                 /* If we already know where to stop the log traversal,
 316                  * check right now that we haven't gone past the end of
 317                  * the log. */
 318
 319                 if (pass != PASS_SCAN)
 320                         if (tid_geq(next_commit_ID, info->end_transaction))
 321                                 break;
 322
 323                 jfs_debug(2, "Scanning for sequence ID %u at %lu/%lu\n",
 324                           next_commit_ID, next_log_block, journal->j_last);
 325
 326                 /* Skip over each chunk of the transaction looking
 327                  * either the next descriptor block or the final commit
 328                  * record. */
 329
 330                 jfs_debug(3, "JFS: checking block %ld\n", next_log_block);
 331                 err = jread(&bh, journal, next_log_block);
 332                 if (err)
 333                         goto failed;
 334
 335                 next_log_block++;
 336                 wrap(journal, next_log_block);
 337
 338                 /* What kind of buffer is it?
 339                  *
 340                  * If it is a descriptor block, check that it has the
 341                  * expected sequence number.  Otherwise, we're all done
 342                  * here. */
 343
 344                 tmp = (journal_header_t *) bh->b_data;
 345
 346                 if (tmp->h_magic != htonl(JFS_MAGIC_NUMBER)) {
 347                         brelse(bh);
 348                         break;
 349                 }
 350
 351                 blocktype = ntohl(tmp->h_blocktype);
 352                 sequence = ntohl(tmp->h_sequence);
 353                 jfs_debug(3, "Found magic %d, sequence %d\n",
 354                           blocktype, sequence);
 355
 356                 if (sequence != next_commit_ID) {
 357                         brelse(bh);
 358                         break;
 359                 }
 360
 361                 /* OK, we have a valid descriptor block which matches
 362                  * all of the sequence number checks.  What are we going
 363                  * to do with it?  That depends on the pass... */
 364
 365                 switch(blocktype) {
 366                 case JFS_DESCRIPTOR_BLOCK:
 367                         /* If it is a valid descriptor block, replay it
 368                          * in pass REPLAY; otherwise, just skip over the
 369                          * blocks it describes. */
 370                         if (pass != PASS_REPLAY) {
 371                                 next_log_block += count_tags(bh, journal->j_blocksize);
 372                                 wrap(journal, next_log_block);
 373                                 brelse(bh);
 374                                 continue;
 375                         }
 376
 377                         /* A descriptor block: we can now write all of
 378                          * the data blocks.  Yay, useful work is finally
 379                          * getting done here! */
 380
 381                         tagp = &bh->b_data[sizeof(journal_header_t)];
 382                         while ((tagp - bh->b_data +sizeof(journal_block_tag_t))
 383                                <= journal->j_blocksize) {
 384                                 unsigned long io_block;
 385
 386                                 tag = (journal_block_tag_t *) tagp;
 387                                 flags = ntohl(tag->t_flags);
 388
 389                                 io_block = next_log_block++;
 390                                 wrap(journal, next_log_block);
 391                                 err = jread(&obh, journal, io_block);
 392                                 if (err) {
 393                                         /* Recover what we can, but
 394                                          * report failure at the end. */
 395                                         success = err;
 396                                         printk (KERN_ERR
 397                                                 "JFS: IO error %d recovering "
 398                                                 "block %ld in log\n",
 399                                                 err, io_block);
 400                                 } else {
 401                                         unsigned long blocknr;
 402
 403                                         J_ASSERT(obh != NULL);
 404                                         blocknr = ntohl(tag->t_blocknr);
 405
 406                                         /* If the block has been
 407                                          * revoked, then we're all done
 408                                          * here. */
 409                                         if (journal_test_revoke
 410                                             (journal, blocknr,
 411                                              next_commit_ID)) {
 412                                                 brelse(obh);
 413                                                 ++info->nr_revoke_hits;
 414                                                 goto skip_write;
 415                                         }
 416
 417                                         /* Find a buffer for the new
 418                                          * data being restored */
 419                                         nbh = getblk(journal->j_dev, blocknr,
 420                                                      journal->j_blocksize);
 421                                         if (nbh == NULL) {
 422                                                 printk(KERN_ERR
 423                                                        "JFS: Out of memory "
 424                                                        "during recovery.\n");
 425                                                 err = -ENOMEM;
 426                                                 brelse(bh);
 427                                                 brelse(obh);
 428                                                 goto failed;
 429                                         }
 430
 431                                         memcpy(nbh->b_data, obh->b_data,
 432                                                journal->j_blocksize);
 433                                         if (flags & JFS_FLAG_ESCAPE) {
 434                                                 * ((unsigned int *) bh->b_data) = htonl(JFS_MAGIC_NUMBER);
 435                                         }
 436
 437                                         mark_buffer_dirty(nbh, 1);
 438                                         ++info->nr_replays;
 439                                         // ll_rw_block(WRITE, 1, &nbh);
 440                                         brelse(obh);
 441                                         brelse(nbh);
 442                                 }
 443
 444                         skip_write:
 445                                 tagp += sizeof(journal_block_tag_t);
 446                                 if (!(flags & JFS_FLAG_SAME_UUID))
 447                                         tagp += 16;
 448
 449                                 if (flags & JFS_FLAG_LAST_TAG)
 450                                         break;
 451                         }
 452
 453                         brelse(bh);
 454                         continue;
 455
 456                 case JFS_COMMIT_BLOCK:
 457                         /* Found an expected commit block: not much to
 458                          * do other than move on to the next sequence
 459                          * number. */
 460                         brelse(bh);
 461                         next_commit_ID++;
 462                         continue;
 463
 464                 case JFS_REVOKE_BLOCK:
 465                         /* If we aren't in the REVOKE pass, then we can
 466                          * just skip over this block. */
 467                         if (pass != PASS_REVOKE) {
 468                                 brelse(bh);
 469                                 continue;
 470                         }
 471
 472                         err = scan_revoke_records(journal, bh,
 473                                                   next_commit_ID, info);
 474                         brelse(bh);
 475                         if (err)
 476                                 goto failed;
 477                         continue;
 478
 479                 default:
 480                         jfs_debug(3, "Unrecognised magic %d, end of scan.\n",
 481                                   blocktype);
 482                         goto done;
 483                 }
 484         }
 485
 486  done:
 487         /*
 488          * We broke out of the log scan loop: either we came to the
 489          * known end of the log or we found an unexpected block in the
 490          * log.  If the latter happened, then we know that the "current"
 491          * transaction marks the end of the valid log.
 492          */
 493
 494         if (pass == PASS_SCAN)
 495                 info->end_transaction = next_commit_ID;
 496         else {
 497                 /* It's really bad news if different passes end up at
 498                  * different places (but possible due to IO errors). */
 499                 if (info->end_transaction != next_commit_ID) {
 500                         printk (KERN_ERR "JFS: recovery pass %d ended at "
 501                                 "transaction %u, expected %u\n",
 502                                 pass, next_commit_ID, info->end_transaction);
 503                         if (!success)
 504                                 success = -EIO;
 505                 }
 506         }
 507
 508         return success;
 509
 510  failed:
 511         return err;
 512 }
 513
 514
 515 /* Scan a revoke record, marking all blocks mentioned as revoked. */
 516
 517 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh,
 518                                tid_t sequence, struct recovery_info *info)
 519 {
 520         journal_revoke_header_t *header;
 521         int offset, max;
 522
 523         header = (journal_revoke_header_t *) bh->b_data;
 524         offset = sizeof(journal_revoke_header_t);
 525         max = ntohl(header->r_count);
 526
 527         while (offset < max) {
 528                 unsigned long blocknr;
 529                 int err;
 530
 531                 blocknr = * ((unsigned int *) bh->b_data+offset);
 532                 offset += 4;
 533                 err = journal_set_revoke(journal, blocknr, sequence);
 534                 if (err)
 535                         return err;
 536                 ++info->nr_revokes;
 537         }
 538         return 0;
 539 }