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/slab.h> 24 #include <linux/crc32.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 "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 long blocknr; 137 struct buffer_head *bh; 138 139 *bhp = NULL; 140 141 if (offset >= journal->j_maxlen) { 142 printk(KERN_ERR "JBD: corrupted journal superblock\n"); 143 return -EIO; 144 } 145 146 err = jbd2_journal_bmap(journal, offset, &blocknr); 147 148 if (err) { 149 printk (KERN_ERR "JBD: bad block at offset %u\n", 150 offset); 151 return err; 152 } 153 154 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize); 155 if (!bh) 156 return -ENOMEM; 157 158 if (!buffer_uptodate(bh)) { 159 /* If this is a brand new buffer, start readahead. 160 Otherwise, we assume we are already reading it. */ 161 if (!buffer_req(bh)) 162 do_readahead(journal, offset); 163 wait_on_buffer(bh); 164 } 165 166 if (!buffer_uptodate(bh)) { 167 printk (KERN_ERR "JBD: Failed to read block at offset %u\n", 168 offset); 169 brelse(bh); 170 return -EIO; 171 } 172 173 *bhp = bh; 174 return 0; 175 } 176 177 178 /* 179 * Count the number of in-use tags in a journal descriptor block. 180 */ 181 182 static int count_tags(journal_t *journal, struct buffer_head *bh) 183 { 184 char * tagp; 185 journal_block_tag_t * tag; 186 int nr = 0, size = journal->j_blocksize; 187 int tag_bytes = journal_tag_bytes(journal); 188 189 tagp = &bh->b_data[sizeof(journal_header_t)]; 190 191 while ((tagp - bh->b_data + tag_bytes) <= size) { 192 tag = (journal_block_tag_t *) tagp; 193 194 nr++; 195 tagp += tag_bytes; 196 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID))) 197 tagp += 16; 198 199 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG)) 200 break; 201 } 202 203 return nr; 204 } 205 206 207 /* Make sure we wrap around the log correctly! */ 208 #define wrap(journal, var) \ 209 do { \ 210 if (var >= (journal)->j_last) \ 211 var -= ((journal)->j_last - (journal)->j_first); \ 212 } while (0) 213 214 /** 215 * jbd2_journal_recover - recovers a on-disk journal 216 * @journal: the journal to recover 217 * 218 * The primary function for recovering the log contents when mounting a 219 * journaled device. 220 * 221 * Recovery is done in three passes. In the first pass, we look for the 222 * end of the log. In the second, we assemble the list of revoke 223 * blocks. In the third and final pass, we replay any un-revoked blocks 224 * in the log. 225 */ 226 int jbd2_journal_recover(journal_t *journal) 227 { 228 int err; 229 journal_superblock_t * sb; 230 231 struct recovery_info info; 232 233 memset(&info, 0, sizeof(info)); 234 sb = journal->j_superblock; 235 236 /* 237 * The journal superblock's s_start field (the current log head) 238 * is always zero if, and only if, the journal was cleanly 239 * unmounted. 240 */ 241 242 if (!sb->s_start) { 243 jbd_debug(1, "No recovery required, last transaction %d\n", 244 be32_to_cpu(sb->s_sequence)); 245 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1; 246 return 0; 247 } 248 249 err = do_one_pass(journal, &info, PASS_SCAN); 250 if (!err) 251 err = do_one_pass(journal, &info, PASS_REVOKE); 252 if (!err) 253 err = do_one_pass(journal, &info, PASS_REPLAY); 254 255 jbd_debug(1, "JBD: recovery, exit status %d, " 256 "recovered transactions %u to %u\n", 257 err, info.start_transaction, info.end_transaction); 258 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n", 259 info.nr_replays, info.nr_revoke_hits, info.nr_revokes); 260 261 /* Restart the log at the next transaction ID, thus invalidating 262 * any existing commit records in the log. */ 263 journal->j_transaction_sequence = ++info.end_transaction; 264 265 jbd2_journal_clear_revoke(journal); 266 sync_blockdev(journal->j_fs_dev); 267 return err; 268 } 269 270 /** 271 * jbd2_journal_skip_recovery - Start journal and wipe exiting records 272 * @journal: journal to startup 273 * 274 * Locate any valid recovery information from the journal and set up the 275 * journal structures in memory to ignore it (presumably because the 276 * caller has evidence that it is out of date). 277 * This function does'nt appear to be exorted.. 278 * 279 * We perform one pass over the journal to allow us to tell the user how 280 * much recovery information is being erased, and to let us initialise 281 * the journal transaction sequence numbers to the next unused ID. 282 */ 283 int jbd2_journal_skip_recovery(journal_t *journal) 284 { 285 int err; 286 journal_superblock_t * sb; 287 288 struct recovery_info info; 289 290 memset (&info, 0, sizeof(info)); 291 sb = journal->j_superblock; 292 293 err = do_one_pass(journal, &info, PASS_SCAN); 294 295 if (err) { 296 printk(KERN_ERR "JBD: error %d scanning journal\n", err); 297 ++journal->j_transaction_sequence; 298 } else { 299 #ifdef CONFIG_JBD2_DEBUG 300 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence); 301 #endif 302 jbd_debug(1, 303 "JBD: ignoring %d transaction%s from the journal.\n", 304 dropped, (dropped == 1) ? "" : "s"); 305 journal->j_transaction_sequence = ++info.end_transaction; 306 } 307 308 journal->j_tail = 0; 309 return err; 310 } 311 312 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 313 { 314 unsigned long long block = be32_to_cpu(tag->t_blocknr); 315 if (tag_bytes > JBD2_TAG_SIZE32) 316 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32; 317 return block; 318 } 319 320 /* 321 * calc_chksums calculates the checksums for the blocks described in the 322 * descriptor block. 323 */ 324 static int calc_chksums(journal_t *journal, struct buffer_head *bh, 325 unsigned long *next_log_block, __u32 *crc32_sum) 326 { 327 int i, num_blks, err; 328 unsigned long io_block; 329 struct buffer_head *obh; 330 331 num_blks = count_tags(journal, bh); 332 /* Calculate checksum of the descriptor block. */ 333 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 334 335 for (i = 0; i < num_blks; i++) { 336 io_block = (*next_log_block)++; 337 wrap(journal, *next_log_block); 338 err = jread(&obh, journal, io_block); 339 if (err) { 340 printk(KERN_ERR "JBD: IO error %d recovering block " 341 "%lu in log\n", err, io_block); 342 return 1; 343 } else { 344 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 345 obh->b_size); 346 } 347 put_bh(obh); 348 } 349 return 0; 350 } 351 352 static int do_one_pass(journal_t *journal, 353 struct recovery_info *info, enum passtype pass) 354 { 355 unsigned int first_commit_ID, next_commit_ID; 356 unsigned long next_log_block; 357 int err, success = 0; 358 journal_superblock_t * sb; 359 journal_header_t * tmp; 360 struct buffer_head * bh; 361 unsigned int sequence; 362 int blocktype; 363 int tag_bytes = journal_tag_bytes(journal); 364 __u32 crc32_sum = ~0; /* Transactional Checksums */ 365 366 /* Precompute the maximum metadata descriptors in a descriptor block */ 367 int MAX_BLOCKS_PER_DESC; 368 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t)) 369 / tag_bytes); 370 371 /* 372 * First thing is to establish what we expect to find in the log 373 * (in terms of transaction IDs), and where (in terms of log 374 * block offsets): query the superblock. 375 */ 376 377 sb = journal->j_superblock; 378 next_commit_ID = be32_to_cpu(sb->s_sequence); 379 next_log_block = be32_to_cpu(sb->s_start); 380 381 first_commit_ID = next_commit_ID; 382 if (pass == PASS_SCAN) 383 info->start_transaction = first_commit_ID; 384 385 jbd_debug(1, "Starting recovery pass %d\n", pass); 386 387 /* 388 * Now we walk through the log, transaction by transaction, 389 * making sure that each transaction has a commit block in the 390 * expected place. Each complete transaction gets replayed back 391 * into the main filesystem. 392 */ 393 394 while (1) { 395 int flags; 396 char * tagp; 397 journal_block_tag_t * tag; 398 struct buffer_head * obh; 399 struct buffer_head * nbh; 400 401 cond_resched(); 402 403 /* If we already know where to stop the log traversal, 404 * check right now that we haven't gone past the end of 405 * the log. */ 406 407 if (pass != PASS_SCAN) 408 if (tid_geq(next_commit_ID, info->end_transaction)) 409 break; 410 411 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n", 412 next_commit_ID, next_log_block, journal->j_last); 413 414 /* Skip over each chunk of the transaction looking 415 * either the next descriptor block or the final commit 416 * record. */ 417 418 jbd_debug(3, "JBD: checking block %ld\n", next_log_block); 419 err = jread(&bh, journal, next_log_block); 420 if (err) 421 goto failed; 422 423 next_log_block++; 424 wrap(journal, next_log_block); 425 426 /* What kind of buffer is it? 427 * 428 * If it is a descriptor block, check that it has the 429 * expected sequence number. Otherwise, we're all done 430 * here. */ 431 432 tmp = (journal_header_t *)bh->b_data; 433 434 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) { 435 brelse(bh); 436 break; 437 } 438 439 blocktype = be32_to_cpu(tmp->h_blocktype); 440 sequence = be32_to_cpu(tmp->h_sequence); 441 jbd_debug(3, "Found magic %d, sequence %d\n", 442 blocktype, sequence); 443 444 if (sequence != next_commit_ID) { 445 brelse(bh); 446 break; 447 } 448 449 /* OK, we have a valid descriptor block which matches 450 * all of the sequence number checks. What are we going 451 * to do with it? That depends on the pass... */ 452 453 switch(blocktype) { 454 case JBD2_DESCRIPTOR_BLOCK: 455 /* If it is a valid descriptor block, replay it 456 * in pass REPLAY; if journal_checksums enabled, then 457 * calculate checksums in PASS_SCAN, otherwise, 458 * just skip over the blocks it describes. */ 459 if (pass != PASS_REPLAY) { 460 if (pass == PASS_SCAN && 461 JBD2_HAS_COMPAT_FEATURE(journal, 462 JBD2_FEATURE_COMPAT_CHECKSUM) && 463 !info->end_transaction) { 464 if (calc_chksums(journal, bh, 465 &next_log_block, 466 &crc32_sum)) { 467 put_bh(bh); 468 break; 469 } 470 put_bh(bh); 471 continue; 472 } 473 next_log_block += count_tags(journal, bh); 474 wrap(journal, next_log_block); 475 put_bh(bh); 476 continue; 477 } 478 479 /* A descriptor block: we can now write all of 480 * the data blocks. Yay, useful work is finally 481 * getting done here! */ 482 483 tagp = &bh->b_data[sizeof(journal_header_t)]; 484 while ((tagp - bh->b_data + tag_bytes) 485 <= journal->j_blocksize) { 486 unsigned long io_block; 487 488 tag = (journal_block_tag_t *) tagp; 489 flags = be32_to_cpu(tag->t_flags); 490 491 io_block = next_log_block++; 492 wrap(journal, next_log_block); 493 err = jread(&obh, journal, io_block); 494 if (err) { 495 /* Recover what we can, but 496 * report failure at the end. */ 497 success = err; 498 printk (KERN_ERR 499 "JBD: IO error %d recovering " 500 "block %ld in log\n", 501 err, io_block); 502 } else { 503 unsigned long long blocknr; 504 505 J_ASSERT(obh != NULL); 506 blocknr = read_tag_block(tag_bytes, 507 tag); 508 509 /* If the block has been 510 * revoked, then we're all done 511 * here. */ 512 if (jbd2_journal_test_revoke 513 (journal, blocknr, 514 next_commit_ID)) { 515 brelse(obh); 516 ++info->nr_revoke_hits; 517 goto skip_write; 518 } 519 520 /* Find a buffer for the new 521 * data being restored */ 522 nbh = __getblk(journal->j_fs_dev, 523 blocknr, 524 journal->j_blocksize); 525 if (nbh == NULL) { 526 printk(KERN_ERR 527 "JBD: Out of memory " 528 "during recovery.\n"); 529 err = -ENOMEM; 530 brelse(bh); 531 brelse(obh); 532 goto failed; 533 } 534 535 lock_buffer(nbh); 536 memcpy(nbh->b_data, obh->b_data, 537 journal->j_blocksize); 538 if (flags & JBD2_FLAG_ESCAPE) { 539 *((__be32 *)nbh->b_data) = 540 cpu_to_be32(JBD2_MAGIC_NUMBER); 541 } 542 543 BUFFER_TRACE(nbh, "marking dirty"); 544 set_buffer_uptodate(nbh); 545 mark_buffer_dirty(nbh); 546 BUFFER_TRACE(nbh, "marking uptodate"); 547 ++info->nr_replays; 548 /* ll_rw_block(WRITE, 1, &nbh); */ 549 unlock_buffer(nbh); 550 brelse(obh); 551 brelse(nbh); 552 } 553 554 skip_write: 555 tagp += tag_bytes; 556 if (!(flags & JBD2_FLAG_SAME_UUID)) 557 tagp += 16; 558 559 if (flags & JBD2_FLAG_LAST_TAG) 560 break; 561 } 562 563 brelse(bh); 564 continue; 565 566 case JBD2_COMMIT_BLOCK: 567 /* How to differentiate between interrupted commit 568 * and journal corruption ? 569 * 570 * {nth transaction} 571 * Checksum Verification Failed 572 * | 573 * ____________________ 574 * | | 575 * async_commit sync_commit 576 * | | 577 * | GO TO NEXT "Journal Corruption" 578 * | TRANSACTION 579 * | 580 * {(n+1)th transanction} 581 * | 582 * _______|______________ 583 * | | 584 * Commit block found Commit block not found 585 * | | 586 * "Journal Corruption" | 587 * _____________|_________ 588 * | | 589 * nth trans corrupt OR nth trans 590 * and (n+1)th interrupted interrupted 591 * before commit block 592 * could reach the disk. 593 * (Cannot find the difference in above 594 * mentioned conditions. Hence assume 595 * "Interrupted Commit".) 596 */ 597 598 /* Found an expected commit block: if checksums 599 * are present verify them in PASS_SCAN; else not 600 * much to do other than move on to the next sequence 601 * number. */ 602 if (pass == PASS_SCAN && 603 JBD2_HAS_COMPAT_FEATURE(journal, 604 JBD2_FEATURE_COMPAT_CHECKSUM)) { 605 int chksum_err, chksum_seen; 606 struct commit_header *cbh = 607 (struct commit_header *)bh->b_data; 608 unsigned found_chksum = 609 be32_to_cpu(cbh->h_chksum[0]); 610 611 chksum_err = chksum_seen = 0; 612 613 if (info->end_transaction) { 614 journal->j_failed_commit = 615 info->end_transaction; 616 brelse(bh); 617 break; 618 } 619 620 if (crc32_sum == found_chksum && 621 cbh->h_chksum_type == JBD2_CRC32_CHKSUM && 622 cbh->h_chksum_size == 623 JBD2_CRC32_CHKSUM_SIZE) 624 chksum_seen = 1; 625 else if (!(cbh->h_chksum_type == 0 && 626 cbh->h_chksum_size == 0 && 627 found_chksum == 0 && 628 !chksum_seen)) 629 /* 630 * If fs is mounted using an old kernel and then 631 * kernel with journal_chksum is used then we 632 * get a situation where the journal flag has 633 * checksum flag set but checksums are not 634 * present i.e chksum = 0, in the individual 635 * commit blocks. 636 * Hence to avoid checksum failures, in this 637 * situation, this extra check is added. 638 */ 639 chksum_err = 1; 640 641 if (chksum_err) { 642 info->end_transaction = next_commit_ID; 643 644 if (!JBD2_HAS_INCOMPAT_FEATURE(journal, 645 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 646 journal->j_failed_commit = 647 next_commit_ID; 648 brelse(bh); 649 break; 650 } 651 } 652 crc32_sum = ~0; 653 } 654 brelse(bh); 655 next_commit_ID++; 656 continue; 657 658 case JBD2_REVOKE_BLOCK: 659 /* If we aren't in the REVOKE pass, then we can 660 * just skip over this block. */ 661 if (pass != PASS_REVOKE) { 662 brelse(bh); 663 continue; 664 } 665 666 err = scan_revoke_records(journal, bh, 667 next_commit_ID, info); 668 brelse(bh); 669 if (err) 670 goto failed; 671 continue; 672 673 default: 674 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 675 blocktype); 676 brelse(bh); 677 goto done; 678 } 679 } 680 681 done: 682 /* 683 * We broke out of the log scan loop: either we came to the 684 * known end of the log or we found an unexpected block in the 685 * log. If the latter happened, then we know that the "current" 686 * transaction marks the end of the valid log. 687 */ 688 689 if (pass == PASS_SCAN) { 690 if (!info->end_transaction) 691 info->end_transaction = next_commit_ID; 692 } else { 693 /* It's really bad news if different passes end up at 694 * different places (but possible due to IO errors). */ 695 if (info->end_transaction != next_commit_ID) { 696 printk (KERN_ERR "JBD: recovery pass %d ended at " 697 "transaction %u, expected %u\n", 698 pass, next_commit_ID, info->end_transaction); 699 if (!success) 700 success = -EIO; 701 } 702 } 703 704 return success; 705 706 failed: 707 return err; 708 } 709 710 711 /* Scan a revoke record, marking all blocks mentioned as revoked. */ 712 713 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 714 tid_t sequence, struct recovery_info *info) 715 { 716 jbd2_journal_revoke_header_t *header; 717 int offset, max; 718 int record_len = 4; 719 720 header = (jbd2_journal_revoke_header_t *) bh->b_data; 721 offset = sizeof(jbd2_journal_revoke_header_t); 722 max = be32_to_cpu(header->r_count); 723 724 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) 725 record_len = 8; 726 727 while (offset + record_len <= max) { 728 unsigned long long blocknr; 729 int err; 730 731 if (record_len == 4) 732 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset))); 733 else 734 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset))); 735 offset += record_len; 736 err = jbd2_journal_set_revoke(journal, blocknr, sequence); 737 if (err) 738 return err; 739 ++info->nr_revokes; 740 } 741 return 0; 742 } 743