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, err2; 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 err2 = sync_blockdev(journal->j_fs_dev); 267 if (!err) 268 err = err2; 269 270 return err; 271 } 272 273 /** 274 * jbd2_journal_skip_recovery - Start journal and wipe exiting records 275 * @journal: journal to startup 276 * 277 * Locate any valid recovery information from the journal and set up the 278 * journal structures in memory to ignore it (presumably because the 279 * caller has evidence that it is out of date). 280 * This function does'nt appear to be exorted.. 281 * 282 * We perform one pass over the journal to allow us to tell the user how 283 * much recovery information is being erased, and to let us initialise 284 * the journal transaction sequence numbers to the next unused ID. 285 */ 286 int jbd2_journal_skip_recovery(journal_t *journal) 287 { 288 int err; 289 journal_superblock_t * sb; 290 291 struct recovery_info info; 292 293 memset (&info, 0, sizeof(info)); 294 sb = journal->j_superblock; 295 296 err = do_one_pass(journal, &info, PASS_SCAN); 297 298 if (err) { 299 printk(KERN_ERR "JBD: error %d scanning journal\n", err); 300 ++journal->j_transaction_sequence; 301 } else { 302 #ifdef CONFIG_JBD2_DEBUG 303 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence); 304 #endif 305 jbd_debug(1, 306 "JBD: ignoring %d transaction%s from the journal.\n", 307 dropped, (dropped == 1) ? "" : "s"); 308 journal->j_transaction_sequence = ++info.end_transaction; 309 } 310 311 journal->j_tail = 0; 312 return err; 313 } 314 315 static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 316 { 317 unsigned long long block = be32_to_cpu(tag->t_blocknr); 318 if (tag_bytes > JBD2_TAG_SIZE32) 319 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32; 320 return block; 321 } 322 323 /* 324 * calc_chksums calculates the checksums for the blocks described in the 325 * descriptor block. 326 */ 327 static int calc_chksums(journal_t *journal, struct buffer_head *bh, 328 unsigned long *next_log_block, __u32 *crc32_sum) 329 { 330 int i, num_blks, err; 331 unsigned long io_block; 332 struct buffer_head *obh; 333 334 num_blks = count_tags(journal, bh); 335 /* Calculate checksum of the descriptor block. */ 336 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 337 338 for (i = 0; i < num_blks; i++) { 339 io_block = (*next_log_block)++; 340 wrap(journal, *next_log_block); 341 err = jread(&obh, journal, io_block); 342 if (err) { 343 printk(KERN_ERR "JBD: IO error %d recovering block " 344 "%lu in log\n", err, io_block); 345 return 1; 346 } else { 347 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 348 obh->b_size); 349 } 350 put_bh(obh); 351 } 352 return 0; 353 } 354 355 static int do_one_pass(journal_t *journal, 356 struct recovery_info *info, enum passtype pass) 357 { 358 unsigned int first_commit_ID, next_commit_ID; 359 unsigned long next_log_block; 360 int err, success = 0; 361 journal_superblock_t * sb; 362 journal_header_t * tmp; 363 struct buffer_head * bh; 364 unsigned int sequence; 365 int blocktype; 366 int tag_bytes = journal_tag_bytes(journal); 367 __u32 crc32_sum = ~0; /* Transactional Checksums */ 368 369 /* Precompute the maximum metadata descriptors in a descriptor block */ 370 int MAX_BLOCKS_PER_DESC; 371 MAX_BLOCKS_PER_DESC = ((journal->j_blocksize-sizeof(journal_header_t)) 372 / tag_bytes); 373 374 /* 375 * First thing is to establish what we expect to find in the log 376 * (in terms of transaction IDs), and where (in terms of log 377 * block offsets): query the superblock. 378 */ 379 380 sb = journal->j_superblock; 381 next_commit_ID = be32_to_cpu(sb->s_sequence); 382 next_log_block = be32_to_cpu(sb->s_start); 383 384 first_commit_ID = next_commit_ID; 385 if (pass == PASS_SCAN) 386 info->start_transaction = first_commit_ID; 387 388 jbd_debug(1, "Starting recovery pass %d\n", pass); 389 390 /* 391 * Now we walk through the log, transaction by transaction, 392 * making sure that each transaction has a commit block in the 393 * expected place. Each complete transaction gets replayed back 394 * into the main filesystem. 395 */ 396 397 while (1) { 398 int flags; 399 char * tagp; 400 journal_block_tag_t * tag; 401 struct buffer_head * obh; 402 struct buffer_head * nbh; 403 404 cond_resched(); 405 406 /* If we already know where to stop the log traversal, 407 * check right now that we haven't gone past the end of 408 * the log. */ 409 410 if (pass != PASS_SCAN) 411 if (tid_geq(next_commit_ID, info->end_transaction)) 412 break; 413 414 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n", 415 next_commit_ID, next_log_block, journal->j_last); 416 417 /* Skip over each chunk of the transaction looking 418 * either the next descriptor block or the final commit 419 * record. */ 420 421 jbd_debug(3, "JBD: checking block %ld\n", next_log_block); 422 err = jread(&bh, journal, next_log_block); 423 if (err) 424 goto failed; 425 426 next_log_block++; 427 wrap(journal, next_log_block); 428 429 /* What kind of buffer is it? 430 * 431 * If it is a descriptor block, check that it has the 432 * expected sequence number. Otherwise, we're all done 433 * here. */ 434 435 tmp = (journal_header_t *)bh->b_data; 436 437 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) { 438 brelse(bh); 439 break; 440 } 441 442 blocktype = be32_to_cpu(tmp->h_blocktype); 443 sequence = be32_to_cpu(tmp->h_sequence); 444 jbd_debug(3, "Found magic %d, sequence %d\n", 445 blocktype, sequence); 446 447 if (sequence != next_commit_ID) { 448 brelse(bh); 449 break; 450 } 451 452 /* OK, we have a valid descriptor block which matches 453 * all of the sequence number checks. What are we going 454 * to do with it? That depends on the pass... */ 455 456 switch(blocktype) { 457 case JBD2_DESCRIPTOR_BLOCK: 458 /* If it is a valid descriptor block, replay it 459 * in pass REPLAY; if journal_checksums enabled, then 460 * calculate checksums in PASS_SCAN, otherwise, 461 * just skip over the blocks it describes. */ 462 if (pass != PASS_REPLAY) { 463 if (pass == PASS_SCAN && 464 JBD2_HAS_COMPAT_FEATURE(journal, 465 JBD2_FEATURE_COMPAT_CHECKSUM) && 466 !info->end_transaction) { 467 if (calc_chksums(journal, bh, 468 &next_log_block, 469 &crc32_sum)) { 470 put_bh(bh); 471 break; 472 } 473 put_bh(bh); 474 continue; 475 } 476 next_log_block += count_tags(journal, bh); 477 wrap(journal, next_log_block); 478 put_bh(bh); 479 continue; 480 } 481 482 /* A descriptor block: we can now write all of 483 * the data blocks. Yay, useful work is finally 484 * getting done here! */ 485 486 tagp = &bh->b_data[sizeof(journal_header_t)]; 487 while ((tagp - bh->b_data + tag_bytes) 488 <= journal->j_blocksize) { 489 unsigned long io_block; 490 491 tag = (journal_block_tag_t *) tagp; 492 flags = be32_to_cpu(tag->t_flags); 493 494 io_block = next_log_block++; 495 wrap(journal, next_log_block); 496 err = jread(&obh, journal, io_block); 497 if (err) { 498 /* Recover what we can, but 499 * report failure at the end. */ 500 success = err; 501 printk (KERN_ERR 502 "JBD: IO error %d recovering " 503 "block %ld in log\n", 504 err, io_block); 505 } else { 506 unsigned long long blocknr; 507 508 J_ASSERT(obh != NULL); 509 blocknr = read_tag_block(tag_bytes, 510 tag); 511 512 /* If the block has been 513 * revoked, then we're all done 514 * here. */ 515 if (jbd2_journal_test_revoke 516 (journal, blocknr, 517 next_commit_ID)) { 518 brelse(obh); 519 ++info->nr_revoke_hits; 520 goto skip_write; 521 } 522 523 /* Find a buffer for the new 524 * data being restored */ 525 nbh = __getblk(journal->j_fs_dev, 526 blocknr, 527 journal->j_blocksize); 528 if (nbh == NULL) { 529 printk(KERN_ERR 530 "JBD: Out of memory " 531 "during recovery.\n"); 532 err = -ENOMEM; 533 brelse(bh); 534 brelse(obh); 535 goto failed; 536 } 537 538 lock_buffer(nbh); 539 memcpy(nbh->b_data, obh->b_data, 540 journal->j_blocksize); 541 if (flags & JBD2_FLAG_ESCAPE) { 542 *((__be32 *)nbh->b_data) = 543 cpu_to_be32(JBD2_MAGIC_NUMBER); 544 } 545 546 BUFFER_TRACE(nbh, "marking dirty"); 547 set_buffer_uptodate(nbh); 548 mark_buffer_dirty(nbh); 549 BUFFER_TRACE(nbh, "marking uptodate"); 550 ++info->nr_replays; 551 /* ll_rw_block(WRITE, 1, &nbh); */ 552 unlock_buffer(nbh); 553 brelse(obh); 554 brelse(nbh); 555 } 556 557 skip_write: 558 tagp += tag_bytes; 559 if (!(flags & JBD2_FLAG_SAME_UUID)) 560 tagp += 16; 561 562 if (flags & JBD2_FLAG_LAST_TAG) 563 break; 564 } 565 566 brelse(bh); 567 continue; 568 569 case JBD2_COMMIT_BLOCK: 570 /* How to differentiate between interrupted commit 571 * and journal corruption ? 572 * 573 * {nth transaction} 574 * Checksum Verification Failed 575 * | 576 * ____________________ 577 * | | 578 * async_commit sync_commit 579 * | | 580 * | GO TO NEXT "Journal Corruption" 581 * | TRANSACTION 582 * | 583 * {(n+1)th transanction} 584 * | 585 * _______|______________ 586 * | | 587 * Commit block found Commit block not found 588 * | | 589 * "Journal Corruption" | 590 * _____________|_________ 591 * | | 592 * nth trans corrupt OR nth trans 593 * and (n+1)th interrupted interrupted 594 * before commit block 595 * could reach the disk. 596 * (Cannot find the difference in above 597 * mentioned conditions. Hence assume 598 * "Interrupted Commit".) 599 */ 600 601 /* Found an expected commit block: if checksums 602 * are present verify them in PASS_SCAN; else not 603 * much to do other than move on to the next sequence 604 * number. */ 605 if (pass == PASS_SCAN && 606 JBD2_HAS_COMPAT_FEATURE(journal, 607 JBD2_FEATURE_COMPAT_CHECKSUM)) { 608 int chksum_err, chksum_seen; 609 struct commit_header *cbh = 610 (struct commit_header *)bh->b_data; 611 unsigned found_chksum = 612 be32_to_cpu(cbh->h_chksum[0]); 613 614 chksum_err = chksum_seen = 0; 615 616 if (info->end_transaction) { 617 journal->j_failed_commit = 618 info->end_transaction; 619 brelse(bh); 620 break; 621 } 622 623 if (crc32_sum == found_chksum && 624 cbh->h_chksum_type == JBD2_CRC32_CHKSUM && 625 cbh->h_chksum_size == 626 JBD2_CRC32_CHKSUM_SIZE) 627 chksum_seen = 1; 628 else if (!(cbh->h_chksum_type == 0 && 629 cbh->h_chksum_size == 0 && 630 found_chksum == 0 && 631 !chksum_seen)) 632 /* 633 * If fs is mounted using an old kernel and then 634 * kernel with journal_chksum is used then we 635 * get a situation where the journal flag has 636 * checksum flag set but checksums are not 637 * present i.e chksum = 0, in the individual 638 * commit blocks. 639 * Hence to avoid checksum failures, in this 640 * situation, this extra check is added. 641 */ 642 chksum_err = 1; 643 644 if (chksum_err) { 645 info->end_transaction = next_commit_ID; 646 647 if (!JBD2_HAS_INCOMPAT_FEATURE(journal, 648 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 649 journal->j_failed_commit = 650 next_commit_ID; 651 brelse(bh); 652 break; 653 } 654 } 655 crc32_sum = ~0; 656 } 657 brelse(bh); 658 next_commit_ID++; 659 continue; 660 661 case JBD2_REVOKE_BLOCK: 662 /* If we aren't in the REVOKE pass, then we can 663 * just skip over this block. */ 664 if (pass != PASS_REVOKE) { 665 brelse(bh); 666 continue; 667 } 668 669 err = scan_revoke_records(journal, bh, 670 next_commit_ID, info); 671 brelse(bh); 672 if (err) 673 goto failed; 674 continue; 675 676 default: 677 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 678 blocktype); 679 brelse(bh); 680 goto done; 681 } 682 } 683 684 done: 685 /* 686 * We broke out of the log scan loop: either we came to the 687 * known end of the log or we found an unexpected block in the 688 * log. If the latter happened, then we know that the "current" 689 * transaction marks the end of the valid log. 690 */ 691 692 if (pass == PASS_SCAN) { 693 if (!info->end_transaction) 694 info->end_transaction = next_commit_ID; 695 } else { 696 /* It's really bad news if different passes end up at 697 * different places (but possible due to IO errors). */ 698 if (info->end_transaction != next_commit_ID) { 699 printk (KERN_ERR "JBD: recovery pass %d ended at " 700 "transaction %u, expected %u\n", 701 pass, next_commit_ID, info->end_transaction); 702 if (!success) 703 success = -EIO; 704 } 705 } 706 707 return success; 708 709 failed: 710 return err; 711 } 712 713 714 /* Scan a revoke record, marking all blocks mentioned as revoked. */ 715 716 static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 717 tid_t sequence, struct recovery_info *info) 718 { 719 jbd2_journal_revoke_header_t *header; 720 int offset, max; 721 int record_len = 4; 722 723 header = (jbd2_journal_revoke_header_t *) bh->b_data; 724 offset = sizeof(jbd2_journal_revoke_header_t); 725 max = be32_to_cpu(header->r_count); 726 727 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) 728 record_len = 8; 729 730 while (offset + record_len <= max) { 731 unsigned long long blocknr; 732 int err; 733 734 if (record_len == 4) 735 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset))); 736 else 737 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset))); 738 offset += record_len; 739 err = jbd2_journal_set_revoke(journal, blocknr, sequence); 740 if (err) 741 return err; 742 ++info->nr_revokes; 743 } 744 return 0; 745 } 746