recovery.c revision e3507739e4185bdb2394928eb6479e48f4e690a8
1/* 2 * linux/fs/jbd/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 "config.h" 18#include "jfs_user.h" 19#else 20#include <linux/time.h> 21#include <linux/fs.h> 22#include <linux/jbd.h> 23#include <linux/errno.h> 24#include <linux/slab.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 */ 31struct 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 41enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY}; 42static int do_one_pass(journal_t *journal, 43 struct recovery_info *info, enum passtype pass); 44static 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 */ 50static 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 70static 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 = 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 119failed: 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 132static 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 = 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 182static 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(JFS_FLAG_SAME_UUID))) 197 tagp += 16; 198 199 if (tag->t_flags & cpu_to_be32(JFS_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) \ 209do { \ 210 if (var >= (journal)->j_last) \ 211 var -= ((journal)->j_last - (journal)->j_first); \ 212} while (0) 213 214/** 215 * 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 */ 226int 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 journal_clear_revoke(journal); 266 sync_blockdev(journal->j_fs_dev); 267 return err; 268} 269 270/** 271 * 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 */ 283int journal_skip_recovery(journal_t *journal) 284{ 285 int err; 286 struct recovery_info info; 287 288 memset (&info, 0, sizeof(info)); 289 290 err = do_one_pass(journal, &info, PASS_SCAN); 291 292 if (err) { 293 printk(KERN_ERR "JBD: error %d scanning journal\n", err); 294 ++journal->j_transaction_sequence; 295 } else { 296#ifdef CONFIG_JBD_DEBUG 297 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence); 298#endif 299 jbd_debug(1, 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 return err; 307} 308 309static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 310{ 311 unsigned long long block = be32_to_cpu(tag->t_blocknr); 312 if (tag_bytes > JBD_TAG_SIZE32) 313 block |= (__u64)be32_to_cpu(tag->t_blocknr_high) << 32; 314 return block; 315} 316 317/* 318 * calc_chksums calculates the checksums for the blocks described in the 319 * descriptor block. 320 */ 321static int calc_chksums(journal_t *journal, struct buffer_head *bh, 322 unsigned long long *next_log_block, __u32 *crc32_sum) 323{ 324 int i, num_blks, err; 325 unsigned long long io_block; 326 struct buffer_head *obh; 327 328 num_blks = count_tags(journal, bh); 329 /* Calculate checksum of the descriptor block. */ 330 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 331 332 for (i = 0; i < num_blks; i++) { 333 io_block = (*next_log_block)++; 334 wrap(journal, *next_log_block); 335 err = jread(&obh, journal, io_block); 336 if (err) { 337 printk(KERN_ERR "JBD: IO error %d recovering block " 338 "%llu in log\n", err, io_block); 339 return 1; 340 } else { 341 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 342 obh->b_size); 343 } 344 brelse(obh); 345 } 346 return 0; 347} 348 349static int do_one_pass(journal_t *journal, 350 struct recovery_info *info, enum passtype pass) 351{ 352 unsigned int first_commit_ID, next_commit_ID; 353 unsigned long long next_log_block; 354 int err, success = 0; 355 journal_superblock_t * sb; 356 journal_header_t * tmp; 357 struct buffer_head * bh; 358 unsigned int sequence; 359 int blocktype; 360 int tag_bytes = journal_tag_bytes(journal); 361 __u32 crc32_sum = ~0; /* Transactional Checksums */ 362 363 /* 364 * First thing is to establish what we expect to find in the log 365 * (in terms of transaction IDs), and where (in terms of log 366 * block offsets): query the superblock. 367 */ 368 369 sb = journal->j_superblock; 370 next_commit_ID = be32_to_cpu(sb->s_sequence); 371 next_log_block = be32_to_cpu(sb->s_start); 372 373 first_commit_ID = next_commit_ID; 374 if (pass == PASS_SCAN) 375 info->start_transaction = first_commit_ID; 376 377 jbd_debug(1, "Starting recovery pass %d\n", pass); 378 379 /* 380 * Now we walk through the log, transaction by transaction, 381 * making sure that each transaction has a commit block in the 382 * expected place. Each complete transaction gets replayed back 383 * into the main filesystem. 384 */ 385 386 while (1) { 387 int flags; 388 char * tagp; 389 journal_block_tag_t * tag; 390 struct buffer_head * obh; 391 struct buffer_head * nbh; 392 393 cond_resched(); 394 395 /* If we already know where to stop the log traversal, 396 * check right now that we haven't gone past the end of 397 * the log. */ 398 399 if (pass != PASS_SCAN) 400 if (tid_geq(next_commit_ID, info->end_transaction)) 401 break; 402 403 jbd_debug(2, "Scanning for sequence ID %u at %llu/%lu\n", 404 next_commit_ID, next_log_block, journal->j_last); 405 406 /* Skip over each chunk of the transaction looking 407 * either the next descriptor block or the final commit 408 * record. */ 409 410 jbd_debug(3, "JBD: checking block %llu\n", next_log_block); 411 err = jread(&bh, journal, next_log_block); 412 if (err) 413 goto failed; 414 415 next_log_block++; 416 wrap(journal, next_log_block); 417 418 /* What kind of buffer is it? 419 * 420 * If it is a descriptor block, check that it has the 421 * expected sequence number. Otherwise, we're all done 422 * here. */ 423 424 tmp = (journal_header_t *)bh->b_data; 425 426 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) { 427 brelse(bh); 428 break; 429 } 430 431 blocktype = be32_to_cpu(tmp->h_blocktype); 432 sequence = be32_to_cpu(tmp->h_sequence); 433 jbd_debug(3, "Found magic %d, sequence %d\n", 434 blocktype, sequence); 435 436 if (sequence != next_commit_ID) { 437 brelse(bh); 438 break; 439 } 440 441 /* OK, we have a valid descriptor block which matches 442 * all of the sequence number checks. What are we going 443 * to do with it? That depends on the pass... */ 444 445 switch(blocktype) { 446 case JFS_DESCRIPTOR_BLOCK: 447 /* If it is a valid descriptor block, replay it 448 * in pass REPLAY; if journal_checksums enabled, then 449 * calculate checksums in PASS_SCAN, otherwise, 450 * just skip over the blocks it describes. */ 451 if (pass != PASS_REPLAY) { 452 if (pass == PASS_SCAN && 453 JFS_HAS_COMPAT_FEATURE(journal, 454 JFS_FEATURE_COMPAT_CHECKSUM) && 455 !info->end_transaction) { 456 if (calc_chksums(journal, bh, 457 &next_log_block, 458 &crc32_sum)) { 459 brelse(bh); 460 break; 461 } 462 brelse(bh); 463 continue; 464 } 465 next_log_block += count_tags(journal, bh); 466 wrap(journal, next_log_block); 467 brelse(bh); 468 continue; 469 } 470 471 /* A descriptor block: we can now write all of 472 * the data blocks. Yay, useful work is finally 473 * getting done here! */ 474 475 tagp = &bh->b_data[sizeof(journal_header_t)]; 476 while ((tagp - bh->b_data + tag_bytes) 477 <= journal->j_blocksize) { 478 unsigned long long io_block; 479 480 tag = (journal_block_tag_t *) tagp; 481 flags = be32_to_cpu(tag->t_flags); 482 483 io_block = next_log_block++; 484 wrap(journal, next_log_block); 485 err = jread(&obh, journal, io_block); 486 if (err) { 487 /* Recover what we can, but 488 * report failure at the end. */ 489 success = err; 490 printk (KERN_ERR 491 "JBD: IO error %d recovering " 492 "block %llu in log\n", 493 err, io_block); 494 } else { 495 unsigned long long blocknr; 496 497 J_ASSERT(obh != NULL); 498 blocknr = read_tag_block(tag_bytes, 499 tag); 500 501 /* If the block has been 502 * revoked, then we're all done 503 * here. */ 504 if (journal_test_revoke 505 (journal, blocknr, 506 next_commit_ID)) { 507 brelse(obh); 508 ++info->nr_revoke_hits; 509 goto skip_write; 510 } 511 512 /* Find a buffer for the new 513 * data being restored */ 514 nbh = __getblk(journal->j_fs_dev, 515 blocknr, 516 journal->j_blocksize); 517 if (nbh == NULL) { 518 printk(KERN_ERR 519 "JBD: Out of memory " 520 "during recovery.\n"); 521 err = -ENOMEM; 522 brelse(bh); 523 brelse(obh); 524 goto failed; 525 } 526 527 lock_buffer(nbh); 528 memcpy(nbh->b_data, obh->b_data, 529 journal->j_blocksize); 530 if (flags & JFS_FLAG_ESCAPE) { 531 journal_header_t *header; 532 533 header = (journal_header_t *) &nbh->b_data[0]; 534 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER); 535 } 536 537 BUFFER_TRACE(nbh, "marking dirty"); 538 set_buffer_uptodate(nbh); 539 mark_buffer_dirty(nbh); 540 BUFFER_TRACE(nbh, "marking uptodate"); 541 ++info->nr_replays; 542 /* ll_rw_block(WRITE, 1, &nbh); */ 543 unlock_buffer(nbh); 544 brelse(obh); 545 brelse(nbh); 546 } 547 548 skip_write: 549 tagp += tag_bytes; 550 if (!(flags & JFS_FLAG_SAME_UUID)) 551 tagp += 16; 552 553 if (flags & JFS_FLAG_LAST_TAG) 554 break; 555 } 556 557 brelse(bh); 558 continue; 559 560 case JFS_COMMIT_BLOCK: 561 jbd_debug(3, "Commit block for #%u found\n", 562 next_commit_ID); 563 /* How to differentiate between interrupted commit 564 * and journal corruption ? 565 * 566 * {nth transaction} 567 * Checksum Verification Failed 568 * | 569 * ____________________ 570 * | | 571 * async_commit sync_commit 572 * | | 573 * | GO TO NEXT "Journal Corruption" 574 * | TRANSACTION 575 * | 576 * {(n+1)th transanction} 577 * | 578 * _______|______________ 579 * | | 580 * Commit block found Commit block not found 581 * | | 582 * "Journal Corruption" | 583 * _____________|_________ 584 * | | 585 * nth trans corrupt OR nth trans 586 * and (n+1)th interrupted interrupted 587 * before commit block 588 * could reach the disk. 589 * (Cannot find the difference in above 590 * mentioned conditions. Hence assume 591 * "Interrupted Commit".) 592 */ 593 594 /* Found an expected commit block: if checksums 595 * are present verify them in PASS_SCAN; else not 596 * much to do other than move on to the next sequence 597 * number. */ 598 if (pass == PASS_SCAN && 599 JFS_HAS_COMPAT_FEATURE(journal, 600 JFS_FEATURE_COMPAT_CHECKSUM)) { 601 int chksum_err, chksum_seen; 602 struct commit_header *cbh = 603 (struct commit_header *)bh->b_data; 604 unsigned found_chksum = 605 be32_to_cpu(cbh->h_chksum[0]); 606 607 chksum_err = chksum_seen = 0; 608 609 jbd_debug(3, "Checksums %x %x\n", 610 crc32_sum, found_chksum); 611 if (info->end_transaction) { 612 journal->j_failed_commit = 613 info->end_transaction; 614 brelse(bh); 615 break; 616 } 617 618 if (crc32_sum == found_chksum && 619 cbh->h_chksum_type == JBD2_CRC32_CHKSUM && 620 cbh->h_chksum_size == 621 JBD2_CRC32_CHKSUM_SIZE) 622 chksum_seen = 1; 623 else if (!(cbh->h_chksum_type == 0 && 624 cbh->h_chksum_size == 0 && 625 found_chksum == 0 && 626 !chksum_seen)) 627 /* 628 * If fs is mounted using an old kernel and then 629 * kernel with journal_chksum is used then we 630 * get a situation where the journal flag has 631 * checksum flag set but checksums are not 632 * present i.e chksum = 0, in the individual 633 * commit blocks. 634 * Hence to avoid checksum failures, in this 635 * situation, this extra check is added. 636 */ 637 chksum_err = 1; 638 639 if (chksum_err) { 640 info->end_transaction = next_commit_ID; 641 jbd_debug(1, "Checksum_err %x %x\n", 642 crc32_sum, found_chksum); 643 if (!JFS_HAS_INCOMPAT_FEATURE(journal, 644 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 645 journal->j_failed_commit = 646 next_commit_ID; 647 brelse(bh); 648 break; 649 } 650 } 651 crc32_sum = ~0; 652 } 653 brelse(bh); 654 next_commit_ID++; 655 continue; 656 657 case JFS_REVOKE_BLOCK: 658 /* If we aren't in the REVOKE pass, then we can 659 * just skip over this block. */ 660 if (pass != PASS_REVOKE) { 661 brelse(bh); 662 continue; 663 } 664 665 err = scan_revoke_records(journal, bh, 666 next_commit_ID, info); 667 brelse(bh); 668 if (err) 669 goto failed; 670 continue; 671 672 default: 673 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 674 blocktype); 675 brelse(bh); 676 goto done; 677 } 678 } 679 680 done: 681 /* 682 * We broke out of the log scan loop: either we came to the 683 * known end of the log or we found an unexpected block in the 684 * log. If the latter happened, then we know that the "current" 685 * transaction marks the end of the valid log. 686 */ 687 688 if (pass == PASS_SCAN) { 689 if (!info->end_transaction) 690 info->end_transaction = next_commit_ID; 691 } else { 692 /* It's really bad news if different passes end up at 693 * different places (but possible due to IO errors). */ 694 if (info->end_transaction != next_commit_ID) { 695 printk (KERN_ERR "JBD: recovery pass %d ended at " 696 "transaction %u, expected %u\n", 697 pass, next_commit_ID, info->end_transaction); 698 if (!success) 699 success = -EIO; 700 } 701 } 702 703 return success; 704 705 failed: 706 return err; 707} 708 709 710/* Scan a revoke record, marking all blocks mentioned as revoked. */ 711 712static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 713 tid_t sequence, struct recovery_info *info) 714{ 715 journal_revoke_header_t *header; 716 int offset, max; 717 int record_len = 4; 718 719 header = (journal_revoke_header_t *) bh->b_data; 720 offset = sizeof(journal_revoke_header_t); 721 max = be32_to_cpu(header->r_count); 722 723 if (JFS_HAS_INCOMPAT_FEATURE(journal, JFS_FEATURE_INCOMPAT_64BIT)) 724 record_len = 8; 725 726 while (offset < max) { 727 unsigned long long blocknr; 728 int err; 729 730 if (record_len == 4) 731 blocknr = ext2fs_be32_to_cpu(*((__be32 *)(bh->b_data + 732 offset))); 733 else 734 blocknr = ext2fs_be64_to_cpu(*((__be64 *)(bh->b_data + 735 offset))); 736 offset += record_len; 737 err = journal_set_revoke(journal, blocknr, sequence); 738 if (err) 739 return err; 740 ++info->nr_revokes; 741 } 742 return 0; 743} 744