recovery.c revision d3f58ef166dea70f1e3e7a09b19c3ec964eb1730
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 journal_superblock_t *sb = journal->j_superblock; 298 299 int dropped = info.end_transaction - be32_to_cpu(sb->s_sequence); 300#endif 301 jbd_debug(1, 302 "JBD: ignoring %d transaction%s from the journal.\n", 303 dropped, (dropped == 1) ? "" : "s"); 304 journal->j_transaction_sequence = ++info.end_transaction; 305 } 306 307 journal->j_tail = 0; 308 return err; 309} 310 311static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 312{ 313 unsigned long long block = be32_to_cpu(tag->t_blocknr); 314 if (tag_bytes > JBD_TAG_SIZE32) 315 block |= (__u64)be32_to_cpu(tag->t_blocknr_high) << 32; 316 return block; 317} 318 319/* 320 * calc_chksums calculates the checksums for the blocks described in the 321 * descriptor block. 322 */ 323static int calc_chksums(journal_t *journal, struct buffer_head *bh, 324 unsigned long long *next_log_block, __u32 *crc32_sum) 325{ 326 int i, num_blks, err; 327 unsigned long long io_block; 328 struct buffer_head *obh; 329 330 num_blks = count_tags(journal, bh); 331 /* Calculate checksum of the descriptor block. */ 332 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 333 334 for (i = 0; i < num_blks; i++) { 335 io_block = (*next_log_block)++; 336 wrap(journal, *next_log_block); 337 err = jread(&obh, journal, io_block); 338 if (err) { 339 printk(KERN_ERR "JBD: IO error %d recovering block " 340 "%llu in log\n", err, io_block); 341 return 1; 342 } else { 343 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 344 obh->b_size); 345 } 346 brelse(obh); 347 } 348 return 0; 349} 350 351static int do_one_pass(journal_t *journal, 352 struct recovery_info *info, enum passtype pass) 353{ 354 unsigned int first_commit_ID, next_commit_ID; 355 unsigned long long next_log_block; 356 int err, success = 0; 357 journal_superblock_t * sb; 358 journal_header_t * tmp; 359 struct buffer_head * bh; 360 unsigned int sequence; 361 int blocktype; 362 int tag_bytes = journal_tag_bytes(journal); 363 __u32 crc32_sum = ~0; /* Transactional Checksums */ 364 365 /* 366 * First thing is to establish what we expect to find in the log 367 * (in terms of transaction IDs), and where (in terms of log 368 * block offsets): query the superblock. 369 */ 370 371 sb = journal->j_superblock; 372 next_commit_ID = be32_to_cpu(sb->s_sequence); 373 next_log_block = be32_to_cpu(sb->s_start); 374 375 first_commit_ID = next_commit_ID; 376 if (pass == PASS_SCAN) 377 info->start_transaction = first_commit_ID; 378 379 jbd_debug(1, "Starting recovery pass %d\n", pass); 380 381 /* 382 * Now we walk through the log, transaction by transaction, 383 * making sure that each transaction has a commit block in the 384 * expected place. Each complete transaction gets replayed back 385 * into the main filesystem. 386 */ 387 388 while (1) { 389 int flags; 390 char * tagp; 391 journal_block_tag_t * tag; 392 struct buffer_head * obh; 393 struct buffer_head * nbh; 394 395 cond_resched(); 396 397 /* If we already know where to stop the log traversal, 398 * check right now that we haven't gone past the end of 399 * the log. */ 400 401 if (pass != PASS_SCAN) 402 if (tid_geq(next_commit_ID, info->end_transaction)) 403 break; 404 405 jbd_debug(2, "Scanning for sequence ID %u at %llu/%lu\n", 406 next_commit_ID, next_log_block, journal->j_last); 407 408 /* Skip over each chunk of the transaction looking 409 * either the next descriptor block or the final commit 410 * record. */ 411 412 jbd_debug(3, "JBD: checking block %llu\n", next_log_block); 413 err = jread(&bh, journal, next_log_block); 414 if (err) 415 goto failed; 416 417 next_log_block++; 418 wrap(journal, next_log_block); 419 420 /* What kind of buffer is it? 421 * 422 * If it is a descriptor block, check that it has the 423 * expected sequence number. Otherwise, we're all done 424 * here. */ 425 426 tmp = (journal_header_t *)bh->b_data; 427 428 if (tmp->h_magic != cpu_to_be32(JFS_MAGIC_NUMBER)) { 429 brelse(bh); 430 break; 431 } 432 433 blocktype = be32_to_cpu(tmp->h_blocktype); 434 sequence = be32_to_cpu(tmp->h_sequence); 435 jbd_debug(3, "Found magic %d, sequence %d\n", 436 blocktype, sequence); 437 438 if (sequence != next_commit_ID) { 439 brelse(bh); 440 break; 441 } 442 443 /* OK, we have a valid descriptor block which matches 444 * all of the sequence number checks. What are we going 445 * to do with it? That depends on the pass... */ 446 447 switch(blocktype) { 448 case JFS_DESCRIPTOR_BLOCK: 449 /* If it is a valid descriptor block, replay it 450 * in pass REPLAY; if journal_checksums enabled, then 451 * calculate checksums in PASS_SCAN, otherwise, 452 * just skip over the blocks it describes. */ 453 if (pass != PASS_REPLAY) { 454 if (pass == PASS_SCAN && 455 JFS_HAS_COMPAT_FEATURE(journal, 456 JFS_FEATURE_COMPAT_CHECKSUM) && 457 !info->end_transaction) { 458 if (calc_chksums(journal, bh, 459 &next_log_block, 460 &crc32_sum)) { 461 brelse(bh); 462 break; 463 } 464 brelse(bh); 465 continue; 466 } 467 next_log_block += count_tags(journal, bh); 468 wrap(journal, next_log_block); 469 brelse(bh); 470 continue; 471 } 472 473 /* A descriptor block: we can now write all of 474 * the data blocks. Yay, useful work is finally 475 * getting done here! */ 476 477 tagp = &bh->b_data[sizeof(journal_header_t)]; 478 while ((tagp - bh->b_data + tag_bytes) 479 <= journal->j_blocksize) { 480 unsigned long long io_block; 481 482 tag = (journal_block_tag_t *) tagp; 483 flags = be32_to_cpu(tag->t_flags); 484 485 io_block = next_log_block++; 486 wrap(journal, next_log_block); 487 err = jread(&obh, journal, io_block); 488 if (err) { 489 /* Recover what we can, but 490 * report failure at the end. */ 491 success = err; 492 printk (KERN_ERR 493 "JBD: IO error %d recovering " 494 "block %llu in log\n", 495 err, io_block); 496 } else { 497 unsigned long long blocknr; 498 499 J_ASSERT(obh != NULL); 500 blocknr = read_tag_block(tag_bytes, 501 tag); 502 503 /* If the block has been 504 * revoked, then we're all done 505 * here. */ 506 if (journal_test_revoke 507 (journal, blocknr, 508 next_commit_ID)) { 509 brelse(obh); 510 ++info->nr_revoke_hits; 511 goto skip_write; 512 } 513 514 /* Find a buffer for the new 515 * data being restored */ 516 nbh = __getblk(journal->j_fs_dev, 517 blocknr, 518 journal->j_blocksize); 519 if (nbh == NULL) { 520 printk(KERN_ERR 521 "JBD: Out of memory " 522 "during recovery.\n"); 523 err = -ENOMEM; 524 brelse(bh); 525 brelse(obh); 526 goto failed; 527 } 528 529 lock_buffer(nbh); 530 memcpy(nbh->b_data, obh->b_data, 531 journal->j_blocksize); 532 if (flags & JFS_FLAG_ESCAPE) { 533 journal_header_t *header; 534 535 header = (journal_header_t *) &nbh->b_data[0]; 536 header->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER); 537 } 538 539 BUFFER_TRACE(nbh, "marking dirty"); 540 set_buffer_uptodate(nbh); 541 mark_buffer_dirty(nbh); 542 BUFFER_TRACE(nbh, "marking uptodate"); 543 ++info->nr_replays; 544 /* ll_rw_block(WRITE, 1, &nbh); */ 545 unlock_buffer(nbh); 546 brelse(obh); 547 brelse(nbh); 548 } 549 550 skip_write: 551 tagp += tag_bytes; 552 if (!(flags & JFS_FLAG_SAME_UUID)) 553 tagp += 16; 554 555 if (flags & JFS_FLAG_LAST_TAG) 556 break; 557 } 558 559 brelse(bh); 560 continue; 561 562 case JFS_COMMIT_BLOCK: 563 jbd_debug(3, "Commit block for #%u found\n", 564 next_commit_ID); 565 /* How to differentiate between interrupted commit 566 * and journal corruption ? 567 * 568 * {nth transaction} 569 * Checksum Verification Failed 570 * | 571 * ____________________ 572 * | | 573 * async_commit sync_commit 574 * | | 575 * | GO TO NEXT "Journal Corruption" 576 * | TRANSACTION 577 * | 578 * {(n+1)th transanction} 579 * | 580 * _______|______________ 581 * | | 582 * Commit block found Commit block not found 583 * | | 584 * "Journal Corruption" | 585 * _____________|_________ 586 * | | 587 * nth trans corrupt OR nth trans 588 * and (n+1)th interrupted interrupted 589 * before commit block 590 * could reach the disk. 591 * (Cannot find the difference in above 592 * mentioned conditions. Hence assume 593 * "Interrupted Commit".) 594 */ 595 596 /* Found an expected commit block: if checksums 597 * are present verify them in PASS_SCAN; else not 598 * much to do other than move on to the next sequence 599 * number. */ 600 if (pass == PASS_SCAN && 601 JFS_HAS_COMPAT_FEATURE(journal, 602 JFS_FEATURE_COMPAT_CHECKSUM)) { 603 int chksum_err, chksum_seen; 604 struct commit_header *cbh = 605 (struct commit_header *)bh->b_data; 606 unsigned found_chksum = 607 be32_to_cpu(cbh->h_chksum[0]); 608 609 chksum_err = chksum_seen = 0; 610 611 jbd_debug(3, "Checksums %x %x\n", 612 crc32_sum, found_chksum); 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 jbd_debug(1, "Checksum_err %x %x\n", 644 crc32_sum, found_chksum); 645 if (!JFS_HAS_INCOMPAT_FEATURE(journal, 646 JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 647 journal->j_failed_commit = 648 next_commit_ID; 649 brelse(bh); 650 break; 651 } 652 } 653 crc32_sum = ~0; 654 } 655 brelse(bh); 656 next_commit_ID++; 657 continue; 658 659 case JFS_REVOKE_BLOCK: 660 /* If we aren't in the REVOKE pass, then we can 661 * just skip over this block. */ 662 if (pass != PASS_REVOKE) { 663 brelse(bh); 664 continue; 665 } 666 667 err = scan_revoke_records(journal, bh, 668 next_commit_ID, info); 669 brelse(bh); 670 if (err) 671 goto failed; 672 continue; 673 674 default: 675 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 676 blocktype); 677 brelse(bh); 678 goto done; 679 } 680 } 681 682 done: 683 /* 684 * We broke out of the log scan loop: either we came to the 685 * known end of the log or we found an unexpected block in the 686 * log. If the latter happened, then we know that the "current" 687 * transaction marks the end of the valid log. 688 */ 689 690 if (pass == PASS_SCAN) { 691 if (!info->end_transaction) 692 info->end_transaction = next_commit_ID; 693 } else { 694 /* It's really bad news if different passes end up at 695 * different places (but possible due to IO errors). */ 696 if (info->end_transaction != next_commit_ID) { 697 printk (KERN_ERR "JBD: recovery pass %d ended at " 698 "transaction %u, expected %u\n", 699 pass, next_commit_ID, info->end_transaction); 700 if (!success) 701 success = -EIO; 702 } 703 } 704 705 return success; 706 707 failed: 708 return err; 709} 710 711 712/* Scan a revoke record, marking all blocks mentioned as revoked. */ 713 714static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 715 tid_t sequence, struct recovery_info *info) 716{ 717 journal_revoke_header_t *header; 718 int offset, max; 719 int record_len = 4; 720 721 header = (journal_revoke_header_t *) bh->b_data; 722 offset = sizeof(journal_revoke_header_t); 723 max = be32_to_cpu(header->r_count); 724 725 if (JFS_HAS_INCOMPAT_FEATURE(journal, JFS_FEATURE_INCOMPAT_64BIT)) 726 record_len = 8; 727 728 while (offset < max) { 729 unsigned long long blocknr; 730 int err; 731 732 if (record_len == 4) 733 blocknr = ext2fs_be32_to_cpu(*((__be32 *)(bh->b_data + 734 offset))); 735 else 736 blocknr = ext2fs_be64_to_cpu(*((__be64 *)(bh->b_data + 737 offset))); 738 offset += record_len; 739 err = journal_set_revoke(journal, blocknr, sequence); 740 if (err) 741 return err; 742 ++info->nr_revokes; 743 } 744 return 0; 745} 746