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/crc32.h> 24#endif 25 26/* 27 * Maintain information about the progress of the recovery job, so that 28 * the different passes can carry information between them. 29 */ 30struct recovery_info 31{ 32 tid_t start_transaction; 33 tid_t end_transaction; 34 35 int nr_replays; 36 int nr_revokes; 37 int nr_revoke_hits; 38}; 39 40enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY}; 41static int do_one_pass(journal_t *journal, 42 struct recovery_info *info, enum passtype pass); 43static int scan_revoke_records(journal_t *, struct buffer_head *, 44 tid_t, struct recovery_info *); 45 46#ifdef __KERNEL__ 47 48/* Release readahead buffers after use */ 49static void journal_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#define MAXBUF 8 69static int do_readahead(journal_t *journal, unsigned int start) 70{ 71 int err; 72 unsigned int max, nbufs, next; 73 unsigned long long blocknr; 74 struct buffer_head *bh; 75 76 struct buffer_head * bufs[MAXBUF]; 77 78 /* Do up to 128K of readahead */ 79 max = start + (128 * 1024 / journal->j_blocksize); 80 if (max > journal->j_maxlen) 81 max = journal->j_maxlen; 82 83 /* Do the readahead itself. We'll submit MAXBUF buffer_heads at 84 * a time to the block device IO layer. */ 85 86 nbufs = 0; 87 88 for (next = start; next < max; next++) { 89 err = jbd2_journal_bmap(journal, next, &blocknr); 90 91 if (err) { 92 printk (KERN_ERR "JBD: bad block at offset %u\n", 93 next); 94 goto failed; 95 } 96 97 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize); 98 if (!bh) { 99 err = -ENOMEM; 100 goto failed; 101 } 102 103 if (!buffer_uptodate(bh) && !buffer_locked(bh)) { 104 bufs[nbufs++] = bh; 105 if (nbufs == MAXBUF) { 106 ll_rw_block(READ, nbufs, bufs); 107 journal_brelse_array(bufs, nbufs); 108 nbufs = 0; 109 } 110 } else 111 brelse(bh); 112 } 113 114 if (nbufs) 115 ll_rw_block(READ, nbufs, bufs); 116 err = 0; 117 118failed: 119 if (nbufs) 120 journal_brelse_array(bufs, nbufs); 121 return err; 122} 123 124#endif /* __KERNEL__ */ 125 126 127/* 128 * Read a block from the journal 129 */ 130 131static int jread(struct buffer_head **bhp, journal_t *journal, 132 unsigned int offset) 133{ 134 int err; 135 unsigned long long blocknr; 136 struct buffer_head *bh; 137 138 *bhp = NULL; 139 140 if (offset >= journal->j_maxlen) { 141 printk(KERN_ERR "JBD: corrupted journal superblock\n"); 142 return -EIO; 143 } 144 145 err = jbd2_journal_bmap(journal, offset, &blocknr); 146 147 if (err) { 148 printk (KERN_ERR "JBD: bad block at offset %u\n", 149 offset); 150 return err; 151 } 152 153 bh = __getblk(journal->j_dev, blocknr, journal->j_blocksize); 154 if (!bh) 155 return -ENOMEM; 156 157 if (!buffer_uptodate(bh)) { 158 /* If this is a brand new buffer, start readahead. 159 Otherwise, we assume we are already reading it. */ 160 if (!buffer_req(bh)) 161 do_readahead(journal, offset); 162 wait_on_buffer(bh); 163 } 164 165 if (!buffer_uptodate(bh)) { 166 printk (KERN_ERR "JBD: Failed to read block at offset %u\n", 167 offset); 168 brelse(bh); 169 return -EIO; 170 } 171 172 *bhp = bh; 173 return 0; 174} 175 176 177/* 178 * Count the number of in-use tags in a journal descriptor block. 179 */ 180 181static int count_tags(journal_t *journal, struct buffer_head *bh) 182{ 183 char * tagp; 184 journal_block_tag_t * tag; 185 int nr = 0, size = journal->j_blocksize; 186 int tag_bytes = journal_tag_bytes(journal); 187 188 tagp = &bh->b_data[sizeof(journal_header_t)]; 189 190 while ((tagp - bh->b_data + tag_bytes) <= size) { 191 tag = (journal_block_tag_t *) tagp; 192 193 nr++; 194 tagp += tag_bytes; 195 if (!(tag->t_flags & cpu_to_be32(JBD2_FLAG_SAME_UUID))) 196 tagp += 16; 197 198 if (tag->t_flags & cpu_to_be32(JBD2_FLAG_LAST_TAG)) 199 break; 200 } 201 202 return nr; 203} 204 205 206/* Make sure we wrap around the log correctly! */ 207#define wrap(journal, var) \ 208do { \ 209 if (var >= (journal)->j_last) \ 210 var -= ((journal)->j_last - (journal)->j_first); \ 211} while (0) 212 213/** 214 * jbd2_journal_recover - recovers a on-disk journal 215 * @journal: the journal to recover 216 * 217 * The primary function for recovering the log contents when mounting a 218 * journaled device. 219 * 220 * Recovery is done in three passes. In the first pass, we look for the 221 * end of the log. In the second, we assemble the list of revoke 222 * blocks. In the third and final pass, we replay any un-revoked blocks 223 * in the log. 224 */ 225int jbd2_journal_recover(journal_t *journal) 226{ 227 int err, err2; 228 journal_superblock_t * sb; 229 230 struct recovery_info info; 231 232 memset(&info, 0, sizeof(info)); 233 sb = journal->j_superblock; 234 235 /* 236 * The journal superblock's s_start field (the current log head) 237 * is always zero if, and only if, the journal was cleanly 238 * unmounted. 239 */ 240 241 if (!sb->s_start) { 242 jbd_debug(1, "No recovery required, last transaction %d\n", 243 be32_to_cpu(sb->s_sequence)); 244 journal->j_transaction_sequence = be32_to_cpu(sb->s_sequence) + 1; 245 return 0; 246 } 247 248 err = do_one_pass(journal, &info, PASS_SCAN); 249 if (!err) 250 err = do_one_pass(journal, &info, PASS_REVOKE); 251 if (!err) 252 err = do_one_pass(journal, &info, PASS_REPLAY); 253 254 jbd_debug(1, "JBD: recovery, exit status %d, " 255 "recovered transactions %u to %u\n", 256 err, info.start_transaction, info.end_transaction); 257 jbd_debug(1, "JBD: Replayed %d and revoked %d/%d blocks\n", 258 info.nr_replays, info.nr_revoke_hits, info.nr_revokes); 259 260 /* Restart the log at the next transaction ID, thus invalidating 261 * any existing commit records in the log. */ 262 journal->j_transaction_sequence = ++info.end_transaction; 263 264 jbd2_journal_clear_revoke(journal); 265 err2 = sync_blockdev(journal->j_fs_dev); 266 if (!err) 267 err = err2; 268 269 return err; 270} 271 272/** 273 * jbd2_journal_skip_recovery - Start journal and wipe exiting records 274 * @journal: journal to startup 275 * 276 * Locate any valid recovery information from the journal and set up the 277 * journal structures in memory to ignore it (presumably because the 278 * caller has evidence that it is out of date). 279 * This function does'nt appear to be exorted.. 280 * 281 * We perform one pass over the journal to allow us to tell the user how 282 * much recovery information is being erased, and to let us initialise 283 * the journal transaction sequence numbers to the next unused ID. 284 */ 285int jbd2_journal_skip_recovery(journal_t *journal) 286{ 287 int err; 288 289 struct recovery_info info; 290 291 memset (&info, 0, sizeof(info)); 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 - 301 be32_to_cpu(journal->j_superblock->s_sequence); 302#endif 303 jbd_debug(1, 304 "JBD: ignoring %d transaction%s from the journal.\n", 305 dropped, (dropped == 1) ? "" : "s"); 306 journal->j_transaction_sequence = ++info.end_transaction; 307 } 308 309 journal->j_tail = 0; 310 return err; 311} 312 313static inline unsigned long long read_tag_block(int tag_bytes, journal_block_tag_t *tag) 314{ 315 unsigned long long block = be32_to_cpu(tag->t_blocknr); 316 if (tag_bytes > JBD2_TAG_SIZE32) 317 block |= (u64)be32_to_cpu(tag->t_blocknr_high) << 32; 318 return block; 319} 320 321/* 322 * calc_chksums calculates the checksums for the blocks described in the 323 * descriptor block. 324 */ 325static int calc_chksums(journal_t *journal, struct buffer_head *bh, 326 unsigned long *next_log_block, __u32 *crc32_sum) 327{ 328 int i, num_blks, err; 329 unsigned long io_block; 330 struct buffer_head *obh; 331 332 num_blks = count_tags(journal, bh); 333 /* Calculate checksum of the descriptor block. */ 334 *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size); 335 336 for (i = 0; i < num_blks; i++) { 337 io_block = (*next_log_block)++; 338 wrap(journal, *next_log_block); 339 err = jread(&obh, journal, io_block); 340 if (err) { 341 printk(KERN_ERR "JBD: IO error %d recovering block " 342 "%lu in log\n", err, io_block); 343 return 1; 344 } else { 345 *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data, 346 obh->b_size); 347 } 348 put_bh(obh); 349 } 350 return 0; 351} 352 353static int do_one_pass(journal_t *journal, 354 struct recovery_info *info, enum passtype pass) 355{ 356 unsigned int first_commit_ID, next_commit_ID; 357 unsigned long next_log_block; 358 int err, success = 0; 359 journal_superblock_t * sb; 360 journal_header_t * tmp; 361 struct buffer_head * bh; 362 unsigned int sequence; 363 int blocktype; 364 int tag_bytes = journal_tag_bytes(journal); 365 __u32 crc32_sum = ~0; /* Transactional Checksums */ 366 367 /* 368 * First thing is to establish what we expect to find in the log 369 * (in terms of transaction IDs), and where (in terms of log 370 * block offsets): query the superblock. 371 */ 372 373 sb = journal->j_superblock; 374 next_commit_ID = be32_to_cpu(sb->s_sequence); 375 next_log_block = be32_to_cpu(sb->s_start); 376 377 first_commit_ID = next_commit_ID; 378 if (pass == PASS_SCAN) 379 info->start_transaction = first_commit_ID; 380 381 jbd_debug(1, "Starting recovery pass %d\n", pass); 382 383 /* 384 * Now we walk through the log, transaction by transaction, 385 * making sure that each transaction has a commit block in the 386 * expected place. Each complete transaction gets replayed back 387 * into the main filesystem. 388 */ 389 390 while (1) { 391 int flags; 392 char * tagp; 393 journal_block_tag_t * tag; 394 struct buffer_head * obh; 395 struct buffer_head * nbh; 396 397 cond_resched(); 398 399 /* If we already know where to stop the log traversal, 400 * check right now that we haven't gone past the end of 401 * the log. */ 402 403 if (pass != PASS_SCAN) 404 if (tid_geq(next_commit_ID, info->end_transaction)) 405 break; 406 407 jbd_debug(2, "Scanning for sequence ID %u at %lu/%lu\n", 408 next_commit_ID, next_log_block, journal->j_last); 409 410 /* Skip over each chunk of the transaction looking 411 * either the next descriptor block or the final commit 412 * record. */ 413 414 jbd_debug(3, "JBD: checking block %ld\n", next_log_block); 415 err = jread(&bh, journal, next_log_block); 416 if (err) 417 goto failed; 418 419 next_log_block++; 420 wrap(journal, next_log_block); 421 422 /* What kind of buffer is it? 423 * 424 * If it is a descriptor block, check that it has the 425 * expected sequence number. Otherwise, we're all done 426 * here. */ 427 428 tmp = (journal_header_t *)bh->b_data; 429 430 if (tmp->h_magic != cpu_to_be32(JBD2_MAGIC_NUMBER)) { 431 brelse(bh); 432 break; 433 } 434 435 blocktype = be32_to_cpu(tmp->h_blocktype); 436 sequence = be32_to_cpu(tmp->h_sequence); 437 jbd_debug(3, "Found magic %d, sequence %d\n", 438 blocktype, sequence); 439 440 if (sequence != next_commit_ID) { 441 brelse(bh); 442 break; 443 } 444 445 /* OK, we have a valid descriptor block which matches 446 * all of the sequence number checks. What are we going 447 * to do with it? That depends on the pass... */ 448 449 switch(blocktype) { 450 case JBD2_DESCRIPTOR_BLOCK: 451 /* If it is a valid descriptor block, replay it 452 * in pass REPLAY; if journal_checksums enabled, then 453 * calculate checksums in PASS_SCAN, otherwise, 454 * just skip over the blocks it describes. */ 455 if (pass != PASS_REPLAY) { 456 if (pass == PASS_SCAN && 457 JBD2_HAS_COMPAT_FEATURE(journal, 458 JBD2_FEATURE_COMPAT_CHECKSUM) && 459 !info->end_transaction) { 460 if (calc_chksums(journal, bh, 461 &next_log_block, 462 &crc32_sum)) { 463 put_bh(bh); 464 break; 465 } 466 put_bh(bh); 467 continue; 468 } 469 next_log_block += count_tags(journal, bh); 470 wrap(journal, next_log_block); 471 put_bh(bh); 472 continue; 473 } 474 475 /* A descriptor block: we can now write all of 476 * the data blocks. Yay, useful work is finally 477 * getting done here! */ 478 479 tagp = &bh->b_data[sizeof(journal_header_t)]; 480 while ((tagp - bh->b_data + tag_bytes) 481 <= journal->j_blocksize) { 482 unsigned long io_block; 483 484 tag = (journal_block_tag_t *) tagp; 485 flags = be32_to_cpu(tag->t_flags); 486 487 io_block = next_log_block++; 488 wrap(journal, next_log_block); 489 err = jread(&obh, journal, io_block); 490 if (err) { 491 /* Recover what we can, but 492 * report failure at the end. */ 493 success = err; 494 printk (KERN_ERR 495 "JBD: IO error %d recovering " 496 "block %ld in log\n", 497 err, io_block); 498 } else { 499 unsigned long long blocknr; 500 501 J_ASSERT(obh != NULL); 502 blocknr = read_tag_block(tag_bytes, 503 tag); 504 505 /* If the block has been 506 * revoked, then we're all done 507 * here. */ 508 if (jbd2_journal_test_revoke 509 (journal, blocknr, 510 next_commit_ID)) { 511 brelse(obh); 512 ++info->nr_revoke_hits; 513 goto skip_write; 514 } 515 516 /* Find a buffer for the new 517 * data being restored */ 518 nbh = __getblk(journal->j_fs_dev, 519 blocknr, 520 journal->j_blocksize); 521 if (nbh == NULL) { 522 printk(KERN_ERR 523 "JBD: Out of memory " 524 "during recovery.\n"); 525 err = -ENOMEM; 526 brelse(bh); 527 brelse(obh); 528 goto failed; 529 } 530 531 lock_buffer(nbh); 532 memcpy(nbh->b_data, obh->b_data, 533 journal->j_blocksize); 534 if (flags & JBD2_FLAG_ESCAPE) { 535 *((__be32 *)nbh->b_data) = 536 cpu_to_be32(JBD2_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 & JBD2_FLAG_SAME_UUID)) 553 tagp += 16; 554 555 if (flags & JBD2_FLAG_LAST_TAG) 556 break; 557 } 558 559 brelse(bh); 560 continue; 561 562 case JBD2_COMMIT_BLOCK: 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 JBD2_HAS_COMPAT_FEATURE(journal, 600 JBD2_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 if (info->end_transaction) { 610 journal->j_failed_commit = 611 info->end_transaction; 612 brelse(bh); 613 break; 614 } 615 616 if (crc32_sum == found_chksum && 617 cbh->h_chksum_type == JBD2_CRC32_CHKSUM && 618 cbh->h_chksum_size == 619 JBD2_CRC32_CHKSUM_SIZE) 620 chksum_seen = 1; 621 else if (!(cbh->h_chksum_type == 0 && 622 cbh->h_chksum_size == 0 && 623 found_chksum == 0 && 624 !chksum_seen)) 625 /* 626 * If fs is mounted using an old kernel and then 627 * kernel with journal_chksum is used then we 628 * get a situation where the journal flag has 629 * checksum flag set but checksums are not 630 * present i.e chksum = 0, in the individual 631 * commit blocks. 632 * Hence to avoid checksum failures, in this 633 * situation, this extra check is added. 634 */ 635 chksum_err = 1; 636 637 if (chksum_err) { 638 info->end_transaction = next_commit_ID; 639 640 if (!JBD2_HAS_INCOMPAT_FEATURE(journal, 641 JBD2_FEATURE_INCOMPAT_ASYNC_COMMIT)){ 642 journal->j_failed_commit = 643 next_commit_ID; 644 brelse(bh); 645 break; 646 } 647 } 648 crc32_sum = ~0; 649 } 650 brelse(bh); 651 next_commit_ID++; 652 continue; 653 654 case JBD2_REVOKE_BLOCK: 655 /* If we aren't in the REVOKE pass, then we can 656 * just skip over this block. */ 657 if (pass != PASS_REVOKE) { 658 brelse(bh); 659 continue; 660 } 661 662 err = scan_revoke_records(journal, bh, 663 next_commit_ID, info); 664 brelse(bh); 665 if (err) 666 goto failed; 667 continue; 668 669 default: 670 jbd_debug(3, "Unrecognised magic %d, end of scan.\n", 671 blocktype); 672 brelse(bh); 673 goto done; 674 } 675 } 676 677 done: 678 /* 679 * We broke out of the log scan loop: either we came to the 680 * known end of the log or we found an unexpected block in the 681 * log. If the latter happened, then we know that the "current" 682 * transaction marks the end of the valid log. 683 */ 684 685 if (pass == PASS_SCAN) { 686 if (!info->end_transaction) 687 info->end_transaction = next_commit_ID; 688 } else { 689 /* It's really bad news if different passes end up at 690 * different places (but possible due to IO errors). */ 691 if (info->end_transaction != next_commit_ID) { 692 printk (KERN_ERR "JBD: recovery pass %d ended at " 693 "transaction %u, expected %u\n", 694 pass, next_commit_ID, info->end_transaction); 695 if (!success) 696 success = -EIO; 697 } 698 } 699 700 return success; 701 702 failed: 703 return err; 704} 705 706 707/* Scan a revoke record, marking all blocks mentioned as revoked. */ 708 709static int scan_revoke_records(journal_t *journal, struct buffer_head *bh, 710 tid_t sequence, struct recovery_info *info) 711{ 712 jbd2_journal_revoke_header_t *header; 713 int offset, max; 714 int record_len = 4; 715 716 header = (jbd2_journal_revoke_header_t *) bh->b_data; 717 offset = sizeof(jbd2_journal_revoke_header_t); 718 max = be32_to_cpu(header->r_count); 719 720 if (JBD2_HAS_INCOMPAT_FEATURE(journal, JBD2_FEATURE_INCOMPAT_64BIT)) 721 record_len = 8; 722 723 while (offset + record_len <= max) { 724 unsigned long long blocknr; 725 int err; 726 727 if (record_len == 4) 728 blocknr = be32_to_cpu(* ((__be32 *) (bh->b_data+offset))); 729 else 730 blocknr = be64_to_cpu(* ((__be64 *) (bh->b_data+offset))); 731 offset += record_len; 732 err = jbd2_journal_set_revoke(journal, blocknr, sequence); 733 if (err) 734 return err; 735 ++info->nr_revokes; 736 } 737 return 0; 738} 739