scrub.c revision aa1b8cd409f05e1489ec77ff219eff6ed4b801b8
1/* 2 * Copyright (C) 2011, 2012 STRATO. All rights reserved. 3 * 4 * This program is free software; you can redistribute it and/or 5 * modify it under the terms of the GNU General Public 6 * License v2 as published by the Free Software Foundation. 7 * 8 * This program is distributed in the hope that it will be useful, 9 * but WITHOUT ANY WARRANTY; without even the implied warranty of 10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 11 * General Public License for more details. 12 * 13 * You should have received a copy of the GNU General Public 14 * License along with this program; if not, write to the 15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 16 * Boston, MA 021110-1307, USA. 17 */ 18 19#include <linux/blkdev.h> 20#include <linux/ratelimit.h> 21#include "ctree.h" 22#include "volumes.h" 23#include "disk-io.h" 24#include "ordered-data.h" 25#include "transaction.h" 26#include "backref.h" 27#include "extent_io.h" 28#include "check-integrity.h" 29#include "rcu-string.h" 30 31/* 32 * This is only the first step towards a full-features scrub. It reads all 33 * extent and super block and verifies the checksums. In case a bad checksum 34 * is found or the extent cannot be read, good data will be written back if 35 * any can be found. 36 * 37 * Future enhancements: 38 * - In case an unrepairable extent is encountered, track which files are 39 * affected and report them 40 * - track and record media errors, throw out bad devices 41 * - add a mode to also read unallocated space 42 */ 43 44struct scrub_block; 45struct scrub_ctx; 46 47#define SCRUB_PAGES_PER_BIO 16 /* 64k per bio */ 48#define SCRUB_BIOS_PER_CTX 16 /* 1 MB per device in flight */ 49 50/* 51 * the following value times PAGE_SIZE needs to be large enough to match the 52 * largest node/leaf/sector size that shall be supported. 53 * Values larger than BTRFS_STRIPE_LEN are not supported. 54 */ 55#define SCRUB_MAX_PAGES_PER_BLOCK 16 /* 64k per node/leaf/sector */ 56 57struct scrub_page { 58 struct scrub_block *sblock; 59 struct page *page; 60 struct btrfs_device *dev; 61 u64 flags; /* extent flags */ 62 u64 generation; 63 u64 logical; 64 u64 physical; 65 atomic_t ref_count; 66 struct { 67 unsigned int mirror_num:8; 68 unsigned int have_csum:1; 69 unsigned int io_error:1; 70 }; 71 u8 csum[BTRFS_CSUM_SIZE]; 72}; 73 74struct scrub_bio { 75 int index; 76 struct scrub_ctx *sctx; 77 struct btrfs_device *dev; 78 struct bio *bio; 79 int err; 80 u64 logical; 81 u64 physical; 82 struct scrub_page *pagev[SCRUB_PAGES_PER_BIO]; 83 int page_count; 84 int next_free; 85 struct btrfs_work work; 86}; 87 88struct scrub_block { 89 struct scrub_page *pagev[SCRUB_MAX_PAGES_PER_BLOCK]; 90 int page_count; 91 atomic_t outstanding_pages; 92 atomic_t ref_count; /* free mem on transition to zero */ 93 struct scrub_ctx *sctx; 94 struct { 95 unsigned int header_error:1; 96 unsigned int checksum_error:1; 97 unsigned int no_io_error_seen:1; 98 unsigned int generation_error:1; /* also sets header_error */ 99 }; 100}; 101 102struct scrub_ctx { 103 struct scrub_bio *bios[SCRUB_BIOS_PER_CTX]; 104 struct btrfs_root *dev_root; 105 int first_free; 106 int curr; 107 atomic_t bios_in_flight; 108 atomic_t workers_pending; 109 spinlock_t list_lock; 110 wait_queue_head_t list_wait; 111 u16 csum_size; 112 struct list_head csum_list; 113 atomic_t cancel_req; 114 int readonly; 115 int pages_per_bio; /* <= SCRUB_PAGES_PER_BIO */ 116 u32 sectorsize; 117 u32 nodesize; 118 u32 leafsize; 119 /* 120 * statistics 121 */ 122 struct btrfs_scrub_progress stat; 123 spinlock_t stat_lock; 124}; 125 126struct scrub_fixup_nodatasum { 127 struct scrub_ctx *sctx; 128 struct btrfs_device *dev; 129 u64 logical; 130 struct btrfs_root *root; 131 struct btrfs_work work; 132 int mirror_num; 133}; 134 135struct scrub_warning { 136 struct btrfs_path *path; 137 u64 extent_item_size; 138 char *scratch_buf; 139 char *msg_buf; 140 const char *errstr; 141 sector_t sector; 142 u64 logical; 143 struct btrfs_device *dev; 144 int msg_bufsize; 145 int scratch_bufsize; 146}; 147 148 149static void scrub_pending_bio_inc(struct scrub_ctx *sctx); 150static void scrub_pending_bio_dec(struct scrub_ctx *sctx); 151static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx); 152static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx); 153static int scrub_handle_errored_block(struct scrub_block *sblock_to_check); 154static int scrub_setup_recheck_block(struct scrub_ctx *sctx, 155 struct btrfs_fs_info *fs_info, 156 u64 length, u64 logical, 157 struct scrub_block *sblock); 158static void scrub_recheck_block(struct btrfs_fs_info *fs_info, 159 struct scrub_block *sblock, int is_metadata, 160 int have_csum, u8 *csum, u64 generation, 161 u16 csum_size); 162static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info, 163 struct scrub_block *sblock, 164 int is_metadata, int have_csum, 165 const u8 *csum, u64 generation, 166 u16 csum_size); 167static void scrub_complete_bio_end_io(struct bio *bio, int err); 168static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad, 169 struct scrub_block *sblock_good, 170 int force_write); 171static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad, 172 struct scrub_block *sblock_good, 173 int page_num, int force_write); 174static int scrub_checksum_data(struct scrub_block *sblock); 175static int scrub_checksum_tree_block(struct scrub_block *sblock); 176static int scrub_checksum_super(struct scrub_block *sblock); 177static void scrub_block_get(struct scrub_block *sblock); 178static void scrub_block_put(struct scrub_block *sblock); 179static void scrub_page_get(struct scrub_page *spage); 180static void scrub_page_put(struct scrub_page *spage); 181static int scrub_add_page_to_bio(struct scrub_ctx *sctx, 182 struct scrub_page *spage); 183static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len, 184 u64 physical, struct btrfs_device *dev, u64 flags, 185 u64 gen, int mirror_num, u8 *csum, int force); 186static void scrub_bio_end_io(struct bio *bio, int err); 187static void scrub_bio_end_io_worker(struct btrfs_work *work); 188static void scrub_block_complete(struct scrub_block *sblock); 189 190 191static void scrub_pending_bio_inc(struct scrub_ctx *sctx) 192{ 193 atomic_inc(&sctx->bios_in_flight); 194} 195 196static void scrub_pending_bio_dec(struct scrub_ctx *sctx) 197{ 198 atomic_dec(&sctx->bios_in_flight); 199 wake_up(&sctx->list_wait); 200} 201 202/* 203 * used for workers that require transaction commits (i.e., for the 204 * NOCOW case) 205 */ 206static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx) 207{ 208 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info; 209 210 /* 211 * increment scrubs_running to prevent cancel requests from 212 * completing as long as a worker is running. we must also 213 * increment scrubs_paused to prevent deadlocking on pause 214 * requests used for transactions commits (as the worker uses a 215 * transaction context). it is safe to regard the worker 216 * as paused for all matters practical. effectively, we only 217 * avoid cancellation requests from completing. 218 */ 219 mutex_lock(&fs_info->scrub_lock); 220 atomic_inc(&fs_info->scrubs_running); 221 atomic_inc(&fs_info->scrubs_paused); 222 mutex_unlock(&fs_info->scrub_lock); 223 atomic_inc(&sctx->workers_pending); 224} 225 226/* used for workers that require transaction commits */ 227static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx) 228{ 229 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info; 230 231 /* 232 * see scrub_pending_trans_workers_inc() why we're pretending 233 * to be paused in the scrub counters 234 */ 235 mutex_lock(&fs_info->scrub_lock); 236 atomic_dec(&fs_info->scrubs_running); 237 atomic_dec(&fs_info->scrubs_paused); 238 mutex_unlock(&fs_info->scrub_lock); 239 atomic_dec(&sctx->workers_pending); 240 wake_up(&fs_info->scrub_pause_wait); 241 wake_up(&sctx->list_wait); 242} 243 244static void scrub_free_csums(struct scrub_ctx *sctx) 245{ 246 while (!list_empty(&sctx->csum_list)) { 247 struct btrfs_ordered_sum *sum; 248 sum = list_first_entry(&sctx->csum_list, 249 struct btrfs_ordered_sum, list); 250 list_del(&sum->list); 251 kfree(sum); 252 } 253} 254 255static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx) 256{ 257 int i; 258 259 if (!sctx) 260 return; 261 262 /* this can happen when scrub is cancelled */ 263 if (sctx->curr != -1) { 264 struct scrub_bio *sbio = sctx->bios[sctx->curr]; 265 266 for (i = 0; i < sbio->page_count; i++) { 267 BUG_ON(!sbio->pagev[i]); 268 BUG_ON(!sbio->pagev[i]->page); 269 scrub_block_put(sbio->pagev[i]->sblock); 270 } 271 bio_put(sbio->bio); 272 } 273 274 for (i = 0; i < SCRUB_BIOS_PER_CTX; ++i) { 275 struct scrub_bio *sbio = sctx->bios[i]; 276 277 if (!sbio) 278 break; 279 kfree(sbio); 280 } 281 282 scrub_free_csums(sctx); 283 kfree(sctx); 284} 285 286static noinline_for_stack 287struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev) 288{ 289 struct scrub_ctx *sctx; 290 int i; 291 struct btrfs_fs_info *fs_info = dev->dev_root->fs_info; 292 int pages_per_bio; 293 294 pages_per_bio = min_t(int, SCRUB_PAGES_PER_BIO, 295 bio_get_nr_vecs(dev->bdev)); 296 sctx = kzalloc(sizeof(*sctx), GFP_NOFS); 297 if (!sctx) 298 goto nomem; 299 sctx->pages_per_bio = pages_per_bio; 300 sctx->curr = -1; 301 sctx->dev_root = dev->dev_root; 302 for (i = 0; i < SCRUB_BIOS_PER_CTX; ++i) { 303 struct scrub_bio *sbio; 304 305 sbio = kzalloc(sizeof(*sbio), GFP_NOFS); 306 if (!sbio) 307 goto nomem; 308 sctx->bios[i] = sbio; 309 310 sbio->index = i; 311 sbio->sctx = sctx; 312 sbio->page_count = 0; 313 sbio->work.func = scrub_bio_end_io_worker; 314 315 if (i != SCRUB_BIOS_PER_CTX - 1) 316 sctx->bios[i]->next_free = i + 1; 317 else 318 sctx->bios[i]->next_free = -1; 319 } 320 sctx->first_free = 0; 321 sctx->nodesize = dev->dev_root->nodesize; 322 sctx->leafsize = dev->dev_root->leafsize; 323 sctx->sectorsize = dev->dev_root->sectorsize; 324 atomic_set(&sctx->bios_in_flight, 0); 325 atomic_set(&sctx->workers_pending, 0); 326 atomic_set(&sctx->cancel_req, 0); 327 sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy); 328 INIT_LIST_HEAD(&sctx->csum_list); 329 330 spin_lock_init(&sctx->list_lock); 331 spin_lock_init(&sctx->stat_lock); 332 init_waitqueue_head(&sctx->list_wait); 333 return sctx; 334 335nomem: 336 scrub_free_ctx(sctx); 337 return ERR_PTR(-ENOMEM); 338} 339 340static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root, void *ctx) 341{ 342 u64 isize; 343 u32 nlink; 344 int ret; 345 int i; 346 struct extent_buffer *eb; 347 struct btrfs_inode_item *inode_item; 348 struct scrub_warning *swarn = ctx; 349 struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info; 350 struct inode_fs_paths *ipath = NULL; 351 struct btrfs_root *local_root; 352 struct btrfs_key root_key; 353 354 root_key.objectid = root; 355 root_key.type = BTRFS_ROOT_ITEM_KEY; 356 root_key.offset = (u64)-1; 357 local_root = btrfs_read_fs_root_no_name(fs_info, &root_key); 358 if (IS_ERR(local_root)) { 359 ret = PTR_ERR(local_root); 360 goto err; 361 } 362 363 ret = inode_item_info(inum, 0, local_root, swarn->path); 364 if (ret) { 365 btrfs_release_path(swarn->path); 366 goto err; 367 } 368 369 eb = swarn->path->nodes[0]; 370 inode_item = btrfs_item_ptr(eb, swarn->path->slots[0], 371 struct btrfs_inode_item); 372 isize = btrfs_inode_size(eb, inode_item); 373 nlink = btrfs_inode_nlink(eb, inode_item); 374 btrfs_release_path(swarn->path); 375 376 ipath = init_ipath(4096, local_root, swarn->path); 377 if (IS_ERR(ipath)) { 378 ret = PTR_ERR(ipath); 379 ipath = NULL; 380 goto err; 381 } 382 ret = paths_from_inode(inum, ipath); 383 384 if (ret < 0) 385 goto err; 386 387 /* 388 * we deliberately ignore the bit ipath might have been too small to 389 * hold all of the paths here 390 */ 391 for (i = 0; i < ipath->fspath->elem_cnt; ++i) 392 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev " 393 "%s, sector %llu, root %llu, inode %llu, offset %llu, " 394 "length %llu, links %u (path: %s)\n", swarn->errstr, 395 swarn->logical, rcu_str_deref(swarn->dev->name), 396 (unsigned long long)swarn->sector, root, inum, offset, 397 min(isize - offset, (u64)PAGE_SIZE), nlink, 398 (char *)(unsigned long)ipath->fspath->val[i]); 399 400 free_ipath(ipath); 401 return 0; 402 403err: 404 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev " 405 "%s, sector %llu, root %llu, inode %llu, offset %llu: path " 406 "resolving failed with ret=%d\n", swarn->errstr, 407 swarn->logical, rcu_str_deref(swarn->dev->name), 408 (unsigned long long)swarn->sector, root, inum, offset, ret); 409 410 free_ipath(ipath); 411 return 0; 412} 413 414static void scrub_print_warning(const char *errstr, struct scrub_block *sblock) 415{ 416 struct btrfs_device *dev; 417 struct btrfs_fs_info *fs_info; 418 struct btrfs_path *path; 419 struct btrfs_key found_key; 420 struct extent_buffer *eb; 421 struct btrfs_extent_item *ei; 422 struct scrub_warning swarn; 423 unsigned long ptr = 0; 424 u64 extent_item_pos; 425 u64 flags = 0; 426 u64 ref_root; 427 u32 item_size; 428 u8 ref_level; 429 const int bufsize = 4096; 430 int ret; 431 432 WARN_ON(sblock->page_count < 1); 433 dev = sblock->pagev[0]->dev; 434 fs_info = sblock->sctx->dev_root->fs_info; 435 436 path = btrfs_alloc_path(); 437 438 swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS); 439 swarn.msg_buf = kmalloc(bufsize, GFP_NOFS); 440 swarn.sector = (sblock->pagev[0]->physical) >> 9; 441 swarn.logical = sblock->pagev[0]->logical; 442 swarn.errstr = errstr; 443 swarn.dev = NULL; 444 swarn.msg_bufsize = bufsize; 445 swarn.scratch_bufsize = bufsize; 446 447 if (!path || !swarn.scratch_buf || !swarn.msg_buf) 448 goto out; 449 450 ret = extent_from_logical(fs_info, swarn.logical, path, &found_key, 451 &flags); 452 if (ret < 0) 453 goto out; 454 455 extent_item_pos = swarn.logical - found_key.objectid; 456 swarn.extent_item_size = found_key.offset; 457 458 eb = path->nodes[0]; 459 ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item); 460 item_size = btrfs_item_size_nr(eb, path->slots[0]); 461 btrfs_release_path(path); 462 463 if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 464 do { 465 ret = tree_backref_for_extent(&ptr, eb, ei, item_size, 466 &ref_root, &ref_level); 467 printk_in_rcu(KERN_WARNING 468 "btrfs: %s at logical %llu on dev %s, " 469 "sector %llu: metadata %s (level %d) in tree " 470 "%llu\n", errstr, swarn.logical, 471 rcu_str_deref(dev->name), 472 (unsigned long long)swarn.sector, 473 ref_level ? "node" : "leaf", 474 ret < 0 ? -1 : ref_level, 475 ret < 0 ? -1 : ref_root); 476 } while (ret != 1); 477 } else { 478 swarn.path = path; 479 swarn.dev = dev; 480 iterate_extent_inodes(fs_info, found_key.objectid, 481 extent_item_pos, 1, 482 scrub_print_warning_inode, &swarn); 483 } 484 485out: 486 btrfs_free_path(path); 487 kfree(swarn.scratch_buf); 488 kfree(swarn.msg_buf); 489} 490 491static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *ctx) 492{ 493 struct page *page = NULL; 494 unsigned long index; 495 struct scrub_fixup_nodatasum *fixup = ctx; 496 int ret; 497 int corrected = 0; 498 struct btrfs_key key; 499 struct inode *inode = NULL; 500 u64 end = offset + PAGE_SIZE - 1; 501 struct btrfs_root *local_root; 502 503 key.objectid = root; 504 key.type = BTRFS_ROOT_ITEM_KEY; 505 key.offset = (u64)-1; 506 local_root = btrfs_read_fs_root_no_name(fixup->root->fs_info, &key); 507 if (IS_ERR(local_root)) 508 return PTR_ERR(local_root); 509 510 key.type = BTRFS_INODE_ITEM_KEY; 511 key.objectid = inum; 512 key.offset = 0; 513 inode = btrfs_iget(fixup->root->fs_info->sb, &key, local_root, NULL); 514 if (IS_ERR(inode)) 515 return PTR_ERR(inode); 516 517 index = offset >> PAGE_CACHE_SHIFT; 518 519 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS); 520 if (!page) { 521 ret = -ENOMEM; 522 goto out; 523 } 524 525 if (PageUptodate(page)) { 526 struct btrfs_fs_info *fs_info; 527 if (PageDirty(page)) { 528 /* 529 * we need to write the data to the defect sector. the 530 * data that was in that sector is not in memory, 531 * because the page was modified. we must not write the 532 * modified page to that sector. 533 * 534 * TODO: what could be done here: wait for the delalloc 535 * runner to write out that page (might involve 536 * COW) and see whether the sector is still 537 * referenced afterwards. 538 * 539 * For the meantime, we'll treat this error 540 * incorrectable, although there is a chance that a 541 * later scrub will find the bad sector again and that 542 * there's no dirty page in memory, then. 543 */ 544 ret = -EIO; 545 goto out; 546 } 547 fs_info = BTRFS_I(inode)->root->fs_info; 548 ret = repair_io_failure(fs_info, offset, PAGE_SIZE, 549 fixup->logical, page, 550 fixup->mirror_num); 551 unlock_page(page); 552 corrected = !ret; 553 } else { 554 /* 555 * we need to get good data first. the general readpage path 556 * will call repair_io_failure for us, we just have to make 557 * sure we read the bad mirror. 558 */ 559 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end, 560 EXTENT_DAMAGED, GFP_NOFS); 561 if (ret) { 562 /* set_extent_bits should give proper error */ 563 WARN_ON(ret > 0); 564 if (ret > 0) 565 ret = -EFAULT; 566 goto out; 567 } 568 569 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page, 570 btrfs_get_extent, 571 fixup->mirror_num); 572 wait_on_page_locked(page); 573 574 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset, 575 end, EXTENT_DAMAGED, 0, NULL); 576 if (!corrected) 577 clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end, 578 EXTENT_DAMAGED, GFP_NOFS); 579 } 580 581out: 582 if (page) 583 put_page(page); 584 if (inode) 585 iput(inode); 586 587 if (ret < 0) 588 return ret; 589 590 if (ret == 0 && corrected) { 591 /* 592 * we only need to call readpage for one of the inodes belonging 593 * to this extent. so make iterate_extent_inodes stop 594 */ 595 return 1; 596 } 597 598 return -EIO; 599} 600 601static void scrub_fixup_nodatasum(struct btrfs_work *work) 602{ 603 int ret; 604 struct scrub_fixup_nodatasum *fixup; 605 struct scrub_ctx *sctx; 606 struct btrfs_trans_handle *trans = NULL; 607 struct btrfs_fs_info *fs_info; 608 struct btrfs_path *path; 609 int uncorrectable = 0; 610 611 fixup = container_of(work, struct scrub_fixup_nodatasum, work); 612 sctx = fixup->sctx; 613 fs_info = fixup->root->fs_info; 614 615 path = btrfs_alloc_path(); 616 if (!path) { 617 spin_lock(&sctx->stat_lock); 618 ++sctx->stat.malloc_errors; 619 spin_unlock(&sctx->stat_lock); 620 uncorrectable = 1; 621 goto out; 622 } 623 624 trans = btrfs_join_transaction(fixup->root); 625 if (IS_ERR(trans)) { 626 uncorrectable = 1; 627 goto out; 628 } 629 630 /* 631 * the idea is to trigger a regular read through the standard path. we 632 * read a page from the (failed) logical address by specifying the 633 * corresponding copynum of the failed sector. thus, that readpage is 634 * expected to fail. 635 * that is the point where on-the-fly error correction will kick in 636 * (once it's finished) and rewrite the failed sector if a good copy 637 * can be found. 638 */ 639 ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info, 640 path, scrub_fixup_readpage, 641 fixup); 642 if (ret < 0) { 643 uncorrectable = 1; 644 goto out; 645 } 646 WARN_ON(ret != 1); 647 648 spin_lock(&sctx->stat_lock); 649 ++sctx->stat.corrected_errors; 650 spin_unlock(&sctx->stat_lock); 651 652out: 653 if (trans && !IS_ERR(trans)) 654 btrfs_end_transaction(trans, fixup->root); 655 if (uncorrectable) { 656 spin_lock(&sctx->stat_lock); 657 ++sctx->stat.uncorrectable_errors; 658 spin_unlock(&sctx->stat_lock); 659 660 printk_ratelimited_in_rcu(KERN_ERR 661 "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n", 662 (unsigned long long)fixup->logical, 663 rcu_str_deref(fixup->dev->name)); 664 } 665 666 btrfs_free_path(path); 667 kfree(fixup); 668 669 scrub_pending_trans_workers_dec(sctx); 670} 671 672/* 673 * scrub_handle_errored_block gets called when either verification of the 674 * pages failed or the bio failed to read, e.g. with EIO. In the latter 675 * case, this function handles all pages in the bio, even though only one 676 * may be bad. 677 * The goal of this function is to repair the errored block by using the 678 * contents of one of the mirrors. 679 */ 680static int scrub_handle_errored_block(struct scrub_block *sblock_to_check) 681{ 682 struct scrub_ctx *sctx = sblock_to_check->sctx; 683 struct btrfs_device *dev; 684 struct btrfs_fs_info *fs_info; 685 u64 length; 686 u64 logical; 687 u64 generation; 688 unsigned int failed_mirror_index; 689 unsigned int is_metadata; 690 unsigned int have_csum; 691 u8 *csum; 692 struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */ 693 struct scrub_block *sblock_bad; 694 int ret; 695 int mirror_index; 696 int page_num; 697 int success; 698 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL, 699 DEFAULT_RATELIMIT_BURST); 700 701 BUG_ON(sblock_to_check->page_count < 1); 702 fs_info = sctx->dev_root->fs_info; 703 length = sblock_to_check->page_count * PAGE_SIZE; 704 logical = sblock_to_check->pagev[0]->logical; 705 generation = sblock_to_check->pagev[0]->generation; 706 BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1); 707 failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1; 708 is_metadata = !(sblock_to_check->pagev[0]->flags & 709 BTRFS_EXTENT_FLAG_DATA); 710 have_csum = sblock_to_check->pagev[0]->have_csum; 711 csum = sblock_to_check->pagev[0]->csum; 712 dev = sblock_to_check->pagev[0]->dev; 713 714 /* 715 * read all mirrors one after the other. This includes to 716 * re-read the extent or metadata block that failed (that was 717 * the cause that this fixup code is called) another time, 718 * page by page this time in order to know which pages 719 * caused I/O errors and which ones are good (for all mirrors). 720 * It is the goal to handle the situation when more than one 721 * mirror contains I/O errors, but the errors do not 722 * overlap, i.e. the data can be repaired by selecting the 723 * pages from those mirrors without I/O error on the 724 * particular pages. One example (with blocks >= 2 * PAGE_SIZE) 725 * would be that mirror #1 has an I/O error on the first page, 726 * the second page is good, and mirror #2 has an I/O error on 727 * the second page, but the first page is good. 728 * Then the first page of the first mirror can be repaired by 729 * taking the first page of the second mirror, and the 730 * second page of the second mirror can be repaired by 731 * copying the contents of the 2nd page of the 1st mirror. 732 * One more note: if the pages of one mirror contain I/O 733 * errors, the checksum cannot be verified. In order to get 734 * the best data for repairing, the first attempt is to find 735 * a mirror without I/O errors and with a validated checksum. 736 * Only if this is not possible, the pages are picked from 737 * mirrors with I/O errors without considering the checksum. 738 * If the latter is the case, at the end, the checksum of the 739 * repaired area is verified in order to correctly maintain 740 * the statistics. 741 */ 742 743 sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS * 744 sizeof(*sblocks_for_recheck), 745 GFP_NOFS); 746 if (!sblocks_for_recheck) { 747 spin_lock(&sctx->stat_lock); 748 sctx->stat.malloc_errors++; 749 sctx->stat.read_errors++; 750 sctx->stat.uncorrectable_errors++; 751 spin_unlock(&sctx->stat_lock); 752 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS); 753 goto out; 754 } 755 756 /* setup the context, map the logical blocks and alloc the pages */ 757 ret = scrub_setup_recheck_block(sctx, fs_info, length, 758 logical, sblocks_for_recheck); 759 if (ret) { 760 spin_lock(&sctx->stat_lock); 761 sctx->stat.read_errors++; 762 sctx->stat.uncorrectable_errors++; 763 spin_unlock(&sctx->stat_lock); 764 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS); 765 goto out; 766 } 767 BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS); 768 sblock_bad = sblocks_for_recheck + failed_mirror_index; 769 770 /* build and submit the bios for the failed mirror, check checksums */ 771 scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum, 772 csum, generation, sctx->csum_size); 773 774 if (!sblock_bad->header_error && !sblock_bad->checksum_error && 775 sblock_bad->no_io_error_seen) { 776 /* 777 * the error disappeared after reading page by page, or 778 * the area was part of a huge bio and other parts of the 779 * bio caused I/O errors, or the block layer merged several 780 * read requests into one and the error is caused by a 781 * different bio (usually one of the two latter cases is 782 * the cause) 783 */ 784 spin_lock(&sctx->stat_lock); 785 sctx->stat.unverified_errors++; 786 spin_unlock(&sctx->stat_lock); 787 788 goto out; 789 } 790 791 if (!sblock_bad->no_io_error_seen) { 792 spin_lock(&sctx->stat_lock); 793 sctx->stat.read_errors++; 794 spin_unlock(&sctx->stat_lock); 795 if (__ratelimit(&_rs)) 796 scrub_print_warning("i/o error", sblock_to_check); 797 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS); 798 } else if (sblock_bad->checksum_error) { 799 spin_lock(&sctx->stat_lock); 800 sctx->stat.csum_errors++; 801 spin_unlock(&sctx->stat_lock); 802 if (__ratelimit(&_rs)) 803 scrub_print_warning("checksum error", sblock_to_check); 804 btrfs_dev_stat_inc_and_print(dev, 805 BTRFS_DEV_STAT_CORRUPTION_ERRS); 806 } else if (sblock_bad->header_error) { 807 spin_lock(&sctx->stat_lock); 808 sctx->stat.verify_errors++; 809 spin_unlock(&sctx->stat_lock); 810 if (__ratelimit(&_rs)) 811 scrub_print_warning("checksum/header error", 812 sblock_to_check); 813 if (sblock_bad->generation_error) 814 btrfs_dev_stat_inc_and_print(dev, 815 BTRFS_DEV_STAT_GENERATION_ERRS); 816 else 817 btrfs_dev_stat_inc_and_print(dev, 818 BTRFS_DEV_STAT_CORRUPTION_ERRS); 819 } 820 821 if (sctx->readonly) 822 goto did_not_correct_error; 823 824 if (!is_metadata && !have_csum) { 825 struct scrub_fixup_nodatasum *fixup_nodatasum; 826 827 /* 828 * !is_metadata and !have_csum, this means that the data 829 * might not be COW'ed, that it might be modified 830 * concurrently. The general strategy to work on the 831 * commit root does not help in the case when COW is not 832 * used. 833 */ 834 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS); 835 if (!fixup_nodatasum) 836 goto did_not_correct_error; 837 fixup_nodatasum->sctx = sctx; 838 fixup_nodatasum->dev = dev; 839 fixup_nodatasum->logical = logical; 840 fixup_nodatasum->root = fs_info->extent_root; 841 fixup_nodatasum->mirror_num = failed_mirror_index + 1; 842 scrub_pending_trans_workers_inc(sctx); 843 fixup_nodatasum->work.func = scrub_fixup_nodatasum; 844 btrfs_queue_worker(&fs_info->scrub_workers, 845 &fixup_nodatasum->work); 846 goto out; 847 } 848 849 /* 850 * now build and submit the bios for the other mirrors, check 851 * checksums. 852 * First try to pick the mirror which is completely without I/O 853 * errors and also does not have a checksum error. 854 * If one is found, and if a checksum is present, the full block 855 * that is known to contain an error is rewritten. Afterwards 856 * the block is known to be corrected. 857 * If a mirror is found which is completely correct, and no 858 * checksum is present, only those pages are rewritten that had 859 * an I/O error in the block to be repaired, since it cannot be 860 * determined, which copy of the other pages is better (and it 861 * could happen otherwise that a correct page would be 862 * overwritten by a bad one). 863 */ 864 for (mirror_index = 0; 865 mirror_index < BTRFS_MAX_MIRRORS && 866 sblocks_for_recheck[mirror_index].page_count > 0; 867 mirror_index++) { 868 struct scrub_block *sblock_other; 869 870 if (mirror_index == failed_mirror_index) 871 continue; 872 sblock_other = sblocks_for_recheck + mirror_index; 873 874 /* build and submit the bios, check checksums */ 875 scrub_recheck_block(fs_info, sblock_other, is_metadata, 876 have_csum, csum, generation, 877 sctx->csum_size); 878 879 if (!sblock_other->header_error && 880 !sblock_other->checksum_error && 881 sblock_other->no_io_error_seen) { 882 int force_write = is_metadata || have_csum; 883 884 ret = scrub_repair_block_from_good_copy(sblock_bad, 885 sblock_other, 886 force_write); 887 if (0 == ret) 888 goto corrected_error; 889 } 890 } 891 892 /* 893 * in case of I/O errors in the area that is supposed to be 894 * repaired, continue by picking good copies of those pages. 895 * Select the good pages from mirrors to rewrite bad pages from 896 * the area to fix. Afterwards verify the checksum of the block 897 * that is supposed to be repaired. This verification step is 898 * only done for the purpose of statistic counting and for the 899 * final scrub report, whether errors remain. 900 * A perfect algorithm could make use of the checksum and try 901 * all possible combinations of pages from the different mirrors 902 * until the checksum verification succeeds. For example, when 903 * the 2nd page of mirror #1 faces I/O errors, and the 2nd page 904 * of mirror #2 is readable but the final checksum test fails, 905 * then the 2nd page of mirror #3 could be tried, whether now 906 * the final checksum succeedes. But this would be a rare 907 * exception and is therefore not implemented. At least it is 908 * avoided that the good copy is overwritten. 909 * A more useful improvement would be to pick the sectors 910 * without I/O error based on sector sizes (512 bytes on legacy 911 * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one 912 * mirror could be repaired by taking 512 byte of a different 913 * mirror, even if other 512 byte sectors in the same PAGE_SIZE 914 * area are unreadable. 915 */ 916 917 /* can only fix I/O errors from here on */ 918 if (sblock_bad->no_io_error_seen) 919 goto did_not_correct_error; 920 921 success = 1; 922 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) { 923 struct scrub_page *page_bad = sblock_bad->pagev[page_num]; 924 925 if (!page_bad->io_error) 926 continue; 927 928 for (mirror_index = 0; 929 mirror_index < BTRFS_MAX_MIRRORS && 930 sblocks_for_recheck[mirror_index].page_count > 0; 931 mirror_index++) { 932 struct scrub_block *sblock_other = sblocks_for_recheck + 933 mirror_index; 934 struct scrub_page *page_other = sblock_other->pagev[ 935 page_num]; 936 937 if (!page_other->io_error) { 938 ret = scrub_repair_page_from_good_copy( 939 sblock_bad, sblock_other, page_num, 0); 940 if (0 == ret) { 941 page_bad->io_error = 0; 942 break; /* succeeded for this page */ 943 } 944 } 945 } 946 947 if (page_bad->io_error) { 948 /* did not find a mirror to copy the page from */ 949 success = 0; 950 } 951 } 952 953 if (success) { 954 if (is_metadata || have_csum) { 955 /* 956 * need to verify the checksum now that all 957 * sectors on disk are repaired (the write 958 * request for data to be repaired is on its way). 959 * Just be lazy and use scrub_recheck_block() 960 * which re-reads the data before the checksum 961 * is verified, but most likely the data comes out 962 * of the page cache. 963 */ 964 scrub_recheck_block(fs_info, sblock_bad, 965 is_metadata, have_csum, csum, 966 generation, sctx->csum_size); 967 if (!sblock_bad->header_error && 968 !sblock_bad->checksum_error && 969 sblock_bad->no_io_error_seen) 970 goto corrected_error; 971 else 972 goto did_not_correct_error; 973 } else { 974corrected_error: 975 spin_lock(&sctx->stat_lock); 976 sctx->stat.corrected_errors++; 977 spin_unlock(&sctx->stat_lock); 978 printk_ratelimited_in_rcu(KERN_ERR 979 "btrfs: fixed up error at logical %llu on dev %s\n", 980 (unsigned long long)logical, 981 rcu_str_deref(dev->name)); 982 } 983 } else { 984did_not_correct_error: 985 spin_lock(&sctx->stat_lock); 986 sctx->stat.uncorrectable_errors++; 987 spin_unlock(&sctx->stat_lock); 988 printk_ratelimited_in_rcu(KERN_ERR 989 "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n", 990 (unsigned long long)logical, 991 rcu_str_deref(dev->name)); 992 } 993 994out: 995 if (sblocks_for_recheck) { 996 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS; 997 mirror_index++) { 998 struct scrub_block *sblock = sblocks_for_recheck + 999 mirror_index; 1000 int page_index; 1001 1002 for (page_index = 0; page_index < sblock->page_count; 1003 page_index++) { 1004 sblock->pagev[page_index]->sblock = NULL; 1005 scrub_page_put(sblock->pagev[page_index]); 1006 } 1007 } 1008 kfree(sblocks_for_recheck); 1009 } 1010 1011 return 0; 1012} 1013 1014static int scrub_setup_recheck_block(struct scrub_ctx *sctx, 1015 struct btrfs_fs_info *fs_info, 1016 u64 length, u64 logical, 1017 struct scrub_block *sblocks_for_recheck) 1018{ 1019 int page_index; 1020 int mirror_index; 1021 int ret; 1022 1023 /* 1024 * note: the two members ref_count and outstanding_pages 1025 * are not used (and not set) in the blocks that are used for 1026 * the recheck procedure 1027 */ 1028 1029 page_index = 0; 1030 while (length > 0) { 1031 u64 sublen = min_t(u64, length, PAGE_SIZE); 1032 u64 mapped_length = sublen; 1033 struct btrfs_bio *bbio = NULL; 1034 1035 /* 1036 * with a length of PAGE_SIZE, each returned stripe 1037 * represents one mirror 1038 */ 1039 ret = btrfs_map_block(fs_info, WRITE, logical, &mapped_length, 1040 &bbio, 0); 1041 if (ret || !bbio || mapped_length < sublen) { 1042 kfree(bbio); 1043 return -EIO; 1044 } 1045 1046 BUG_ON(page_index >= SCRUB_PAGES_PER_BIO); 1047 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes; 1048 mirror_index++) { 1049 struct scrub_block *sblock; 1050 struct scrub_page *page; 1051 1052 if (mirror_index >= BTRFS_MAX_MIRRORS) 1053 continue; 1054 1055 sblock = sblocks_for_recheck + mirror_index; 1056 sblock->sctx = sctx; 1057 page = kzalloc(sizeof(*page), GFP_NOFS); 1058 if (!page) { 1059leave_nomem: 1060 spin_lock(&sctx->stat_lock); 1061 sctx->stat.malloc_errors++; 1062 spin_unlock(&sctx->stat_lock); 1063 kfree(bbio); 1064 return -ENOMEM; 1065 } 1066 scrub_page_get(page); 1067 sblock->pagev[page_index] = page; 1068 page->logical = logical; 1069 page->physical = bbio->stripes[mirror_index].physical; 1070 /* for missing devices, dev->bdev is NULL */ 1071 page->dev = bbio->stripes[mirror_index].dev; 1072 page->mirror_num = mirror_index + 1; 1073 sblock->page_count++; 1074 page->page = alloc_page(GFP_NOFS); 1075 if (!page->page) 1076 goto leave_nomem; 1077 } 1078 kfree(bbio); 1079 length -= sublen; 1080 logical += sublen; 1081 page_index++; 1082 } 1083 1084 return 0; 1085} 1086 1087/* 1088 * this function will check the on disk data for checksum errors, header 1089 * errors and read I/O errors. If any I/O errors happen, the exact pages 1090 * which are errored are marked as being bad. The goal is to enable scrub 1091 * to take those pages that are not errored from all the mirrors so that 1092 * the pages that are errored in the just handled mirror can be repaired. 1093 */ 1094static void scrub_recheck_block(struct btrfs_fs_info *fs_info, 1095 struct scrub_block *sblock, int is_metadata, 1096 int have_csum, u8 *csum, u64 generation, 1097 u16 csum_size) 1098{ 1099 int page_num; 1100 1101 sblock->no_io_error_seen = 1; 1102 sblock->header_error = 0; 1103 sblock->checksum_error = 0; 1104 1105 for (page_num = 0; page_num < sblock->page_count; page_num++) { 1106 struct bio *bio; 1107 struct scrub_page *page = sblock->pagev[page_num]; 1108 DECLARE_COMPLETION_ONSTACK(complete); 1109 1110 if (page->dev->bdev == NULL) { 1111 page->io_error = 1; 1112 sblock->no_io_error_seen = 0; 1113 continue; 1114 } 1115 1116 WARN_ON(!page->page); 1117 bio = bio_alloc(GFP_NOFS, 1); 1118 if (!bio) { 1119 page->io_error = 1; 1120 sblock->no_io_error_seen = 0; 1121 continue; 1122 } 1123 bio->bi_bdev = page->dev->bdev; 1124 bio->bi_sector = page->physical >> 9; 1125 bio->bi_end_io = scrub_complete_bio_end_io; 1126 bio->bi_private = &complete; 1127 1128 bio_add_page(bio, page->page, PAGE_SIZE, 0); 1129 btrfsic_submit_bio(READ, bio); 1130 1131 /* this will also unplug the queue */ 1132 wait_for_completion(&complete); 1133 1134 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags); 1135 if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) 1136 sblock->no_io_error_seen = 0; 1137 bio_put(bio); 1138 } 1139 1140 if (sblock->no_io_error_seen) 1141 scrub_recheck_block_checksum(fs_info, sblock, is_metadata, 1142 have_csum, csum, generation, 1143 csum_size); 1144 1145 return; 1146} 1147 1148static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info, 1149 struct scrub_block *sblock, 1150 int is_metadata, int have_csum, 1151 const u8 *csum, u64 generation, 1152 u16 csum_size) 1153{ 1154 int page_num; 1155 u8 calculated_csum[BTRFS_CSUM_SIZE]; 1156 u32 crc = ~(u32)0; 1157 struct btrfs_root *root = fs_info->extent_root; 1158 void *mapped_buffer; 1159 1160 WARN_ON(!sblock->pagev[0]->page); 1161 if (is_metadata) { 1162 struct btrfs_header *h; 1163 1164 mapped_buffer = kmap_atomic(sblock->pagev[0]->page); 1165 h = (struct btrfs_header *)mapped_buffer; 1166 1167 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) || 1168 memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) || 1169 memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid, 1170 BTRFS_UUID_SIZE)) { 1171 sblock->header_error = 1; 1172 } else if (generation != le64_to_cpu(h->generation)) { 1173 sblock->header_error = 1; 1174 sblock->generation_error = 1; 1175 } 1176 csum = h->csum; 1177 } else { 1178 if (!have_csum) 1179 return; 1180 1181 mapped_buffer = kmap_atomic(sblock->pagev[0]->page); 1182 } 1183 1184 for (page_num = 0;;) { 1185 if (page_num == 0 && is_metadata) 1186 crc = btrfs_csum_data(root, 1187 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE, 1188 crc, PAGE_SIZE - BTRFS_CSUM_SIZE); 1189 else 1190 crc = btrfs_csum_data(root, mapped_buffer, crc, 1191 PAGE_SIZE); 1192 1193 kunmap_atomic(mapped_buffer); 1194 page_num++; 1195 if (page_num >= sblock->page_count) 1196 break; 1197 WARN_ON(!sblock->pagev[page_num]->page); 1198 1199 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page); 1200 } 1201 1202 btrfs_csum_final(crc, calculated_csum); 1203 if (memcmp(calculated_csum, csum, csum_size)) 1204 sblock->checksum_error = 1; 1205} 1206 1207static void scrub_complete_bio_end_io(struct bio *bio, int err) 1208{ 1209 complete((struct completion *)bio->bi_private); 1210} 1211 1212static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad, 1213 struct scrub_block *sblock_good, 1214 int force_write) 1215{ 1216 int page_num; 1217 int ret = 0; 1218 1219 for (page_num = 0; page_num < sblock_bad->page_count; page_num++) { 1220 int ret_sub; 1221 1222 ret_sub = scrub_repair_page_from_good_copy(sblock_bad, 1223 sblock_good, 1224 page_num, 1225 force_write); 1226 if (ret_sub) 1227 ret = ret_sub; 1228 } 1229 1230 return ret; 1231} 1232 1233static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad, 1234 struct scrub_block *sblock_good, 1235 int page_num, int force_write) 1236{ 1237 struct scrub_page *page_bad = sblock_bad->pagev[page_num]; 1238 struct scrub_page *page_good = sblock_good->pagev[page_num]; 1239 1240 BUG_ON(page_bad->page == NULL); 1241 BUG_ON(page_good->page == NULL); 1242 if (force_write || sblock_bad->header_error || 1243 sblock_bad->checksum_error || page_bad->io_error) { 1244 struct bio *bio; 1245 int ret; 1246 DECLARE_COMPLETION_ONSTACK(complete); 1247 1248 bio = bio_alloc(GFP_NOFS, 1); 1249 if (!bio) 1250 return -EIO; 1251 bio->bi_bdev = page_bad->dev->bdev; 1252 bio->bi_sector = page_bad->physical >> 9; 1253 bio->bi_end_io = scrub_complete_bio_end_io; 1254 bio->bi_private = &complete; 1255 1256 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0); 1257 if (PAGE_SIZE != ret) { 1258 bio_put(bio); 1259 return -EIO; 1260 } 1261 btrfsic_submit_bio(WRITE, bio); 1262 1263 /* this will also unplug the queue */ 1264 wait_for_completion(&complete); 1265 if (!bio_flagged(bio, BIO_UPTODATE)) { 1266 btrfs_dev_stat_inc_and_print(page_bad->dev, 1267 BTRFS_DEV_STAT_WRITE_ERRS); 1268 bio_put(bio); 1269 return -EIO; 1270 } 1271 bio_put(bio); 1272 } 1273 1274 return 0; 1275} 1276 1277static void scrub_checksum(struct scrub_block *sblock) 1278{ 1279 u64 flags; 1280 int ret; 1281 1282 WARN_ON(sblock->page_count < 1); 1283 flags = sblock->pagev[0]->flags; 1284 ret = 0; 1285 if (flags & BTRFS_EXTENT_FLAG_DATA) 1286 ret = scrub_checksum_data(sblock); 1287 else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) 1288 ret = scrub_checksum_tree_block(sblock); 1289 else if (flags & BTRFS_EXTENT_FLAG_SUPER) 1290 (void)scrub_checksum_super(sblock); 1291 else 1292 WARN_ON(1); 1293 if (ret) 1294 scrub_handle_errored_block(sblock); 1295} 1296 1297static int scrub_checksum_data(struct scrub_block *sblock) 1298{ 1299 struct scrub_ctx *sctx = sblock->sctx; 1300 u8 csum[BTRFS_CSUM_SIZE]; 1301 u8 *on_disk_csum; 1302 struct page *page; 1303 void *buffer; 1304 u32 crc = ~(u32)0; 1305 int fail = 0; 1306 struct btrfs_root *root = sctx->dev_root; 1307 u64 len; 1308 int index; 1309 1310 BUG_ON(sblock->page_count < 1); 1311 if (!sblock->pagev[0]->have_csum) 1312 return 0; 1313 1314 on_disk_csum = sblock->pagev[0]->csum; 1315 page = sblock->pagev[0]->page; 1316 buffer = kmap_atomic(page); 1317 1318 len = sctx->sectorsize; 1319 index = 0; 1320 for (;;) { 1321 u64 l = min_t(u64, len, PAGE_SIZE); 1322 1323 crc = btrfs_csum_data(root, buffer, crc, l); 1324 kunmap_atomic(buffer); 1325 len -= l; 1326 if (len == 0) 1327 break; 1328 index++; 1329 BUG_ON(index >= sblock->page_count); 1330 BUG_ON(!sblock->pagev[index]->page); 1331 page = sblock->pagev[index]->page; 1332 buffer = kmap_atomic(page); 1333 } 1334 1335 btrfs_csum_final(crc, csum); 1336 if (memcmp(csum, on_disk_csum, sctx->csum_size)) 1337 fail = 1; 1338 1339 return fail; 1340} 1341 1342static int scrub_checksum_tree_block(struct scrub_block *sblock) 1343{ 1344 struct scrub_ctx *sctx = sblock->sctx; 1345 struct btrfs_header *h; 1346 struct btrfs_root *root = sctx->dev_root; 1347 struct btrfs_fs_info *fs_info = root->fs_info; 1348 u8 calculated_csum[BTRFS_CSUM_SIZE]; 1349 u8 on_disk_csum[BTRFS_CSUM_SIZE]; 1350 struct page *page; 1351 void *mapped_buffer; 1352 u64 mapped_size; 1353 void *p; 1354 u32 crc = ~(u32)0; 1355 int fail = 0; 1356 int crc_fail = 0; 1357 u64 len; 1358 int index; 1359 1360 BUG_ON(sblock->page_count < 1); 1361 page = sblock->pagev[0]->page; 1362 mapped_buffer = kmap_atomic(page); 1363 h = (struct btrfs_header *)mapped_buffer; 1364 memcpy(on_disk_csum, h->csum, sctx->csum_size); 1365 1366 /* 1367 * we don't use the getter functions here, as we 1368 * a) don't have an extent buffer and 1369 * b) the page is already kmapped 1370 */ 1371 1372 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr)) 1373 ++fail; 1374 1375 if (sblock->pagev[0]->generation != le64_to_cpu(h->generation)) 1376 ++fail; 1377 1378 if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE)) 1379 ++fail; 1380 1381 if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid, 1382 BTRFS_UUID_SIZE)) 1383 ++fail; 1384 1385 BUG_ON(sctx->nodesize != sctx->leafsize); 1386 len = sctx->nodesize - BTRFS_CSUM_SIZE; 1387 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE; 1388 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE; 1389 index = 0; 1390 for (;;) { 1391 u64 l = min_t(u64, len, mapped_size); 1392 1393 crc = btrfs_csum_data(root, p, crc, l); 1394 kunmap_atomic(mapped_buffer); 1395 len -= l; 1396 if (len == 0) 1397 break; 1398 index++; 1399 BUG_ON(index >= sblock->page_count); 1400 BUG_ON(!sblock->pagev[index]->page); 1401 page = sblock->pagev[index]->page; 1402 mapped_buffer = kmap_atomic(page); 1403 mapped_size = PAGE_SIZE; 1404 p = mapped_buffer; 1405 } 1406 1407 btrfs_csum_final(crc, calculated_csum); 1408 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size)) 1409 ++crc_fail; 1410 1411 return fail || crc_fail; 1412} 1413 1414static int scrub_checksum_super(struct scrub_block *sblock) 1415{ 1416 struct btrfs_super_block *s; 1417 struct scrub_ctx *sctx = sblock->sctx; 1418 struct btrfs_root *root = sctx->dev_root; 1419 struct btrfs_fs_info *fs_info = root->fs_info; 1420 u8 calculated_csum[BTRFS_CSUM_SIZE]; 1421 u8 on_disk_csum[BTRFS_CSUM_SIZE]; 1422 struct page *page; 1423 void *mapped_buffer; 1424 u64 mapped_size; 1425 void *p; 1426 u32 crc = ~(u32)0; 1427 int fail_gen = 0; 1428 int fail_cor = 0; 1429 u64 len; 1430 int index; 1431 1432 BUG_ON(sblock->page_count < 1); 1433 page = sblock->pagev[0]->page; 1434 mapped_buffer = kmap_atomic(page); 1435 s = (struct btrfs_super_block *)mapped_buffer; 1436 memcpy(on_disk_csum, s->csum, sctx->csum_size); 1437 1438 if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr)) 1439 ++fail_cor; 1440 1441 if (sblock->pagev[0]->generation != le64_to_cpu(s->generation)) 1442 ++fail_gen; 1443 1444 if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE)) 1445 ++fail_cor; 1446 1447 len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE; 1448 mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE; 1449 p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE; 1450 index = 0; 1451 for (;;) { 1452 u64 l = min_t(u64, len, mapped_size); 1453 1454 crc = btrfs_csum_data(root, p, crc, l); 1455 kunmap_atomic(mapped_buffer); 1456 len -= l; 1457 if (len == 0) 1458 break; 1459 index++; 1460 BUG_ON(index >= sblock->page_count); 1461 BUG_ON(!sblock->pagev[index]->page); 1462 page = sblock->pagev[index]->page; 1463 mapped_buffer = kmap_atomic(page); 1464 mapped_size = PAGE_SIZE; 1465 p = mapped_buffer; 1466 } 1467 1468 btrfs_csum_final(crc, calculated_csum); 1469 if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size)) 1470 ++fail_cor; 1471 1472 if (fail_cor + fail_gen) { 1473 /* 1474 * if we find an error in a super block, we just report it. 1475 * They will get written with the next transaction commit 1476 * anyway 1477 */ 1478 spin_lock(&sctx->stat_lock); 1479 ++sctx->stat.super_errors; 1480 spin_unlock(&sctx->stat_lock); 1481 if (fail_cor) 1482 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev, 1483 BTRFS_DEV_STAT_CORRUPTION_ERRS); 1484 else 1485 btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev, 1486 BTRFS_DEV_STAT_GENERATION_ERRS); 1487 } 1488 1489 return fail_cor + fail_gen; 1490} 1491 1492static void scrub_block_get(struct scrub_block *sblock) 1493{ 1494 atomic_inc(&sblock->ref_count); 1495} 1496 1497static void scrub_block_put(struct scrub_block *sblock) 1498{ 1499 if (atomic_dec_and_test(&sblock->ref_count)) { 1500 int i; 1501 1502 for (i = 0; i < sblock->page_count; i++) 1503 scrub_page_put(sblock->pagev[i]); 1504 kfree(sblock); 1505 } 1506} 1507 1508static void scrub_page_get(struct scrub_page *spage) 1509{ 1510 atomic_inc(&spage->ref_count); 1511} 1512 1513static void scrub_page_put(struct scrub_page *spage) 1514{ 1515 if (atomic_dec_and_test(&spage->ref_count)) { 1516 if (spage->page) 1517 __free_page(spage->page); 1518 kfree(spage); 1519 } 1520} 1521 1522static void scrub_submit(struct scrub_ctx *sctx) 1523{ 1524 struct scrub_bio *sbio; 1525 1526 if (sctx->curr == -1) 1527 return; 1528 1529 sbio = sctx->bios[sctx->curr]; 1530 sctx->curr = -1; 1531 scrub_pending_bio_inc(sctx); 1532 1533 btrfsic_submit_bio(READ, sbio->bio); 1534} 1535 1536static int scrub_add_page_to_bio(struct scrub_ctx *sctx, 1537 struct scrub_page *spage) 1538{ 1539 struct scrub_block *sblock = spage->sblock; 1540 struct scrub_bio *sbio; 1541 int ret; 1542 1543again: 1544 /* 1545 * grab a fresh bio or wait for one to become available 1546 */ 1547 while (sctx->curr == -1) { 1548 spin_lock(&sctx->list_lock); 1549 sctx->curr = sctx->first_free; 1550 if (sctx->curr != -1) { 1551 sctx->first_free = sctx->bios[sctx->curr]->next_free; 1552 sctx->bios[sctx->curr]->next_free = -1; 1553 sctx->bios[sctx->curr]->page_count = 0; 1554 spin_unlock(&sctx->list_lock); 1555 } else { 1556 spin_unlock(&sctx->list_lock); 1557 wait_event(sctx->list_wait, sctx->first_free != -1); 1558 } 1559 } 1560 sbio = sctx->bios[sctx->curr]; 1561 if (sbio->page_count == 0) { 1562 struct bio *bio; 1563 1564 sbio->physical = spage->physical; 1565 sbio->logical = spage->logical; 1566 sbio->dev = spage->dev; 1567 bio = sbio->bio; 1568 if (!bio) { 1569 bio = bio_alloc(GFP_NOFS, sctx->pages_per_bio); 1570 if (!bio) 1571 return -ENOMEM; 1572 sbio->bio = bio; 1573 } 1574 1575 bio->bi_private = sbio; 1576 bio->bi_end_io = scrub_bio_end_io; 1577 bio->bi_bdev = sbio->dev->bdev; 1578 bio->bi_sector = sbio->physical >> 9; 1579 sbio->err = 0; 1580 } else if (sbio->physical + sbio->page_count * PAGE_SIZE != 1581 spage->physical || 1582 sbio->logical + sbio->page_count * PAGE_SIZE != 1583 spage->logical || 1584 sbio->dev != spage->dev) { 1585 scrub_submit(sctx); 1586 goto again; 1587 } 1588 1589 sbio->pagev[sbio->page_count] = spage; 1590 ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0); 1591 if (ret != PAGE_SIZE) { 1592 if (sbio->page_count < 1) { 1593 bio_put(sbio->bio); 1594 sbio->bio = NULL; 1595 return -EIO; 1596 } 1597 scrub_submit(sctx); 1598 goto again; 1599 } 1600 1601 scrub_block_get(sblock); /* one for the added page */ 1602 atomic_inc(&sblock->outstanding_pages); 1603 sbio->page_count++; 1604 if (sbio->page_count == sctx->pages_per_bio) 1605 scrub_submit(sctx); 1606 1607 return 0; 1608} 1609 1610static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len, 1611 u64 physical, struct btrfs_device *dev, u64 flags, 1612 u64 gen, int mirror_num, u8 *csum, int force) 1613{ 1614 struct scrub_block *sblock; 1615 int index; 1616 1617 sblock = kzalloc(sizeof(*sblock), GFP_NOFS); 1618 if (!sblock) { 1619 spin_lock(&sctx->stat_lock); 1620 sctx->stat.malloc_errors++; 1621 spin_unlock(&sctx->stat_lock); 1622 return -ENOMEM; 1623 } 1624 1625 /* one ref inside this function, plus one for each page added to 1626 * a bio later on */ 1627 atomic_set(&sblock->ref_count, 1); 1628 sblock->sctx = sctx; 1629 sblock->no_io_error_seen = 1; 1630 1631 for (index = 0; len > 0; index++) { 1632 struct scrub_page *spage; 1633 u64 l = min_t(u64, len, PAGE_SIZE); 1634 1635 spage = kzalloc(sizeof(*spage), GFP_NOFS); 1636 if (!spage) { 1637leave_nomem: 1638 spin_lock(&sctx->stat_lock); 1639 sctx->stat.malloc_errors++; 1640 spin_unlock(&sctx->stat_lock); 1641 scrub_block_put(sblock); 1642 return -ENOMEM; 1643 } 1644 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK); 1645 scrub_page_get(spage); 1646 sblock->pagev[index] = spage; 1647 spage->sblock = sblock; 1648 spage->dev = dev; 1649 spage->flags = flags; 1650 spage->generation = gen; 1651 spage->logical = logical; 1652 spage->physical = physical; 1653 spage->mirror_num = mirror_num; 1654 if (csum) { 1655 spage->have_csum = 1; 1656 memcpy(spage->csum, csum, sctx->csum_size); 1657 } else { 1658 spage->have_csum = 0; 1659 } 1660 sblock->page_count++; 1661 spage->page = alloc_page(GFP_NOFS); 1662 if (!spage->page) 1663 goto leave_nomem; 1664 len -= l; 1665 logical += l; 1666 physical += l; 1667 } 1668 1669 WARN_ON(sblock->page_count == 0); 1670 for (index = 0; index < sblock->page_count; index++) { 1671 struct scrub_page *spage = sblock->pagev[index]; 1672 int ret; 1673 1674 ret = scrub_add_page_to_bio(sctx, spage); 1675 if (ret) { 1676 scrub_block_put(sblock); 1677 return ret; 1678 } 1679 } 1680 1681 if (force) 1682 scrub_submit(sctx); 1683 1684 /* last one frees, either here or in bio completion for last page */ 1685 scrub_block_put(sblock); 1686 return 0; 1687} 1688 1689static void scrub_bio_end_io(struct bio *bio, int err) 1690{ 1691 struct scrub_bio *sbio = bio->bi_private; 1692 struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info; 1693 1694 sbio->err = err; 1695 sbio->bio = bio; 1696 1697 btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work); 1698} 1699 1700static void scrub_bio_end_io_worker(struct btrfs_work *work) 1701{ 1702 struct scrub_bio *sbio = container_of(work, struct scrub_bio, work); 1703 struct scrub_ctx *sctx = sbio->sctx; 1704 int i; 1705 1706 BUG_ON(sbio->page_count > SCRUB_PAGES_PER_BIO); 1707 if (sbio->err) { 1708 for (i = 0; i < sbio->page_count; i++) { 1709 struct scrub_page *spage = sbio->pagev[i]; 1710 1711 spage->io_error = 1; 1712 spage->sblock->no_io_error_seen = 0; 1713 } 1714 } 1715 1716 /* now complete the scrub_block items that have all pages completed */ 1717 for (i = 0; i < sbio->page_count; i++) { 1718 struct scrub_page *spage = sbio->pagev[i]; 1719 struct scrub_block *sblock = spage->sblock; 1720 1721 if (atomic_dec_and_test(&sblock->outstanding_pages)) 1722 scrub_block_complete(sblock); 1723 scrub_block_put(sblock); 1724 } 1725 1726 bio_put(sbio->bio); 1727 sbio->bio = NULL; 1728 spin_lock(&sctx->list_lock); 1729 sbio->next_free = sctx->first_free; 1730 sctx->first_free = sbio->index; 1731 spin_unlock(&sctx->list_lock); 1732 scrub_pending_bio_dec(sctx); 1733} 1734 1735static void scrub_block_complete(struct scrub_block *sblock) 1736{ 1737 if (!sblock->no_io_error_seen) 1738 scrub_handle_errored_block(sblock); 1739 else 1740 scrub_checksum(sblock); 1741} 1742 1743static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len, 1744 u8 *csum) 1745{ 1746 struct btrfs_ordered_sum *sum = NULL; 1747 int ret = 0; 1748 unsigned long i; 1749 unsigned long num_sectors; 1750 1751 while (!list_empty(&sctx->csum_list)) { 1752 sum = list_first_entry(&sctx->csum_list, 1753 struct btrfs_ordered_sum, list); 1754 if (sum->bytenr > logical) 1755 return 0; 1756 if (sum->bytenr + sum->len > logical) 1757 break; 1758 1759 ++sctx->stat.csum_discards; 1760 list_del(&sum->list); 1761 kfree(sum); 1762 sum = NULL; 1763 } 1764 if (!sum) 1765 return 0; 1766 1767 num_sectors = sum->len / sctx->sectorsize; 1768 for (i = 0; i < num_sectors; ++i) { 1769 if (sum->sums[i].bytenr == logical) { 1770 memcpy(csum, &sum->sums[i].sum, sctx->csum_size); 1771 ret = 1; 1772 break; 1773 } 1774 } 1775 if (ret && i == num_sectors - 1) { 1776 list_del(&sum->list); 1777 kfree(sum); 1778 } 1779 return ret; 1780} 1781 1782/* scrub extent tries to collect up to 64 kB for each bio */ 1783static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len, 1784 u64 physical, struct btrfs_device *dev, u64 flags, 1785 u64 gen, int mirror_num) 1786{ 1787 int ret; 1788 u8 csum[BTRFS_CSUM_SIZE]; 1789 u32 blocksize; 1790 1791 if (flags & BTRFS_EXTENT_FLAG_DATA) { 1792 blocksize = sctx->sectorsize; 1793 spin_lock(&sctx->stat_lock); 1794 sctx->stat.data_extents_scrubbed++; 1795 sctx->stat.data_bytes_scrubbed += len; 1796 spin_unlock(&sctx->stat_lock); 1797 } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) { 1798 BUG_ON(sctx->nodesize != sctx->leafsize); 1799 blocksize = sctx->nodesize; 1800 spin_lock(&sctx->stat_lock); 1801 sctx->stat.tree_extents_scrubbed++; 1802 sctx->stat.tree_bytes_scrubbed += len; 1803 spin_unlock(&sctx->stat_lock); 1804 } else { 1805 blocksize = sctx->sectorsize; 1806 BUG_ON(1); 1807 } 1808 1809 while (len) { 1810 u64 l = min_t(u64, len, blocksize); 1811 int have_csum = 0; 1812 1813 if (flags & BTRFS_EXTENT_FLAG_DATA) { 1814 /* push csums to sbio */ 1815 have_csum = scrub_find_csum(sctx, logical, l, csum); 1816 if (have_csum == 0) 1817 ++sctx->stat.no_csum; 1818 } 1819 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen, 1820 mirror_num, have_csum ? csum : NULL, 0); 1821 if (ret) 1822 return ret; 1823 len -= l; 1824 logical += l; 1825 physical += l; 1826 } 1827 return 0; 1828} 1829 1830static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx, 1831 struct map_lookup *map, 1832 struct btrfs_device *scrub_dev, 1833 int num, u64 base, u64 length) 1834{ 1835 struct btrfs_path *path; 1836 struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info; 1837 struct btrfs_root *root = fs_info->extent_root; 1838 struct btrfs_root *csum_root = fs_info->csum_root; 1839 struct btrfs_extent_item *extent; 1840 struct blk_plug plug; 1841 u64 flags; 1842 int ret; 1843 int slot; 1844 int i; 1845 u64 nstripes; 1846 struct extent_buffer *l; 1847 struct btrfs_key key; 1848 u64 physical; 1849 u64 logical; 1850 u64 generation; 1851 int mirror_num; 1852 struct reada_control *reada1; 1853 struct reada_control *reada2; 1854 struct btrfs_key key_start; 1855 struct btrfs_key key_end; 1856 u64 increment = map->stripe_len; 1857 u64 offset; 1858 1859 nstripes = length; 1860 offset = 0; 1861 do_div(nstripes, map->stripe_len); 1862 if (map->type & BTRFS_BLOCK_GROUP_RAID0) { 1863 offset = map->stripe_len * num; 1864 increment = map->stripe_len * map->num_stripes; 1865 mirror_num = 1; 1866 } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) { 1867 int factor = map->num_stripes / map->sub_stripes; 1868 offset = map->stripe_len * (num / map->sub_stripes); 1869 increment = map->stripe_len * factor; 1870 mirror_num = num % map->sub_stripes + 1; 1871 } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) { 1872 increment = map->stripe_len; 1873 mirror_num = num % map->num_stripes + 1; 1874 } else if (map->type & BTRFS_BLOCK_GROUP_DUP) { 1875 increment = map->stripe_len; 1876 mirror_num = num % map->num_stripes + 1; 1877 } else { 1878 increment = map->stripe_len; 1879 mirror_num = 1; 1880 } 1881 1882 path = btrfs_alloc_path(); 1883 if (!path) 1884 return -ENOMEM; 1885 1886 /* 1887 * work on commit root. The related disk blocks are static as 1888 * long as COW is applied. This means, it is save to rewrite 1889 * them to repair disk errors without any race conditions 1890 */ 1891 path->search_commit_root = 1; 1892 path->skip_locking = 1; 1893 1894 /* 1895 * trigger the readahead for extent tree csum tree and wait for 1896 * completion. During readahead, the scrub is officially paused 1897 * to not hold off transaction commits 1898 */ 1899 logical = base + offset; 1900 1901 wait_event(sctx->list_wait, 1902 atomic_read(&sctx->bios_in_flight) == 0); 1903 atomic_inc(&fs_info->scrubs_paused); 1904 wake_up(&fs_info->scrub_pause_wait); 1905 1906 /* FIXME it might be better to start readahead at commit root */ 1907 key_start.objectid = logical; 1908 key_start.type = BTRFS_EXTENT_ITEM_KEY; 1909 key_start.offset = (u64)0; 1910 key_end.objectid = base + offset + nstripes * increment; 1911 key_end.type = BTRFS_EXTENT_ITEM_KEY; 1912 key_end.offset = (u64)0; 1913 reada1 = btrfs_reada_add(root, &key_start, &key_end); 1914 1915 key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID; 1916 key_start.type = BTRFS_EXTENT_CSUM_KEY; 1917 key_start.offset = logical; 1918 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID; 1919 key_end.type = BTRFS_EXTENT_CSUM_KEY; 1920 key_end.offset = base + offset + nstripes * increment; 1921 reada2 = btrfs_reada_add(csum_root, &key_start, &key_end); 1922 1923 if (!IS_ERR(reada1)) 1924 btrfs_reada_wait(reada1); 1925 if (!IS_ERR(reada2)) 1926 btrfs_reada_wait(reada2); 1927 1928 mutex_lock(&fs_info->scrub_lock); 1929 while (atomic_read(&fs_info->scrub_pause_req)) { 1930 mutex_unlock(&fs_info->scrub_lock); 1931 wait_event(fs_info->scrub_pause_wait, 1932 atomic_read(&fs_info->scrub_pause_req) == 0); 1933 mutex_lock(&fs_info->scrub_lock); 1934 } 1935 atomic_dec(&fs_info->scrubs_paused); 1936 mutex_unlock(&fs_info->scrub_lock); 1937 wake_up(&fs_info->scrub_pause_wait); 1938 1939 /* 1940 * collect all data csums for the stripe to avoid seeking during 1941 * the scrub. This might currently (crc32) end up to be about 1MB 1942 */ 1943 blk_start_plug(&plug); 1944 1945 /* 1946 * now find all extents for each stripe and scrub them 1947 */ 1948 logical = base + offset; 1949 physical = map->stripes[num].physical; 1950 ret = 0; 1951 for (i = 0; i < nstripes; ++i) { 1952 /* 1953 * canceled? 1954 */ 1955 if (atomic_read(&fs_info->scrub_cancel_req) || 1956 atomic_read(&sctx->cancel_req)) { 1957 ret = -ECANCELED; 1958 goto out; 1959 } 1960 /* 1961 * check to see if we have to pause 1962 */ 1963 if (atomic_read(&fs_info->scrub_pause_req)) { 1964 /* push queued extents */ 1965 scrub_submit(sctx); 1966 wait_event(sctx->list_wait, 1967 atomic_read(&sctx->bios_in_flight) == 0); 1968 atomic_inc(&fs_info->scrubs_paused); 1969 wake_up(&fs_info->scrub_pause_wait); 1970 mutex_lock(&fs_info->scrub_lock); 1971 while (atomic_read(&fs_info->scrub_pause_req)) { 1972 mutex_unlock(&fs_info->scrub_lock); 1973 wait_event(fs_info->scrub_pause_wait, 1974 atomic_read(&fs_info->scrub_pause_req) == 0); 1975 mutex_lock(&fs_info->scrub_lock); 1976 } 1977 atomic_dec(&fs_info->scrubs_paused); 1978 mutex_unlock(&fs_info->scrub_lock); 1979 wake_up(&fs_info->scrub_pause_wait); 1980 } 1981 1982 ret = btrfs_lookup_csums_range(csum_root, logical, 1983 logical + map->stripe_len - 1, 1984 &sctx->csum_list, 1); 1985 if (ret) 1986 goto out; 1987 1988 key.objectid = logical; 1989 key.type = BTRFS_EXTENT_ITEM_KEY; 1990 key.offset = (u64)0; 1991 1992 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 1993 if (ret < 0) 1994 goto out; 1995 if (ret > 0) { 1996 ret = btrfs_previous_item(root, path, 0, 1997 BTRFS_EXTENT_ITEM_KEY); 1998 if (ret < 0) 1999 goto out; 2000 if (ret > 0) { 2001 /* there's no smaller item, so stick with the 2002 * larger one */ 2003 btrfs_release_path(path); 2004 ret = btrfs_search_slot(NULL, root, &key, 2005 path, 0, 0); 2006 if (ret < 0) 2007 goto out; 2008 } 2009 } 2010 2011 while (1) { 2012 l = path->nodes[0]; 2013 slot = path->slots[0]; 2014 if (slot >= btrfs_header_nritems(l)) { 2015 ret = btrfs_next_leaf(root, path); 2016 if (ret == 0) 2017 continue; 2018 if (ret < 0) 2019 goto out; 2020 2021 break; 2022 } 2023 btrfs_item_key_to_cpu(l, &key, slot); 2024 2025 if (key.objectid + key.offset <= logical) 2026 goto next; 2027 2028 if (key.objectid >= logical + map->stripe_len) 2029 break; 2030 2031 if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY) 2032 goto next; 2033 2034 extent = btrfs_item_ptr(l, slot, 2035 struct btrfs_extent_item); 2036 flags = btrfs_extent_flags(l, extent); 2037 generation = btrfs_extent_generation(l, extent); 2038 2039 if (key.objectid < logical && 2040 (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) { 2041 printk(KERN_ERR 2042 "btrfs scrub: tree block %llu spanning " 2043 "stripes, ignored. logical=%llu\n", 2044 (unsigned long long)key.objectid, 2045 (unsigned long long)logical); 2046 goto next; 2047 } 2048 2049 /* 2050 * trim extent to this stripe 2051 */ 2052 if (key.objectid < logical) { 2053 key.offset -= logical - key.objectid; 2054 key.objectid = logical; 2055 } 2056 if (key.objectid + key.offset > 2057 logical + map->stripe_len) { 2058 key.offset = logical + map->stripe_len - 2059 key.objectid; 2060 } 2061 2062 ret = scrub_extent(sctx, key.objectid, key.offset, 2063 key.objectid - logical + physical, 2064 scrub_dev, flags, generation, 2065 mirror_num); 2066 if (ret) 2067 goto out; 2068 2069next: 2070 path->slots[0]++; 2071 } 2072 btrfs_release_path(path); 2073 logical += increment; 2074 physical += map->stripe_len; 2075 spin_lock(&sctx->stat_lock); 2076 sctx->stat.last_physical = physical; 2077 spin_unlock(&sctx->stat_lock); 2078 } 2079 /* push queued extents */ 2080 scrub_submit(sctx); 2081 2082out: 2083 blk_finish_plug(&plug); 2084 btrfs_free_path(path); 2085 return ret < 0 ? ret : 0; 2086} 2087 2088static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx, 2089 struct btrfs_device *scrub_dev, 2090 u64 chunk_tree, u64 chunk_objectid, 2091 u64 chunk_offset, u64 length, 2092 u64 dev_offset) 2093{ 2094 struct btrfs_mapping_tree *map_tree = 2095 &sctx->dev_root->fs_info->mapping_tree; 2096 struct map_lookup *map; 2097 struct extent_map *em; 2098 int i; 2099 int ret = -EINVAL; 2100 2101 read_lock(&map_tree->map_tree.lock); 2102 em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1); 2103 read_unlock(&map_tree->map_tree.lock); 2104 2105 if (!em) 2106 return -EINVAL; 2107 2108 map = (struct map_lookup *)em->bdev; 2109 if (em->start != chunk_offset) 2110 goto out; 2111 2112 if (em->len < length) 2113 goto out; 2114 2115 for (i = 0; i < map->num_stripes; ++i) { 2116 if (map->stripes[i].dev->bdev == scrub_dev->bdev && 2117 map->stripes[i].physical == dev_offset) { 2118 ret = scrub_stripe(sctx, map, scrub_dev, i, 2119 chunk_offset, length); 2120 if (ret) 2121 goto out; 2122 } 2123 } 2124out: 2125 free_extent_map(em); 2126 2127 return ret; 2128} 2129 2130static noinline_for_stack 2131int scrub_enumerate_chunks(struct scrub_ctx *sctx, 2132 struct btrfs_device *scrub_dev, u64 start, u64 end) 2133{ 2134 struct btrfs_dev_extent *dev_extent = NULL; 2135 struct btrfs_path *path; 2136 struct btrfs_root *root = sctx->dev_root; 2137 struct btrfs_fs_info *fs_info = root->fs_info; 2138 u64 length; 2139 u64 chunk_tree; 2140 u64 chunk_objectid; 2141 u64 chunk_offset; 2142 int ret; 2143 int slot; 2144 struct extent_buffer *l; 2145 struct btrfs_key key; 2146 struct btrfs_key found_key; 2147 struct btrfs_block_group_cache *cache; 2148 2149 path = btrfs_alloc_path(); 2150 if (!path) 2151 return -ENOMEM; 2152 2153 path->reada = 2; 2154 path->search_commit_root = 1; 2155 path->skip_locking = 1; 2156 2157 key.objectid = scrub_dev->devid; 2158 key.offset = 0ull; 2159 key.type = BTRFS_DEV_EXTENT_KEY; 2160 2161 while (1) { 2162 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0); 2163 if (ret < 0) 2164 break; 2165 if (ret > 0) { 2166 if (path->slots[0] >= 2167 btrfs_header_nritems(path->nodes[0])) { 2168 ret = btrfs_next_leaf(root, path); 2169 if (ret) 2170 break; 2171 } 2172 } 2173 2174 l = path->nodes[0]; 2175 slot = path->slots[0]; 2176 2177 btrfs_item_key_to_cpu(l, &found_key, slot); 2178 2179 if (found_key.objectid != scrub_dev->devid) 2180 break; 2181 2182 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY) 2183 break; 2184 2185 if (found_key.offset >= end) 2186 break; 2187 2188 if (found_key.offset < key.offset) 2189 break; 2190 2191 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent); 2192 length = btrfs_dev_extent_length(l, dev_extent); 2193 2194 if (found_key.offset + length <= start) { 2195 key.offset = found_key.offset + length; 2196 btrfs_release_path(path); 2197 continue; 2198 } 2199 2200 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent); 2201 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent); 2202 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent); 2203 2204 /* 2205 * get a reference on the corresponding block group to prevent 2206 * the chunk from going away while we scrub it 2207 */ 2208 cache = btrfs_lookup_block_group(fs_info, chunk_offset); 2209 if (!cache) { 2210 ret = -ENOENT; 2211 break; 2212 } 2213 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid, 2214 chunk_offset, length, found_key.offset); 2215 btrfs_put_block_group(cache); 2216 if (ret) 2217 break; 2218 2219 key.offset = found_key.offset + length; 2220 btrfs_release_path(path); 2221 } 2222 2223 btrfs_free_path(path); 2224 2225 /* 2226 * ret can still be 1 from search_slot or next_leaf, 2227 * that's not an error 2228 */ 2229 return ret < 0 ? ret : 0; 2230} 2231 2232static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx, 2233 struct btrfs_device *scrub_dev) 2234{ 2235 int i; 2236 u64 bytenr; 2237 u64 gen; 2238 int ret; 2239 struct btrfs_root *root = sctx->dev_root; 2240 2241 if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) 2242 return -EIO; 2243 2244 gen = root->fs_info->last_trans_committed; 2245 2246 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) { 2247 bytenr = btrfs_sb_offset(i); 2248 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes) 2249 break; 2250 2251 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr, 2252 scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i, 2253 NULL, 1); 2254 if (ret) 2255 return ret; 2256 } 2257 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0); 2258 2259 return 0; 2260} 2261 2262/* 2263 * get a reference count on fs_info->scrub_workers. start worker if necessary 2264 */ 2265static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info) 2266{ 2267 int ret = 0; 2268 2269 mutex_lock(&fs_info->scrub_lock); 2270 if (fs_info->scrub_workers_refcnt == 0) { 2271 btrfs_init_workers(&fs_info->scrub_workers, "scrub", 2272 fs_info->thread_pool_size, &fs_info->generic_worker); 2273 fs_info->scrub_workers.idle_thresh = 4; 2274 ret = btrfs_start_workers(&fs_info->scrub_workers); 2275 if (ret) 2276 goto out; 2277 } 2278 ++fs_info->scrub_workers_refcnt; 2279out: 2280 mutex_unlock(&fs_info->scrub_lock); 2281 2282 return ret; 2283} 2284 2285static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info) 2286{ 2287 mutex_lock(&fs_info->scrub_lock); 2288 if (--fs_info->scrub_workers_refcnt == 0) 2289 btrfs_stop_workers(&fs_info->scrub_workers); 2290 WARN_ON(fs_info->scrub_workers_refcnt < 0); 2291 mutex_unlock(&fs_info->scrub_lock); 2292} 2293 2294int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start, 2295 u64 end, struct btrfs_scrub_progress *progress, 2296 int readonly) 2297{ 2298 struct scrub_ctx *sctx; 2299 int ret; 2300 struct btrfs_device *dev; 2301 2302 if (btrfs_fs_closing(fs_info)) 2303 return -EINVAL; 2304 2305 /* 2306 * check some assumptions 2307 */ 2308 if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) { 2309 printk(KERN_ERR 2310 "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n", 2311 fs_info->chunk_root->nodesize, 2312 fs_info->chunk_root->leafsize); 2313 return -EINVAL; 2314 } 2315 2316 if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) { 2317 /* 2318 * in this case scrub is unable to calculate the checksum 2319 * the way scrub is implemented. Do not handle this 2320 * situation at all because it won't ever happen. 2321 */ 2322 printk(KERN_ERR 2323 "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n", 2324 fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN); 2325 return -EINVAL; 2326 } 2327 2328 if (fs_info->chunk_root->sectorsize != PAGE_SIZE) { 2329 /* not supported for data w/o checksums */ 2330 printk(KERN_ERR 2331 "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n", 2332 fs_info->chunk_root->sectorsize, 2333 (unsigned long long)PAGE_SIZE); 2334 return -EINVAL; 2335 } 2336 2337 if (fs_info->chunk_root->nodesize > 2338 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK || 2339 fs_info->chunk_root->sectorsize > 2340 PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) { 2341 /* 2342 * would exhaust the array bounds of pagev member in 2343 * struct scrub_block 2344 */ 2345 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n", 2346 fs_info->chunk_root->nodesize, 2347 SCRUB_MAX_PAGES_PER_BLOCK, 2348 fs_info->chunk_root->sectorsize, 2349 SCRUB_MAX_PAGES_PER_BLOCK); 2350 return -EINVAL; 2351 } 2352 2353 ret = scrub_workers_get(fs_info); 2354 if (ret) 2355 return ret; 2356 2357 mutex_lock(&fs_info->fs_devices->device_list_mutex); 2358 dev = btrfs_find_device(fs_info, devid, NULL, NULL); 2359 if (!dev || dev->missing) { 2360 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2361 scrub_workers_put(fs_info); 2362 return -ENODEV; 2363 } 2364 mutex_lock(&fs_info->scrub_lock); 2365 2366 if (!dev->in_fs_metadata) { 2367 mutex_unlock(&fs_info->scrub_lock); 2368 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2369 scrub_workers_put(fs_info); 2370 return -EIO; 2371 } 2372 2373 if (dev->scrub_device) { 2374 mutex_unlock(&fs_info->scrub_lock); 2375 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2376 scrub_workers_put(fs_info); 2377 return -EINPROGRESS; 2378 } 2379 sctx = scrub_setup_ctx(dev); 2380 if (IS_ERR(sctx)) { 2381 mutex_unlock(&fs_info->scrub_lock); 2382 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2383 scrub_workers_put(fs_info); 2384 return PTR_ERR(sctx); 2385 } 2386 sctx->readonly = readonly; 2387 dev->scrub_device = sctx; 2388 2389 atomic_inc(&fs_info->scrubs_running); 2390 mutex_unlock(&fs_info->scrub_lock); 2391 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2392 2393 down_read(&fs_info->scrub_super_lock); 2394 ret = scrub_supers(sctx, dev); 2395 up_read(&fs_info->scrub_super_lock); 2396 2397 if (!ret) 2398 ret = scrub_enumerate_chunks(sctx, dev, start, end); 2399 2400 wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0); 2401 atomic_dec(&fs_info->scrubs_running); 2402 wake_up(&fs_info->scrub_pause_wait); 2403 2404 wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0); 2405 2406 if (progress) 2407 memcpy(progress, &sctx->stat, sizeof(*progress)); 2408 2409 mutex_lock(&fs_info->scrub_lock); 2410 dev->scrub_device = NULL; 2411 mutex_unlock(&fs_info->scrub_lock); 2412 2413 scrub_free_ctx(sctx); 2414 scrub_workers_put(fs_info); 2415 2416 return ret; 2417} 2418 2419void btrfs_scrub_pause(struct btrfs_root *root) 2420{ 2421 struct btrfs_fs_info *fs_info = root->fs_info; 2422 2423 mutex_lock(&fs_info->scrub_lock); 2424 atomic_inc(&fs_info->scrub_pause_req); 2425 while (atomic_read(&fs_info->scrubs_paused) != 2426 atomic_read(&fs_info->scrubs_running)) { 2427 mutex_unlock(&fs_info->scrub_lock); 2428 wait_event(fs_info->scrub_pause_wait, 2429 atomic_read(&fs_info->scrubs_paused) == 2430 atomic_read(&fs_info->scrubs_running)); 2431 mutex_lock(&fs_info->scrub_lock); 2432 } 2433 mutex_unlock(&fs_info->scrub_lock); 2434} 2435 2436void btrfs_scrub_continue(struct btrfs_root *root) 2437{ 2438 struct btrfs_fs_info *fs_info = root->fs_info; 2439 2440 atomic_dec(&fs_info->scrub_pause_req); 2441 wake_up(&fs_info->scrub_pause_wait); 2442} 2443 2444void btrfs_scrub_pause_super(struct btrfs_root *root) 2445{ 2446 down_write(&root->fs_info->scrub_super_lock); 2447} 2448 2449void btrfs_scrub_continue_super(struct btrfs_root *root) 2450{ 2451 up_write(&root->fs_info->scrub_super_lock); 2452} 2453 2454int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info) 2455{ 2456 mutex_lock(&fs_info->scrub_lock); 2457 if (!atomic_read(&fs_info->scrubs_running)) { 2458 mutex_unlock(&fs_info->scrub_lock); 2459 return -ENOTCONN; 2460 } 2461 2462 atomic_inc(&fs_info->scrub_cancel_req); 2463 while (atomic_read(&fs_info->scrubs_running)) { 2464 mutex_unlock(&fs_info->scrub_lock); 2465 wait_event(fs_info->scrub_pause_wait, 2466 atomic_read(&fs_info->scrubs_running) == 0); 2467 mutex_lock(&fs_info->scrub_lock); 2468 } 2469 atomic_dec(&fs_info->scrub_cancel_req); 2470 mutex_unlock(&fs_info->scrub_lock); 2471 2472 return 0; 2473} 2474 2475int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info, 2476 struct btrfs_device *dev) 2477{ 2478 struct scrub_ctx *sctx; 2479 2480 mutex_lock(&fs_info->scrub_lock); 2481 sctx = dev->scrub_device; 2482 if (!sctx) { 2483 mutex_unlock(&fs_info->scrub_lock); 2484 return -ENOTCONN; 2485 } 2486 atomic_inc(&sctx->cancel_req); 2487 while (dev->scrub_device) { 2488 mutex_unlock(&fs_info->scrub_lock); 2489 wait_event(fs_info->scrub_pause_wait, 2490 dev->scrub_device == NULL); 2491 mutex_lock(&fs_info->scrub_lock); 2492 } 2493 mutex_unlock(&fs_info->scrub_lock); 2494 2495 return 0; 2496} 2497 2498int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid) 2499{ 2500 struct btrfs_fs_info *fs_info = root->fs_info; 2501 struct btrfs_device *dev; 2502 int ret; 2503 2504 /* 2505 * we have to hold the device_list_mutex here so the device 2506 * does not go away in cancel_dev. FIXME: find a better solution 2507 */ 2508 mutex_lock(&fs_info->fs_devices->device_list_mutex); 2509 dev = btrfs_find_device(fs_info, devid, NULL, NULL); 2510 if (!dev) { 2511 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2512 return -ENODEV; 2513 } 2514 ret = btrfs_scrub_cancel_dev(fs_info, dev); 2515 mutex_unlock(&fs_info->fs_devices->device_list_mutex); 2516 2517 return ret; 2518} 2519 2520int btrfs_scrub_progress(struct btrfs_root *root, u64 devid, 2521 struct btrfs_scrub_progress *progress) 2522{ 2523 struct btrfs_device *dev; 2524 struct scrub_ctx *sctx = NULL; 2525 2526 mutex_lock(&root->fs_info->fs_devices->device_list_mutex); 2527 dev = btrfs_find_device(root->fs_info, devid, NULL, NULL); 2528 if (dev) 2529 sctx = dev->scrub_device; 2530 if (sctx) 2531 memcpy(progress, &sctx->stat, sizeof(*progress)); 2532 mutex_unlock(&root->fs_info->fs_devices->device_list_mutex); 2533 2534 return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV; 2535} 2536