fs/btrfs/scrub.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/blkdev.h>
   7 #include <linux/ratelimit.h>
   8 #include <linux/sched/mm.h>
   9 #include <crypto/hash.h>
  10 #include "ctree.h"
  11 #include "discard.h"
  12 #include "volumes.h"
  13 #include "disk-io.h"
  14 #include "ordered-data.h"
  15 #include "transaction.h"
  16 #include "backref.h"
  17 #include "extent_io.h"
  18 #include "dev-replace.h"
  19 #include "check-integrity.h"
  20 #include "raid56.h"
  21 #include "block-group.h"
  22 #include "zoned.h"
  23 #include "fs.h"
  24 #include "accessors.h"
  25 #include "file-item.h"
  26 #include "scrub.h"
  27
  28 /*
  29  * This is only the first step towards a full-features scrub. It reads all
  30  * extent and super block and verifies the checksums. In case a bad checksum
  31  * is found or the extent cannot be read, good data will be written back if
  32  * any can be found.
  33  *
  34  * Future enhancements:
  35  *  - In case an unrepairable extent is encountered, track which files are
  36  *    affected and report them
  37  *  - track and record media errors, throw out bad devices
  38  *  - add a mode to also read unallocated space
  39  */
  40
  41 struct scrub_ctx;
  42
  43 /*
  44  * The following value only influences the performance.
  45  *
  46  * This detemines how many stripes would be submitted in one go,
  47  * which is 512KiB (BTRFS_STRIPE_LEN * SCRUB_STRIPES_PER_GROUP).
  48  */
  49 #define SCRUB_STRIPES_PER_GROUP         8
  50
  51 /*
  52  * How many groups we have for each sctx.
  53  *
  54  * This would be 8M per device, the same value as the old scrub in-flight bios
  55  * size limit.
  56  */
  57 #define SCRUB_GROUPS_PER_SCTX           16
  58
  59 #define SCRUB_TOTAL_STRIPES             (SCRUB_GROUPS_PER_SCTX * SCRUB_STRIPES_PER_GROUP)
  60
  61 /*
  62  * The following value times PAGE_SIZE needs to be large enough to match the
  63  * largest node/leaf/sector size that shall be supported.
  64  */
  65 #define SCRUB_MAX_SECTORS_PER_BLOCK     (BTRFS_MAX_METADATA_BLOCKSIZE / SZ_4K)
  66
  67 /* Represent one sector and its needed info to verify the content. */
  68 struct scrub_sector_verification {
  69         bool is_metadata;
  70
  71         union {
  72                 /*
  73                  * Csum pointer for data csum verification.  Should point to a
  74                  * sector csum inside scrub_stripe::csums.
  75                  *
  76                  * NULL if this data sector has no csum.
  77                  */
  78                 u8 *csum;
  79
  80                 /*
  81                  * Extra info for metadata verification.  All sectors inside a
  82                  * tree block share the same generation.
  83                  */
  84                 u64 generation;
  85         };
  86 };
  87
  88 enum scrub_stripe_flags {
  89         /* Set when @mirror_num, @dev, @physical and @logical are set. */
  90         SCRUB_STRIPE_FLAG_INITIALIZED,
  91
  92         /* Set when the read-repair is finished. */
  93         SCRUB_STRIPE_FLAG_REPAIR_DONE,
  94
  95         /*
  96          * Set for data stripes if it's triggered from P/Q stripe.
  97          * During such scrub, we should not report errors in data stripes, nor
  98          * update the accounting.
  99          */
 100         SCRUB_STRIPE_FLAG_NO_REPORT,
 101 };
 102
 103 #define SCRUB_STRIPE_PAGES              (BTRFS_STRIPE_LEN / PAGE_SIZE)
 104
 105 /*
 106  * Represent one contiguous range with a length of BTRFS_STRIPE_LEN.
 107  */
 108 struct scrub_stripe {
 109         struct scrub_ctx *sctx;
 110         struct btrfs_block_group *bg;
 111
 112         struct page *pages[SCRUB_STRIPE_PAGES];
 113         struct scrub_sector_verification *sectors;
 114
 115         struct btrfs_device *dev;
 116         u64 logical;
 117         u64 physical;
 118
 119         u16 mirror_num;
 120
 121         /* Should be BTRFS_STRIPE_LEN / sectorsize. */
 122         u16 nr_sectors;
 123
 124         /*
 125          * How many data/meta extents are in this stripe.  Only for scrub status
 126          * reporting purposes.
 127          */
 128         u16 nr_data_extents;
 129         u16 nr_meta_extents;
 130
 131         atomic_t pending_io;
 132         wait_queue_head_t io_wait;
 133         wait_queue_head_t repair_wait;
 134
 135         /*
 136          * Indicate the states of the stripe.  Bits are defined in
 137          * scrub_stripe_flags enum.
 138          */
 139         unsigned long state;
 140
 141         /* Indicate which sectors are covered by extent items. */
 142         unsigned long extent_sector_bitmap;
 143
 144         /*
 145          * The errors hit during the initial read of the stripe.
 146          *
 147          * Would be utilized for error reporting and repair.
 148          *
 149          * The remaining init_nr_* records the number of errors hit, only used
 150          * by error reporting.
 151          */
 152         unsigned long init_error_bitmap;
 153         unsigned int init_nr_io_errors;
 154         unsigned int init_nr_csum_errors;
 155         unsigned int init_nr_meta_errors;
 156
 157         /*
 158          * The following error bitmaps are all for the current status.
 159          * Every time we submit a new read, these bitmaps may be updated.
 160          *
 161          * error_bitmap = io_error_bitmap | csum_error_bitmap | meta_error_bitmap;
 162          *
 163          * IO and csum errors can happen for both metadata and data.
 164          */
 165         unsigned long error_bitmap;
 166         unsigned long io_error_bitmap;
 167         unsigned long csum_error_bitmap;
 168         unsigned long meta_error_bitmap;
 169
 170         /* For writeback (repair or replace) error reporting. */
 171         unsigned long write_error_bitmap;
 172
 173         /* Writeback can be concurrent, thus we need to protect the bitmap. */
 174         spinlock_t write_error_lock;
 175
 176         /*
 177          * Checksum for the whole stripe if this stripe is inside a data block
 178          * group.
 179          */
 180         u8 *csums;
 181
 182         struct work_struct work;
 183 };
 184
 185 struct scrub_ctx {
 186         struct scrub_stripe     stripes[SCRUB_TOTAL_STRIPES];
 187         struct scrub_stripe     *raid56_data_stripes;
 188         struct btrfs_fs_info    *fs_info;
 189         struct btrfs_path       extent_path;
 190         struct btrfs_path       csum_path;
 191         int                     first_free;
 192         int                     cur_stripe;
 193         atomic_t                cancel_req;
 194         int                     readonly;
 195         int                     sectors_per_bio;
 196
 197         /* State of IO submission throttling affecting the associated device */
 198         ktime_t                 throttle_deadline;
 199         u64                     throttle_sent;
 200
 201         int                     is_dev_replace;
 202         u64                     write_pointer;
 203
 204         struct mutex            wr_lock;
 205         struct btrfs_device     *wr_tgtdev;
 206
 207         /*
 208          * statistics
 209          */
 210         struct btrfs_scrub_progress stat;
 211         spinlock_t              stat_lock;
 212
 213         /*
 214          * Use a ref counter to avoid use-after-free issues. Scrub workers
 215          * decrement bios_in_flight and workers_pending and then do a wakeup
 216          * on the list_wait wait queue. We must ensure the main scrub task
 217          * doesn't free the scrub context before or while the workers are
 218          * doing the wakeup() call.
 219          */
 220         refcount_t              refs;
 221 };
 222
 223 struct scrub_warning {
 224         struct btrfs_path       *path;
 225         u64                     extent_item_size;
 226         const char              *errstr;
 227         u64                     physical;
 228         u64                     logical;
 229         struct btrfs_device     *dev;
 230 };
 231
 232 static void release_scrub_stripe(struct scrub_stripe *stripe)
 233 {
 234         if (!stripe)
 235                 return;
 236
 237         for (int i = 0; i < SCRUB_STRIPE_PAGES; i++) {
 238                 if (stripe->pages[i])
 239                         __free_page(stripe->pages[i]);
 240                 stripe->pages[i] = NULL;
 241         }
 242         kfree(stripe->sectors);
 243         kfree(stripe->csums);
 244         stripe->sectors = NULL;
 245         stripe->csums = NULL;
 246         stripe->sctx = NULL;
 247         stripe->state = 0;
 248 }
 249
 250 static int init_scrub_stripe(struct btrfs_fs_info *fs_info,
 251                              struct scrub_stripe *stripe)
 252 {
 253         int ret;
 254
 255         memset(stripe, 0, sizeof(*stripe));
 256
 257         stripe->nr_sectors = BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits;
 258         stripe->state = 0;
 259
 260         init_waitqueue_head(&stripe->io_wait);
 261         init_waitqueue_head(&stripe->repair_wait);
 262         atomic_set(&stripe->pending_io, 0);
 263         spin_lock_init(&stripe->write_error_lock);
 264
 265         ret = btrfs_alloc_page_array(SCRUB_STRIPE_PAGES, stripe->pages);
 266         if (ret < 0)
 267                 goto error;
 268
 269         stripe->sectors = kcalloc(stripe->nr_sectors,
 270                                   sizeof(struct scrub_sector_verification),
 271                                   GFP_KERNEL);
 272         if (!stripe->sectors)
 273                 goto error;
 274
 275         stripe->csums = kcalloc(BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits,
 276                                 fs_info->csum_size, GFP_KERNEL);
 277         if (!stripe->csums)
 278                 goto error;
 279         return 0;
 280 error:
 281         release_scrub_stripe(stripe);
 282         return -ENOMEM;
 283 }
 284
 285 static void wait_scrub_stripe_io(struct scrub_stripe *stripe)
 286 {
 287         wait_event(stripe->io_wait, atomic_read(&stripe->pending_io) == 0);
 288 }
 289
 290 static void scrub_put_ctx(struct scrub_ctx *sctx);
 291
 292 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 293 {
 294         while (atomic_read(&fs_info->scrub_pause_req)) {
 295                 mutex_unlock(&fs_info->scrub_lock);
 296                 wait_event(fs_info->scrub_pause_wait,
 297                    atomic_read(&fs_info->scrub_pause_req) == 0);
 298                 mutex_lock(&fs_info->scrub_lock);
 299         }
 300 }
 301
 302 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 303 {
 304         atomic_inc(&fs_info->scrubs_paused);
 305         wake_up(&fs_info->scrub_pause_wait);
 306 }
 307
 308 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 309 {
 310         mutex_lock(&fs_info->scrub_lock);
 311         __scrub_blocked_if_needed(fs_info);
 312         atomic_dec(&fs_info->scrubs_paused);
 313         mutex_unlock(&fs_info->scrub_lock);
 314
 315         wake_up(&fs_info->scrub_pause_wait);
 316 }
 317
 318 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 319 {
 320         scrub_pause_on(fs_info);
 321         scrub_pause_off(fs_info);
 322 }
 323
 324 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 325 {
 326         int i;
 327
 328         if (!sctx)
 329                 return;
 330
 331         for (i = 0; i < SCRUB_TOTAL_STRIPES; i++)
 332                 release_scrub_stripe(&sctx->stripes[i]);
 333
 334         kvfree(sctx);
 335 }
 336
 337 static void scrub_put_ctx(struct scrub_ctx *sctx)
 338 {
 339         if (refcount_dec_and_test(&sctx->refs))
 340                 scrub_free_ctx(sctx);
 341 }
 342
 343 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 344                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 345 {
 346         struct scrub_ctx *sctx;
 347         int             i;
 348
 349         /* Since sctx has inline 128 stripes, it can go beyond 64K easily.  Use
 350          * kvzalloc().
 351          */
 352         sctx = kvzalloc(sizeof(*sctx), GFP_KERNEL);
 353         if (!sctx)
 354                 goto nomem;
 355         refcount_set(&sctx->refs, 1);
 356         sctx->is_dev_replace = is_dev_replace;
 357         sctx->fs_info = fs_info;
 358         sctx->extent_path.search_commit_root = 1;
 359         sctx->extent_path.skip_locking = 1;
 360         sctx->csum_path.search_commit_root = 1;
 361         sctx->csum_path.skip_locking = 1;
 362         for (i = 0; i < SCRUB_TOTAL_STRIPES; i++) {
 363                 int ret;
 364
 365                 ret = init_scrub_stripe(fs_info, &sctx->stripes[i]);
 366                 if (ret < 0)
 367                         goto nomem;
 368                 sctx->stripes[i].sctx = sctx;
 369         }
 370         sctx->first_free = 0;
 371         atomic_set(&sctx->cancel_req, 0);
 372
 373         spin_lock_init(&sctx->stat_lock);
 374         sctx->throttle_deadline = 0;
 375
 376         mutex_init(&sctx->wr_lock);
 377         if (is_dev_replace) {
 378                 WARN_ON(!fs_info->dev_replace.tgtdev);
 379                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 380         }
 381
 382         return sctx;
 383
 384 nomem:
 385         scrub_free_ctx(sctx);
 386         return ERR_PTR(-ENOMEM);
 387 }
 388
 389 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 num_bytes,
 390                                      u64 root, void *warn_ctx)
 391 {
 392         u32 nlink;
 393         int ret;
 394         int i;
 395         unsigned nofs_flag;
 396         struct extent_buffer *eb;
 397         struct btrfs_inode_item *inode_item;
 398         struct scrub_warning *swarn = warn_ctx;
 399         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 400         struct inode_fs_paths *ipath = NULL;
 401         struct btrfs_root *local_root;
 402         struct btrfs_key key;
 403
 404         local_root = btrfs_get_fs_root(fs_info, root, true);
 405         if (IS_ERR(local_root)) {
 406                 ret = PTR_ERR(local_root);
 407                 goto err;
 408         }
 409
 410         /*
 411          * this makes the path point to (inum INODE_ITEM ioff)
 412          */
 413         key.objectid = inum;
 414         key.type = BTRFS_INODE_ITEM_KEY;
 415         key.offset = 0;
 416
 417         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 418         if (ret) {
 419                 btrfs_put_root(local_root);
 420                 btrfs_release_path(swarn->path);
 421                 goto err;
 422         }
 423
 424         eb = swarn->path->nodes[0];
 425         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 426                                         struct btrfs_inode_item);
 427         nlink = btrfs_inode_nlink(eb, inode_item);
 428         btrfs_release_path(swarn->path);
 429
 430         /*
 431          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 432          * uses GFP_NOFS in this context, so we keep it consistent but it does
 433          * not seem to be strictly necessary.
 434          */
 435         nofs_flag = memalloc_nofs_save();
 436         ipath = init_ipath(4096, local_root, swarn->path);
 437         memalloc_nofs_restore(nofs_flag);
 438         if (IS_ERR(ipath)) {
 439                 btrfs_put_root(local_root);
 440                 ret = PTR_ERR(ipath);
 441                 ipath = NULL;
 442                 goto err;
 443         }
 444         ret = paths_from_inode(inum, ipath);
 445
 446         if (ret < 0)
 447                 goto err;
 448
 449         /*
 450          * we deliberately ignore the bit ipath might have been too small to
 451          * hold all of the paths here
 452          */
 453         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 454                 btrfs_warn_in_rcu(fs_info,
 455 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %u, links %u (path: %s)",
 456                                   swarn->errstr, swarn->logical,
 457                                   btrfs_dev_name(swarn->dev),
 458                                   swarn->physical,
 459                                   root, inum, offset,
 460                                   fs_info->sectorsize, nlink,
 461                                   (char *)(unsigned long)ipath->fspath->val[i]);
 462
 463         btrfs_put_root(local_root);
 464         free_ipath(ipath);
 465         return 0;
 466
 467 err:
 468         btrfs_warn_in_rcu(fs_info,
 469                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 470                           swarn->errstr, swarn->logical,
 471                           btrfs_dev_name(swarn->dev),
 472                           swarn->physical,
 473                           root, inum, offset, ret);
 474
 475         free_ipath(ipath);
 476         return 0;
 477 }
 478
 479 static void scrub_print_common_warning(const char *errstr, struct btrfs_device *dev,
 480                                        bool is_super, u64 logical, u64 physical)
 481 {
 482         struct btrfs_fs_info *fs_info = dev->fs_info;
 483         struct btrfs_path *path;
 484         struct btrfs_key found_key;
 485         struct extent_buffer *eb;
 486         struct btrfs_extent_item *ei;
 487         struct scrub_warning swarn;
 488         u64 flags = 0;
 489         u32 item_size;
 490         int ret;
 491
 492         /* Super block error, no need to search extent tree. */
 493         if (is_super) {
 494                 btrfs_warn_in_rcu(fs_info, "%s on device %s, physical %llu",
 495                                   errstr, btrfs_dev_name(dev), physical);
 496                 return;
 497         }
 498         path = btrfs_alloc_path();
 499         if (!path)
 500                 return;
 501
 502         swarn.physical = physical;
 503         swarn.logical = logical;
 504         swarn.errstr = errstr;
 505         swarn.dev = NULL;
 506
 507         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 508                                   &flags);
 509         if (ret < 0)
 510                 goto out;
 511
 512         swarn.extent_item_size = found_key.offset;
 513
 514         eb = path->nodes[0];
 515         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 516         item_size = btrfs_item_size(eb, path->slots[0]);
 517
 518         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 519                 unsigned long ptr = 0;
 520                 u8 ref_level;
 521                 u64 ref_root;
 522
 523                 while (true) {
 524                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 525                                                       item_size, &ref_root,
 526                                                       &ref_level);
 527                         if (ret < 0) {
 528                                 btrfs_warn(fs_info,
 529                                 "failed to resolve tree backref for logical %llu: %d",
 530                                                   swarn.logical, ret);
 531                                 break;
 532                         }
 533                         if (ret > 0)
 534                                 break;
 535                         btrfs_warn_in_rcu(fs_info,
 536 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 537                                 errstr, swarn.logical, btrfs_dev_name(dev),
 538                                 swarn.physical, (ref_level ? "node" : "leaf"),
 539                                 ref_level, ref_root);
 540                 }
 541                 btrfs_release_path(path);
 542         } else {
 543                 struct btrfs_backref_walk_ctx ctx = { 0 };
 544
 545                 btrfs_release_path(path);
 546
 547                 ctx.bytenr = found_key.objectid;
 548                 ctx.extent_item_pos = swarn.logical - found_key.objectid;
 549                 ctx.fs_info = fs_info;
 550
 551                 swarn.path = path;
 552                 swarn.dev = dev;
 553
 554                 iterate_extent_inodes(&ctx, true, scrub_print_warning_inode, &swarn);
 555         }
 556
 557 out:
 558         btrfs_free_path(path);
 559 }
 560
 561 static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
 562 {
 563         int ret = 0;
 564         u64 length;
 565
 566         if (!btrfs_is_zoned(sctx->fs_info))
 567                 return 0;
 568
 569         if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
 570                 return 0;
 571
 572         if (sctx->write_pointer < physical) {
 573                 length = physical - sctx->write_pointer;
 574
 575                 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
 576                                                 sctx->write_pointer, length);
 577                 if (!ret)
 578                         sctx->write_pointer = physical;
 579         }
 580         return ret;
 581 }
 582
 583 static struct page *scrub_stripe_get_page(struct scrub_stripe *stripe, int sector_nr)
 584 {
 585         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 586         int page_index = (sector_nr << fs_info->sectorsize_bits) >> PAGE_SHIFT;
 587
 588         return stripe->pages[page_index];
 589 }
 590
 591 static unsigned int scrub_stripe_get_page_offset(struct scrub_stripe *stripe,
 592                                                  int sector_nr)
 593 {
 594         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 595
 596         return offset_in_page(sector_nr << fs_info->sectorsize_bits);
 597 }
 598
 599 static void scrub_verify_one_metadata(struct scrub_stripe *stripe, int sector_nr)
 600 {
 601         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 602         const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
 603         const u64 logical = stripe->logical + (sector_nr << fs_info->sectorsize_bits);
 604         const struct page *first_page = scrub_stripe_get_page(stripe, sector_nr);
 605         const unsigned int first_off = scrub_stripe_get_page_offset(stripe, sector_nr);
 606         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
 607         u8 on_disk_csum[BTRFS_CSUM_SIZE];
 608         u8 calculated_csum[BTRFS_CSUM_SIZE];
 609         struct btrfs_header *header;
 610
 611         /*
 612          * Here we don't have a good way to attach the pages (and subpages)
 613          * to a dummy extent buffer, thus we have to directly grab the members
 614          * from pages.
 615          */
 616         header = (struct btrfs_header *)(page_address(first_page) + first_off);
 617         memcpy(on_disk_csum, header->csum, fs_info->csum_size);
 618
 619         if (logical != btrfs_stack_header_bytenr(header)) {
 620                 bitmap_set(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
 621                 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 622                 btrfs_warn_rl(fs_info,
 623                 "tree block %llu mirror %u has bad bytenr, has %llu want %llu",
 624                               logical, stripe->mirror_num,
 625                               btrfs_stack_header_bytenr(header), logical);
 626                 return;
 627         }
 628         if (memcmp(header->fsid, fs_info->fs_devices->metadata_uuid,
 629                    BTRFS_FSID_SIZE) != 0) {
 630                 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
 631                 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 632                 btrfs_warn_rl(fs_info,
 633                 "tree block %llu mirror %u has bad fsid, has %pU want %pU",
 634                               logical, stripe->mirror_num,
 635                               header->fsid, fs_info->fs_devices->fsid);
 636                 return;
 637         }
 638         if (memcmp(header->chunk_tree_uuid, fs_info->chunk_tree_uuid,
 639                    BTRFS_UUID_SIZE) != 0) {
 640                 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
 641                 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 642                 btrfs_warn_rl(fs_info,
 643                 "tree block %llu mirror %u has bad chunk tree uuid, has %pU want %pU",
 644                               logical, stripe->mirror_num,
 645                               header->chunk_tree_uuid, fs_info->chunk_tree_uuid);
 646                 return;
 647         }
 648
 649         /* Now check tree block csum. */
 650         shash->tfm = fs_info->csum_shash;
 651         crypto_shash_init(shash);
 652         crypto_shash_update(shash, page_address(first_page) + first_off +
 653                             BTRFS_CSUM_SIZE, fs_info->sectorsize - BTRFS_CSUM_SIZE);
 654
 655         for (int i = sector_nr + 1; i < sector_nr + sectors_per_tree; i++) {
 656                 struct page *page = scrub_stripe_get_page(stripe, i);
 657                 unsigned int page_off = scrub_stripe_get_page_offset(stripe, i);
 658
 659                 crypto_shash_update(shash, page_address(page) + page_off,
 660                                     fs_info->sectorsize);
 661         }
 662
 663         crypto_shash_final(shash, calculated_csum);
 664         if (memcmp(calculated_csum, on_disk_csum, fs_info->csum_size) != 0) {
 665                 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
 666                 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 667                 btrfs_warn_rl(fs_info,
 668                 "tree block %llu mirror %u has bad csum, has " CSUM_FMT " want " CSUM_FMT,
 669                               logical, stripe->mirror_num,
 670                               CSUM_FMT_VALUE(fs_info->csum_size, on_disk_csum),
 671                               CSUM_FMT_VALUE(fs_info->csum_size, calculated_csum));
 672                 return;
 673         }
 674         if (stripe->sectors[sector_nr].generation !=
 675             btrfs_stack_header_generation(header)) {
 676                 bitmap_set(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
 677                 bitmap_set(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 678                 btrfs_warn_rl(fs_info,
 679                 "tree block %llu mirror %u has bad generation, has %llu want %llu",
 680                               logical, stripe->mirror_num,
 681                               btrfs_stack_header_generation(header),
 682                               stripe->sectors[sector_nr].generation);
 683                 return;
 684         }
 685         bitmap_clear(&stripe->error_bitmap, sector_nr, sectors_per_tree);
 686         bitmap_clear(&stripe->csum_error_bitmap, sector_nr, sectors_per_tree);
 687         bitmap_clear(&stripe->meta_error_bitmap, sector_nr, sectors_per_tree);
 688 }
 689
 690 static void scrub_verify_one_sector(struct scrub_stripe *stripe, int sector_nr)
 691 {
 692         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 693         struct scrub_sector_verification *sector = &stripe->sectors[sector_nr];
 694         const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
 695         struct page *page = scrub_stripe_get_page(stripe, sector_nr);
 696         unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
 697         u8 csum_buf[BTRFS_CSUM_SIZE];
 698         int ret;
 699
 700         ASSERT(sector_nr >= 0 && sector_nr < stripe->nr_sectors);
 701
 702         /* Sector not utilized, skip it. */
 703         if (!test_bit(sector_nr, &stripe->extent_sector_bitmap))
 704                 return;
 705
 706         /* IO error, no need to check. */
 707         if (test_bit(sector_nr, &stripe->io_error_bitmap))
 708                 return;
 709
 710         /* Metadata, verify the full tree block. */
 711         if (sector->is_metadata) {
 712                 /*
 713                  * Check if the tree block crosses the stripe boudary.  If
 714                  * crossed the boundary, we cannot verify it but only give a
 715                  * warning.
 716                  *
 717                  * This can only happen on a very old filesystem where chunks
 718                  * are not ensured to be stripe aligned.
 719                  */
 720                 if (unlikely(sector_nr + sectors_per_tree > stripe->nr_sectors)) {
 721                         btrfs_warn_rl(fs_info,
 722                         "tree block at %llu crosses stripe boundary %llu",
 723                                       stripe->logical +
 724                                       (sector_nr << fs_info->sectorsize_bits),
 725                                       stripe->logical);
 726                         return;
 727                 }
 728                 scrub_verify_one_metadata(stripe, sector_nr);
 729                 return;
 730         }
 731
 732         /*
 733          * Data is easier, we just verify the data csum (if we have it).  For
 734          * cases without csum, we have no other choice but to trust it.
 735          */
 736         if (!sector->csum) {
 737                 clear_bit(sector_nr, &stripe->error_bitmap);
 738                 return;
 739         }
 740
 741         ret = btrfs_check_sector_csum(fs_info, page, pgoff, csum_buf, sector->csum);
 742         if (ret < 0) {
 743                 set_bit(sector_nr, &stripe->csum_error_bitmap);
 744                 set_bit(sector_nr, &stripe->error_bitmap);
 745         } else {
 746                 clear_bit(sector_nr, &stripe->csum_error_bitmap);
 747                 clear_bit(sector_nr, &stripe->error_bitmap);
 748         }
 749 }
 750
 751 /* Verify specified sectors of a stripe. */
 752 static void scrub_verify_one_stripe(struct scrub_stripe *stripe, unsigned long bitmap)
 753 {
 754         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 755         const u32 sectors_per_tree = fs_info->nodesize >> fs_info->sectorsize_bits;
 756         int sector_nr;
 757
 758         for_each_set_bit(sector_nr, &bitmap, stripe->nr_sectors) {
 759                 scrub_verify_one_sector(stripe, sector_nr);
 760                 if (stripe->sectors[sector_nr].is_metadata)
 761                         sector_nr += sectors_per_tree - 1;
 762         }
 763 }
 764
 765 static int calc_sector_number(struct scrub_stripe *stripe, struct bio_vec *first_bvec)
 766 {
 767         int i;
 768
 769         for (i = 0; i < stripe->nr_sectors; i++) {
 770                 if (scrub_stripe_get_page(stripe, i) == first_bvec->bv_page &&
 771                     scrub_stripe_get_page_offset(stripe, i) == first_bvec->bv_offset)
 772                         break;
 773         }
 774         ASSERT(i < stripe->nr_sectors);
 775         return i;
 776 }
 777
 778 /*
 779  * Repair read is different to the regular read:
 780  *
 781  * - Only reads the failed sectors
 782  * - May have extra blocksize limits
 783  */
 784 static void scrub_repair_read_endio(struct btrfs_bio *bbio)
 785 {
 786         struct scrub_stripe *stripe = bbio->private;
 787         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 788         struct bio_vec *bvec;
 789         int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
 790         u32 bio_size = 0;
 791         int i;
 792
 793         ASSERT(sector_nr < stripe->nr_sectors);
 794
 795         bio_for_each_bvec_all(bvec, &bbio->bio, i)
 796                 bio_size += bvec->bv_len;
 797
 798         if (bbio->bio.bi_status) {
 799                 bitmap_set(&stripe->io_error_bitmap, sector_nr,
 800                            bio_size >> fs_info->sectorsize_bits);
 801                 bitmap_set(&stripe->error_bitmap, sector_nr,
 802                            bio_size >> fs_info->sectorsize_bits);
 803         } else {
 804                 bitmap_clear(&stripe->io_error_bitmap, sector_nr,
 805                              bio_size >> fs_info->sectorsize_bits);
 806         }
 807         bio_put(&bbio->bio);
 808         if (atomic_dec_and_test(&stripe->pending_io))
 809                 wake_up(&stripe->io_wait);
 810 }
 811
 812 static int calc_next_mirror(int mirror, int num_copies)
 813 {
 814         ASSERT(mirror <= num_copies);
 815         return (mirror + 1 > num_copies) ? 1 : mirror + 1;
 816 }
 817
 818 static void scrub_stripe_submit_repair_read(struct scrub_stripe *stripe,
 819                                             int mirror, int blocksize, bool wait)
 820 {
 821         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
 822         struct btrfs_bio *bbio = NULL;
 823         const unsigned long old_error_bitmap = stripe->error_bitmap;
 824         int i;
 825
 826         ASSERT(stripe->mirror_num >= 1);
 827         ASSERT(atomic_read(&stripe->pending_io) == 0);
 828
 829         for_each_set_bit(i, &old_error_bitmap, stripe->nr_sectors) {
 830                 struct page *page;
 831                 int pgoff;
 832                 int ret;
 833
 834                 page = scrub_stripe_get_page(stripe, i);
 835                 pgoff = scrub_stripe_get_page_offset(stripe, i);
 836
 837                 /* The current sector cannot be merged, submit the bio. */
 838                 if (bbio && ((i > 0 && !test_bit(i - 1, &stripe->error_bitmap)) ||
 839                              bbio->bio.bi_iter.bi_size >= blocksize)) {
 840                         ASSERT(bbio->bio.bi_iter.bi_size);
 841                         atomic_inc(&stripe->pending_io);
 842                         btrfs_submit_bio(bbio, mirror);
 843                         if (wait)
 844                                 wait_scrub_stripe_io(stripe);
 845                         bbio = NULL;
 846                 }
 847
 848                 if (!bbio) {
 849                         bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_READ,
 850                                 fs_info, scrub_repair_read_endio, stripe);
 851                         bbio->bio.bi_iter.bi_sector = (stripe->logical +
 852                                 (i << fs_info->sectorsize_bits)) >> SECTOR_SHIFT;
 853                 }
 854
 855                 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
 856                 ASSERT(ret == fs_info->sectorsize);
 857         }
 858         if (bbio) {
 859                 ASSERT(bbio->bio.bi_iter.bi_size);
 860                 atomic_inc(&stripe->pending_io);
 861                 btrfs_submit_bio(bbio, mirror);
 862                 if (wait)
 863                         wait_scrub_stripe_io(stripe);
 864         }
 865 }
 866
 867 static void scrub_stripe_report_errors(struct scrub_ctx *sctx,
 868                                        struct scrub_stripe *stripe)
 869 {
 870         static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
 871                                       DEFAULT_RATELIMIT_BURST);
 872         struct btrfs_fs_info *fs_info = sctx->fs_info;
 873         struct btrfs_device *dev = NULL;
 874         u64 physical = 0;
 875         int nr_data_sectors = 0;
 876         int nr_meta_sectors = 0;
 877         int nr_nodatacsum_sectors = 0;
 878         int nr_repaired_sectors = 0;
 879         int sector_nr;
 880
 881         if (test_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state))
 882                 return;
 883
 884         /*
 885          * Init needed infos for error reporting.
 886          *
 887          * Although our scrub_stripe infrastucture is mostly based on btrfs_submit_bio()
 888          * thus no need for dev/physical, error reporting still needs dev and physical.
 889          */
 890         if (!bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors)) {
 891                 u64 mapped_len = fs_info->sectorsize;
 892                 struct btrfs_io_context *bioc = NULL;
 893                 int stripe_index = stripe->mirror_num - 1;
 894                 int ret;
 895
 896                 /* For scrub, our mirror_num should always start at 1. */
 897                 ASSERT(stripe->mirror_num >= 1);
 898                 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
 899                                       stripe->logical, &mapped_len, &bioc,
 900                                       NULL, NULL, 1);
 901                 /*
 902                  * If we failed, dev will be NULL, and later detailed reports
 903                  * will just be skipped.
 904                  */
 905                 if (ret < 0)
 906                         goto skip;
 907                 physical = bioc->stripes[stripe_index].physical;
 908                 dev = bioc->stripes[stripe_index].dev;
 909                 btrfs_put_bioc(bioc);
 910         }
 911
 912 skip:
 913         for_each_set_bit(sector_nr, &stripe->extent_sector_bitmap, stripe->nr_sectors) {
 914                 bool repaired = false;
 915
 916                 if (stripe->sectors[sector_nr].is_metadata) {
 917                         nr_meta_sectors++;
 918                 } else {
 919                         nr_data_sectors++;
 920                         if (!stripe->sectors[sector_nr].csum)
 921                                 nr_nodatacsum_sectors++;
 922                 }
 923
 924                 if (test_bit(sector_nr, &stripe->init_error_bitmap) &&
 925                     !test_bit(sector_nr, &stripe->error_bitmap)) {
 926                         nr_repaired_sectors++;
 927                         repaired = true;
 928                 }
 929
 930                 /* Good sector from the beginning, nothing need to be done. */
 931                 if (!test_bit(sector_nr, &stripe->init_error_bitmap))
 932                         continue;
 933
 934                 /*
 935                  * Report error for the corrupted sectors.  If repaired, just
 936                  * output the message of repaired message.
 937                  */
 938                 if (repaired) {
 939                         if (dev) {
 940                                 btrfs_err_rl_in_rcu(fs_info,
 941                         "fixed up error at logical %llu on dev %s physical %llu",
 942                                             stripe->logical, btrfs_dev_name(dev),
 943                                             physical);
 944                         } else {
 945                                 btrfs_err_rl_in_rcu(fs_info,
 946                         "fixed up error at logical %llu on mirror %u",
 947                                             stripe->logical, stripe->mirror_num);
 948                         }
 949                         continue;
 950                 }
 951
 952                 /* The remaining are all for unrepaired. */
 953                 if (dev) {
 954                         btrfs_err_rl_in_rcu(fs_info,
 955         "unable to fixup (regular) error at logical %llu on dev %s physical %llu",
 956                                             stripe->logical, btrfs_dev_name(dev),
 957                                             physical);
 958                 } else {
 959                         btrfs_err_rl_in_rcu(fs_info,
 960         "unable to fixup (regular) error at logical %llu on mirror %u",
 961                                             stripe->logical, stripe->mirror_num);
 962                 }
 963
 964                 if (test_bit(sector_nr, &stripe->io_error_bitmap))
 965                         if (__ratelimit(&rs) && dev)
 966                                 scrub_print_common_warning("i/o error", dev, false,
 967                                                      stripe->logical, physical);
 968                 if (test_bit(sector_nr, &stripe->csum_error_bitmap))
 969                         if (__ratelimit(&rs) && dev)
 970                                 scrub_print_common_warning("checksum error", dev, false,
 971                                                      stripe->logical, physical);
 972                 if (test_bit(sector_nr, &stripe->meta_error_bitmap))
 973                         if (__ratelimit(&rs) && dev)
 974                                 scrub_print_common_warning("header error", dev, false,
 975                                                      stripe->logical, physical);
 976         }
 977
 978         spin_lock(&sctx->stat_lock);
 979         sctx->stat.data_extents_scrubbed += stripe->nr_data_extents;
 980         sctx->stat.tree_extents_scrubbed += stripe->nr_meta_extents;
 981         sctx->stat.data_bytes_scrubbed += nr_data_sectors << fs_info->sectorsize_bits;
 982         sctx->stat.tree_bytes_scrubbed += nr_meta_sectors << fs_info->sectorsize_bits;
 983         sctx->stat.no_csum += nr_nodatacsum_sectors;
 984         sctx->stat.read_errors += stripe->init_nr_io_errors;
 985         sctx->stat.csum_errors += stripe->init_nr_csum_errors;
 986         sctx->stat.verify_errors += stripe->init_nr_meta_errors;
 987         sctx->stat.uncorrectable_errors +=
 988                 bitmap_weight(&stripe->error_bitmap, stripe->nr_sectors);
 989         sctx->stat.corrected_errors += nr_repaired_sectors;
 990         spin_unlock(&sctx->stat_lock);
 991 }
 992
 993 static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
 994                                 unsigned long write_bitmap, bool dev_replace);
 995
 996 /*
 997  * The main entrance for all read related scrub work, including:
 998  *
 999  * - Wait for the initial read to finish
1000  * - Verify and locate any bad sectors
1001  * - Go through the remaining mirrors and try to read as large blocksize as
1002  *   possible
1003  * - Go through all mirrors (including the failed mirror) sector-by-sector
1004  * - Submit writeback for repaired sectors
1005  *
1006  * Writeback for dev-replace does not happen here, it needs extra
1007  * synchronization for zoned devices.
1008  */
1009 static void scrub_stripe_read_repair_worker(struct work_struct *work)
1010 {
1011         struct scrub_stripe *stripe = container_of(work, struct scrub_stripe, work);
1012         struct scrub_ctx *sctx = stripe->sctx;
1013         struct btrfs_fs_info *fs_info = sctx->fs_info;
1014         int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1015                                           stripe->bg->length);
1016         int mirror;
1017         int i;
1018
1019         ASSERT(stripe->mirror_num > 0);
1020
1021         wait_scrub_stripe_io(stripe);
1022         scrub_verify_one_stripe(stripe, stripe->extent_sector_bitmap);
1023         /* Save the initial failed bitmap for later repair and report usage. */
1024         stripe->init_error_bitmap = stripe->error_bitmap;
1025         stripe->init_nr_io_errors = bitmap_weight(&stripe->io_error_bitmap,
1026                                                   stripe->nr_sectors);
1027         stripe->init_nr_csum_errors = bitmap_weight(&stripe->csum_error_bitmap,
1028                                                     stripe->nr_sectors);
1029         stripe->init_nr_meta_errors = bitmap_weight(&stripe->meta_error_bitmap,
1030                                                     stripe->nr_sectors);
1031
1032         if (bitmap_empty(&stripe->init_error_bitmap, stripe->nr_sectors))
1033                 goto out;
1034
1035         /*
1036          * Try all remaining mirrors.
1037          *
1038          * Here we still try to read as large block as possible, as this is
1039          * faster and we have extra safety nets to rely on.
1040          */
1041         for (mirror = calc_next_mirror(stripe->mirror_num, num_copies);
1042              mirror != stripe->mirror_num;
1043              mirror = calc_next_mirror(mirror, num_copies)) {
1044                 const unsigned long old_error_bitmap = stripe->error_bitmap;
1045
1046                 scrub_stripe_submit_repair_read(stripe, mirror,
1047                                                 BTRFS_STRIPE_LEN, false);
1048                 wait_scrub_stripe_io(stripe);
1049                 scrub_verify_one_stripe(stripe, old_error_bitmap);
1050                 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1051                         goto out;
1052         }
1053
1054         /*
1055          * Last safety net, try re-checking all mirrors, including the failed
1056          * one, sector-by-sector.
1057          *
1058          * As if one sector failed the drive's internal csum, the whole read
1059          * containing the offending sector would be marked as error.
1060          * Thus here we do sector-by-sector read.
1061          *
1062          * This can be slow, thus we only try it as the last resort.
1063          */
1064
1065         for (i = 0, mirror = stripe->mirror_num;
1066              i < num_copies;
1067              i++, mirror = calc_next_mirror(mirror, num_copies)) {
1068                 const unsigned long old_error_bitmap = stripe->error_bitmap;
1069
1070                 scrub_stripe_submit_repair_read(stripe, mirror,
1071                                                 fs_info->sectorsize, true);
1072                 wait_scrub_stripe_io(stripe);
1073                 scrub_verify_one_stripe(stripe, old_error_bitmap);
1074                 if (bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1075                         goto out;
1076         }
1077 out:
1078         /*
1079          * Submit the repaired sectors.  For zoned case, we cannot do repair
1080          * in-place, but queue the bg to be relocated.
1081          */
1082         if (btrfs_is_zoned(fs_info)) {
1083                 if (!bitmap_empty(&stripe->error_bitmap, stripe->nr_sectors))
1084                         btrfs_repair_one_zone(fs_info, sctx->stripes[0].bg->start);
1085         } else if (!sctx->readonly) {
1086                 unsigned long repaired;
1087
1088                 bitmap_andnot(&repaired, &stripe->init_error_bitmap,
1089                               &stripe->error_bitmap, stripe->nr_sectors);
1090                 scrub_write_sectors(sctx, stripe, repaired, false);
1091                 wait_scrub_stripe_io(stripe);
1092         }
1093
1094         scrub_stripe_report_errors(sctx, stripe);
1095         set_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state);
1096         wake_up(&stripe->repair_wait);
1097 }
1098
1099 static void scrub_read_endio(struct btrfs_bio *bbio)
1100 {
1101         struct scrub_stripe *stripe = bbio->private;
1102
1103         if (bbio->bio.bi_status) {
1104                 bitmap_set(&stripe->io_error_bitmap, 0, stripe->nr_sectors);
1105                 bitmap_set(&stripe->error_bitmap, 0, stripe->nr_sectors);
1106         } else {
1107                 bitmap_clear(&stripe->io_error_bitmap, 0, stripe->nr_sectors);
1108         }
1109         bio_put(&bbio->bio);
1110         if (atomic_dec_and_test(&stripe->pending_io)) {
1111                 wake_up(&stripe->io_wait);
1112                 INIT_WORK(&stripe->work, scrub_stripe_read_repair_worker);
1113                 queue_work(stripe->bg->fs_info->scrub_workers, &stripe->work);
1114         }
1115 }
1116
1117 static void scrub_write_endio(struct btrfs_bio *bbio)
1118 {
1119         struct scrub_stripe *stripe = bbio->private;
1120         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1121         struct bio_vec *bvec;
1122         int sector_nr = calc_sector_number(stripe, bio_first_bvec_all(&bbio->bio));
1123         u32 bio_size = 0;
1124         int i;
1125
1126         bio_for_each_bvec_all(bvec, &bbio->bio, i)
1127                 bio_size += bvec->bv_len;
1128
1129         if (bbio->bio.bi_status) {
1130                 unsigned long flags;
1131
1132                 spin_lock_irqsave(&stripe->write_error_lock, flags);
1133                 bitmap_set(&stripe->write_error_bitmap, sector_nr,
1134                            bio_size >> fs_info->sectorsize_bits);
1135                 spin_unlock_irqrestore(&stripe->write_error_lock, flags);
1136         }
1137         bio_put(&bbio->bio);
1138
1139         if (atomic_dec_and_test(&stripe->pending_io))
1140                 wake_up(&stripe->io_wait);
1141 }
1142
1143 static void scrub_submit_write_bio(struct scrub_ctx *sctx,
1144                                    struct scrub_stripe *stripe,
1145                                    struct btrfs_bio *bbio, bool dev_replace)
1146 {
1147         struct btrfs_fs_info *fs_info = sctx->fs_info;
1148         u32 bio_len = bbio->bio.bi_iter.bi_size;
1149         u32 bio_off = (bbio->bio.bi_iter.bi_sector << SECTOR_SHIFT) -
1150                       stripe->logical;
1151
1152         fill_writer_pointer_gap(sctx, stripe->physical + bio_off);
1153         atomic_inc(&stripe->pending_io);
1154         btrfs_submit_repair_write(bbio, stripe->mirror_num, dev_replace);
1155         if (!btrfs_is_zoned(fs_info))
1156                 return;
1157         /*
1158          * For zoned writeback, queue depth must be 1, thus we must wait for
1159          * the write to finish before the next write.
1160          */
1161         wait_scrub_stripe_io(stripe);
1162
1163         /*
1164          * And also need to update the write pointer if write finished
1165          * successfully.
1166          */
1167         if (!test_bit(bio_off >> fs_info->sectorsize_bits,
1168                       &stripe->write_error_bitmap))
1169                 sctx->write_pointer += bio_len;
1170 }
1171
1172 /*
1173  * Submit the write bio(s) for the sectors specified by @write_bitmap.
1174  *
1175  * Here we utilize btrfs_submit_repair_write(), which has some extra benefits:
1176  *
1177  * - Only needs logical bytenr and mirror_num
1178  *   Just like the scrub read path
1179  *
1180  * - Would only result in writes to the specified mirror
1181  *   Unlike the regular writeback path, which would write back to all stripes
1182  *
1183  * - Handle dev-replace and read-repair writeback differently
1184  */
1185 static void scrub_write_sectors(struct scrub_ctx *sctx, struct scrub_stripe *stripe,
1186                                 unsigned long write_bitmap, bool dev_replace)
1187 {
1188         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1189         struct btrfs_bio *bbio = NULL;
1190         int sector_nr;
1191
1192         for_each_set_bit(sector_nr, &write_bitmap, stripe->nr_sectors) {
1193                 struct page *page = scrub_stripe_get_page(stripe, sector_nr);
1194                 unsigned int pgoff = scrub_stripe_get_page_offset(stripe, sector_nr);
1195                 int ret;
1196
1197                 /* We should only writeback sectors covered by an extent. */
1198                 ASSERT(test_bit(sector_nr, &stripe->extent_sector_bitmap));
1199
1200                 /* Cannot merge with previous sector, submit the current one. */
1201                 if (bbio && sector_nr && !test_bit(sector_nr - 1, &write_bitmap)) {
1202                         scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1203                         bbio = NULL;
1204                 }
1205                 if (!bbio) {
1206                         bbio = btrfs_bio_alloc(stripe->nr_sectors, REQ_OP_WRITE,
1207                                                fs_info, scrub_write_endio, stripe);
1208                         bbio->bio.bi_iter.bi_sector = (stripe->logical +
1209                                 (sector_nr << fs_info->sectorsize_bits)) >>
1210                                 SECTOR_SHIFT;
1211                 }
1212                 ret = bio_add_page(&bbio->bio, page, fs_info->sectorsize, pgoff);
1213                 ASSERT(ret == fs_info->sectorsize);
1214         }
1215         if (bbio)
1216                 scrub_submit_write_bio(sctx, stripe, bbio, dev_replace);
1217 }
1218
1219 /*
1220  * Throttling of IO submission, bandwidth-limit based, the timeslice is 1
1221  * second.  Limit can be set via /sys/fs/UUID/devinfo/devid/scrub_speed_max.
1222  */
1223 static void scrub_throttle_dev_io(struct scrub_ctx *sctx, struct btrfs_device *device,
1224                                   unsigned int bio_size)
1225 {
1226         const int time_slice = 1000;
1227         s64 delta;
1228         ktime_t now;
1229         u32 div;
1230         u64 bwlimit;
1231
1232         bwlimit = READ_ONCE(device->scrub_speed_max);
1233         if (bwlimit == 0)
1234                 return;
1235
1236         /*
1237          * Slice is divided into intervals when the IO is submitted, adjust by
1238          * bwlimit and maximum of 64 intervals.
1239          */
1240         div = max_t(u32, 1, (u32)(bwlimit / (16 * 1024 * 1024)));
1241         div = min_t(u32, 64, div);
1242
1243         /* Start new epoch, set deadline */
1244         now = ktime_get();
1245         if (sctx->throttle_deadline == 0) {
1246                 sctx->throttle_deadline = ktime_add_ms(now, time_slice / div);
1247                 sctx->throttle_sent = 0;
1248         }
1249
1250         /* Still in the time to send? */
1251         if (ktime_before(now, sctx->throttle_deadline)) {
1252                 /* If current bio is within the limit, send it */
1253                 sctx->throttle_sent += bio_size;
1254                 if (sctx->throttle_sent <= div_u64(bwlimit, div))
1255                         return;
1256
1257                 /* We're over the limit, sleep until the rest of the slice */
1258                 delta = ktime_ms_delta(sctx->throttle_deadline, now);
1259         } else {
1260                 /* New request after deadline, start new epoch */
1261                 delta = 0;
1262         }
1263
1264         if (delta) {
1265                 long timeout;
1266
1267                 timeout = div_u64(delta * HZ, 1000);
1268                 schedule_timeout_interruptible(timeout);
1269         }
1270
1271         /* Next call will start the deadline period */
1272         sctx->throttle_deadline = 0;
1273 }
1274
1275 /*
1276  * Given a physical address, this will calculate it's
1277  * logical offset. if this is a parity stripe, it will return
1278  * the most left data stripe's logical offset.
1279  *
1280  * return 0 if it is a data stripe, 1 means parity stripe.
1281  */
1282 static int get_raid56_logic_offset(u64 physical, int num,
1283                                    struct map_lookup *map, u64 *offset,
1284                                    u64 *stripe_start)
1285 {
1286         int i;
1287         int j = 0;
1288         u64 last_offset;
1289         const int data_stripes = nr_data_stripes(map);
1290
1291         last_offset = (physical - map->stripes[num].physical) * data_stripes;
1292         if (stripe_start)
1293                 *stripe_start = last_offset;
1294
1295         *offset = last_offset;
1296         for (i = 0; i < data_stripes; i++) {
1297                 u32 stripe_nr;
1298                 u32 stripe_index;
1299                 u32 rot;
1300
1301                 *offset = last_offset + btrfs_stripe_nr_to_offset(i);
1302
1303                 stripe_nr = (u32)(*offset >> BTRFS_STRIPE_LEN_SHIFT) / data_stripes;
1304
1305                 /* Work out the disk rotation on this stripe-set */
1306                 rot = stripe_nr % map->num_stripes;
1307                 /* calculate which stripe this data locates */
1308                 rot += i;
1309                 stripe_index = rot % map->num_stripes;
1310                 if (stripe_index == num)
1311                         return 0;
1312                 if (stripe_index < num)
1313                         j++;
1314         }
1315         *offset = last_offset + btrfs_stripe_nr_to_offset(j);
1316         return 1;
1317 }
1318
1319 /*
1320  * Return 0 if the extent item range covers any byte of the range.
1321  * Return <0 if the extent item is before @search_start.
1322  * Return >0 if the extent item is after @start_start + @search_len.
1323  */
1324 static int compare_extent_item_range(struct btrfs_path *path,
1325                                      u64 search_start, u64 search_len)
1326 {
1327         struct btrfs_fs_info *fs_info = path->nodes[0]->fs_info;
1328         u64 len;
1329         struct btrfs_key key;
1330
1331         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1332         ASSERT(key.type == BTRFS_EXTENT_ITEM_KEY ||
1333                key.type == BTRFS_METADATA_ITEM_KEY);
1334         if (key.type == BTRFS_METADATA_ITEM_KEY)
1335                 len = fs_info->nodesize;
1336         else
1337                 len = key.offset;
1338
1339         if (key.objectid + len <= search_start)
1340                 return -1;
1341         if (key.objectid >= search_start + search_len)
1342                 return 1;
1343         return 0;
1344 }
1345
1346 /*
1347  * Locate one extent item which covers any byte in range
1348  * [@search_start, @search_start + @search_length)
1349  *
1350  * If the path is not initialized, we will initialize the search by doing
1351  * a btrfs_search_slot().
1352  * If the path is already initialized, we will use the path as the initial
1353  * slot, to avoid duplicated btrfs_search_slot() calls.
1354  *
1355  * NOTE: If an extent item starts before @search_start, we will still
1356  * return the extent item. This is for data extent crossing stripe boundary.
1357  *
1358  * Return 0 if we found such extent item, and @path will point to the extent item.
1359  * Return >0 if no such extent item can be found, and @path will be released.
1360  * Return <0 if hit fatal error, and @path will be released.
1361  */
1362 static int find_first_extent_item(struct btrfs_root *extent_root,
1363                                   struct btrfs_path *path,
1364                                   u64 search_start, u64 search_len)
1365 {
1366         struct btrfs_fs_info *fs_info = extent_root->fs_info;
1367         struct btrfs_key key;
1368         int ret;
1369
1370         /* Continue using the existing path */
1371         if (path->nodes[0])
1372                 goto search_forward;
1373
1374         if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
1375                 key.type = BTRFS_METADATA_ITEM_KEY;
1376         else
1377                 key.type = BTRFS_EXTENT_ITEM_KEY;
1378         key.objectid = search_start;
1379         key.offset = (u64)-1;
1380
1381         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
1382         if (ret < 0)
1383                 return ret;
1384
1385         ASSERT(ret > 0);
1386         /*
1387          * Here we intentionally pass 0 as @min_objectid, as there could be
1388          * an extent item starting before @search_start.
1389          */
1390         ret = btrfs_previous_extent_item(extent_root, path, 0);
1391         if (ret < 0)
1392                 return ret;
1393         /*
1394          * No matter whether we have found an extent item, the next loop will
1395          * properly do every check on the key.
1396          */
1397 search_forward:
1398         while (true) {
1399                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1400                 if (key.objectid >= search_start + search_len)
1401                         break;
1402                 if (key.type != BTRFS_METADATA_ITEM_KEY &&
1403                     key.type != BTRFS_EXTENT_ITEM_KEY)
1404                         goto next;
1405
1406                 ret = compare_extent_item_range(path, search_start, search_len);
1407                 if (ret == 0)
1408                         return ret;
1409                 if (ret > 0)
1410                         break;
1411 next:
1412                 path->slots[0]++;
1413                 if (path->slots[0] >= btrfs_header_nritems(path->nodes[0])) {
1414                         ret = btrfs_next_leaf(extent_root, path);
1415                         if (ret) {
1416                                 /* Either no more item or fatal error */
1417                                 btrfs_release_path(path);
1418                                 return ret;
1419                         }
1420                 }
1421         }
1422         btrfs_release_path(path);
1423         return 1;
1424 }
1425
1426 static void get_extent_info(struct btrfs_path *path, u64 *extent_start_ret,
1427                             u64 *size_ret, u64 *flags_ret, u64 *generation_ret)
1428 {
1429         struct btrfs_key key;
1430         struct btrfs_extent_item *ei;
1431
1432         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1433         ASSERT(key.type == BTRFS_METADATA_ITEM_KEY ||
1434                key.type == BTRFS_EXTENT_ITEM_KEY);
1435         *extent_start_ret = key.objectid;
1436         if (key.type == BTRFS_METADATA_ITEM_KEY)
1437                 *size_ret = path->nodes[0]->fs_info->nodesize;
1438         else
1439                 *size_ret = key.offset;
1440         ei = btrfs_item_ptr(path->nodes[0], path->slots[0], struct btrfs_extent_item);
1441         *flags_ret = btrfs_extent_flags(path->nodes[0], ei);
1442         *generation_ret = btrfs_extent_generation(path->nodes[0], ei);
1443 }
1444
1445 static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
1446                                         u64 physical, u64 physical_end)
1447 {
1448         struct btrfs_fs_info *fs_info = sctx->fs_info;
1449         int ret = 0;
1450
1451         if (!btrfs_is_zoned(fs_info))
1452                 return 0;
1453
1454         mutex_lock(&sctx->wr_lock);
1455         if (sctx->write_pointer < physical_end) {
1456                 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
1457                                                     physical,
1458                                                     sctx->write_pointer);
1459                 if (ret)
1460                         btrfs_err(fs_info,
1461                                   "zoned: failed to recover write pointer");
1462         }
1463         mutex_unlock(&sctx->wr_lock);
1464         btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
1465
1466         return ret;
1467 }
1468
1469 static void fill_one_extent_info(struct btrfs_fs_info *fs_info,
1470                                  struct scrub_stripe *stripe,
1471                                  u64 extent_start, u64 extent_len,
1472                                  u64 extent_flags, u64 extent_gen)
1473 {
1474         for (u64 cur_logical = max(stripe->logical, extent_start);
1475              cur_logical < min(stripe->logical + BTRFS_STRIPE_LEN,
1476                                extent_start + extent_len);
1477              cur_logical += fs_info->sectorsize) {
1478                 const int nr_sector = (cur_logical - stripe->logical) >>
1479                                       fs_info->sectorsize_bits;
1480                 struct scrub_sector_verification *sector =
1481                                                 &stripe->sectors[nr_sector];
1482
1483                 set_bit(nr_sector, &stripe->extent_sector_bitmap);
1484                 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1485                         sector->is_metadata = true;
1486                         sector->generation = extent_gen;
1487                 }
1488         }
1489 }
1490
1491 static void scrub_stripe_reset_bitmaps(struct scrub_stripe *stripe)
1492 {
1493         stripe->extent_sector_bitmap = 0;
1494         stripe->init_error_bitmap = 0;
1495         stripe->init_nr_io_errors = 0;
1496         stripe->init_nr_csum_errors = 0;
1497         stripe->init_nr_meta_errors = 0;
1498         stripe->error_bitmap = 0;
1499         stripe->io_error_bitmap = 0;
1500         stripe->csum_error_bitmap = 0;
1501         stripe->meta_error_bitmap = 0;
1502 }
1503
1504 /*
1505  * Locate one stripe which has at least one extent in its range.
1506  *
1507  * Return 0 if found such stripe, and store its info into @stripe.
1508  * Return >0 if there is no such stripe in the specified range.
1509  * Return <0 for error.
1510  */
1511 static int scrub_find_fill_first_stripe(struct btrfs_block_group *bg,
1512                                         struct btrfs_path *extent_path,
1513                                         struct btrfs_path *csum_path,
1514                                         struct btrfs_device *dev, u64 physical,
1515                                         int mirror_num, u64 logical_start,
1516                                         u32 logical_len,
1517                                         struct scrub_stripe *stripe)
1518 {
1519         struct btrfs_fs_info *fs_info = bg->fs_info;
1520         struct btrfs_root *extent_root = btrfs_extent_root(fs_info, bg->start);
1521         struct btrfs_root *csum_root = btrfs_csum_root(fs_info, bg->start);
1522         const u64 logical_end = logical_start + logical_len;
1523         u64 cur_logical = logical_start;
1524         u64 stripe_end;
1525         u64 extent_start;
1526         u64 extent_len;
1527         u64 extent_flags;
1528         u64 extent_gen;
1529         int ret;
1530
1531         memset(stripe->sectors, 0, sizeof(struct scrub_sector_verification) *
1532                                    stripe->nr_sectors);
1533         scrub_stripe_reset_bitmaps(stripe);
1534
1535         /* The range must be inside the bg. */
1536         ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
1537
1538         ret = find_first_extent_item(extent_root, extent_path, logical_start,
1539                                      logical_len);
1540         /* Either error or not found. */
1541         if (ret)
1542                 goto out;
1543         get_extent_info(extent_path, &extent_start, &extent_len, &extent_flags,
1544                         &extent_gen);
1545         if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1546                 stripe->nr_meta_extents++;
1547         if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1548                 stripe->nr_data_extents++;
1549         cur_logical = max(extent_start, cur_logical);
1550
1551         /*
1552          * Round down to stripe boundary.
1553          *
1554          * The extra calculation against bg->start is to handle block groups
1555          * whose logical bytenr is not BTRFS_STRIPE_LEN aligned.
1556          */
1557         stripe->logical = round_down(cur_logical - bg->start, BTRFS_STRIPE_LEN) +
1558                           bg->start;
1559         stripe->physical = physical + stripe->logical - logical_start;
1560         stripe->dev = dev;
1561         stripe->bg = bg;
1562         stripe->mirror_num = mirror_num;
1563         stripe_end = stripe->logical + BTRFS_STRIPE_LEN - 1;
1564
1565         /* Fill the first extent info into stripe->sectors[] array. */
1566         fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1567                              extent_flags, extent_gen);
1568         cur_logical = extent_start + extent_len;
1569
1570         /* Fill the extent info for the remaining sectors. */
1571         while (cur_logical <= stripe_end) {
1572                 ret = find_first_extent_item(extent_root, extent_path, cur_logical,
1573                                              stripe_end - cur_logical + 1);
1574                 if (ret < 0)
1575                         goto out;
1576                 if (ret > 0) {
1577                         ret = 0;
1578                         break;
1579                 }
1580                 get_extent_info(extent_path, &extent_start, &extent_len,
1581                                 &extent_flags, &extent_gen);
1582                 if (extent_flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1583                         stripe->nr_meta_extents++;
1584                 if (extent_flags & BTRFS_EXTENT_FLAG_DATA)
1585                         stripe->nr_data_extents++;
1586                 fill_one_extent_info(fs_info, stripe, extent_start, extent_len,
1587                                      extent_flags, extent_gen);
1588                 cur_logical = extent_start + extent_len;
1589         }
1590
1591         /* Now fill the data csum. */
1592         if (bg->flags & BTRFS_BLOCK_GROUP_DATA) {
1593                 int sector_nr;
1594                 unsigned long csum_bitmap = 0;
1595
1596                 /* Csum space should have already been allocated. */
1597                 ASSERT(stripe->csums);
1598
1599                 /*
1600                  * Our csum bitmap should be large enough, as BTRFS_STRIPE_LEN
1601                  * should contain at most 16 sectors.
1602                  */
1603                 ASSERT(BITS_PER_LONG >= BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1604
1605                 ret = btrfs_lookup_csums_bitmap(csum_root, csum_path,
1606                                                 stripe->logical, stripe_end,
1607                                                 stripe->csums, &csum_bitmap);
1608                 if (ret < 0)
1609                         goto out;
1610                 if (ret > 0)
1611                         ret = 0;
1612
1613                 for_each_set_bit(sector_nr, &csum_bitmap, stripe->nr_sectors) {
1614                         stripe->sectors[sector_nr].csum = stripe->csums +
1615                                 sector_nr * fs_info->csum_size;
1616                 }
1617         }
1618         set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1619 out:
1620         return ret;
1621 }
1622
1623 static void scrub_reset_stripe(struct scrub_stripe *stripe)
1624 {
1625         scrub_stripe_reset_bitmaps(stripe);
1626
1627         stripe->nr_meta_extents = 0;
1628         stripe->nr_data_extents = 0;
1629         stripe->state = 0;
1630
1631         for (int i = 0; i < stripe->nr_sectors; i++) {
1632                 stripe->sectors[i].is_metadata = false;
1633                 stripe->sectors[i].csum = NULL;
1634                 stripe->sectors[i].generation = 0;
1635         }
1636 }
1637
1638 static void scrub_submit_initial_read(struct scrub_ctx *sctx,
1639                                       struct scrub_stripe *stripe)
1640 {
1641         struct btrfs_fs_info *fs_info = sctx->fs_info;
1642         struct btrfs_bio *bbio;
1643         int mirror = stripe->mirror_num;
1644
1645         ASSERT(stripe->bg);
1646         ASSERT(stripe->mirror_num > 0);
1647         ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1648
1649         bbio = btrfs_bio_alloc(SCRUB_STRIPE_PAGES, REQ_OP_READ, fs_info,
1650                                scrub_read_endio, stripe);
1651
1652         /* Read the whole stripe. */
1653         bbio->bio.bi_iter.bi_sector = stripe->logical >> SECTOR_SHIFT;
1654         for (int i = 0; i < BTRFS_STRIPE_LEN >> PAGE_SHIFT; i++) {
1655                 int ret;
1656
1657                 ret = bio_add_page(&bbio->bio, stripe->pages[i], PAGE_SIZE, 0);
1658                 /* We should have allocated enough bio vectors. */
1659                 ASSERT(ret == PAGE_SIZE);
1660         }
1661         atomic_inc(&stripe->pending_io);
1662
1663         /*
1664          * For dev-replace, either user asks to avoid the source dev, or
1665          * the device is missing, we try the next mirror instead.
1666          */
1667         if (sctx->is_dev_replace &&
1668             (fs_info->dev_replace.cont_reading_from_srcdev_mode ==
1669              BTRFS_DEV_REPLACE_ITEM_CONT_READING_FROM_SRCDEV_MODE_AVOID ||
1670              !stripe->dev->bdev)) {
1671                 int num_copies = btrfs_num_copies(fs_info, stripe->bg->start,
1672                                                   stripe->bg->length);
1673
1674                 mirror = calc_next_mirror(mirror, num_copies);
1675         }
1676         btrfs_submit_bio(bbio, mirror);
1677 }
1678
1679 static bool stripe_has_metadata_error(struct scrub_stripe *stripe)
1680 {
1681         int i;
1682
1683         for_each_set_bit(i, &stripe->error_bitmap, stripe->nr_sectors) {
1684                 if (stripe->sectors[i].is_metadata) {
1685                         struct btrfs_fs_info *fs_info = stripe->bg->fs_info;
1686
1687                         btrfs_err(fs_info,
1688                         "stripe %llu has unrepaired metadata sector at %llu",
1689                                   stripe->logical,
1690                                   stripe->logical + (i << fs_info->sectorsize_bits));
1691                         return true;
1692                 }
1693         }
1694         return false;
1695 }
1696
1697 static void submit_initial_group_read(struct scrub_ctx *sctx,
1698                                       unsigned int first_slot,
1699                                       unsigned int nr_stripes)
1700 {
1701         struct blk_plug plug;
1702
1703         ASSERT(first_slot < SCRUB_TOTAL_STRIPES);
1704         ASSERT(first_slot + nr_stripes <= SCRUB_TOTAL_STRIPES);
1705
1706         scrub_throttle_dev_io(sctx, sctx->stripes[0].dev,
1707                               btrfs_stripe_nr_to_offset(nr_stripes));
1708         blk_start_plug(&plug);
1709         for (int i = 0; i < nr_stripes; i++) {
1710                 struct scrub_stripe *stripe = &sctx->stripes[first_slot + i];
1711
1712                 /* Those stripes should be initialized. */
1713                 ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state));
1714                 scrub_submit_initial_read(sctx, stripe);
1715         }
1716         blk_finish_plug(&plug);
1717 }
1718
1719 static int flush_scrub_stripes(struct scrub_ctx *sctx)
1720 {
1721         struct btrfs_fs_info *fs_info = sctx->fs_info;
1722         struct scrub_stripe *stripe;
1723         const int nr_stripes = sctx->cur_stripe;
1724         int ret = 0;
1725
1726         if (!nr_stripes)
1727                 return 0;
1728
1729         ASSERT(test_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &sctx->stripes[0].state));
1730
1731         /* Submit the stripes which are populated but not submitted. */
1732         if (nr_stripes % SCRUB_STRIPES_PER_GROUP) {
1733                 const int first_slot = round_down(nr_stripes, SCRUB_STRIPES_PER_GROUP);
1734
1735                 submit_initial_group_read(sctx, first_slot, nr_stripes - first_slot);
1736         }
1737
1738         for (int i = 0; i < nr_stripes; i++) {
1739                 stripe = &sctx->stripes[i];
1740
1741                 wait_event(stripe->repair_wait,
1742                            test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1743         }
1744
1745         /* Submit for dev-replace. */
1746         if (sctx->is_dev_replace) {
1747                 /*
1748                  * For dev-replace, if we know there is something wrong with
1749                  * metadata, we should immedately abort.
1750                  */
1751                 for (int i = 0; i < nr_stripes; i++) {
1752                         if (stripe_has_metadata_error(&sctx->stripes[i])) {
1753                                 ret = -EIO;
1754                                 goto out;
1755                         }
1756                 }
1757                 for (int i = 0; i < nr_stripes; i++) {
1758                         unsigned long good;
1759
1760                         stripe = &sctx->stripes[i];
1761
1762                         ASSERT(stripe->dev == fs_info->dev_replace.srcdev);
1763
1764                         bitmap_andnot(&good, &stripe->extent_sector_bitmap,
1765                                       &stripe->error_bitmap, stripe->nr_sectors);
1766                         scrub_write_sectors(sctx, stripe, good, true);
1767                 }
1768         }
1769
1770         /* Wait for the above writebacks to finish. */
1771         for (int i = 0; i < nr_stripes; i++) {
1772                 stripe = &sctx->stripes[i];
1773
1774                 wait_scrub_stripe_io(stripe);
1775                 scrub_reset_stripe(stripe);
1776         }
1777 out:
1778         sctx->cur_stripe = 0;
1779         return ret;
1780 }
1781
1782 static void raid56_scrub_wait_endio(struct bio *bio)
1783 {
1784         complete(bio->bi_private);
1785 }
1786
1787 static int queue_scrub_stripe(struct scrub_ctx *sctx, struct btrfs_block_group *bg,
1788                               struct btrfs_device *dev, int mirror_num,
1789                               u64 logical, u32 length, u64 physical,
1790                               u64 *found_logical_ret)
1791 {
1792         struct scrub_stripe *stripe;
1793         int ret;
1794
1795         /*
1796          * There should always be one slot left, as caller filling the last
1797          * slot should flush them all.
1798          */
1799         ASSERT(sctx->cur_stripe < SCRUB_TOTAL_STRIPES);
1800
1801         stripe = &sctx->stripes[sctx->cur_stripe];
1802         scrub_reset_stripe(stripe);
1803         ret = scrub_find_fill_first_stripe(bg, &sctx->extent_path,
1804                                            &sctx->csum_path, dev, physical,
1805                                            mirror_num, logical, length, stripe);
1806         /* Either >0 as no more extents or <0 for error. */
1807         if (ret)
1808                 return ret;
1809         if (found_logical_ret)
1810                 *found_logical_ret = stripe->logical;
1811         sctx->cur_stripe++;
1812
1813         /* We filled one group, submit it. */
1814         if (sctx->cur_stripe % SCRUB_STRIPES_PER_GROUP == 0) {
1815                 const int first_slot = sctx->cur_stripe - SCRUB_STRIPES_PER_GROUP;
1816
1817                 submit_initial_group_read(sctx, first_slot, SCRUB_STRIPES_PER_GROUP);
1818         }
1819
1820         /* Last slot used, flush them all. */
1821         if (sctx->cur_stripe == SCRUB_TOTAL_STRIPES)
1822                 return flush_scrub_stripes(sctx);
1823         return 0;
1824 }
1825
1826 static int scrub_raid56_parity_stripe(struct scrub_ctx *sctx,
1827                                       struct btrfs_device *scrub_dev,
1828                                       struct btrfs_block_group *bg,
1829                                       struct map_lookup *map,
1830                                       u64 full_stripe_start)
1831 {
1832         DECLARE_COMPLETION_ONSTACK(io_done);
1833         struct btrfs_fs_info *fs_info = sctx->fs_info;
1834         struct btrfs_raid_bio *rbio;
1835         struct btrfs_io_context *bioc = NULL;
1836         struct btrfs_path extent_path = { 0 };
1837         struct btrfs_path csum_path = { 0 };
1838         struct bio *bio;
1839         struct scrub_stripe *stripe;
1840         bool all_empty = true;
1841         const int data_stripes = nr_data_stripes(map);
1842         unsigned long extent_bitmap = 0;
1843         u64 length = btrfs_stripe_nr_to_offset(data_stripes);
1844         int ret;
1845
1846         ASSERT(sctx->raid56_data_stripes);
1847
1848         /*
1849          * For data stripe search, we cannot re-use the same extent/csum paths,
1850          * as the data stripe bytenr may be smaller than previous extent.  Thus
1851          * we have to use our own extent/csum paths.
1852          */
1853         extent_path.search_commit_root = 1;
1854         extent_path.skip_locking = 1;
1855         csum_path.search_commit_root = 1;
1856         csum_path.skip_locking = 1;
1857
1858         for (int i = 0; i < data_stripes; i++) {
1859                 int stripe_index;
1860                 int rot;
1861                 u64 physical;
1862
1863                 stripe = &sctx->raid56_data_stripes[i];
1864                 rot = div_u64(full_stripe_start - bg->start,
1865                               data_stripes) >> BTRFS_STRIPE_LEN_SHIFT;
1866                 stripe_index = (i + rot) % map->num_stripes;
1867                 physical = map->stripes[stripe_index].physical +
1868                            btrfs_stripe_nr_to_offset(rot);
1869
1870                 scrub_reset_stripe(stripe);
1871                 set_bit(SCRUB_STRIPE_FLAG_NO_REPORT, &stripe->state);
1872                 ret = scrub_find_fill_first_stripe(bg, &extent_path, &csum_path,
1873                                 map->stripes[stripe_index].dev, physical, 1,
1874                                 full_stripe_start + btrfs_stripe_nr_to_offset(i),
1875                                 BTRFS_STRIPE_LEN, stripe);
1876                 if (ret < 0)
1877                         goto out;
1878                 /*
1879                  * No extent in this data stripe, need to manually mark them
1880                  * initialized to make later read submission happy.
1881                  */
1882                 if (ret > 0) {
1883                         stripe->logical = full_stripe_start +
1884                                           btrfs_stripe_nr_to_offset(i);
1885                         stripe->dev = map->stripes[stripe_index].dev;
1886                         stripe->mirror_num = 1;
1887                         set_bit(SCRUB_STRIPE_FLAG_INITIALIZED, &stripe->state);
1888                 }
1889         }
1890
1891         /* Check if all data stripes are empty. */
1892         for (int i = 0; i < data_stripes; i++) {
1893                 stripe = &sctx->raid56_data_stripes[i];
1894                 if (!bitmap_empty(&stripe->extent_sector_bitmap, stripe->nr_sectors)) {
1895                         all_empty = false;
1896                         break;
1897                 }
1898         }
1899         if (all_empty) {
1900                 ret = 0;
1901                 goto out;
1902         }
1903
1904         for (int i = 0; i < data_stripes; i++) {
1905                 stripe = &sctx->raid56_data_stripes[i];
1906                 scrub_submit_initial_read(sctx, stripe);
1907         }
1908         for (int i = 0; i < data_stripes; i++) {
1909                 stripe = &sctx->raid56_data_stripes[i];
1910
1911                 wait_event(stripe->repair_wait,
1912                            test_bit(SCRUB_STRIPE_FLAG_REPAIR_DONE, &stripe->state));
1913         }
1914         /* For now, no zoned support for RAID56. */
1915         ASSERT(!btrfs_is_zoned(sctx->fs_info));
1916
1917         /*
1918          * Now all data stripes are properly verified. Check if we have any
1919          * unrepaired, if so abort immediately or we could further corrupt the
1920          * P/Q stripes.
1921          *
1922          * During the loop, also populate extent_bitmap.
1923          */
1924         for (int i = 0; i < data_stripes; i++) {
1925                 unsigned long error;
1926
1927                 stripe = &sctx->raid56_data_stripes[i];
1928
1929                 /*
1930                  * We should only check the errors where there is an extent.
1931                  * As we may hit an empty data stripe while it's missing.
1932                  */
1933                 bitmap_and(&error, &stripe->error_bitmap,
1934                            &stripe->extent_sector_bitmap, stripe->nr_sectors);
1935                 if (!bitmap_empty(&error, stripe->nr_sectors)) {
1936                         btrfs_err(fs_info,
1937 "unrepaired sectors detected, full stripe %llu data stripe %u errors %*pbl",
1938                                   full_stripe_start, i, stripe->nr_sectors,
1939                                   &error);
1940                         ret = -EIO;
1941                         goto out;
1942                 }
1943                 bitmap_or(&extent_bitmap, &extent_bitmap,
1944                           &stripe->extent_sector_bitmap, stripe->nr_sectors);
1945         }
1946
1947         /* Now we can check and regenerate the P/Q stripe. */
1948         bio = bio_alloc(NULL, 1, REQ_OP_READ, GFP_NOFS);
1949         bio->bi_iter.bi_sector = full_stripe_start >> SECTOR_SHIFT;
1950         bio->bi_private = &io_done;
1951         bio->bi_end_io = raid56_scrub_wait_endio;
1952
1953         btrfs_bio_counter_inc_blocked(fs_info);
1954         ret = btrfs_map_block(fs_info, BTRFS_MAP_WRITE, full_stripe_start,
1955                               &length, &bioc, NULL, NULL, 1);
1956         if (ret < 0) {
1957                 btrfs_put_bioc(bioc);
1958                 btrfs_bio_counter_dec(fs_info);
1959                 goto out;
1960         }
1961         rbio = raid56_parity_alloc_scrub_rbio(bio, bioc, scrub_dev, &extent_bitmap,
1962                                 BTRFS_STRIPE_LEN >> fs_info->sectorsize_bits);
1963         btrfs_put_bioc(bioc);
1964         if (!rbio) {
1965                 ret = -ENOMEM;
1966                 btrfs_bio_counter_dec(fs_info);
1967                 goto out;
1968         }
1969         /* Use the recovered stripes as cache to avoid read them from disk again. */
1970         for (int i = 0; i < data_stripes; i++) {
1971                 stripe = &sctx->raid56_data_stripes[i];
1972
1973                 raid56_parity_cache_data_pages(rbio, stripe->pages,
1974                                 full_stripe_start + (i << BTRFS_STRIPE_LEN_SHIFT));
1975         }
1976         raid56_parity_submit_scrub_rbio(rbio);
1977         wait_for_completion_io(&io_done);
1978         ret = blk_status_to_errno(bio->bi_status);
1979         bio_put(bio);
1980         btrfs_bio_counter_dec(fs_info);
1981
1982         btrfs_release_path(&extent_path);
1983         btrfs_release_path(&csum_path);
1984 out:
1985         return ret;
1986 }
1987
1988 /*
1989  * Scrub one range which can only has simple mirror based profile.
1990  * (Including all range in SINGLE/DUP/RAID1/RAID1C*, and each stripe in
1991  *  RAID0/RAID10).
1992  *
1993  * Since we may need to handle a subset of block group, we need @logical_start
1994  * and @logical_length parameter.
1995  */
1996 static int scrub_simple_mirror(struct scrub_ctx *sctx,
1997                                struct btrfs_block_group *bg,
1998                                struct map_lookup *map,
1999                                u64 logical_start, u64 logical_length,
2000                                struct btrfs_device *device,
2001                                u64 physical, int mirror_num)
2002 {
2003         struct btrfs_fs_info *fs_info = sctx->fs_info;
2004         const u64 logical_end = logical_start + logical_length;
2005         u64 cur_logical = logical_start;
2006         int ret;
2007
2008         /* The range must be inside the bg */
2009         ASSERT(logical_start >= bg->start && logical_end <= bg->start + bg->length);
2010
2011         /* Go through each extent items inside the logical range */
2012         while (cur_logical < logical_end) {
2013                 u64 found_logical;
2014                 u64 cur_physical = physical + cur_logical - logical_start;
2015
2016                 /* Canceled? */
2017                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2018                     atomic_read(&sctx->cancel_req)) {
2019                         ret = -ECANCELED;
2020                         break;
2021                 }
2022                 /* Paused? */
2023                 if (atomic_read(&fs_info->scrub_pause_req)) {
2024                         /* Push queued extents */
2025                         scrub_blocked_if_needed(fs_info);
2026                 }
2027                 /* Block group removed? */
2028                 spin_lock(&bg->lock);
2029                 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags)) {
2030                         spin_unlock(&bg->lock);
2031                         ret = 0;
2032                         break;
2033                 }
2034                 spin_unlock(&bg->lock);
2035
2036                 ret = queue_scrub_stripe(sctx, bg, device, mirror_num,
2037                                          cur_logical, logical_end - cur_logical,
2038                                          cur_physical, &found_logical);
2039                 if (ret > 0) {
2040                         /* No more extent, just update the accounting */
2041                         sctx->stat.last_physical = physical + logical_length;
2042                         ret = 0;
2043                         break;
2044                 }
2045                 if (ret < 0)
2046                         break;
2047
2048                 cur_logical = found_logical + BTRFS_STRIPE_LEN;
2049
2050                 /* Don't hold CPU for too long time */
2051                 cond_resched();
2052         }
2053         return ret;
2054 }
2055
2056 /* Calculate the full stripe length for simple stripe based profiles */
2057 static u64 simple_stripe_full_stripe_len(const struct map_lookup *map)
2058 {
2059         ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2060                             BTRFS_BLOCK_GROUP_RAID10));
2061
2062         return btrfs_stripe_nr_to_offset(map->num_stripes / map->sub_stripes);
2063 }
2064
2065 /* Get the logical bytenr for the stripe */
2066 static u64 simple_stripe_get_logical(struct map_lookup *map,
2067                                      struct btrfs_block_group *bg,
2068                                      int stripe_index)
2069 {
2070         ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2071                             BTRFS_BLOCK_GROUP_RAID10));
2072         ASSERT(stripe_index < map->num_stripes);
2073
2074         /*
2075          * (stripe_index / sub_stripes) gives how many data stripes we need to
2076          * skip.
2077          */
2078         return btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes) +
2079                bg->start;
2080 }
2081
2082 /* Get the mirror number for the stripe */
2083 static int simple_stripe_mirror_num(struct map_lookup *map, int stripe_index)
2084 {
2085         ASSERT(map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2086                             BTRFS_BLOCK_GROUP_RAID10));
2087         ASSERT(stripe_index < map->num_stripes);
2088
2089         /* For RAID0, it's fixed to 1, for RAID10 it's 0,1,0,1... */
2090         return stripe_index % map->sub_stripes + 1;
2091 }
2092
2093 static int scrub_simple_stripe(struct scrub_ctx *sctx,
2094                                struct btrfs_block_group *bg,
2095                                struct map_lookup *map,
2096                                struct btrfs_device *device,
2097                                int stripe_index)
2098 {
2099         const u64 logical_increment = simple_stripe_full_stripe_len(map);
2100         const u64 orig_logical = simple_stripe_get_logical(map, bg, stripe_index);
2101         const u64 orig_physical = map->stripes[stripe_index].physical;
2102         const int mirror_num = simple_stripe_mirror_num(map, stripe_index);
2103         u64 cur_logical = orig_logical;
2104         u64 cur_physical = orig_physical;
2105         int ret = 0;
2106
2107         while (cur_logical < bg->start + bg->length) {
2108                 /*
2109                  * Inside each stripe, RAID0 is just SINGLE, and RAID10 is
2110                  * just RAID1, so we can reuse scrub_simple_mirror() to scrub
2111                  * this stripe.
2112                  */
2113                 ret = scrub_simple_mirror(sctx, bg, map, cur_logical,
2114                                           BTRFS_STRIPE_LEN, device, cur_physical,
2115                                           mirror_num);
2116                 if (ret)
2117                         return ret;
2118                 /* Skip to next stripe which belongs to the target device */
2119                 cur_logical += logical_increment;
2120                 /* For physical offset, we just go to next stripe */
2121                 cur_physical += BTRFS_STRIPE_LEN;
2122         }
2123         return ret;
2124 }
2125
2126 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2127                                            struct btrfs_block_group *bg,
2128                                            struct extent_map *em,
2129                                            struct btrfs_device *scrub_dev,
2130                                            int stripe_index)
2131 {
2132         struct btrfs_fs_info *fs_info = sctx->fs_info;
2133         struct map_lookup *map = em->map_lookup;
2134         const u64 profile = map->type & BTRFS_BLOCK_GROUP_PROFILE_MASK;
2135         const u64 chunk_logical = bg->start;
2136         int ret;
2137         int ret2;
2138         u64 physical = map->stripes[stripe_index].physical;
2139         const u64 dev_stripe_len = btrfs_calc_stripe_length(em);
2140         const u64 physical_end = physical + dev_stripe_len;
2141         u64 logical;
2142         u64 logic_end;
2143         /* The logical increment after finishing one stripe */
2144         u64 increment;
2145         /* Offset inside the chunk */
2146         u64 offset;
2147         u64 stripe_logical;
2148         int stop_loop = 0;
2149
2150         /* Extent_path should be released by now. */
2151         ASSERT(sctx->extent_path.nodes[0] == NULL);
2152
2153         scrub_blocked_if_needed(fs_info);
2154
2155         if (sctx->is_dev_replace &&
2156             btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
2157                 mutex_lock(&sctx->wr_lock);
2158                 sctx->write_pointer = physical;
2159                 mutex_unlock(&sctx->wr_lock);
2160         }
2161
2162         /* Prepare the extra data stripes used by RAID56. */
2163         if (profile & BTRFS_BLOCK_GROUP_RAID56_MASK) {
2164                 ASSERT(sctx->raid56_data_stripes == NULL);
2165
2166                 sctx->raid56_data_stripes = kcalloc(nr_data_stripes(map),
2167                                                     sizeof(struct scrub_stripe),
2168                                                     GFP_KERNEL);
2169                 if (!sctx->raid56_data_stripes) {
2170                         ret = -ENOMEM;
2171                         goto out;
2172                 }
2173                 for (int i = 0; i < nr_data_stripes(map); i++) {
2174                         ret = init_scrub_stripe(fs_info,
2175                                                 &sctx->raid56_data_stripes[i]);
2176                         if (ret < 0)
2177                                 goto out;
2178                         sctx->raid56_data_stripes[i].bg = bg;
2179                         sctx->raid56_data_stripes[i].sctx = sctx;
2180                 }
2181         }
2182         /*
2183          * There used to be a big double loop to handle all profiles using the
2184          * same routine, which grows larger and more gross over time.
2185          *
2186          * So here we handle each profile differently, so simpler profiles
2187          * have simpler scrubbing function.
2188          */
2189         if (!(profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10 |
2190                          BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2191                 /*
2192                  * Above check rules out all complex profile, the remaining
2193                  * profiles are SINGLE|DUP|RAID1|RAID1C*, which is simple
2194                  * mirrored duplication without stripe.
2195                  *
2196                  * Only @physical and @mirror_num needs to calculated using
2197                  * @stripe_index.
2198                  */
2199                 ret = scrub_simple_mirror(sctx, bg, map, bg->start, bg->length,
2200                                 scrub_dev, map->stripes[stripe_index].physical,
2201                                 stripe_index + 1);
2202                 offset = 0;
2203                 goto out;
2204         }
2205         if (profile & (BTRFS_BLOCK_GROUP_RAID0 | BTRFS_BLOCK_GROUP_RAID10)) {
2206                 ret = scrub_simple_stripe(sctx, bg, map, scrub_dev, stripe_index);
2207                 offset = btrfs_stripe_nr_to_offset(stripe_index / map->sub_stripes);
2208                 goto out;
2209         }
2210
2211         /* Only RAID56 goes through the old code */
2212         ASSERT(map->type & BTRFS_BLOCK_GROUP_RAID56_MASK);
2213         ret = 0;
2214
2215         /* Calculate the logical end of the stripe */
2216         get_raid56_logic_offset(physical_end, stripe_index,
2217                                 map, &logic_end, NULL);
2218         logic_end += chunk_logical;
2219
2220         /* Initialize @offset in case we need to go to out: label */
2221         get_raid56_logic_offset(physical, stripe_index, map, &offset, NULL);
2222         increment = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2223
2224         /*
2225          * Due to the rotation, for RAID56 it's better to iterate each stripe
2226          * using their physical offset.
2227          */
2228         while (physical < physical_end) {
2229                 ret = get_raid56_logic_offset(physical, stripe_index, map,
2230                                               &logical, &stripe_logical);
2231                 logical += chunk_logical;
2232                 if (ret) {
2233                         /* it is parity strip */
2234                         stripe_logical += chunk_logical;
2235                         ret = scrub_raid56_parity_stripe(sctx, scrub_dev, bg,
2236                                                          map, stripe_logical);
2237                         if (ret)
2238                                 goto out;
2239                         goto next;
2240                 }
2241
2242                 /*
2243                  * Now we're at a data stripe, scrub each extents in the range.
2244                  *
2245                  * At this stage, if we ignore the repair part, inside each data
2246                  * stripe it is no different than SINGLE profile.
2247                  * We can reuse scrub_simple_mirror() here, as the repair part
2248                  * is still based on @mirror_num.
2249                  */
2250                 ret = scrub_simple_mirror(sctx, bg, map, logical, BTRFS_STRIPE_LEN,
2251                                           scrub_dev, physical, 1);
2252                 if (ret < 0)
2253                         goto out;
2254 next:
2255                 logical += increment;
2256                 physical += BTRFS_STRIPE_LEN;
2257                 spin_lock(&sctx->stat_lock);
2258                 if (stop_loop)
2259                         sctx->stat.last_physical =
2260                                 map->stripes[stripe_index].physical + dev_stripe_len;
2261                 else
2262                         sctx->stat.last_physical = physical;
2263                 spin_unlock(&sctx->stat_lock);
2264                 if (stop_loop)
2265                         break;
2266         }
2267 out:
2268         ret2 = flush_scrub_stripes(sctx);
2269         if (!ret)
2270                 ret = ret2;
2271         btrfs_release_path(&sctx->extent_path);
2272         btrfs_release_path(&sctx->csum_path);
2273
2274         if (sctx->raid56_data_stripes) {
2275                 for (int i = 0; i < nr_data_stripes(map); i++)
2276                         release_scrub_stripe(&sctx->raid56_data_stripes[i]);
2277                 kfree(sctx->raid56_data_stripes);
2278                 sctx->raid56_data_stripes = NULL;
2279         }
2280
2281         if (sctx->is_dev_replace && ret >= 0) {
2282                 int ret2;
2283
2284                 ret2 = sync_write_pointer_for_zoned(sctx,
2285                                 chunk_logical + offset,
2286                                 map->stripes[stripe_index].physical,
2287                                 physical_end);
2288                 if (ret2)
2289                         ret = ret2;
2290         }
2291
2292         return ret < 0 ? ret : 0;
2293 }
2294
2295 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2296                                           struct btrfs_block_group *bg,
2297                                           struct btrfs_device *scrub_dev,
2298                                           u64 dev_offset,
2299                                           u64 dev_extent_len)
2300 {
2301         struct btrfs_fs_info *fs_info = sctx->fs_info;
2302         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2303         struct map_lookup *map;
2304         struct extent_map *em;
2305         int i;
2306         int ret = 0;
2307
2308         read_lock(&map_tree->lock);
2309         em = lookup_extent_mapping(map_tree, bg->start, bg->length);
2310         read_unlock(&map_tree->lock);
2311
2312         if (!em) {
2313                 /*
2314                  * Might have been an unused block group deleted by the cleaner
2315                  * kthread or relocation.
2316                  */
2317                 spin_lock(&bg->lock);
2318                 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &bg->runtime_flags))
2319                         ret = -EINVAL;
2320                 spin_unlock(&bg->lock);
2321
2322                 return ret;
2323         }
2324         if (em->start != bg->start)
2325                 goto out;
2326         if (em->len < dev_extent_len)
2327                 goto out;
2328
2329         map = em->map_lookup;
2330         for (i = 0; i < map->num_stripes; ++i) {
2331                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2332                     map->stripes[i].physical == dev_offset) {
2333                         ret = scrub_stripe(sctx, bg, em, scrub_dev, i);
2334                         if (ret)
2335                                 goto out;
2336                 }
2337         }
2338 out:
2339         free_extent_map(em);
2340
2341         return ret;
2342 }
2343
2344 static int finish_extent_writes_for_zoned(struct btrfs_root *root,
2345                                           struct btrfs_block_group *cache)
2346 {
2347         struct btrfs_fs_info *fs_info = cache->fs_info;
2348         struct btrfs_trans_handle *trans;
2349
2350         if (!btrfs_is_zoned(fs_info))
2351                 return 0;
2352
2353         btrfs_wait_block_group_reservations(cache);
2354         btrfs_wait_nocow_writers(cache);
2355         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
2356
2357         trans = btrfs_join_transaction(root);
2358         if (IS_ERR(trans))
2359                 return PTR_ERR(trans);
2360         return btrfs_commit_transaction(trans);
2361 }
2362
2363 static noinline_for_stack
2364 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2365                            struct btrfs_device *scrub_dev, u64 start, u64 end)
2366 {
2367         struct btrfs_dev_extent *dev_extent = NULL;
2368         struct btrfs_path *path;
2369         struct btrfs_fs_info *fs_info = sctx->fs_info;
2370         struct btrfs_root *root = fs_info->dev_root;
2371         u64 chunk_offset;
2372         int ret = 0;
2373         int ro_set;
2374         int slot;
2375         struct extent_buffer *l;
2376         struct btrfs_key key;
2377         struct btrfs_key found_key;
2378         struct btrfs_block_group *cache;
2379         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2380
2381         path = btrfs_alloc_path();
2382         if (!path)
2383                 return -ENOMEM;
2384
2385         path->reada = READA_FORWARD;
2386         path->search_commit_root = 1;
2387         path->skip_locking = 1;
2388
2389         key.objectid = scrub_dev->devid;
2390         key.offset = 0ull;
2391         key.type = BTRFS_DEV_EXTENT_KEY;
2392
2393         while (1) {
2394                 u64 dev_extent_len;
2395
2396                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2397                 if (ret < 0)
2398                         break;
2399                 if (ret > 0) {
2400                         if (path->slots[0] >=
2401                             btrfs_header_nritems(path->nodes[0])) {
2402                                 ret = btrfs_next_leaf(root, path);
2403                                 if (ret < 0)
2404                                         break;
2405                                 if (ret > 0) {
2406                                         ret = 0;
2407                                         break;
2408                                 }
2409                         } else {
2410                                 ret = 0;
2411                         }
2412                 }
2413
2414                 l = path->nodes[0];
2415                 slot = path->slots[0];
2416
2417                 btrfs_item_key_to_cpu(l, &found_key, slot);
2418
2419                 if (found_key.objectid != scrub_dev->devid)
2420                         break;
2421
2422                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2423                         break;
2424
2425                 if (found_key.offset >= end)
2426                         break;
2427
2428                 if (found_key.offset < key.offset)
2429                         break;
2430
2431                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2432                 dev_extent_len = btrfs_dev_extent_length(l, dev_extent);
2433
2434                 if (found_key.offset + dev_extent_len <= start)
2435                         goto skip;
2436
2437                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2438
2439                 /*
2440                  * get a reference on the corresponding block group to prevent
2441                  * the chunk from going away while we scrub it
2442                  */
2443                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2444
2445                 /* some chunks are removed but not committed to disk yet,
2446                  * continue scrubbing */
2447                 if (!cache)
2448                         goto skip;
2449
2450                 ASSERT(cache->start <= chunk_offset);
2451                 /*
2452                  * We are using the commit root to search for device extents, so
2453                  * that means we could have found a device extent item from a
2454                  * block group that was deleted in the current transaction. The
2455                  * logical start offset of the deleted block group, stored at
2456                  * @chunk_offset, might be part of the logical address range of
2457                  * a new block group (which uses different physical extents).
2458                  * In this case btrfs_lookup_block_group() has returned the new
2459                  * block group, and its start address is less than @chunk_offset.
2460                  *
2461                  * We skip such new block groups, because it's pointless to
2462                  * process them, as we won't find their extents because we search
2463                  * for them using the commit root of the extent tree. For a device
2464                  * replace it's also fine to skip it, we won't miss copying them
2465                  * to the target device because we have the write duplication
2466                  * setup through the regular write path (by btrfs_map_block()),
2467                  * and we have committed a transaction when we started the device
2468                  * replace, right after setting up the device replace state.
2469                  */
2470                 if (cache->start < chunk_offset) {
2471                         btrfs_put_block_group(cache);
2472                         goto skip;
2473                 }
2474
2475                 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
2476                         if (!test_bit(BLOCK_GROUP_FLAG_TO_COPY, &cache->runtime_flags)) {
2477                                 btrfs_put_block_group(cache);
2478                                 goto skip;
2479                         }
2480                 }
2481
2482                 /*
2483                  * Make sure that while we are scrubbing the corresponding block
2484                  * group doesn't get its logical address and its device extents
2485                  * reused for another block group, which can possibly be of a
2486                  * different type and different profile. We do this to prevent
2487                  * false error detections and crashes due to bogus attempts to
2488                  * repair extents.
2489                  */
2490                 spin_lock(&cache->lock);
2491                 if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags)) {
2492                         spin_unlock(&cache->lock);
2493                         btrfs_put_block_group(cache);
2494                         goto skip;
2495                 }
2496                 btrfs_freeze_block_group(cache);
2497                 spin_unlock(&cache->lock);
2498
2499                 /*
2500                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
2501                  * to avoid deadlock caused by:
2502                  * btrfs_inc_block_group_ro()
2503                  * -> btrfs_wait_for_commit()
2504                  * -> btrfs_commit_transaction()
2505                  * -> btrfs_scrub_pause()
2506                  */
2507                 scrub_pause_on(fs_info);
2508
2509                 /*
2510                  * Don't do chunk preallocation for scrub.
2511                  *
2512                  * This is especially important for SYSTEM bgs, or we can hit
2513                  * -EFBIG from btrfs_finish_chunk_alloc() like:
2514                  * 1. The only SYSTEM bg is marked RO.
2515                  *    Since SYSTEM bg is small, that's pretty common.
2516                  * 2. New SYSTEM bg will be allocated
2517                  *    Due to regular version will allocate new chunk.
2518                  * 3. New SYSTEM bg is empty and will get cleaned up
2519                  *    Before cleanup really happens, it's marked RO again.
2520                  * 4. Empty SYSTEM bg get scrubbed
2521                  *    We go back to 2.
2522                  *
2523                  * This can easily boost the amount of SYSTEM chunks if cleaner
2524                  * thread can't be triggered fast enough, and use up all space
2525                  * of btrfs_super_block::sys_chunk_array
2526                  *
2527                  * While for dev replace, we need to try our best to mark block
2528                  * group RO, to prevent race between:
2529                  * - Write duplication
2530                  *   Contains latest data
2531                  * - Scrub copy
2532                  *   Contains data from commit tree
2533                  *
2534                  * If target block group is not marked RO, nocow writes can
2535                  * be overwritten by scrub copy, causing data corruption.
2536                  * So for dev-replace, it's not allowed to continue if a block
2537                  * group is not RO.
2538                  */
2539                 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
2540                 if (!ret && sctx->is_dev_replace) {
2541                         ret = finish_extent_writes_for_zoned(root, cache);
2542                         if (ret) {
2543                                 btrfs_dec_block_group_ro(cache);
2544                                 scrub_pause_off(fs_info);
2545                                 btrfs_put_block_group(cache);
2546                                 break;
2547                         }
2548                 }
2549
2550                 if (ret == 0) {
2551                         ro_set = 1;
2552                 } else if (ret == -ENOSPC && !sctx->is_dev_replace &&
2553                            !(cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK)) {
2554                         /*
2555                          * btrfs_inc_block_group_ro return -ENOSPC when it
2556                          * failed in creating new chunk for metadata.
2557                          * It is not a problem for scrub, because
2558                          * metadata are always cowed, and our scrub paused
2559                          * commit_transactions.
2560                          *
2561                          * For RAID56 chunks, we have to mark them read-only
2562                          * for scrub, as later we would use our own cache
2563                          * out of RAID56 realm.
2564                          * Thus we want the RAID56 bg to be marked RO to
2565                          * prevent RMW from screwing up out cache.
2566                          */
2567                         ro_set = 0;
2568                 } else if (ret == -ETXTBSY) {
2569                         btrfs_warn(fs_info,
2570                    "skipping scrub of block group %llu due to active swapfile",
2571                                    cache->start);
2572                         scrub_pause_off(fs_info);
2573                         ret = 0;
2574                         goto skip_unfreeze;
2575                 } else {
2576                         btrfs_warn(fs_info,
2577                                    "failed setting block group ro: %d", ret);
2578                         btrfs_unfreeze_block_group(cache);
2579                         btrfs_put_block_group(cache);
2580                         scrub_pause_off(fs_info);
2581                         break;
2582                 }
2583
2584                 /*
2585                  * Now the target block is marked RO, wait for nocow writes to
2586                  * finish before dev-replace.
2587                  * COW is fine, as COW never overwrites extents in commit tree.
2588                  */
2589                 if (sctx->is_dev_replace) {
2590                         btrfs_wait_nocow_writers(cache);
2591                         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
2592                                         cache->length);
2593                 }
2594
2595                 scrub_pause_off(fs_info);
2596                 down_write(&dev_replace->rwsem);
2597                 dev_replace->cursor_right = found_key.offset + dev_extent_len;
2598                 dev_replace->cursor_left = found_key.offset;
2599                 dev_replace->item_needs_writeback = 1;
2600                 up_write(&dev_replace->rwsem);
2601
2602                 ret = scrub_chunk(sctx, cache, scrub_dev, found_key.offset,
2603                                   dev_extent_len);
2604                 if (sctx->is_dev_replace &&
2605                     !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
2606                                                       cache, found_key.offset))
2607                         ro_set = 0;
2608
2609                 down_write(&dev_replace->rwsem);
2610                 dev_replace->cursor_left = dev_replace->cursor_right;
2611                 dev_replace->item_needs_writeback = 1;
2612                 up_write(&dev_replace->rwsem);
2613
2614                 if (ro_set)
2615                         btrfs_dec_block_group_ro(cache);
2616
2617                 /*
2618                  * We might have prevented the cleaner kthread from deleting
2619                  * this block group if it was already unused because we raced
2620                  * and set it to RO mode first. So add it back to the unused
2621                  * list, otherwise it might not ever be deleted unless a manual
2622                  * balance is triggered or it becomes used and unused again.
2623                  */
2624                 spin_lock(&cache->lock);
2625                 if (!test_bit(BLOCK_GROUP_FLAG_REMOVED, &cache->runtime_flags) &&
2626                     !cache->ro && cache->reserved == 0 && cache->used == 0) {
2627                         spin_unlock(&cache->lock);
2628                         if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
2629                                 btrfs_discard_queue_work(&fs_info->discard_ctl,
2630                                                          cache);
2631                         else
2632                                 btrfs_mark_bg_unused(cache);
2633                 } else {
2634                         spin_unlock(&cache->lock);
2635                 }
2636 skip_unfreeze:
2637                 btrfs_unfreeze_block_group(cache);
2638                 btrfs_put_block_group(cache);
2639                 if (ret)
2640                         break;
2641                 if (sctx->is_dev_replace &&
2642                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2643                         ret = -EIO;
2644                         break;
2645                 }
2646                 if (sctx->stat.malloc_errors > 0) {
2647                         ret = -ENOMEM;
2648                         break;
2649                 }
2650 skip:
2651                 key.offset = found_key.offset + dev_extent_len;
2652                 btrfs_release_path(path);
2653         }
2654
2655         btrfs_free_path(path);
2656
2657         return ret;
2658 }
2659
2660 static int scrub_one_super(struct scrub_ctx *sctx, struct btrfs_device *dev,
2661                            struct page *page, u64 physical, u64 generation)
2662 {
2663         struct btrfs_fs_info *fs_info = sctx->fs_info;
2664         struct bio_vec bvec;
2665         struct bio bio;
2666         struct btrfs_super_block *sb = page_address(page);
2667         int ret;
2668
2669         bio_init(&bio, dev->bdev, &bvec, 1, REQ_OP_READ);
2670         bio.bi_iter.bi_sector = physical >> SECTOR_SHIFT;
2671         __bio_add_page(&bio, page, BTRFS_SUPER_INFO_SIZE, 0);
2672         ret = submit_bio_wait(&bio);
2673         bio_uninit(&bio);
2674
2675         if (ret < 0)
2676                 return ret;
2677         ret = btrfs_check_super_csum(fs_info, sb);
2678         if (ret != 0) {
2679                 btrfs_err_rl(fs_info,
2680                         "super block at physical %llu devid %llu has bad csum",
2681                         physical, dev->devid);
2682                 return -EIO;
2683         }
2684         if (btrfs_super_generation(sb) != generation) {
2685                 btrfs_err_rl(fs_info,
2686 "super block at physical %llu devid %llu has bad generation %llu expect %llu",
2687                              physical, dev->devid,
2688                              btrfs_super_generation(sb), generation);
2689                 return -EUCLEAN;
2690         }
2691
2692         return btrfs_validate_super(fs_info, sb, -1);
2693 }
2694
2695 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2696                                            struct btrfs_device *scrub_dev)
2697 {
2698         int     i;
2699         u64     bytenr;
2700         u64     gen;
2701         int ret = 0;
2702         struct page *page;
2703         struct btrfs_fs_info *fs_info = sctx->fs_info;
2704
2705         if (BTRFS_FS_ERROR(fs_info))
2706                 return -EROFS;
2707
2708         page = alloc_page(GFP_KERNEL);
2709         if (!page) {
2710                 spin_lock(&sctx->stat_lock);
2711                 sctx->stat.malloc_errors++;
2712                 spin_unlock(&sctx->stat_lock);
2713                 return -ENOMEM;
2714         }
2715
2716         /* Seed devices of a new filesystem has their own generation. */
2717         if (scrub_dev->fs_devices != fs_info->fs_devices)
2718                 gen = scrub_dev->generation;
2719         else
2720                 gen = fs_info->last_trans_committed;
2721
2722         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2723                 bytenr = btrfs_sb_offset(i);
2724                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2725                     scrub_dev->commit_total_bytes)
2726                         break;
2727                 if (!btrfs_check_super_location(scrub_dev, bytenr))
2728                         continue;
2729
2730                 ret = scrub_one_super(sctx, scrub_dev, page, bytenr, gen);
2731                 if (ret) {
2732                         spin_lock(&sctx->stat_lock);
2733                         sctx->stat.super_errors++;
2734                         spin_unlock(&sctx->stat_lock);
2735                 }
2736         }
2737         __free_page(page);
2738         return 0;
2739 }
2740
2741 static void scrub_workers_put(struct btrfs_fs_info *fs_info)
2742 {
2743         if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
2744                                         &fs_info->scrub_lock)) {
2745                 struct workqueue_struct *scrub_workers = fs_info->scrub_workers;
2746
2747                 fs_info->scrub_workers = NULL;
2748                 mutex_unlock(&fs_info->scrub_lock);
2749
2750                 if (scrub_workers)
2751                         destroy_workqueue(scrub_workers);
2752         }
2753 }
2754
2755 /*
2756  * get a reference count on fs_info->scrub_workers. start worker if necessary
2757  */
2758 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info)
2759 {
2760         struct workqueue_struct *scrub_workers = NULL;
2761         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
2762         int max_active = fs_info->thread_pool_size;
2763         int ret = -ENOMEM;
2764
2765         if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
2766                 return 0;
2767
2768         scrub_workers = alloc_workqueue("btrfs-scrub", flags, max_active);
2769         if (!scrub_workers)
2770                 return -ENOMEM;
2771
2772         mutex_lock(&fs_info->scrub_lock);
2773         if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
2774                 ASSERT(fs_info->scrub_workers == NULL);
2775                 fs_info->scrub_workers = scrub_workers;
2776                 refcount_set(&fs_info->scrub_workers_refcnt, 1);
2777                 mutex_unlock(&fs_info->scrub_lock);
2778                 return 0;
2779         }
2780         /* Other thread raced in and created the workers for us */
2781         refcount_inc(&fs_info->scrub_workers_refcnt);
2782         mutex_unlock(&fs_info->scrub_lock);
2783
2784         ret = 0;
2785
2786         destroy_workqueue(scrub_workers);
2787         return ret;
2788 }
2789
2790 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2791                     u64 end, struct btrfs_scrub_progress *progress,
2792                     int readonly, int is_dev_replace)
2793 {
2794         struct btrfs_dev_lookup_args args = { .devid = devid };
2795         struct scrub_ctx *sctx;
2796         int ret;
2797         struct btrfs_device *dev;
2798         unsigned int nofs_flag;
2799         bool need_commit = false;
2800
2801         if (btrfs_fs_closing(fs_info))
2802                 return -EAGAIN;
2803
2804         /* At mount time we have ensured nodesize is in the range of [4K, 64K]. */
2805         ASSERT(fs_info->nodesize <= BTRFS_STRIPE_LEN);
2806
2807         /*
2808          * SCRUB_MAX_SECTORS_PER_BLOCK is calculated using the largest possible
2809          * value (max nodesize / min sectorsize), thus nodesize should always
2810          * be fine.
2811          */
2812         ASSERT(fs_info->nodesize <=
2813                SCRUB_MAX_SECTORS_PER_BLOCK << fs_info->sectorsize_bits);
2814
2815         /* Allocate outside of device_list_mutex */
2816         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
2817         if (IS_ERR(sctx))
2818                 return PTR_ERR(sctx);
2819
2820         ret = scrub_workers_get(fs_info);
2821         if (ret)
2822                 goto out_free_ctx;
2823
2824         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2825         dev = btrfs_find_device(fs_info->fs_devices, &args);
2826         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
2827                      !is_dev_replace)) {
2828                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2829                 ret = -ENODEV;
2830                 goto out;
2831         }
2832
2833         if (!is_dev_replace && !readonly &&
2834             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
2835                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2836                 btrfs_err_in_rcu(fs_info,
2837                         "scrub on devid %llu: filesystem on %s is not writable",
2838                                  devid, btrfs_dev_name(dev));
2839                 ret = -EROFS;
2840                 goto out;
2841         }
2842
2843         mutex_lock(&fs_info->scrub_lock);
2844         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
2845             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
2846                 mutex_unlock(&fs_info->scrub_lock);
2847                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2848                 ret = -EIO;
2849                 goto out;
2850         }
2851
2852         down_read(&fs_info->dev_replace.rwsem);
2853         if (dev->scrub_ctx ||
2854             (!is_dev_replace &&
2855              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2856                 up_read(&fs_info->dev_replace.rwsem);
2857                 mutex_unlock(&fs_info->scrub_lock);
2858                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2859                 ret = -EINPROGRESS;
2860                 goto out;
2861         }
2862         up_read(&fs_info->dev_replace.rwsem);
2863
2864         sctx->readonly = readonly;
2865         dev->scrub_ctx = sctx;
2866         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2867
2868         /*
2869          * checking @scrub_pause_req here, we can avoid
2870          * race between committing transaction and scrubbing.
2871          */
2872         __scrub_blocked_if_needed(fs_info);
2873         atomic_inc(&fs_info->scrubs_running);
2874         mutex_unlock(&fs_info->scrub_lock);
2875
2876         /*
2877          * In order to avoid deadlock with reclaim when there is a transaction
2878          * trying to pause scrub, make sure we use GFP_NOFS for all the
2879          * allocations done at btrfs_scrub_sectors() and scrub_sectors_for_parity()
2880          * invoked by our callees. The pausing request is done when the
2881          * transaction commit starts, and it blocks the transaction until scrub
2882          * is paused (done at specific points at scrub_stripe() or right above
2883          * before incrementing fs_info->scrubs_running).
2884          */
2885         nofs_flag = memalloc_nofs_save();
2886         if (!is_dev_replace) {
2887                 u64 old_super_errors;
2888
2889                 spin_lock(&sctx->stat_lock);
2890                 old_super_errors = sctx->stat.super_errors;
2891                 spin_unlock(&sctx->stat_lock);
2892
2893                 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
2894                 /*
2895                  * by holding device list mutex, we can
2896                  * kick off writing super in log tree sync.
2897                  */
2898                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2899                 ret = scrub_supers(sctx, dev);
2900                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2901
2902                 spin_lock(&sctx->stat_lock);
2903                 /*
2904                  * Super block errors found, but we can not commit transaction
2905                  * at current context, since btrfs_commit_transaction() needs
2906                  * to pause the current running scrub (hold by ourselves).
2907                  */
2908                 if (sctx->stat.super_errors > old_super_errors && !sctx->readonly)
2909                         need_commit = true;
2910                 spin_unlock(&sctx->stat_lock);
2911         }
2912
2913         if (!ret)
2914                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
2915         memalloc_nofs_restore(nofs_flag);
2916
2917         atomic_dec(&fs_info->scrubs_running);
2918         wake_up(&fs_info->scrub_pause_wait);
2919
2920         if (progress)
2921                 memcpy(progress, &sctx->stat, sizeof(*progress));
2922
2923         if (!is_dev_replace)
2924                 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
2925                         ret ? "not finished" : "finished", devid, ret);
2926
2927         mutex_lock(&fs_info->scrub_lock);
2928         dev->scrub_ctx = NULL;
2929         mutex_unlock(&fs_info->scrub_lock);
2930
2931         scrub_workers_put(fs_info);
2932         scrub_put_ctx(sctx);
2933
2934         /*
2935          * We found some super block errors before, now try to force a
2936          * transaction commit, as scrub has finished.
2937          */
2938         if (need_commit) {
2939                 struct btrfs_trans_handle *trans;
2940
2941                 trans = btrfs_start_transaction(fs_info->tree_root, 0);
2942                 if (IS_ERR(trans)) {
2943                         ret = PTR_ERR(trans);
2944                         btrfs_err(fs_info,
2945         "scrub: failed to start transaction to fix super block errors: %d", ret);
2946                         return ret;
2947                 }
2948                 ret = btrfs_commit_transaction(trans);
2949                 if (ret < 0)
2950                         btrfs_err(fs_info,
2951         "scrub: failed to commit transaction to fix super block errors: %d", ret);
2952         }
2953         return ret;
2954 out:
2955         scrub_workers_put(fs_info);
2956 out_free_ctx:
2957         scrub_free_ctx(sctx);
2958
2959         return ret;
2960 }
2961
2962 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
2963 {
2964         mutex_lock(&fs_info->scrub_lock);
2965         atomic_inc(&fs_info->scrub_pause_req);
2966         while (atomic_read(&fs_info->scrubs_paused) !=
2967                atomic_read(&fs_info->scrubs_running)) {
2968                 mutex_unlock(&fs_info->scrub_lock);
2969                 wait_event(fs_info->scrub_pause_wait,
2970                            atomic_read(&fs_info->scrubs_paused) ==
2971                            atomic_read(&fs_info->scrubs_running));
2972                 mutex_lock(&fs_info->scrub_lock);
2973         }
2974         mutex_unlock(&fs_info->scrub_lock);
2975 }
2976
2977 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
2978 {
2979         atomic_dec(&fs_info->scrub_pause_req);
2980         wake_up(&fs_info->scrub_pause_wait);
2981 }
2982
2983 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2984 {
2985         mutex_lock(&fs_info->scrub_lock);
2986         if (!atomic_read(&fs_info->scrubs_running)) {
2987                 mutex_unlock(&fs_info->scrub_lock);
2988                 return -ENOTCONN;
2989         }
2990
2991         atomic_inc(&fs_info->scrub_cancel_req);
2992         while (atomic_read(&fs_info->scrubs_running)) {
2993                 mutex_unlock(&fs_info->scrub_lock);
2994                 wait_event(fs_info->scrub_pause_wait,
2995                            atomic_read(&fs_info->scrubs_running) == 0);
2996                 mutex_lock(&fs_info->scrub_lock);
2997         }
2998         atomic_dec(&fs_info->scrub_cancel_req);
2999         mutex_unlock(&fs_info->scrub_lock);
3000
3001         return 0;
3002 }
3003
3004 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
3005 {
3006         struct btrfs_fs_info *fs_info = dev->fs_info;
3007         struct scrub_ctx *sctx;
3008
3009         mutex_lock(&fs_info->scrub_lock);
3010         sctx = dev->scrub_ctx;
3011         if (!sctx) {
3012                 mutex_unlock(&fs_info->scrub_lock);
3013                 return -ENOTCONN;
3014         }
3015         atomic_inc(&sctx->cancel_req);
3016         while (dev->scrub_ctx) {
3017                 mutex_unlock(&fs_info->scrub_lock);
3018                 wait_event(fs_info->scrub_pause_wait,
3019                            dev->scrub_ctx == NULL);
3020                 mutex_lock(&fs_info->scrub_lock);
3021         }
3022         mutex_unlock(&fs_info->scrub_lock);
3023
3024         return 0;
3025 }
3026
3027 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
3028                          struct btrfs_scrub_progress *progress)
3029 {
3030         struct btrfs_dev_lookup_args args = { .devid = devid };
3031         struct btrfs_device *dev;
3032         struct scrub_ctx *sctx = NULL;
3033
3034         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3035         dev = btrfs_find_device(fs_info->fs_devices, &args);
3036         if (dev)
3037                 sctx = dev->scrub_ctx;
3038         if (sctx)
3039                 memcpy(progress, &sctx->stat, sizeof(*progress));
3040         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3041
3042         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3043 }