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 "rcu-string.h"
  21 #include "raid56.h"
  22 #include "block-group.h"
  23 #include "zoned.h"
  24
  25 /*
  26  * This is only the first step towards a full-features scrub. It reads all
  27  * extent and super block and verifies the checksums. In case a bad checksum
  28  * is found or the extent cannot be read, good data will be written back if
  29  * any can be found.
  30  *
  31  * Future enhancements:
  32  *  - In case an unrepairable extent is encountered, track which files are
  33  *    affected and report them
  34  *  - track and record media errors, throw out bad devices
  35  *  - add a mode to also read unallocated space
  36  */
  37
  38 struct scrub_block;
  39 struct scrub_ctx;
  40
  41 /*
  42  * the following three values only influence the performance.
  43  * The last one configures the number of parallel and outstanding I/O
  44  * operations. The first two values configure an upper limit for the number
  45  * of (dynamically allocated) pages that are added to a bio.
  46  */
  47 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  48 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  49 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  50
  51 /*
  52  * the following value times PAGE_SIZE needs to be large enough to match the
  53  * largest node/leaf/sector size that shall be supported.
  54  * Values larger than BTRFS_STRIPE_LEN are not supported.
  55  */
  56 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  57
  58 struct scrub_recover {
  59         refcount_t              refs;
  60         struct btrfs_bio        *bbio;
  61         u64                     map_length;
  62 };
  63
  64 struct scrub_page {
  65         struct scrub_block      *sblock;
  66         struct page             *page;
  67         struct btrfs_device     *dev;
  68         struct list_head        list;
  69         u64                     flags;  /* extent flags */
  70         u64                     generation;
  71         u64                     logical;
  72         u64                     physical;
  73         u64                     physical_for_dev_replace;
  74         atomic_t                refs;
  75         u8                      mirror_num;
  76         int                     have_csum:1;
  77         int                     io_error:1;
  78         u8                      csum[BTRFS_CSUM_SIZE];
  79
  80         struct scrub_recover    *recover;
  81 };
  82
  83 struct scrub_bio {
  84         int                     index;
  85         struct scrub_ctx        *sctx;
  86         struct btrfs_device     *dev;
  87         struct bio              *bio;
  88         blk_status_t            status;
  89         u64                     logical;
  90         u64                     physical;
  91 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  92         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  93 #else
  94         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  95 #endif
  96         int                     page_count;
  97         int                     next_free;
  98         struct btrfs_work       work;
  99 };
 100
 101 struct scrub_block {
 102         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 103         int                     page_count;
 104         atomic_t                outstanding_pages;
 105         refcount_t              refs; /* free mem on transition to zero */
 106         struct scrub_ctx        *sctx;
 107         struct scrub_parity     *sparity;
 108         struct {
 109                 unsigned int    header_error:1;
 110                 unsigned int    checksum_error:1;
 111                 unsigned int    no_io_error_seen:1;
 112                 unsigned int    generation_error:1; /* also sets header_error */
 113
 114                 /* The following is for the data used to check parity */
 115                 /* It is for the data with checksum */
 116                 unsigned int    data_corrected:1;
 117         };
 118         struct btrfs_work       work;
 119 };
 120
 121 /* Used for the chunks with parity stripe such RAID5/6 */
 122 struct scrub_parity {
 123         struct scrub_ctx        *sctx;
 124
 125         struct btrfs_device     *scrub_dev;
 126
 127         u64                     logic_start;
 128
 129         u64                     logic_end;
 130
 131         int                     nsectors;
 132
 133         u32                     stripe_len;
 134
 135         refcount_t              refs;
 136
 137         struct list_head        spages;
 138
 139         /* Work of parity check and repair */
 140         struct btrfs_work       work;
 141
 142         /* Mark the parity blocks which have data */
 143         unsigned long           *dbitmap;
 144
 145         /*
 146          * Mark the parity blocks which have data, but errors happen when
 147          * read data or check data
 148          */
 149         unsigned long           *ebitmap;
 150
 151         unsigned long           bitmap[];
 152 };
 153
 154 struct scrub_ctx {
 155         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 156         struct btrfs_fs_info    *fs_info;
 157         int                     first_free;
 158         int                     curr;
 159         atomic_t                bios_in_flight;
 160         atomic_t                workers_pending;
 161         spinlock_t              list_lock;
 162         wait_queue_head_t       list_wait;
 163         struct list_head        csum_list;
 164         atomic_t                cancel_req;
 165         int                     readonly;
 166         int                     pages_per_rd_bio;
 167
 168         int                     is_dev_replace;
 169         u64                     write_pointer;
 170
 171         struct scrub_bio        *wr_curr_bio;
 172         struct mutex            wr_lock;
 173         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 174         struct btrfs_device     *wr_tgtdev;
 175         bool                    flush_all_writes;
 176
 177         /*
 178          * statistics
 179          */
 180         struct btrfs_scrub_progress stat;
 181         spinlock_t              stat_lock;
 182
 183         /*
 184          * Use a ref counter to avoid use-after-free issues. Scrub workers
 185          * decrement bios_in_flight and workers_pending and then do a wakeup
 186          * on the list_wait wait queue. We must ensure the main scrub task
 187          * doesn't free the scrub context before or while the workers are
 188          * doing the wakeup() call.
 189          */
 190         refcount_t              refs;
 191 };
 192
 193 struct scrub_warning {
 194         struct btrfs_path       *path;
 195         u64                     extent_item_size;
 196         const char              *errstr;
 197         u64                     physical;
 198         u64                     logical;
 199         struct btrfs_device     *dev;
 200 };
 201
 202 struct full_stripe_lock {
 203         struct rb_node node;
 204         u64 logical;
 205         u64 refs;
 206         struct mutex mutex;
 207 };
 208
 209 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 210                                      struct scrub_block *sblocks_for_recheck);
 211 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 212                                 struct scrub_block *sblock,
 213                                 int retry_failed_mirror);
 214 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 215 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 216                                              struct scrub_block *sblock_good);
 217 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 218                                             struct scrub_block *sblock_good,
 219                                             int page_num, int force_write);
 220 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 221 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 222                                            int page_num);
 223 static int scrub_checksum_data(struct scrub_block *sblock);
 224 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 225 static int scrub_checksum_super(struct scrub_block *sblock);
 226 static void scrub_block_put(struct scrub_block *sblock);
 227 static void scrub_page_get(struct scrub_page *spage);
 228 static void scrub_page_put(struct scrub_page *spage);
 229 static void scrub_parity_get(struct scrub_parity *sparity);
 230 static void scrub_parity_put(struct scrub_parity *sparity);
 231 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u32 len,
 232                        u64 physical, struct btrfs_device *dev, u64 flags,
 233                        u64 gen, int mirror_num, u8 *csum,
 234                        u64 physical_for_dev_replace);
 235 static void scrub_bio_end_io(struct bio *bio);
 236 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 237 static void scrub_block_complete(struct scrub_block *sblock);
 238 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 239                                u64 extent_logical, u32 extent_len,
 240                                u64 *extent_physical,
 241                                struct btrfs_device **extent_dev,
 242                                int *extent_mirror_num);
 243 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 244                                     struct scrub_page *spage);
 245 static void scrub_wr_submit(struct scrub_ctx *sctx);
 246 static void scrub_wr_bio_end_io(struct bio *bio);
 247 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 248 static void scrub_put_ctx(struct scrub_ctx *sctx);
 249
 250 static inline int scrub_is_page_on_raid56(struct scrub_page *spage)
 251 {
 252         return spage->recover &&
 253                (spage->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 254 }
 255
 256 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 257 {
 258         refcount_inc(&sctx->refs);
 259         atomic_inc(&sctx->bios_in_flight);
 260 }
 261
 262 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 263 {
 264         atomic_dec(&sctx->bios_in_flight);
 265         wake_up(&sctx->list_wait);
 266         scrub_put_ctx(sctx);
 267 }
 268
 269 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 270 {
 271         while (atomic_read(&fs_info->scrub_pause_req)) {
 272                 mutex_unlock(&fs_info->scrub_lock);
 273                 wait_event(fs_info->scrub_pause_wait,
 274                    atomic_read(&fs_info->scrub_pause_req) == 0);
 275                 mutex_lock(&fs_info->scrub_lock);
 276         }
 277 }
 278
 279 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 280 {
 281         atomic_inc(&fs_info->scrubs_paused);
 282         wake_up(&fs_info->scrub_pause_wait);
 283 }
 284
 285 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 286 {
 287         mutex_lock(&fs_info->scrub_lock);
 288         __scrub_blocked_if_needed(fs_info);
 289         atomic_dec(&fs_info->scrubs_paused);
 290         mutex_unlock(&fs_info->scrub_lock);
 291
 292         wake_up(&fs_info->scrub_pause_wait);
 293 }
 294
 295 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 296 {
 297         scrub_pause_on(fs_info);
 298         scrub_pause_off(fs_info);
 299 }
 300
 301 /*
 302  * Insert new full stripe lock into full stripe locks tree
 303  *
 304  * Return pointer to existing or newly inserted full_stripe_lock structure if
 305  * everything works well.
 306  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 307  *
 308  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 309  * function
 310  */
 311 static struct full_stripe_lock *insert_full_stripe_lock(
 312                 struct btrfs_full_stripe_locks_tree *locks_root,
 313                 u64 fstripe_logical)
 314 {
 315         struct rb_node **p;
 316         struct rb_node *parent = NULL;
 317         struct full_stripe_lock *entry;
 318         struct full_stripe_lock *ret;
 319
 320         lockdep_assert_held(&locks_root->lock);
 321
 322         p = &locks_root->root.rb_node;
 323         while (*p) {
 324                 parent = *p;
 325                 entry = rb_entry(parent, struct full_stripe_lock, node);
 326                 if (fstripe_logical < entry->logical) {
 327                         p = &(*p)->rb_left;
 328                 } else if (fstripe_logical > entry->logical) {
 329                         p = &(*p)->rb_right;
 330                 } else {
 331                         entry->refs++;
 332                         return entry;
 333                 }
 334         }
 335
 336         /*
 337          * Insert new lock.
 338          */
 339         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 340         if (!ret)
 341                 return ERR_PTR(-ENOMEM);
 342         ret->logical = fstripe_logical;
 343         ret->refs = 1;
 344         mutex_init(&ret->mutex);
 345
 346         rb_link_node(&ret->node, parent, p);
 347         rb_insert_color(&ret->node, &locks_root->root);
 348         return ret;
 349 }
 350
 351 /*
 352  * Search for a full stripe lock of a block group
 353  *
 354  * Return pointer to existing full stripe lock if found
 355  * Return NULL if not found
 356  */
 357 static struct full_stripe_lock *search_full_stripe_lock(
 358                 struct btrfs_full_stripe_locks_tree *locks_root,
 359                 u64 fstripe_logical)
 360 {
 361         struct rb_node *node;
 362         struct full_stripe_lock *entry;
 363
 364         lockdep_assert_held(&locks_root->lock);
 365
 366         node = locks_root->root.rb_node;
 367         while (node) {
 368                 entry = rb_entry(node, struct full_stripe_lock, node);
 369                 if (fstripe_logical < entry->logical)
 370                         node = node->rb_left;
 371                 else if (fstripe_logical > entry->logical)
 372                         node = node->rb_right;
 373                 else
 374                         return entry;
 375         }
 376         return NULL;
 377 }
 378
 379 /*
 380  * Helper to get full stripe logical from a normal bytenr.
 381  *
 382  * Caller must ensure @cache is a RAID56 block group.
 383  */
 384 static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
 385 {
 386         u64 ret;
 387
 388         /*
 389          * Due to chunk item size limit, full stripe length should not be
 390          * larger than U32_MAX. Just a sanity check here.
 391          */
 392         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 393
 394         /*
 395          * round_down() can only handle power of 2, while RAID56 full
 396          * stripe length can be 64KiB * n, so we need to manually round down.
 397          */
 398         ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 399                         cache->full_stripe_len + cache->start;
 400         return ret;
 401 }
 402
 403 /*
 404  * Lock a full stripe to avoid concurrency of recovery and read
 405  *
 406  * It's only used for profiles with parities (RAID5/6), for other profiles it
 407  * does nothing.
 408  *
 409  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 410  * So caller must call unlock_full_stripe() at the same context.
 411  *
 412  * Return <0 if encounters error.
 413  */
 414 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 415                             bool *locked_ret)
 416 {
 417         struct btrfs_block_group *bg_cache;
 418         struct btrfs_full_stripe_locks_tree *locks_root;
 419         struct full_stripe_lock *existing;
 420         u64 fstripe_start;
 421         int ret = 0;
 422
 423         *locked_ret = false;
 424         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 425         if (!bg_cache) {
 426                 ASSERT(0);
 427                 return -ENOENT;
 428         }
 429
 430         /* Profiles not based on parity don't need full stripe lock */
 431         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 432                 goto out;
 433         locks_root = &bg_cache->full_stripe_locks_root;
 434
 435         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 436
 437         /* Now insert the full stripe lock */
 438         mutex_lock(&locks_root->lock);
 439         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 440         mutex_unlock(&locks_root->lock);
 441         if (IS_ERR(existing)) {
 442                 ret = PTR_ERR(existing);
 443                 goto out;
 444         }
 445         mutex_lock(&existing->mutex);
 446         *locked_ret = true;
 447 out:
 448         btrfs_put_block_group(bg_cache);
 449         return ret;
 450 }
 451
 452 /*
 453  * Unlock a full stripe.
 454  *
 455  * NOTE: Caller must ensure it's the same context calling corresponding
 456  * lock_full_stripe().
 457  *
 458  * Return 0 if we unlock full stripe without problem.
 459  * Return <0 for error
 460  */
 461 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 462                               bool locked)
 463 {
 464         struct btrfs_block_group *bg_cache;
 465         struct btrfs_full_stripe_locks_tree *locks_root;
 466         struct full_stripe_lock *fstripe_lock;
 467         u64 fstripe_start;
 468         bool freeit = false;
 469         int ret = 0;
 470
 471         /* If we didn't acquire full stripe lock, no need to continue */
 472         if (!locked)
 473                 return 0;
 474
 475         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 476         if (!bg_cache) {
 477                 ASSERT(0);
 478                 return -ENOENT;
 479         }
 480         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 481                 goto out;
 482
 483         locks_root = &bg_cache->full_stripe_locks_root;
 484         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 485
 486         mutex_lock(&locks_root->lock);
 487         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 488         /* Unpaired unlock_full_stripe() detected */
 489         if (!fstripe_lock) {
 490                 WARN_ON(1);
 491                 ret = -ENOENT;
 492                 mutex_unlock(&locks_root->lock);
 493                 goto out;
 494         }
 495
 496         if (fstripe_lock->refs == 0) {
 497                 WARN_ON(1);
 498                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 499                         fstripe_lock->logical);
 500         } else {
 501                 fstripe_lock->refs--;
 502         }
 503
 504         if (fstripe_lock->refs == 0) {
 505                 rb_erase(&fstripe_lock->node, &locks_root->root);
 506                 freeit = true;
 507         }
 508         mutex_unlock(&locks_root->lock);
 509
 510         mutex_unlock(&fstripe_lock->mutex);
 511         if (freeit)
 512                 kfree(fstripe_lock);
 513 out:
 514         btrfs_put_block_group(bg_cache);
 515         return ret;
 516 }
 517
 518 static void scrub_free_csums(struct scrub_ctx *sctx)
 519 {
 520         while (!list_empty(&sctx->csum_list)) {
 521                 struct btrfs_ordered_sum *sum;
 522                 sum = list_first_entry(&sctx->csum_list,
 523                                        struct btrfs_ordered_sum, list);
 524                 list_del(&sum->list);
 525                 kfree(sum);
 526         }
 527 }
 528
 529 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 530 {
 531         int i;
 532
 533         if (!sctx)
 534                 return;
 535
 536         /* this can happen when scrub is cancelled */
 537         if (sctx->curr != -1) {
 538                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 539
 540                 for (i = 0; i < sbio->page_count; i++) {
 541                         WARN_ON(!sbio->pagev[i]->page);
 542                         scrub_block_put(sbio->pagev[i]->sblock);
 543                 }
 544                 bio_put(sbio->bio);
 545         }
 546
 547         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 548                 struct scrub_bio *sbio = sctx->bios[i];
 549
 550                 if (!sbio)
 551                         break;
 552                 kfree(sbio);
 553         }
 554
 555         kfree(sctx->wr_curr_bio);
 556         scrub_free_csums(sctx);
 557         kfree(sctx);
 558 }
 559
 560 static void scrub_put_ctx(struct scrub_ctx *sctx)
 561 {
 562         if (refcount_dec_and_test(&sctx->refs))
 563                 scrub_free_ctx(sctx);
 564 }
 565
 566 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 567                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 568 {
 569         struct scrub_ctx *sctx;
 570         int             i;
 571
 572         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 573         if (!sctx)
 574                 goto nomem;
 575         refcount_set(&sctx->refs, 1);
 576         sctx->is_dev_replace = is_dev_replace;
 577         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 578         sctx->curr = -1;
 579         sctx->fs_info = fs_info;
 580         INIT_LIST_HEAD(&sctx->csum_list);
 581         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 582                 struct scrub_bio *sbio;
 583
 584                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 585                 if (!sbio)
 586                         goto nomem;
 587                 sctx->bios[i] = sbio;
 588
 589                 sbio->index = i;
 590                 sbio->sctx = sctx;
 591                 sbio->page_count = 0;
 592                 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
 593                                 NULL);
 594
 595                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 596                         sctx->bios[i]->next_free = i + 1;
 597                 else
 598                         sctx->bios[i]->next_free = -1;
 599         }
 600         sctx->first_free = 0;
 601         atomic_set(&sctx->bios_in_flight, 0);
 602         atomic_set(&sctx->workers_pending, 0);
 603         atomic_set(&sctx->cancel_req, 0);
 604
 605         spin_lock_init(&sctx->list_lock);
 606         spin_lock_init(&sctx->stat_lock);
 607         init_waitqueue_head(&sctx->list_wait);
 608
 609         WARN_ON(sctx->wr_curr_bio != NULL);
 610         mutex_init(&sctx->wr_lock);
 611         sctx->wr_curr_bio = NULL;
 612         if (is_dev_replace) {
 613                 WARN_ON(!fs_info->dev_replace.tgtdev);
 614                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 615                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 616                 sctx->flush_all_writes = false;
 617         }
 618
 619         return sctx;
 620
 621 nomem:
 622         scrub_free_ctx(sctx);
 623         return ERR_PTR(-ENOMEM);
 624 }
 625
 626 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 627                                      void *warn_ctx)
 628 {
 629         u64 isize;
 630         u32 nlink;
 631         int ret;
 632         int i;
 633         unsigned nofs_flag;
 634         struct extent_buffer *eb;
 635         struct btrfs_inode_item *inode_item;
 636         struct scrub_warning *swarn = warn_ctx;
 637         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 638         struct inode_fs_paths *ipath = NULL;
 639         struct btrfs_root *local_root;
 640         struct btrfs_key key;
 641
 642         local_root = btrfs_get_fs_root(fs_info, root, true);
 643         if (IS_ERR(local_root)) {
 644                 ret = PTR_ERR(local_root);
 645                 goto err;
 646         }
 647
 648         /*
 649          * this makes the path point to (inum INODE_ITEM ioff)
 650          */
 651         key.objectid = inum;
 652         key.type = BTRFS_INODE_ITEM_KEY;
 653         key.offset = 0;
 654
 655         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 656         if (ret) {
 657                 btrfs_put_root(local_root);
 658                 btrfs_release_path(swarn->path);
 659                 goto err;
 660         }
 661
 662         eb = swarn->path->nodes[0];
 663         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 664                                         struct btrfs_inode_item);
 665         isize = btrfs_inode_size(eb, inode_item);
 666         nlink = btrfs_inode_nlink(eb, inode_item);
 667         btrfs_release_path(swarn->path);
 668
 669         /*
 670          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 671          * uses GFP_NOFS in this context, so we keep it consistent but it does
 672          * not seem to be strictly necessary.
 673          */
 674         nofs_flag = memalloc_nofs_save();
 675         ipath = init_ipath(4096, local_root, swarn->path);
 676         memalloc_nofs_restore(nofs_flag);
 677         if (IS_ERR(ipath)) {
 678                 btrfs_put_root(local_root);
 679                 ret = PTR_ERR(ipath);
 680                 ipath = NULL;
 681                 goto err;
 682         }
 683         ret = paths_from_inode(inum, ipath);
 684
 685         if (ret < 0)
 686                 goto err;
 687
 688         /*
 689          * we deliberately ignore the bit ipath might have been too small to
 690          * hold all of the paths here
 691          */
 692         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 693                 btrfs_warn_in_rcu(fs_info,
 694 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 695                                   swarn->errstr, swarn->logical,
 696                                   rcu_str_deref(swarn->dev->name),
 697                                   swarn->physical,
 698                                   root, inum, offset,
 699                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
 700                                   (char *)(unsigned long)ipath->fspath->val[i]);
 701
 702         btrfs_put_root(local_root);
 703         free_ipath(ipath);
 704         return 0;
 705
 706 err:
 707         btrfs_warn_in_rcu(fs_info,
 708                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 709                           swarn->errstr, swarn->logical,
 710                           rcu_str_deref(swarn->dev->name),
 711                           swarn->physical,
 712                           root, inum, offset, ret);
 713
 714         free_ipath(ipath);
 715         return 0;
 716 }
 717
 718 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 719 {
 720         struct btrfs_device *dev;
 721         struct btrfs_fs_info *fs_info;
 722         struct btrfs_path *path;
 723         struct btrfs_key found_key;
 724         struct extent_buffer *eb;
 725         struct btrfs_extent_item *ei;
 726         struct scrub_warning swarn;
 727         unsigned long ptr = 0;
 728         u64 extent_item_pos;
 729         u64 flags = 0;
 730         u64 ref_root;
 731         u32 item_size;
 732         u8 ref_level = 0;
 733         int ret;
 734
 735         WARN_ON(sblock->page_count < 1);
 736         dev = sblock->pagev[0]->dev;
 737         fs_info = sblock->sctx->fs_info;
 738
 739         path = btrfs_alloc_path();
 740         if (!path)
 741                 return;
 742
 743         swarn.physical = sblock->pagev[0]->physical;
 744         swarn.logical = sblock->pagev[0]->logical;
 745         swarn.errstr = errstr;
 746         swarn.dev = NULL;
 747
 748         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 749                                   &flags);
 750         if (ret < 0)
 751                 goto out;
 752
 753         extent_item_pos = swarn.logical - found_key.objectid;
 754         swarn.extent_item_size = found_key.offset;
 755
 756         eb = path->nodes[0];
 757         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 758         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 759
 760         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 761                 do {
 762                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 763                                                       item_size, &ref_root,
 764                                                       &ref_level);
 765                         btrfs_warn_in_rcu(fs_info,
 766 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 767                                 errstr, swarn.logical,
 768                                 rcu_str_deref(dev->name),
 769                                 swarn.physical,
 770                                 ref_level ? "node" : "leaf",
 771                                 ret < 0 ? -1 : ref_level,
 772                                 ret < 0 ? -1 : ref_root);
 773                 } while (ret != 1);
 774                 btrfs_release_path(path);
 775         } else {
 776                 btrfs_release_path(path);
 777                 swarn.path = path;
 778                 swarn.dev = dev;
 779                 iterate_extent_inodes(fs_info, found_key.objectid,
 780                                         extent_item_pos, 1,
 781                                         scrub_print_warning_inode, &swarn, false);
 782         }
 783
 784 out:
 785         btrfs_free_path(path);
 786 }
 787
 788 static inline void scrub_get_recover(struct scrub_recover *recover)
 789 {
 790         refcount_inc(&recover->refs);
 791 }
 792
 793 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 794                                      struct scrub_recover *recover)
 795 {
 796         if (refcount_dec_and_test(&recover->refs)) {
 797                 btrfs_bio_counter_dec(fs_info);
 798                 btrfs_put_bbio(recover->bbio);
 799                 kfree(recover);
 800         }
 801 }
 802
 803 /*
 804  * scrub_handle_errored_block gets called when either verification of the
 805  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 806  * case, this function handles all pages in the bio, even though only one
 807  * may be bad.
 808  * The goal of this function is to repair the errored block by using the
 809  * contents of one of the mirrors.
 810  */
 811 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 812 {
 813         struct scrub_ctx *sctx = sblock_to_check->sctx;
 814         struct btrfs_device *dev;
 815         struct btrfs_fs_info *fs_info;
 816         u64 logical;
 817         unsigned int failed_mirror_index;
 818         unsigned int is_metadata;
 819         unsigned int have_csum;
 820         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 821         struct scrub_block *sblock_bad;
 822         int ret;
 823         int mirror_index;
 824         int page_num;
 825         int success;
 826         bool full_stripe_locked;
 827         unsigned int nofs_flag;
 828         static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
 829                                       DEFAULT_RATELIMIT_BURST);
 830
 831         BUG_ON(sblock_to_check->page_count < 1);
 832         fs_info = sctx->fs_info;
 833         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 834                 /*
 835                  * if we find an error in a super block, we just report it.
 836                  * They will get written with the next transaction commit
 837                  * anyway
 838                  */
 839                 spin_lock(&sctx->stat_lock);
 840                 ++sctx->stat.super_errors;
 841                 spin_unlock(&sctx->stat_lock);
 842                 return 0;
 843         }
 844         logical = sblock_to_check->pagev[0]->logical;
 845         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 846         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 847         is_metadata = !(sblock_to_check->pagev[0]->flags &
 848                         BTRFS_EXTENT_FLAG_DATA);
 849         have_csum = sblock_to_check->pagev[0]->have_csum;
 850         dev = sblock_to_check->pagev[0]->dev;
 851
 852         if (btrfs_is_zoned(fs_info) && !sctx->is_dev_replace)
 853                 return btrfs_repair_one_zone(fs_info, logical);
 854
 855         /*
 856          * We must use GFP_NOFS because the scrub task might be waiting for a
 857          * worker task executing this function and in turn a transaction commit
 858          * might be waiting the scrub task to pause (which needs to wait for all
 859          * the worker tasks to complete before pausing).
 860          * We do allocations in the workers through insert_full_stripe_lock()
 861          * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 862          * this function.
 863          */
 864         nofs_flag = memalloc_nofs_save();
 865         /*
 866          * For RAID5/6, race can happen for a different device scrub thread.
 867          * For data corruption, Parity and Data threads will both try
 868          * to recovery the data.
 869          * Race can lead to doubly added csum error, or even unrecoverable
 870          * error.
 871          */
 872         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 873         if (ret < 0) {
 874                 memalloc_nofs_restore(nofs_flag);
 875                 spin_lock(&sctx->stat_lock);
 876                 if (ret == -ENOMEM)
 877                         sctx->stat.malloc_errors++;
 878                 sctx->stat.read_errors++;
 879                 sctx->stat.uncorrectable_errors++;
 880                 spin_unlock(&sctx->stat_lock);
 881                 return ret;
 882         }
 883
 884         /*
 885          * read all mirrors one after the other. This includes to
 886          * re-read the extent or metadata block that failed (that was
 887          * the cause that this fixup code is called) another time,
 888          * page by page this time in order to know which pages
 889          * caused I/O errors and which ones are good (for all mirrors).
 890          * It is the goal to handle the situation when more than one
 891          * mirror contains I/O errors, but the errors do not
 892          * overlap, i.e. the data can be repaired by selecting the
 893          * pages from those mirrors without I/O error on the
 894          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 895          * would be that mirror #1 has an I/O error on the first page,
 896          * the second page is good, and mirror #2 has an I/O error on
 897          * the second page, but the first page is good.
 898          * Then the first page of the first mirror can be repaired by
 899          * taking the first page of the second mirror, and the
 900          * second page of the second mirror can be repaired by
 901          * copying the contents of the 2nd page of the 1st mirror.
 902          * One more note: if the pages of one mirror contain I/O
 903          * errors, the checksum cannot be verified. In order to get
 904          * the best data for repairing, the first attempt is to find
 905          * a mirror without I/O errors and with a validated checksum.
 906          * Only if this is not possible, the pages are picked from
 907          * mirrors with I/O errors without considering the checksum.
 908          * If the latter is the case, at the end, the checksum of the
 909          * repaired area is verified in order to correctly maintain
 910          * the statistics.
 911          */
 912
 913         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 914                                       sizeof(*sblocks_for_recheck), GFP_KERNEL);
 915         if (!sblocks_for_recheck) {
 916                 spin_lock(&sctx->stat_lock);
 917                 sctx->stat.malloc_errors++;
 918                 sctx->stat.read_errors++;
 919                 sctx->stat.uncorrectable_errors++;
 920                 spin_unlock(&sctx->stat_lock);
 921                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 922                 goto out;
 923         }
 924
 925         /* setup the context, map the logical blocks and alloc the pages */
 926         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 927         if (ret) {
 928                 spin_lock(&sctx->stat_lock);
 929                 sctx->stat.read_errors++;
 930                 sctx->stat.uncorrectable_errors++;
 931                 spin_unlock(&sctx->stat_lock);
 932                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 933                 goto out;
 934         }
 935         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 936         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 937
 938         /* build and submit the bios for the failed mirror, check checksums */
 939         scrub_recheck_block(fs_info, sblock_bad, 1);
 940
 941         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 942             sblock_bad->no_io_error_seen) {
 943                 /*
 944                  * the error disappeared after reading page by page, or
 945                  * the area was part of a huge bio and other parts of the
 946                  * bio caused I/O errors, or the block layer merged several
 947                  * read requests into one and the error is caused by a
 948                  * different bio (usually one of the two latter cases is
 949                  * the cause)
 950                  */
 951                 spin_lock(&sctx->stat_lock);
 952                 sctx->stat.unverified_errors++;
 953                 sblock_to_check->data_corrected = 1;
 954                 spin_unlock(&sctx->stat_lock);
 955
 956                 if (sctx->is_dev_replace)
 957                         scrub_write_block_to_dev_replace(sblock_bad);
 958                 goto out;
 959         }
 960
 961         if (!sblock_bad->no_io_error_seen) {
 962                 spin_lock(&sctx->stat_lock);
 963                 sctx->stat.read_errors++;
 964                 spin_unlock(&sctx->stat_lock);
 965                 if (__ratelimit(&rs))
 966                         scrub_print_warning("i/o error", sblock_to_check);
 967                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 968         } else if (sblock_bad->checksum_error) {
 969                 spin_lock(&sctx->stat_lock);
 970                 sctx->stat.csum_errors++;
 971                 spin_unlock(&sctx->stat_lock);
 972                 if (__ratelimit(&rs))
 973                         scrub_print_warning("checksum error", sblock_to_check);
 974                 btrfs_dev_stat_inc_and_print(dev,
 975                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 976         } else if (sblock_bad->header_error) {
 977                 spin_lock(&sctx->stat_lock);
 978                 sctx->stat.verify_errors++;
 979                 spin_unlock(&sctx->stat_lock);
 980                 if (__ratelimit(&rs))
 981                         scrub_print_warning("checksum/header error",
 982                                             sblock_to_check);
 983                 if (sblock_bad->generation_error)
 984                         btrfs_dev_stat_inc_and_print(dev,
 985                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 986                 else
 987                         btrfs_dev_stat_inc_and_print(dev,
 988                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 989         }
 990
 991         if (sctx->readonly) {
 992                 ASSERT(!sctx->is_dev_replace);
 993                 goto out;
 994         }
 995
 996         /*
 997          * now build and submit the bios for the other mirrors, check
 998          * checksums.
 999          * First try to pick the mirror which is completely without I/O
1000          * errors and also does not have a checksum error.
1001          * If one is found, and if a checksum is present, the full block
1002          * that is known to contain an error is rewritten. Afterwards
1003          * the block is known to be corrected.
1004          * If a mirror is found which is completely correct, and no
1005          * checksum is present, only those pages are rewritten that had
1006          * an I/O error in the block to be repaired, since it cannot be
1007          * determined, which copy of the other pages is better (and it
1008          * could happen otherwise that a correct page would be
1009          * overwritten by a bad one).
1010          */
1011         for (mirror_index = 0; ;mirror_index++) {
1012                 struct scrub_block *sblock_other;
1013
1014                 if (mirror_index == failed_mirror_index)
1015                         continue;
1016
1017                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1018                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1019                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1020                                 break;
1021                         if (!sblocks_for_recheck[mirror_index].page_count)
1022                                 break;
1023
1024                         sblock_other = sblocks_for_recheck + mirror_index;
1025                 } else {
1026                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1027                         int max_allowed = r->bbio->num_stripes -
1028                                                 r->bbio->num_tgtdevs;
1029
1030                         if (mirror_index >= max_allowed)
1031                                 break;
1032                         if (!sblocks_for_recheck[1].page_count)
1033                                 break;
1034
1035                         ASSERT(failed_mirror_index == 0);
1036                         sblock_other = sblocks_for_recheck + 1;
1037                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1038                 }
1039
1040                 /* build and submit the bios, check checksums */
1041                 scrub_recheck_block(fs_info, sblock_other, 0);
1042
1043                 if (!sblock_other->header_error &&
1044                     !sblock_other->checksum_error &&
1045                     sblock_other->no_io_error_seen) {
1046                         if (sctx->is_dev_replace) {
1047                                 scrub_write_block_to_dev_replace(sblock_other);
1048                                 goto corrected_error;
1049                         } else {
1050                                 ret = scrub_repair_block_from_good_copy(
1051                                                 sblock_bad, sblock_other);
1052                                 if (!ret)
1053                                         goto corrected_error;
1054                         }
1055                 }
1056         }
1057
1058         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1059                 goto did_not_correct_error;
1060
1061         /*
1062          * In case of I/O errors in the area that is supposed to be
1063          * repaired, continue by picking good copies of those pages.
1064          * Select the good pages from mirrors to rewrite bad pages from
1065          * the area to fix. Afterwards verify the checksum of the block
1066          * that is supposed to be repaired. This verification step is
1067          * only done for the purpose of statistic counting and for the
1068          * final scrub report, whether errors remain.
1069          * A perfect algorithm could make use of the checksum and try
1070          * all possible combinations of pages from the different mirrors
1071          * until the checksum verification succeeds. For example, when
1072          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1073          * of mirror #2 is readable but the final checksum test fails,
1074          * then the 2nd page of mirror #3 could be tried, whether now
1075          * the final checksum succeeds. But this would be a rare
1076          * exception and is therefore not implemented. At least it is
1077          * avoided that the good copy is overwritten.
1078          * A more useful improvement would be to pick the sectors
1079          * without I/O error based on sector sizes (512 bytes on legacy
1080          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1081          * mirror could be repaired by taking 512 byte of a different
1082          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1083          * area are unreadable.
1084          */
1085         success = 1;
1086         for (page_num = 0; page_num < sblock_bad->page_count;
1087              page_num++) {
1088                 struct scrub_page *spage_bad = sblock_bad->pagev[page_num];
1089                 struct scrub_block *sblock_other = NULL;
1090
1091                 /* skip no-io-error page in scrub */
1092                 if (!spage_bad->io_error && !sctx->is_dev_replace)
1093                         continue;
1094
1095                 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1096                         /*
1097                          * In case of dev replace, if raid56 rebuild process
1098                          * didn't work out correct data, then copy the content
1099                          * in sblock_bad to make sure target device is identical
1100                          * to source device, instead of writing garbage data in
1101                          * sblock_for_recheck array to target device.
1102                          */
1103                         sblock_other = NULL;
1104                 } else if (spage_bad->io_error) {
1105                         /* try to find no-io-error page in mirrors */
1106                         for (mirror_index = 0;
1107                              mirror_index < BTRFS_MAX_MIRRORS &&
1108                              sblocks_for_recheck[mirror_index].page_count > 0;
1109                              mirror_index++) {
1110                                 if (!sblocks_for_recheck[mirror_index].
1111                                     pagev[page_num]->io_error) {
1112                                         sblock_other = sblocks_for_recheck +
1113                                                        mirror_index;
1114                                         break;
1115                                 }
1116                         }
1117                         if (!sblock_other)
1118                                 success = 0;
1119                 }
1120
1121                 if (sctx->is_dev_replace) {
1122                         /*
1123                          * did not find a mirror to fetch the page
1124                          * from. scrub_write_page_to_dev_replace()
1125                          * handles this case (page->io_error), by
1126                          * filling the block with zeros before
1127                          * submitting the write request
1128                          */
1129                         if (!sblock_other)
1130                                 sblock_other = sblock_bad;
1131
1132                         if (scrub_write_page_to_dev_replace(sblock_other,
1133                                                             page_num) != 0) {
1134                                 atomic64_inc(
1135                                         &fs_info->dev_replace.num_write_errors);
1136                                 success = 0;
1137                         }
1138                 } else if (sblock_other) {
1139                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1140                                                                sblock_other,
1141                                                                page_num, 0);
1142                         if (0 == ret)
1143                                 spage_bad->io_error = 0;
1144                         else
1145                                 success = 0;
1146                 }
1147         }
1148
1149         if (success && !sctx->is_dev_replace) {
1150                 if (is_metadata || have_csum) {
1151                         /*
1152                          * need to verify the checksum now that all
1153                          * sectors on disk are repaired (the write
1154                          * request for data to be repaired is on its way).
1155                          * Just be lazy and use scrub_recheck_block()
1156                          * which re-reads the data before the checksum
1157                          * is verified, but most likely the data comes out
1158                          * of the page cache.
1159                          */
1160                         scrub_recheck_block(fs_info, sblock_bad, 1);
1161                         if (!sblock_bad->header_error &&
1162                             !sblock_bad->checksum_error &&
1163                             sblock_bad->no_io_error_seen)
1164                                 goto corrected_error;
1165                         else
1166                                 goto did_not_correct_error;
1167                 } else {
1168 corrected_error:
1169                         spin_lock(&sctx->stat_lock);
1170                         sctx->stat.corrected_errors++;
1171                         sblock_to_check->data_corrected = 1;
1172                         spin_unlock(&sctx->stat_lock);
1173                         btrfs_err_rl_in_rcu(fs_info,
1174                                 "fixed up error at logical %llu on dev %s",
1175                                 logical, rcu_str_deref(dev->name));
1176                 }
1177         } else {
1178 did_not_correct_error:
1179                 spin_lock(&sctx->stat_lock);
1180                 sctx->stat.uncorrectable_errors++;
1181                 spin_unlock(&sctx->stat_lock);
1182                 btrfs_err_rl_in_rcu(fs_info,
1183                         "unable to fixup (regular) error at logical %llu on dev %s",
1184                         logical, rcu_str_deref(dev->name));
1185         }
1186
1187 out:
1188         if (sblocks_for_recheck) {
1189                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1190                      mirror_index++) {
1191                         struct scrub_block *sblock = sblocks_for_recheck +
1192                                                      mirror_index;
1193                         struct scrub_recover *recover;
1194                         int page_index;
1195
1196                         for (page_index = 0; page_index < sblock->page_count;
1197                              page_index++) {
1198                                 sblock->pagev[page_index]->sblock = NULL;
1199                                 recover = sblock->pagev[page_index]->recover;
1200                                 if (recover) {
1201                                         scrub_put_recover(fs_info, recover);
1202                                         sblock->pagev[page_index]->recover =
1203                                                                         NULL;
1204                                 }
1205                                 scrub_page_put(sblock->pagev[page_index]);
1206                         }
1207                 }
1208                 kfree(sblocks_for_recheck);
1209         }
1210
1211         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1212         memalloc_nofs_restore(nofs_flag);
1213         if (ret < 0)
1214                 return ret;
1215         return 0;
1216 }
1217
1218 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1219 {
1220         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1221                 return 2;
1222         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1223                 return 3;
1224         else
1225                 return (int)bbio->num_stripes;
1226 }
1227
1228 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1229                                                  u64 *raid_map,
1230                                                  u64 mapped_length,
1231                                                  int nstripes, int mirror,
1232                                                  int *stripe_index,
1233                                                  u64 *stripe_offset)
1234 {
1235         int i;
1236
1237         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1238                 /* RAID5/6 */
1239                 for (i = 0; i < nstripes; i++) {
1240                         if (raid_map[i] == RAID6_Q_STRIPE ||
1241                             raid_map[i] == RAID5_P_STRIPE)
1242                                 continue;
1243
1244                         if (logical >= raid_map[i] &&
1245                             logical < raid_map[i] + mapped_length)
1246                                 break;
1247                 }
1248
1249                 *stripe_index = i;
1250                 *stripe_offset = logical - raid_map[i];
1251         } else {
1252                 /* The other RAID type */
1253                 *stripe_index = mirror;
1254                 *stripe_offset = 0;
1255         }
1256 }
1257
1258 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1259                                      struct scrub_block *sblocks_for_recheck)
1260 {
1261         struct scrub_ctx *sctx = original_sblock->sctx;
1262         struct btrfs_fs_info *fs_info = sctx->fs_info;
1263         u64 length = original_sblock->page_count * PAGE_SIZE;
1264         u64 logical = original_sblock->pagev[0]->logical;
1265         u64 generation = original_sblock->pagev[0]->generation;
1266         u64 flags = original_sblock->pagev[0]->flags;
1267         u64 have_csum = original_sblock->pagev[0]->have_csum;
1268         struct scrub_recover *recover;
1269         struct btrfs_bio *bbio;
1270         u64 sublen;
1271         u64 mapped_length;
1272         u64 stripe_offset;
1273         int stripe_index;
1274         int page_index = 0;
1275         int mirror_index;
1276         int nmirrors;
1277         int ret;
1278
1279         /*
1280          * note: the two members refs and outstanding_pages
1281          * are not used (and not set) in the blocks that are used for
1282          * the recheck procedure
1283          */
1284
1285         while (length > 0) {
1286                 sublen = min_t(u64, length, PAGE_SIZE);
1287                 mapped_length = sublen;
1288                 bbio = NULL;
1289
1290                 /*
1291                  * with a length of PAGE_SIZE, each returned stripe
1292                  * represents one mirror
1293                  */
1294                 btrfs_bio_counter_inc_blocked(fs_info);
1295                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1296                                 logical, &mapped_length, &bbio);
1297                 if (ret || !bbio || mapped_length < sublen) {
1298                         btrfs_put_bbio(bbio);
1299                         btrfs_bio_counter_dec(fs_info);
1300                         return -EIO;
1301                 }
1302
1303                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1304                 if (!recover) {
1305                         btrfs_put_bbio(bbio);
1306                         btrfs_bio_counter_dec(fs_info);
1307                         return -ENOMEM;
1308                 }
1309
1310                 refcount_set(&recover->refs, 1);
1311                 recover->bbio = bbio;
1312                 recover->map_length = mapped_length;
1313
1314                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1315
1316                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1317
1318                 for (mirror_index = 0; mirror_index < nmirrors;
1319                      mirror_index++) {
1320                         struct scrub_block *sblock;
1321                         struct scrub_page *spage;
1322
1323                         sblock = sblocks_for_recheck + mirror_index;
1324                         sblock->sctx = sctx;
1325
1326                         spage = kzalloc(sizeof(*spage), GFP_NOFS);
1327                         if (!spage) {
1328 leave_nomem:
1329                                 spin_lock(&sctx->stat_lock);
1330                                 sctx->stat.malloc_errors++;
1331                                 spin_unlock(&sctx->stat_lock);
1332                                 scrub_put_recover(fs_info, recover);
1333                                 return -ENOMEM;
1334                         }
1335                         scrub_page_get(spage);
1336                         sblock->pagev[page_index] = spage;
1337                         spage->sblock = sblock;
1338                         spage->flags = flags;
1339                         spage->generation = generation;
1340                         spage->logical = logical;
1341                         spage->have_csum = have_csum;
1342                         if (have_csum)
1343                                 memcpy(spage->csum,
1344                                        original_sblock->pagev[0]->csum,
1345                                        sctx->fs_info->csum_size);
1346
1347                         scrub_stripe_index_and_offset(logical,
1348                                                       bbio->map_type,
1349                                                       bbio->raid_map,
1350                                                       mapped_length,
1351                                                       bbio->num_stripes -
1352                                                       bbio->num_tgtdevs,
1353                                                       mirror_index,
1354                                                       &stripe_index,
1355                                                       &stripe_offset);
1356                         spage->physical = bbio->stripes[stripe_index].physical +
1357                                          stripe_offset;
1358                         spage->dev = bbio->stripes[stripe_index].dev;
1359
1360                         BUG_ON(page_index >= original_sblock->page_count);
1361                         spage->physical_for_dev_replace =
1362                                 original_sblock->pagev[page_index]->
1363                                 physical_for_dev_replace;
1364                         /* for missing devices, dev->bdev is NULL */
1365                         spage->mirror_num = mirror_index + 1;
1366                         sblock->page_count++;
1367                         spage->page = alloc_page(GFP_NOFS);
1368                         if (!spage->page)
1369                                 goto leave_nomem;
1370
1371                         scrub_get_recover(recover);
1372                         spage->recover = recover;
1373                 }
1374                 scrub_put_recover(fs_info, recover);
1375                 length -= sublen;
1376                 logical += sublen;
1377                 page_index++;
1378         }
1379
1380         return 0;
1381 }
1382
1383 static void scrub_bio_wait_endio(struct bio *bio)
1384 {
1385         complete(bio->bi_private);
1386 }
1387
1388 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1389                                         struct bio *bio,
1390                                         struct scrub_page *spage)
1391 {
1392         DECLARE_COMPLETION_ONSTACK(done);
1393         int ret;
1394         int mirror_num;
1395
1396         bio->bi_iter.bi_sector = spage->logical >> 9;
1397         bio->bi_private = &done;
1398         bio->bi_end_io = scrub_bio_wait_endio;
1399
1400         mirror_num = spage->sblock->pagev[0]->mirror_num;
1401         ret = raid56_parity_recover(fs_info, bio, spage->recover->bbio,
1402                                     spage->recover->map_length,
1403                                     mirror_num, 0);
1404         if (ret)
1405                 return ret;
1406
1407         wait_for_completion_io(&done);
1408         return blk_status_to_errno(bio->bi_status);
1409 }
1410
1411 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1412                                           struct scrub_block *sblock)
1413 {
1414         struct scrub_page *first_page = sblock->pagev[0];
1415         struct bio *bio;
1416         int page_num;
1417
1418         /* All pages in sblock belong to the same stripe on the same device. */
1419         ASSERT(first_page->dev);
1420         if (!first_page->dev->bdev)
1421                 goto out;
1422
1423         bio = btrfs_io_bio_alloc(BIO_MAX_VECS);
1424         bio_set_dev(bio, first_page->dev->bdev);
1425
1426         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1427                 struct scrub_page *spage = sblock->pagev[page_num];
1428
1429                 WARN_ON(!spage->page);
1430                 bio_add_page(bio, spage->page, PAGE_SIZE, 0);
1431         }
1432
1433         if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1434                 bio_put(bio);
1435                 goto out;
1436         }
1437
1438         bio_put(bio);
1439
1440         scrub_recheck_block_checksum(sblock);
1441
1442         return;
1443 out:
1444         for (page_num = 0; page_num < sblock->page_count; page_num++)
1445                 sblock->pagev[page_num]->io_error = 1;
1446
1447         sblock->no_io_error_seen = 0;
1448 }
1449
1450 /*
1451  * this function will check the on disk data for checksum errors, header
1452  * errors and read I/O errors. If any I/O errors happen, the exact pages
1453  * which are errored are marked as being bad. The goal is to enable scrub
1454  * to take those pages that are not errored from all the mirrors so that
1455  * the pages that are errored in the just handled mirror can be repaired.
1456  */
1457 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1458                                 struct scrub_block *sblock,
1459                                 int retry_failed_mirror)
1460 {
1461         int page_num;
1462
1463         sblock->no_io_error_seen = 1;
1464
1465         /* short cut for raid56 */
1466         if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1467                 return scrub_recheck_block_on_raid56(fs_info, sblock);
1468
1469         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1470                 struct bio *bio;
1471                 struct scrub_page *spage = sblock->pagev[page_num];
1472
1473                 if (spage->dev->bdev == NULL) {
1474                         spage->io_error = 1;
1475                         sblock->no_io_error_seen = 0;
1476                         continue;
1477                 }
1478
1479                 WARN_ON(!spage->page);
1480                 bio = btrfs_io_bio_alloc(1);
1481                 bio_set_dev(bio, spage->dev->bdev);
1482
1483                 bio_add_page(bio, spage->page, PAGE_SIZE, 0);
1484                 bio->bi_iter.bi_sector = spage->physical >> 9;
1485                 bio->bi_opf = REQ_OP_READ;
1486
1487                 if (btrfsic_submit_bio_wait(bio)) {
1488                         spage->io_error = 1;
1489                         sblock->no_io_error_seen = 0;
1490                 }
1491
1492                 bio_put(bio);
1493         }
1494
1495         if (sblock->no_io_error_seen)
1496                 scrub_recheck_block_checksum(sblock);
1497 }
1498
1499 static inline int scrub_check_fsid(u8 fsid[],
1500                                    struct scrub_page *spage)
1501 {
1502         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1503         int ret;
1504
1505         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1506         return !ret;
1507 }
1508
1509 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1510 {
1511         sblock->header_error = 0;
1512         sblock->checksum_error = 0;
1513         sblock->generation_error = 0;
1514
1515         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1516                 scrub_checksum_data(sblock);
1517         else
1518                 scrub_checksum_tree_block(sblock);
1519 }
1520
1521 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1522                                              struct scrub_block *sblock_good)
1523 {
1524         int page_num;
1525         int ret = 0;
1526
1527         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1528                 int ret_sub;
1529
1530                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1531                                                            sblock_good,
1532                                                            page_num, 1);
1533                 if (ret_sub)
1534                         ret = ret_sub;
1535         }
1536
1537         return ret;
1538 }
1539
1540 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1541                                             struct scrub_block *sblock_good,
1542                                             int page_num, int force_write)
1543 {
1544         struct scrub_page *spage_bad = sblock_bad->pagev[page_num];
1545         struct scrub_page *spage_good = sblock_good->pagev[page_num];
1546         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1547
1548         BUG_ON(spage_bad->page == NULL);
1549         BUG_ON(spage_good->page == NULL);
1550         if (force_write || sblock_bad->header_error ||
1551             sblock_bad->checksum_error || spage_bad->io_error) {
1552                 struct bio *bio;
1553                 int ret;
1554
1555                 if (!spage_bad->dev->bdev) {
1556                         btrfs_warn_rl(fs_info,
1557                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1558                         return -EIO;
1559                 }
1560
1561                 bio = btrfs_io_bio_alloc(1);
1562                 bio_set_dev(bio, spage_bad->dev->bdev);
1563                 bio->bi_iter.bi_sector = spage_bad->physical >> 9;
1564                 bio->bi_opf = REQ_OP_WRITE;
1565
1566                 ret = bio_add_page(bio, spage_good->page, PAGE_SIZE, 0);
1567                 if (PAGE_SIZE != ret) {
1568                         bio_put(bio);
1569                         return -EIO;
1570                 }
1571
1572                 if (btrfsic_submit_bio_wait(bio)) {
1573                         btrfs_dev_stat_inc_and_print(spage_bad->dev,
1574                                 BTRFS_DEV_STAT_WRITE_ERRS);
1575                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1576                         bio_put(bio);
1577                         return -EIO;
1578                 }
1579                 bio_put(bio);
1580         }
1581
1582         return 0;
1583 }
1584
1585 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1586 {
1587         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1588         int page_num;
1589
1590         /*
1591          * This block is used for the check of the parity on the source device,
1592          * so the data needn't be written into the destination device.
1593          */
1594         if (sblock->sparity)
1595                 return;
1596
1597         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1598                 int ret;
1599
1600                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1601                 if (ret)
1602                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1603         }
1604 }
1605
1606 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1607                                            int page_num)
1608 {
1609         struct scrub_page *spage = sblock->pagev[page_num];
1610
1611         BUG_ON(spage->page == NULL);
1612         if (spage->io_error)
1613                 clear_page(page_address(spage->page));
1614
1615         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1616 }
1617
1618 static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
1619 {
1620         int ret = 0;
1621         u64 length;
1622
1623         if (!btrfs_is_zoned(sctx->fs_info))
1624                 return 0;
1625
1626         if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1627                 return 0;
1628
1629         if (sctx->write_pointer < physical) {
1630                 length = physical - sctx->write_pointer;
1631
1632                 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1633                                                 sctx->write_pointer, length);
1634                 if (!ret)
1635                         sctx->write_pointer = physical;
1636         }
1637         return ret;
1638 }
1639
1640 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1641                                     struct scrub_page *spage)
1642 {
1643         struct scrub_bio *sbio;
1644         int ret;
1645
1646         mutex_lock(&sctx->wr_lock);
1647 again:
1648         if (!sctx->wr_curr_bio) {
1649                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1650                                               GFP_KERNEL);
1651                 if (!sctx->wr_curr_bio) {
1652                         mutex_unlock(&sctx->wr_lock);
1653                         return -ENOMEM;
1654                 }
1655                 sctx->wr_curr_bio->sctx = sctx;
1656                 sctx->wr_curr_bio->page_count = 0;
1657         }
1658         sbio = sctx->wr_curr_bio;
1659         if (sbio->page_count == 0) {
1660                 struct bio *bio;
1661
1662                 ret = fill_writer_pointer_gap(sctx,
1663                                               spage->physical_for_dev_replace);
1664                 if (ret) {
1665                         mutex_unlock(&sctx->wr_lock);
1666                         return ret;
1667                 }
1668
1669                 sbio->physical = spage->physical_for_dev_replace;
1670                 sbio->logical = spage->logical;
1671                 sbio->dev = sctx->wr_tgtdev;
1672                 bio = sbio->bio;
1673                 if (!bio) {
1674                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1675                         sbio->bio = bio;
1676                 }
1677
1678                 bio->bi_private = sbio;
1679                 bio->bi_end_io = scrub_wr_bio_end_io;
1680                 bio_set_dev(bio, sbio->dev->bdev);
1681                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1682                 bio->bi_opf = REQ_OP_WRITE;
1683                 sbio->status = 0;
1684         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1685                    spage->physical_for_dev_replace ||
1686                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1687                    spage->logical) {
1688                 scrub_wr_submit(sctx);
1689                 goto again;
1690         }
1691
1692         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1693         if (ret != PAGE_SIZE) {
1694                 if (sbio->page_count < 1) {
1695                         bio_put(sbio->bio);
1696                         sbio->bio = NULL;
1697                         mutex_unlock(&sctx->wr_lock);
1698                         return -EIO;
1699                 }
1700                 scrub_wr_submit(sctx);
1701                 goto again;
1702         }
1703
1704         sbio->pagev[sbio->page_count] = spage;
1705         scrub_page_get(spage);
1706         sbio->page_count++;
1707         if (sbio->page_count == sctx->pages_per_wr_bio)
1708                 scrub_wr_submit(sctx);
1709         mutex_unlock(&sctx->wr_lock);
1710
1711         return 0;
1712 }
1713
1714 static void scrub_wr_submit(struct scrub_ctx *sctx)
1715 {
1716         struct scrub_bio *sbio;
1717
1718         if (!sctx->wr_curr_bio)
1719                 return;
1720
1721         sbio = sctx->wr_curr_bio;
1722         sctx->wr_curr_bio = NULL;
1723         WARN_ON(!sbio->bio->bi_bdev);
1724         scrub_pending_bio_inc(sctx);
1725         /* process all writes in a single worker thread. Then the block layer
1726          * orders the requests before sending them to the driver which
1727          * doubled the write performance on spinning disks when measured
1728          * with Linux 3.5 */
1729         btrfsic_submit_bio(sbio->bio);
1730
1731         if (btrfs_is_zoned(sctx->fs_info))
1732                 sctx->write_pointer = sbio->physical + sbio->page_count * PAGE_SIZE;
1733 }
1734
1735 static void scrub_wr_bio_end_io(struct bio *bio)
1736 {
1737         struct scrub_bio *sbio = bio->bi_private;
1738         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1739
1740         sbio->status = bio->bi_status;
1741         sbio->bio = bio;
1742
1743         btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1744         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1745 }
1746
1747 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1748 {
1749         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1750         struct scrub_ctx *sctx = sbio->sctx;
1751         int i;
1752
1753         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1754         if (sbio->status) {
1755                 struct btrfs_dev_replace *dev_replace =
1756                         &sbio->sctx->fs_info->dev_replace;
1757
1758                 for (i = 0; i < sbio->page_count; i++) {
1759                         struct scrub_page *spage = sbio->pagev[i];
1760
1761                         spage->io_error = 1;
1762                         atomic64_inc(&dev_replace->num_write_errors);
1763                 }
1764         }
1765
1766         for (i = 0; i < sbio->page_count; i++)
1767                 scrub_page_put(sbio->pagev[i]);
1768
1769         bio_put(sbio->bio);
1770         kfree(sbio);
1771         scrub_pending_bio_dec(sctx);
1772 }
1773
1774 static int scrub_checksum(struct scrub_block *sblock)
1775 {
1776         u64 flags;
1777         int ret;
1778
1779         /*
1780          * No need to initialize these stats currently,
1781          * because this function only use return value
1782          * instead of these stats value.
1783          *
1784          * Todo:
1785          * always use stats
1786          */
1787         sblock->header_error = 0;
1788         sblock->generation_error = 0;
1789         sblock->checksum_error = 0;
1790
1791         WARN_ON(sblock->page_count < 1);
1792         flags = sblock->pagev[0]->flags;
1793         ret = 0;
1794         if (flags & BTRFS_EXTENT_FLAG_DATA)
1795                 ret = scrub_checksum_data(sblock);
1796         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1797                 ret = scrub_checksum_tree_block(sblock);
1798         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1799                 (void)scrub_checksum_super(sblock);
1800         else
1801                 WARN_ON(1);
1802         if (ret)
1803                 scrub_handle_errored_block(sblock);
1804
1805         return ret;
1806 }
1807
1808 static int scrub_checksum_data(struct scrub_block *sblock)
1809 {
1810         struct scrub_ctx *sctx = sblock->sctx;
1811         struct btrfs_fs_info *fs_info = sctx->fs_info;
1812         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1813         u8 csum[BTRFS_CSUM_SIZE];
1814         struct scrub_page *spage;
1815         char *kaddr;
1816
1817         BUG_ON(sblock->page_count < 1);
1818         spage = sblock->pagev[0];
1819         if (!spage->have_csum)
1820                 return 0;
1821
1822         kaddr = page_address(spage->page);
1823
1824         shash->tfm = fs_info->csum_shash;
1825         crypto_shash_init(shash);
1826
1827         /*
1828          * In scrub_pages() and scrub_pages_for_parity() we ensure each spage
1829          * only contains one sector of data.
1830          */
1831         crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
1832
1833         if (memcmp(csum, spage->csum, fs_info->csum_size))
1834                 sblock->checksum_error = 1;
1835         return sblock->checksum_error;
1836 }
1837
1838 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1839 {
1840         struct scrub_ctx *sctx = sblock->sctx;
1841         struct btrfs_header *h;
1842         struct btrfs_fs_info *fs_info = sctx->fs_info;
1843         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1844         u8 calculated_csum[BTRFS_CSUM_SIZE];
1845         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1846         /*
1847          * This is done in sectorsize steps even for metadata as there's a
1848          * constraint for nodesize to be aligned to sectorsize. This will need
1849          * to change so we don't misuse data and metadata units like that.
1850          */
1851         const u32 sectorsize = sctx->fs_info->sectorsize;
1852         const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
1853         int i;
1854         struct scrub_page *spage;
1855         char *kaddr;
1856
1857         BUG_ON(sblock->page_count < 1);
1858
1859         /* Each member in pagev is just one block, not a full page */
1860         ASSERT(sblock->page_count == num_sectors);
1861
1862         spage = sblock->pagev[0];
1863         kaddr = page_address(spage->page);
1864         h = (struct btrfs_header *)kaddr;
1865         memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
1866
1867         /*
1868          * we don't use the getter functions here, as we
1869          * a) don't have an extent buffer and
1870          * b) the page is already kmapped
1871          */
1872         if (spage->logical != btrfs_stack_header_bytenr(h))
1873                 sblock->header_error = 1;
1874
1875         if (spage->generation != btrfs_stack_header_generation(h)) {
1876                 sblock->header_error = 1;
1877                 sblock->generation_error = 1;
1878         }
1879
1880         if (!scrub_check_fsid(h->fsid, spage))
1881                 sblock->header_error = 1;
1882
1883         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1884                    BTRFS_UUID_SIZE))
1885                 sblock->header_error = 1;
1886
1887         shash->tfm = fs_info->csum_shash;
1888         crypto_shash_init(shash);
1889         crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
1890                             sectorsize - BTRFS_CSUM_SIZE);
1891
1892         for (i = 1; i < num_sectors; i++) {
1893                 kaddr = page_address(sblock->pagev[i]->page);
1894                 crypto_shash_update(shash, kaddr, sectorsize);
1895         }
1896
1897         crypto_shash_final(shash, calculated_csum);
1898         if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size))
1899                 sblock->checksum_error = 1;
1900
1901         return sblock->header_error || sblock->checksum_error;
1902 }
1903
1904 static int scrub_checksum_super(struct scrub_block *sblock)
1905 {
1906         struct btrfs_super_block *s;
1907         struct scrub_ctx *sctx = sblock->sctx;
1908         struct btrfs_fs_info *fs_info = sctx->fs_info;
1909         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1910         u8 calculated_csum[BTRFS_CSUM_SIZE];
1911         struct scrub_page *spage;
1912         char *kaddr;
1913         int fail_gen = 0;
1914         int fail_cor = 0;
1915
1916         BUG_ON(sblock->page_count < 1);
1917         spage = sblock->pagev[0];
1918         kaddr = page_address(spage->page);
1919         s = (struct btrfs_super_block *)kaddr;
1920
1921         if (spage->logical != btrfs_super_bytenr(s))
1922                 ++fail_cor;
1923
1924         if (spage->generation != btrfs_super_generation(s))
1925                 ++fail_gen;
1926
1927         if (!scrub_check_fsid(s->fsid, spage))
1928                 ++fail_cor;
1929
1930         shash->tfm = fs_info->csum_shash;
1931         crypto_shash_init(shash);
1932         crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
1933                         BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
1934
1935         if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
1936                 ++fail_cor;
1937
1938         if (fail_cor + fail_gen) {
1939                 /*
1940                  * if we find an error in a super block, we just report it.
1941                  * They will get written with the next transaction commit
1942                  * anyway
1943                  */
1944                 spin_lock(&sctx->stat_lock);
1945                 ++sctx->stat.super_errors;
1946                 spin_unlock(&sctx->stat_lock);
1947                 if (fail_cor)
1948                         btrfs_dev_stat_inc_and_print(spage->dev,
1949                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1950                 else
1951                         btrfs_dev_stat_inc_and_print(spage->dev,
1952                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1953         }
1954
1955         return fail_cor + fail_gen;
1956 }
1957
1958 static void scrub_block_get(struct scrub_block *sblock)
1959 {
1960         refcount_inc(&sblock->refs);
1961 }
1962
1963 static void scrub_block_put(struct scrub_block *sblock)
1964 {
1965         if (refcount_dec_and_test(&sblock->refs)) {
1966                 int i;
1967
1968                 if (sblock->sparity)
1969                         scrub_parity_put(sblock->sparity);
1970
1971                 for (i = 0; i < sblock->page_count; i++)
1972                         scrub_page_put(sblock->pagev[i]);
1973                 kfree(sblock);
1974         }
1975 }
1976
1977 static void scrub_page_get(struct scrub_page *spage)
1978 {
1979         atomic_inc(&spage->refs);
1980 }
1981
1982 static void scrub_page_put(struct scrub_page *spage)
1983 {
1984         if (atomic_dec_and_test(&spage->refs)) {
1985                 if (spage->page)
1986                         __free_page(spage->page);
1987                 kfree(spage);
1988         }
1989 }
1990
1991 static void scrub_submit(struct scrub_ctx *sctx)
1992 {
1993         struct scrub_bio *sbio;
1994
1995         if (sctx->curr == -1)
1996                 return;
1997
1998         sbio = sctx->bios[sctx->curr];
1999         sctx->curr = -1;
2000         scrub_pending_bio_inc(sctx);
2001         btrfsic_submit_bio(sbio->bio);
2002 }
2003
2004 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2005                                     struct scrub_page *spage)
2006 {
2007         struct scrub_block *sblock = spage->sblock;
2008         struct scrub_bio *sbio;
2009         int ret;
2010
2011 again:
2012         /*
2013          * grab a fresh bio or wait for one to become available
2014          */
2015         while (sctx->curr == -1) {
2016                 spin_lock(&sctx->list_lock);
2017                 sctx->curr = sctx->first_free;
2018                 if (sctx->curr != -1) {
2019                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2020                         sctx->bios[sctx->curr]->next_free = -1;
2021                         sctx->bios[sctx->curr]->page_count = 0;
2022                         spin_unlock(&sctx->list_lock);
2023                 } else {
2024                         spin_unlock(&sctx->list_lock);
2025                         wait_event(sctx->list_wait, sctx->first_free != -1);
2026                 }
2027         }
2028         sbio = sctx->bios[sctx->curr];
2029         if (sbio->page_count == 0) {
2030                 struct bio *bio;
2031
2032                 sbio->physical = spage->physical;
2033                 sbio->logical = spage->logical;
2034                 sbio->dev = spage->dev;
2035                 bio = sbio->bio;
2036                 if (!bio) {
2037                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2038                         sbio->bio = bio;
2039                 }
2040
2041                 bio->bi_private = sbio;
2042                 bio->bi_end_io = scrub_bio_end_io;
2043                 bio_set_dev(bio, sbio->dev->bdev);
2044                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2045                 bio->bi_opf = REQ_OP_READ;
2046                 sbio->status = 0;
2047         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2048                    spage->physical ||
2049                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2050                    spage->logical ||
2051                    sbio->dev != spage->dev) {
2052                 scrub_submit(sctx);
2053                 goto again;
2054         }
2055
2056         sbio->pagev[sbio->page_count] = spage;
2057         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2058         if (ret != PAGE_SIZE) {
2059                 if (sbio->page_count < 1) {
2060                         bio_put(sbio->bio);
2061                         sbio->bio = NULL;
2062                         return -EIO;
2063                 }
2064                 scrub_submit(sctx);
2065                 goto again;
2066         }
2067
2068         scrub_block_get(sblock); /* one for the page added to the bio */
2069         atomic_inc(&sblock->outstanding_pages);
2070         sbio->page_count++;
2071         if (sbio->page_count == sctx->pages_per_rd_bio)
2072                 scrub_submit(sctx);
2073
2074         return 0;
2075 }
2076
2077 static void scrub_missing_raid56_end_io(struct bio *bio)
2078 {
2079         struct scrub_block *sblock = bio->bi_private;
2080         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2081
2082         if (bio->bi_status)
2083                 sblock->no_io_error_seen = 0;
2084
2085         bio_put(bio);
2086
2087         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2088 }
2089
2090 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2091 {
2092         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2093         struct scrub_ctx *sctx = sblock->sctx;
2094         struct btrfs_fs_info *fs_info = sctx->fs_info;
2095         u64 logical;
2096         struct btrfs_device *dev;
2097
2098         logical = sblock->pagev[0]->logical;
2099         dev = sblock->pagev[0]->dev;
2100
2101         if (sblock->no_io_error_seen)
2102                 scrub_recheck_block_checksum(sblock);
2103
2104         if (!sblock->no_io_error_seen) {
2105                 spin_lock(&sctx->stat_lock);
2106                 sctx->stat.read_errors++;
2107                 spin_unlock(&sctx->stat_lock);
2108                 btrfs_err_rl_in_rcu(fs_info,
2109                         "IO error rebuilding logical %llu for dev %s",
2110                         logical, rcu_str_deref(dev->name));
2111         } else if (sblock->header_error || sblock->checksum_error) {
2112                 spin_lock(&sctx->stat_lock);
2113                 sctx->stat.uncorrectable_errors++;
2114                 spin_unlock(&sctx->stat_lock);
2115                 btrfs_err_rl_in_rcu(fs_info,
2116                         "failed to rebuild valid logical %llu for dev %s",
2117                         logical, rcu_str_deref(dev->name));
2118         } else {
2119                 scrub_write_block_to_dev_replace(sblock);
2120         }
2121
2122         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2123                 mutex_lock(&sctx->wr_lock);
2124                 scrub_wr_submit(sctx);
2125                 mutex_unlock(&sctx->wr_lock);
2126         }
2127
2128         scrub_block_put(sblock);
2129         scrub_pending_bio_dec(sctx);
2130 }
2131
2132 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2133 {
2134         struct scrub_ctx *sctx = sblock->sctx;
2135         struct btrfs_fs_info *fs_info = sctx->fs_info;
2136         u64 length = sblock->page_count * PAGE_SIZE;
2137         u64 logical = sblock->pagev[0]->logical;
2138         struct btrfs_bio *bbio = NULL;
2139         struct bio *bio;
2140         struct btrfs_raid_bio *rbio;
2141         int ret;
2142         int i;
2143
2144         btrfs_bio_counter_inc_blocked(fs_info);
2145         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2146                         &length, &bbio);
2147         if (ret || !bbio || !bbio->raid_map)
2148                 goto bbio_out;
2149
2150         if (WARN_ON(!sctx->is_dev_replace ||
2151                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2152                 /*
2153                  * We shouldn't be scrubbing a missing device. Even for dev
2154                  * replace, we should only get here for RAID 5/6. We either
2155                  * managed to mount something with no mirrors remaining or
2156                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2157                  */
2158                 goto bbio_out;
2159         }
2160
2161         bio = btrfs_io_bio_alloc(0);
2162         bio->bi_iter.bi_sector = logical >> 9;
2163         bio->bi_private = sblock;
2164         bio->bi_end_io = scrub_missing_raid56_end_io;
2165
2166         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2167         if (!rbio)
2168                 goto rbio_out;
2169
2170         for (i = 0; i < sblock->page_count; i++) {
2171                 struct scrub_page *spage = sblock->pagev[i];
2172
2173                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2174         }
2175
2176         btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2177         scrub_block_get(sblock);
2178         scrub_pending_bio_inc(sctx);
2179         raid56_submit_missing_rbio(rbio);
2180         return;
2181
2182 rbio_out:
2183         bio_put(bio);
2184 bbio_out:
2185         btrfs_bio_counter_dec(fs_info);
2186         btrfs_put_bbio(bbio);
2187         spin_lock(&sctx->stat_lock);
2188         sctx->stat.malloc_errors++;
2189         spin_unlock(&sctx->stat_lock);
2190 }
2191
2192 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u32 len,
2193                        u64 physical, struct btrfs_device *dev, u64 flags,
2194                        u64 gen, int mirror_num, u8 *csum,
2195                        u64 physical_for_dev_replace)
2196 {
2197         struct scrub_block *sblock;
2198         const u32 sectorsize = sctx->fs_info->sectorsize;
2199         int index;
2200
2201         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2202         if (!sblock) {
2203                 spin_lock(&sctx->stat_lock);
2204                 sctx->stat.malloc_errors++;
2205                 spin_unlock(&sctx->stat_lock);
2206                 return -ENOMEM;
2207         }
2208
2209         /* one ref inside this function, plus one for each page added to
2210          * a bio later on */
2211         refcount_set(&sblock->refs, 1);
2212         sblock->sctx = sctx;
2213         sblock->no_io_error_seen = 1;
2214
2215         for (index = 0; len > 0; index++) {
2216                 struct scrub_page *spage;
2217                 /*
2218                  * Here we will allocate one page for one sector to scrub.
2219                  * This is fine if PAGE_SIZE == sectorsize, but will cost
2220                  * more memory for PAGE_SIZE > sectorsize case.
2221                  */
2222                 u32 l = min(sectorsize, len);
2223
2224                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2225                 if (!spage) {
2226 leave_nomem:
2227                         spin_lock(&sctx->stat_lock);
2228                         sctx->stat.malloc_errors++;
2229                         spin_unlock(&sctx->stat_lock);
2230                         scrub_block_put(sblock);
2231                         return -ENOMEM;
2232                 }
2233                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2234                 scrub_page_get(spage);
2235                 sblock->pagev[index] = spage;
2236                 spage->sblock = sblock;
2237                 spage->dev = dev;
2238                 spage->flags = flags;
2239                 spage->generation = gen;
2240                 spage->logical = logical;
2241                 spage->physical = physical;
2242                 spage->physical_for_dev_replace = physical_for_dev_replace;
2243                 spage->mirror_num = mirror_num;
2244                 if (csum) {
2245                         spage->have_csum = 1;
2246                         memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2247                 } else {
2248                         spage->have_csum = 0;
2249                 }
2250                 sblock->page_count++;
2251                 spage->page = alloc_page(GFP_KERNEL);
2252                 if (!spage->page)
2253                         goto leave_nomem;
2254                 len -= l;
2255                 logical += l;
2256                 physical += l;
2257                 physical_for_dev_replace += l;
2258         }
2259
2260         WARN_ON(sblock->page_count == 0);
2261         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2262                 /*
2263                  * This case should only be hit for RAID 5/6 device replace. See
2264                  * the comment in scrub_missing_raid56_pages() for details.
2265                  */
2266                 scrub_missing_raid56_pages(sblock);
2267         } else {
2268                 for (index = 0; index < sblock->page_count; index++) {
2269                         struct scrub_page *spage = sblock->pagev[index];
2270                         int ret;
2271
2272                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2273                         if (ret) {
2274                                 scrub_block_put(sblock);
2275                                 return ret;
2276                         }
2277                 }
2278
2279                 if (flags & BTRFS_EXTENT_FLAG_SUPER)
2280                         scrub_submit(sctx);
2281         }
2282
2283         /* last one frees, either here or in bio completion for last page */
2284         scrub_block_put(sblock);
2285         return 0;
2286 }
2287
2288 static void scrub_bio_end_io(struct bio *bio)
2289 {
2290         struct scrub_bio *sbio = bio->bi_private;
2291         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2292
2293         sbio->status = bio->bi_status;
2294         sbio->bio = bio;
2295
2296         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2297 }
2298
2299 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2300 {
2301         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2302         struct scrub_ctx *sctx = sbio->sctx;
2303         int i;
2304
2305         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2306         if (sbio->status) {
2307                 for (i = 0; i < sbio->page_count; i++) {
2308                         struct scrub_page *spage = sbio->pagev[i];
2309
2310                         spage->io_error = 1;
2311                         spage->sblock->no_io_error_seen = 0;
2312                 }
2313         }
2314
2315         /* now complete the scrub_block items that have all pages completed */
2316         for (i = 0; i < sbio->page_count; i++) {
2317                 struct scrub_page *spage = sbio->pagev[i];
2318                 struct scrub_block *sblock = spage->sblock;
2319
2320                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2321                         scrub_block_complete(sblock);
2322                 scrub_block_put(sblock);
2323         }
2324
2325         bio_put(sbio->bio);
2326         sbio->bio = NULL;
2327         spin_lock(&sctx->list_lock);
2328         sbio->next_free = sctx->first_free;
2329         sctx->first_free = sbio->index;
2330         spin_unlock(&sctx->list_lock);
2331
2332         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2333                 mutex_lock(&sctx->wr_lock);
2334                 scrub_wr_submit(sctx);
2335                 mutex_unlock(&sctx->wr_lock);
2336         }
2337
2338         scrub_pending_bio_dec(sctx);
2339 }
2340
2341 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2342                                        unsigned long *bitmap,
2343                                        u64 start, u32 len)
2344 {
2345         u64 offset;
2346         u32 nsectors;
2347         u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2348
2349         if (len >= sparity->stripe_len) {
2350                 bitmap_set(bitmap, 0, sparity->nsectors);
2351                 return;
2352         }
2353
2354         start -= sparity->logic_start;
2355         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2356         offset = offset >> sectorsize_bits;
2357         nsectors = len >> sectorsize_bits;
2358
2359         if (offset + nsectors <= sparity->nsectors) {
2360                 bitmap_set(bitmap, offset, nsectors);
2361                 return;
2362         }
2363
2364         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2365         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2366 }
2367
2368 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2369                                                    u64 start, u32 len)
2370 {
2371         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2372 }
2373
2374 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2375                                                   u64 start, u32 len)
2376 {
2377         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2378 }
2379
2380 static void scrub_block_complete(struct scrub_block *sblock)
2381 {
2382         int corrupted = 0;
2383
2384         if (!sblock->no_io_error_seen) {
2385                 corrupted = 1;
2386                 scrub_handle_errored_block(sblock);
2387         } else {
2388                 /*
2389                  * if has checksum error, write via repair mechanism in
2390                  * dev replace case, otherwise write here in dev replace
2391                  * case.
2392                  */
2393                 corrupted = scrub_checksum(sblock);
2394                 if (!corrupted && sblock->sctx->is_dev_replace)
2395                         scrub_write_block_to_dev_replace(sblock);
2396         }
2397
2398         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2399                 u64 start = sblock->pagev[0]->logical;
2400                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2401                           PAGE_SIZE;
2402
2403                 ASSERT(end - start <= U32_MAX);
2404                 scrub_parity_mark_sectors_error(sblock->sparity,
2405                                                 start, end - start);
2406         }
2407 }
2408
2409 static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2410 {
2411         sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2412         list_del(&sum->list);
2413         kfree(sum);
2414 }
2415
2416 /*
2417  * Find the desired csum for range [logical, logical + sectorsize), and store
2418  * the csum into @csum.
2419  *
2420  * The search source is sctx->csum_list, which is a pre-populated list
2421  * storing bytenr ordered csum ranges.  We're reponsible to cleanup any range
2422  * that is before @logical.
2423  *
2424  * Return 0 if there is no csum for the range.
2425  * Return 1 if there is csum for the range and copied to @csum.
2426  */
2427 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2428 {
2429         bool found = false;
2430
2431         while (!list_empty(&sctx->csum_list)) {
2432                 struct btrfs_ordered_sum *sum = NULL;
2433                 unsigned long index;
2434                 unsigned long num_sectors;
2435
2436                 sum = list_first_entry(&sctx->csum_list,
2437                                        struct btrfs_ordered_sum, list);
2438                 /* The current csum range is beyond our range, no csum found */
2439                 if (sum->bytenr > logical)
2440                         break;
2441
2442                 /*
2443                  * The current sum is before our bytenr, since scrub is always
2444                  * done in bytenr order, the csum will never be used anymore,
2445                  * clean it up so that later calls won't bother with the range,
2446                  * and continue search the next range.
2447                  */
2448                 if (sum->bytenr + sum->len <= logical) {
2449                         drop_csum_range(sctx, sum);
2450                         continue;
2451                 }
2452
2453                 /* Now the csum range covers our bytenr, copy the csum */
2454                 found = true;
2455                 index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2456                 num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2457
2458                 memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2459                        sctx->fs_info->csum_size);
2460
2461                 /* Cleanup the range if we're at the end of the csum range */
2462                 if (index == num_sectors - 1)
2463                         drop_csum_range(sctx, sum);
2464                 break;
2465         }
2466         if (!found)
2467                 return 0;
2468         return 1;
2469 }
2470
2471 /* scrub extent tries to collect up to 64 kB for each bio */
2472 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2473                         u64 logical, u32 len,
2474                         u64 physical, struct btrfs_device *dev, u64 flags,
2475                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2476 {
2477         int ret;
2478         u8 csum[BTRFS_CSUM_SIZE];
2479         u32 blocksize;
2480
2481         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2482                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2483                         blocksize = map->stripe_len;
2484                 else
2485                         blocksize = sctx->fs_info->sectorsize;
2486                 spin_lock(&sctx->stat_lock);
2487                 sctx->stat.data_extents_scrubbed++;
2488                 sctx->stat.data_bytes_scrubbed += len;
2489                 spin_unlock(&sctx->stat_lock);
2490         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2491                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2492                         blocksize = map->stripe_len;
2493                 else
2494                         blocksize = sctx->fs_info->nodesize;
2495                 spin_lock(&sctx->stat_lock);
2496                 sctx->stat.tree_extents_scrubbed++;
2497                 sctx->stat.tree_bytes_scrubbed += len;
2498                 spin_unlock(&sctx->stat_lock);
2499         } else {
2500                 blocksize = sctx->fs_info->sectorsize;
2501                 WARN_ON(1);
2502         }
2503
2504         while (len) {
2505                 u32 l = min(len, blocksize);
2506                 int have_csum = 0;
2507
2508                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2509                         /* push csums to sbio */
2510                         have_csum = scrub_find_csum(sctx, logical, csum);
2511                         if (have_csum == 0)
2512                                 ++sctx->stat.no_csum;
2513                 }
2514                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2515                                   mirror_num, have_csum ? csum : NULL,
2516                                   physical_for_dev_replace);
2517                 if (ret)
2518                         return ret;
2519                 len -= l;
2520                 logical += l;
2521                 physical += l;
2522                 physical_for_dev_replace += l;
2523         }
2524         return 0;
2525 }
2526
2527 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2528                                   u64 logical, u32 len,
2529                                   u64 physical, struct btrfs_device *dev,
2530                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2531 {
2532         struct scrub_ctx *sctx = sparity->sctx;
2533         struct scrub_block *sblock;
2534         const u32 sectorsize = sctx->fs_info->sectorsize;
2535         int index;
2536
2537         ASSERT(IS_ALIGNED(len, sectorsize));
2538
2539         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2540         if (!sblock) {
2541                 spin_lock(&sctx->stat_lock);
2542                 sctx->stat.malloc_errors++;
2543                 spin_unlock(&sctx->stat_lock);
2544                 return -ENOMEM;
2545         }
2546
2547         /* one ref inside this function, plus one for each page added to
2548          * a bio later on */
2549         refcount_set(&sblock->refs, 1);
2550         sblock->sctx = sctx;
2551         sblock->no_io_error_seen = 1;
2552         sblock->sparity = sparity;
2553         scrub_parity_get(sparity);
2554
2555         for (index = 0; len > 0; index++) {
2556                 struct scrub_page *spage;
2557
2558                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2559                 if (!spage) {
2560 leave_nomem:
2561                         spin_lock(&sctx->stat_lock);
2562                         sctx->stat.malloc_errors++;
2563                         spin_unlock(&sctx->stat_lock);
2564                         scrub_block_put(sblock);
2565                         return -ENOMEM;
2566                 }
2567                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2568                 /* For scrub block */
2569                 scrub_page_get(spage);
2570                 sblock->pagev[index] = spage;
2571                 /* For scrub parity */
2572                 scrub_page_get(spage);
2573                 list_add_tail(&spage->list, &sparity->spages);
2574                 spage->sblock = sblock;
2575                 spage->dev = dev;
2576                 spage->flags = flags;
2577                 spage->generation = gen;
2578                 spage->logical = logical;
2579                 spage->physical = physical;
2580                 spage->mirror_num = mirror_num;
2581                 if (csum) {
2582                         spage->have_csum = 1;
2583                         memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2584                 } else {
2585                         spage->have_csum = 0;
2586                 }
2587                 sblock->page_count++;
2588                 spage->page = alloc_page(GFP_KERNEL);
2589                 if (!spage->page)
2590                         goto leave_nomem;
2591
2592
2593                 /* Iterate over the stripe range in sectorsize steps */
2594                 len -= sectorsize;
2595                 logical += sectorsize;
2596                 physical += sectorsize;
2597         }
2598
2599         WARN_ON(sblock->page_count == 0);
2600         for (index = 0; index < sblock->page_count; index++) {
2601                 struct scrub_page *spage = sblock->pagev[index];
2602                 int ret;
2603
2604                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2605                 if (ret) {
2606                         scrub_block_put(sblock);
2607                         return ret;
2608                 }
2609         }
2610
2611         /* last one frees, either here or in bio completion for last page */
2612         scrub_block_put(sblock);
2613         return 0;
2614 }
2615
2616 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2617                                    u64 logical, u32 len,
2618                                    u64 physical, struct btrfs_device *dev,
2619                                    u64 flags, u64 gen, int mirror_num)
2620 {
2621         struct scrub_ctx *sctx = sparity->sctx;
2622         int ret;
2623         u8 csum[BTRFS_CSUM_SIZE];
2624         u32 blocksize;
2625
2626         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2627                 scrub_parity_mark_sectors_error(sparity, logical, len);
2628                 return 0;
2629         }
2630
2631         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2632                 blocksize = sparity->stripe_len;
2633         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2634                 blocksize = sparity->stripe_len;
2635         } else {
2636                 blocksize = sctx->fs_info->sectorsize;
2637                 WARN_ON(1);
2638         }
2639
2640         while (len) {
2641                 u32 l = min(len, blocksize);
2642                 int have_csum = 0;
2643
2644                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2645                         /* push csums to sbio */
2646                         have_csum = scrub_find_csum(sctx, logical, csum);
2647                         if (have_csum == 0)
2648                                 goto skip;
2649                 }
2650                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2651                                              flags, gen, mirror_num,
2652                                              have_csum ? csum : NULL);
2653                 if (ret)
2654                         return ret;
2655 skip:
2656                 len -= l;
2657                 logical += l;
2658                 physical += l;
2659         }
2660         return 0;
2661 }
2662
2663 /*
2664  * Given a physical address, this will calculate it's
2665  * logical offset. if this is a parity stripe, it will return
2666  * the most left data stripe's logical offset.
2667  *
2668  * return 0 if it is a data stripe, 1 means parity stripe.
2669  */
2670 static int get_raid56_logic_offset(u64 physical, int num,
2671                                    struct map_lookup *map, u64 *offset,
2672                                    u64 *stripe_start)
2673 {
2674         int i;
2675         int j = 0;
2676         u64 stripe_nr;
2677         u64 last_offset;
2678         u32 stripe_index;
2679         u32 rot;
2680         const int data_stripes = nr_data_stripes(map);
2681
2682         last_offset = (physical - map->stripes[num].physical) * data_stripes;
2683         if (stripe_start)
2684                 *stripe_start = last_offset;
2685
2686         *offset = last_offset;
2687         for (i = 0; i < data_stripes; i++) {
2688                 *offset = last_offset + i * map->stripe_len;
2689
2690                 stripe_nr = div64_u64(*offset, map->stripe_len);
2691                 stripe_nr = div_u64(stripe_nr, data_stripes);
2692
2693                 /* Work out the disk rotation on this stripe-set */
2694                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2695                 /* calculate which stripe this data locates */
2696                 rot += i;
2697                 stripe_index = rot % map->num_stripes;
2698                 if (stripe_index == num)
2699                         return 0;
2700                 if (stripe_index < num)
2701                         j++;
2702         }
2703         *offset = last_offset + j * map->stripe_len;
2704         return 1;
2705 }
2706
2707 static void scrub_free_parity(struct scrub_parity *sparity)
2708 {
2709         struct scrub_ctx *sctx = sparity->sctx;
2710         struct scrub_page *curr, *next;
2711         int nbits;
2712
2713         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2714         if (nbits) {
2715                 spin_lock(&sctx->stat_lock);
2716                 sctx->stat.read_errors += nbits;
2717                 sctx->stat.uncorrectable_errors += nbits;
2718                 spin_unlock(&sctx->stat_lock);
2719         }
2720
2721         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2722                 list_del_init(&curr->list);
2723                 scrub_page_put(curr);
2724         }
2725
2726         kfree(sparity);
2727 }
2728
2729 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2730 {
2731         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2732                                                     work);
2733         struct scrub_ctx *sctx = sparity->sctx;
2734
2735         scrub_free_parity(sparity);
2736         scrub_pending_bio_dec(sctx);
2737 }
2738
2739 static void scrub_parity_bio_endio(struct bio *bio)
2740 {
2741         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2742         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2743
2744         if (bio->bi_status)
2745                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2746                           sparity->nsectors);
2747
2748         bio_put(bio);
2749
2750         btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2751                         NULL);
2752         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2753 }
2754
2755 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2756 {
2757         struct scrub_ctx *sctx = sparity->sctx;
2758         struct btrfs_fs_info *fs_info = sctx->fs_info;
2759         struct bio *bio;
2760         struct btrfs_raid_bio *rbio;
2761         struct btrfs_bio *bbio = NULL;
2762         u64 length;
2763         int ret;
2764
2765         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2766                            sparity->nsectors))
2767                 goto out;
2768
2769         length = sparity->logic_end - sparity->logic_start;
2770
2771         btrfs_bio_counter_inc_blocked(fs_info);
2772         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2773                                &length, &bbio);
2774         if (ret || !bbio || !bbio->raid_map)
2775                 goto bbio_out;
2776
2777         bio = btrfs_io_bio_alloc(0);
2778         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2779         bio->bi_private = sparity;
2780         bio->bi_end_io = scrub_parity_bio_endio;
2781
2782         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2783                                               length, sparity->scrub_dev,
2784                                               sparity->dbitmap,
2785                                               sparity->nsectors);
2786         if (!rbio)
2787                 goto rbio_out;
2788
2789         scrub_pending_bio_inc(sctx);
2790         raid56_parity_submit_scrub_rbio(rbio);
2791         return;
2792
2793 rbio_out:
2794         bio_put(bio);
2795 bbio_out:
2796         btrfs_bio_counter_dec(fs_info);
2797         btrfs_put_bbio(bbio);
2798         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2799                   sparity->nsectors);
2800         spin_lock(&sctx->stat_lock);
2801         sctx->stat.malloc_errors++;
2802         spin_unlock(&sctx->stat_lock);
2803 out:
2804         scrub_free_parity(sparity);
2805 }
2806
2807 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2808 {
2809         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2810 }
2811
2812 static void scrub_parity_get(struct scrub_parity *sparity)
2813 {
2814         refcount_inc(&sparity->refs);
2815 }
2816
2817 static void scrub_parity_put(struct scrub_parity *sparity)
2818 {
2819         if (!refcount_dec_and_test(&sparity->refs))
2820                 return;
2821
2822         scrub_parity_check_and_repair(sparity);
2823 }
2824
2825 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2826                                                   struct map_lookup *map,
2827                                                   struct btrfs_device *sdev,
2828                                                   struct btrfs_path *path,
2829                                                   u64 logic_start,
2830                                                   u64 logic_end)
2831 {
2832         struct btrfs_fs_info *fs_info = sctx->fs_info;
2833         struct btrfs_root *root = fs_info->extent_root;
2834         struct btrfs_root *csum_root = fs_info->csum_root;
2835         struct btrfs_extent_item *extent;
2836         struct btrfs_bio *bbio = NULL;
2837         u64 flags;
2838         int ret;
2839         int slot;
2840         struct extent_buffer *l;
2841         struct btrfs_key key;
2842         u64 generation;
2843         u64 extent_logical;
2844         u64 extent_physical;
2845         /* Check the comment in scrub_stripe() for why u32 is enough here */
2846         u32 extent_len;
2847         u64 mapped_length;
2848         struct btrfs_device *extent_dev;
2849         struct scrub_parity *sparity;
2850         int nsectors;
2851         int bitmap_len;
2852         int extent_mirror_num;
2853         int stop_loop = 0;
2854
2855         ASSERT(map->stripe_len <= U32_MAX);
2856         nsectors = map->stripe_len >> fs_info->sectorsize_bits;
2857         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2858         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2859                           GFP_NOFS);
2860         if (!sparity) {
2861                 spin_lock(&sctx->stat_lock);
2862                 sctx->stat.malloc_errors++;
2863                 spin_unlock(&sctx->stat_lock);
2864                 return -ENOMEM;
2865         }
2866
2867         ASSERT(map->stripe_len <= U32_MAX);
2868         sparity->stripe_len = map->stripe_len;
2869         sparity->nsectors = nsectors;
2870         sparity->sctx = sctx;
2871         sparity->scrub_dev = sdev;
2872         sparity->logic_start = logic_start;
2873         sparity->logic_end = logic_end;
2874         refcount_set(&sparity->refs, 1);
2875         INIT_LIST_HEAD(&sparity->spages);
2876         sparity->dbitmap = sparity->bitmap;
2877         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2878
2879         ret = 0;
2880         while (logic_start < logic_end) {
2881                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2882                         key.type = BTRFS_METADATA_ITEM_KEY;
2883                 else
2884                         key.type = BTRFS_EXTENT_ITEM_KEY;
2885                 key.objectid = logic_start;
2886                 key.offset = (u64)-1;
2887
2888                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2889                 if (ret < 0)
2890                         goto out;
2891
2892                 if (ret > 0) {
2893                         ret = btrfs_previous_extent_item(root, path, 0);
2894                         if (ret < 0)
2895                                 goto out;
2896                         if (ret > 0) {
2897                                 btrfs_release_path(path);
2898                                 ret = btrfs_search_slot(NULL, root, &key,
2899                                                         path, 0, 0);
2900                                 if (ret < 0)
2901                                         goto out;
2902                         }
2903                 }
2904
2905                 stop_loop = 0;
2906                 while (1) {
2907                         u64 bytes;
2908
2909                         l = path->nodes[0];
2910                         slot = path->slots[0];
2911                         if (slot >= btrfs_header_nritems(l)) {
2912                                 ret = btrfs_next_leaf(root, path);
2913                                 if (ret == 0)
2914                                         continue;
2915                                 if (ret < 0)
2916                                         goto out;
2917
2918                                 stop_loop = 1;
2919                                 break;
2920                         }
2921                         btrfs_item_key_to_cpu(l, &key, slot);
2922
2923                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2924                             key.type != BTRFS_METADATA_ITEM_KEY)
2925                                 goto next;
2926
2927                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2928                                 bytes = fs_info->nodesize;
2929                         else
2930                                 bytes = key.offset;
2931
2932                         if (key.objectid + bytes <= logic_start)
2933                                 goto next;
2934
2935                         if (key.objectid >= logic_end) {
2936                                 stop_loop = 1;
2937                                 break;
2938                         }
2939
2940                         while (key.objectid >= logic_start + map->stripe_len)
2941                                 logic_start += map->stripe_len;
2942
2943                         extent = btrfs_item_ptr(l, slot,
2944                                                 struct btrfs_extent_item);
2945                         flags = btrfs_extent_flags(l, extent);
2946                         generation = btrfs_extent_generation(l, extent);
2947
2948                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2949                             (key.objectid < logic_start ||
2950                              key.objectid + bytes >
2951                              logic_start + map->stripe_len)) {
2952                                 btrfs_err(fs_info,
2953                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2954                                           key.objectid, logic_start);
2955                                 spin_lock(&sctx->stat_lock);
2956                                 sctx->stat.uncorrectable_errors++;
2957                                 spin_unlock(&sctx->stat_lock);
2958                                 goto next;
2959                         }
2960 again:
2961                         extent_logical = key.objectid;
2962                         ASSERT(bytes <= U32_MAX);
2963                         extent_len = bytes;
2964
2965                         if (extent_logical < logic_start) {
2966                                 extent_len -= logic_start - extent_logical;
2967                                 extent_logical = logic_start;
2968                         }
2969
2970                         if (extent_logical + extent_len >
2971                             logic_start + map->stripe_len)
2972                                 extent_len = logic_start + map->stripe_len -
2973                                              extent_logical;
2974
2975                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2976                                                        extent_len);
2977
2978                         mapped_length = extent_len;
2979                         bbio = NULL;
2980                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2981                                         extent_logical, &mapped_length, &bbio,
2982                                         0);
2983                         if (!ret) {
2984                                 if (!bbio || mapped_length < extent_len)
2985                                         ret = -EIO;
2986                         }
2987                         if (ret) {
2988                                 btrfs_put_bbio(bbio);
2989                                 goto out;
2990                         }
2991                         extent_physical = bbio->stripes[0].physical;
2992                         extent_mirror_num = bbio->mirror_num;
2993                         extent_dev = bbio->stripes[0].dev;
2994                         btrfs_put_bbio(bbio);
2995
2996                         ret = btrfs_lookup_csums_range(csum_root,
2997                                                 extent_logical,
2998                                                 extent_logical + extent_len - 1,
2999                                                 &sctx->csum_list, 1);
3000                         if (ret)
3001                                 goto out;
3002
3003                         ret = scrub_extent_for_parity(sparity, extent_logical,
3004                                                       extent_len,
3005                                                       extent_physical,
3006                                                       extent_dev, flags,
3007                                                       generation,
3008                                                       extent_mirror_num);
3009
3010                         scrub_free_csums(sctx);
3011
3012                         if (ret)
3013                                 goto out;
3014
3015                         if (extent_logical + extent_len <
3016                             key.objectid + bytes) {
3017                                 logic_start += map->stripe_len;
3018
3019                                 if (logic_start >= logic_end) {
3020                                         stop_loop = 1;
3021                                         break;
3022                                 }
3023
3024                                 if (logic_start < key.objectid + bytes) {
3025                                         cond_resched();
3026                                         goto again;
3027                                 }
3028                         }
3029 next:
3030                         path->slots[0]++;
3031                 }
3032
3033                 btrfs_release_path(path);
3034
3035                 if (stop_loop)
3036                         break;
3037
3038                 logic_start += map->stripe_len;
3039         }
3040 out:
3041         if (ret < 0) {
3042                 ASSERT(logic_end - logic_start <= U32_MAX);
3043                 scrub_parity_mark_sectors_error(sparity, logic_start,
3044                                                 logic_end - logic_start);
3045         }
3046         scrub_parity_put(sparity);
3047         scrub_submit(sctx);
3048         mutex_lock(&sctx->wr_lock);
3049         scrub_wr_submit(sctx);
3050         mutex_unlock(&sctx->wr_lock);
3051
3052         btrfs_release_path(path);
3053         return ret < 0 ? ret : 0;
3054 }
3055
3056 static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3057 {
3058         if (!btrfs_is_zoned(sctx->fs_info))
3059                 return;
3060
3061         sctx->flush_all_writes = true;
3062         scrub_submit(sctx);
3063         mutex_lock(&sctx->wr_lock);
3064         scrub_wr_submit(sctx);
3065         mutex_unlock(&sctx->wr_lock);
3066
3067         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3068 }
3069
3070 static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3071                                         u64 physical, u64 physical_end)
3072 {
3073         struct btrfs_fs_info *fs_info = sctx->fs_info;
3074         int ret = 0;
3075
3076         if (!btrfs_is_zoned(fs_info))
3077                 return 0;
3078
3079         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3080
3081         mutex_lock(&sctx->wr_lock);
3082         if (sctx->write_pointer < physical_end) {
3083                 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3084                                                     physical,
3085                                                     sctx->write_pointer);
3086                 if (ret)
3087                         btrfs_err(fs_info,
3088                                   "zoned: failed to recover write pointer");
3089         }
3090         mutex_unlock(&sctx->wr_lock);
3091         btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3092
3093         return ret;
3094 }
3095
3096 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3097                                            struct map_lookup *map,
3098                                            struct btrfs_device *scrub_dev,
3099                                            int num, u64 base, u64 length,
3100                                            struct btrfs_block_group *cache)
3101 {
3102         struct btrfs_path *path, *ppath;
3103         struct btrfs_fs_info *fs_info = sctx->fs_info;
3104         struct btrfs_root *root = fs_info->extent_root;
3105         struct btrfs_root *csum_root = fs_info->csum_root;
3106         struct btrfs_extent_item *extent;
3107         struct blk_plug plug;
3108         u64 flags;
3109         int ret;
3110         int slot;
3111         u64 nstripes;
3112         struct extent_buffer *l;
3113         u64 physical;
3114         u64 logical;
3115         u64 logic_end;
3116         u64 physical_end;
3117         u64 generation;
3118         int mirror_num;
3119         struct reada_control *reada1;
3120         struct reada_control *reada2;
3121         struct btrfs_key key;
3122         struct btrfs_key key_end;
3123         u64 increment = map->stripe_len;
3124         u64 offset;
3125         u64 extent_logical;
3126         u64 extent_physical;
3127         /*
3128          * Unlike chunk length, extent length should never go beyond
3129          * BTRFS_MAX_EXTENT_SIZE, thus u32 is enough here.
3130          */
3131         u32 extent_len;
3132         u64 stripe_logical;
3133         u64 stripe_end;
3134         struct btrfs_device *extent_dev;
3135         int extent_mirror_num;
3136         int stop_loop = 0;
3137
3138         physical = map->stripes[num].physical;
3139         offset = 0;
3140         nstripes = div64_u64(length, map->stripe_len);
3141         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3142                 offset = map->stripe_len * num;
3143                 increment = map->stripe_len * map->num_stripes;
3144                 mirror_num = 1;
3145         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3146                 int factor = map->num_stripes / map->sub_stripes;
3147                 offset = map->stripe_len * (num / map->sub_stripes);
3148                 increment = map->stripe_len * factor;
3149                 mirror_num = num % map->sub_stripes + 1;
3150         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3151                 increment = map->stripe_len;
3152                 mirror_num = num % map->num_stripes + 1;
3153         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3154                 increment = map->stripe_len;
3155                 mirror_num = num % map->num_stripes + 1;
3156         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3157                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3158                 increment = map->stripe_len * nr_data_stripes(map);
3159                 mirror_num = 1;
3160         } else {
3161                 increment = map->stripe_len;
3162                 mirror_num = 1;
3163         }
3164
3165         path = btrfs_alloc_path();
3166         if (!path)
3167                 return -ENOMEM;
3168
3169         ppath = btrfs_alloc_path();
3170         if (!ppath) {
3171                 btrfs_free_path(path);
3172                 return -ENOMEM;
3173         }
3174
3175         /*
3176          * work on commit root. The related disk blocks are static as
3177          * long as COW is applied. This means, it is save to rewrite
3178          * them to repair disk errors without any race conditions
3179          */
3180         path->search_commit_root = 1;
3181         path->skip_locking = 1;
3182
3183         ppath->search_commit_root = 1;
3184         ppath->skip_locking = 1;
3185         /*
3186          * trigger the readahead for extent tree csum tree and wait for
3187          * completion. During readahead, the scrub is officially paused
3188          * to not hold off transaction commits
3189          */
3190         logical = base + offset;
3191         physical_end = physical + nstripes * map->stripe_len;
3192         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3193                 get_raid56_logic_offset(physical_end, num,
3194                                         map, &logic_end, NULL);
3195                 logic_end += base;
3196         } else {
3197                 logic_end = logical + increment * nstripes;
3198         }
3199         wait_event(sctx->list_wait,
3200                    atomic_read(&sctx->bios_in_flight) == 0);
3201         scrub_blocked_if_needed(fs_info);
3202
3203         /* FIXME it might be better to start readahead at commit root */
3204         key.objectid = logical;
3205         key.type = BTRFS_EXTENT_ITEM_KEY;
3206         key.offset = (u64)0;
3207         key_end.objectid = logic_end;
3208         key_end.type = BTRFS_METADATA_ITEM_KEY;
3209         key_end.offset = (u64)-1;
3210         reada1 = btrfs_reada_add(root, &key, &key_end);
3211
3212         if (cache->flags & BTRFS_BLOCK_GROUP_DATA) {
3213                 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3214                 key.type = BTRFS_EXTENT_CSUM_KEY;
3215                 key.offset = logical;
3216                 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3217                 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3218                 key_end.offset = logic_end;
3219                 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3220         } else {
3221                 reada2 = NULL;
3222         }
3223
3224         if (!IS_ERR(reada1))
3225                 btrfs_reada_wait(reada1);
3226         if (!IS_ERR_OR_NULL(reada2))
3227                 btrfs_reada_wait(reada2);
3228
3229
3230         /*
3231          * collect all data csums for the stripe to avoid seeking during
3232          * the scrub. This might currently (crc32) end up to be about 1MB
3233          */
3234         blk_start_plug(&plug);
3235
3236         if (sctx->is_dev_replace &&
3237             btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3238                 mutex_lock(&sctx->wr_lock);
3239                 sctx->write_pointer = physical;
3240                 mutex_unlock(&sctx->wr_lock);
3241                 sctx->flush_all_writes = true;
3242         }
3243
3244         /*
3245          * now find all extents for each stripe and scrub them
3246          */
3247         ret = 0;
3248         while (physical < physical_end) {
3249                 /*
3250                  * canceled?
3251                  */
3252                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3253                     atomic_read(&sctx->cancel_req)) {
3254                         ret = -ECANCELED;
3255                         goto out;
3256                 }
3257                 /*
3258                  * check to see if we have to pause
3259                  */
3260                 if (atomic_read(&fs_info->scrub_pause_req)) {
3261                         /* push queued extents */
3262                         sctx->flush_all_writes = true;
3263                         scrub_submit(sctx);
3264                         mutex_lock(&sctx->wr_lock);
3265                         scrub_wr_submit(sctx);
3266                         mutex_unlock(&sctx->wr_lock);
3267                         wait_event(sctx->list_wait,
3268                                    atomic_read(&sctx->bios_in_flight) == 0);
3269                         sctx->flush_all_writes = false;
3270                         scrub_blocked_if_needed(fs_info);
3271                 }
3272
3273                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3274                         ret = get_raid56_logic_offset(physical, num, map,
3275                                                       &logical,
3276                                                       &stripe_logical);
3277                         logical += base;
3278                         if (ret) {
3279                                 /* it is parity strip */
3280                                 stripe_logical += base;
3281                                 stripe_end = stripe_logical + increment;
3282                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3283                                                           ppath, stripe_logical,
3284                                                           stripe_end);
3285                                 if (ret)
3286                                         goto out;
3287                                 goto skip;
3288                         }
3289                 }
3290
3291                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3292                         key.type = BTRFS_METADATA_ITEM_KEY;
3293                 else
3294                         key.type = BTRFS_EXTENT_ITEM_KEY;
3295                 key.objectid = logical;
3296                 key.offset = (u64)-1;
3297
3298                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3299                 if (ret < 0)
3300                         goto out;
3301
3302                 if (ret > 0) {
3303                         ret = btrfs_previous_extent_item(root, path, 0);
3304                         if (ret < 0)
3305                                 goto out;
3306                         if (ret > 0) {
3307                                 /* there's no smaller item, so stick with the
3308                                  * larger one */
3309                                 btrfs_release_path(path);
3310                                 ret = btrfs_search_slot(NULL, root, &key,
3311                                                         path, 0, 0);
3312                                 if (ret < 0)
3313                                         goto out;
3314                         }
3315                 }
3316
3317                 stop_loop = 0;
3318                 while (1) {
3319                         u64 bytes;
3320
3321                         l = path->nodes[0];
3322                         slot = path->slots[0];
3323                         if (slot >= btrfs_header_nritems(l)) {
3324                                 ret = btrfs_next_leaf(root, path);
3325                                 if (ret == 0)
3326                                         continue;
3327                                 if (ret < 0)
3328                                         goto out;
3329
3330                                 stop_loop = 1;
3331                                 break;
3332                         }
3333                         btrfs_item_key_to_cpu(l, &key, slot);
3334
3335                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3336                             key.type != BTRFS_METADATA_ITEM_KEY)
3337                                 goto next;
3338
3339                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3340                                 bytes = fs_info->nodesize;
3341                         else
3342                                 bytes = key.offset;
3343
3344                         if (key.objectid + bytes <= logical)
3345                                 goto next;
3346
3347                         if (key.objectid >= logical + map->stripe_len) {
3348                                 /* out of this device extent */
3349                                 if (key.objectid >= logic_end)
3350                                         stop_loop = 1;
3351                                 break;
3352                         }
3353
3354                         /*
3355                          * If our block group was removed in the meanwhile, just
3356                          * stop scrubbing since there is no point in continuing.
3357                          * Continuing would prevent reusing its device extents
3358                          * for new block groups for a long time.
3359                          */
3360                         spin_lock(&cache->lock);
3361                         if (cache->removed) {
3362                                 spin_unlock(&cache->lock);
3363                                 ret = 0;
3364                                 goto out;
3365                         }
3366                         spin_unlock(&cache->lock);
3367
3368                         extent = btrfs_item_ptr(l, slot,
3369                                                 struct btrfs_extent_item);
3370                         flags = btrfs_extent_flags(l, extent);
3371                         generation = btrfs_extent_generation(l, extent);
3372
3373                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3374                             (key.objectid < logical ||
3375                              key.objectid + bytes >
3376                              logical + map->stripe_len)) {
3377                                 btrfs_err(fs_info,
3378                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3379                                        key.objectid, logical);
3380                                 spin_lock(&sctx->stat_lock);
3381                                 sctx->stat.uncorrectable_errors++;
3382                                 spin_unlock(&sctx->stat_lock);
3383                                 goto next;
3384                         }
3385
3386 again:
3387                         extent_logical = key.objectid;
3388                         ASSERT(bytes <= U32_MAX);
3389                         extent_len = bytes;
3390
3391                         /*
3392                          * trim extent to this stripe
3393                          */
3394                         if (extent_logical < logical) {
3395                                 extent_len -= logical - extent_logical;
3396                                 extent_logical = logical;
3397                         }
3398                         if (extent_logical + extent_len >
3399                             logical + map->stripe_len) {
3400                                 extent_len = logical + map->stripe_len -
3401                                              extent_logical;
3402                         }
3403
3404                         extent_physical = extent_logical - logical + physical;
3405                         extent_dev = scrub_dev;
3406                         extent_mirror_num = mirror_num;
3407                         if (sctx->is_dev_replace)
3408                                 scrub_remap_extent(fs_info, extent_logical,
3409                                                    extent_len, &extent_physical,
3410                                                    &extent_dev,
3411                                                    &extent_mirror_num);
3412
3413                         if (flags & BTRFS_EXTENT_FLAG_DATA) {
3414                                 ret = btrfs_lookup_csums_range(csum_root,
3415                                                 extent_logical,
3416                                                 extent_logical + extent_len - 1,
3417                                                 &sctx->csum_list, 1);
3418                                 if (ret)
3419                                         goto out;
3420                         }
3421
3422                         ret = scrub_extent(sctx, map, extent_logical, extent_len,
3423                                            extent_physical, extent_dev, flags,
3424                                            generation, extent_mirror_num,
3425                                            extent_logical - logical + physical);
3426
3427                         scrub_free_csums(sctx);
3428
3429                         if (ret)
3430                                 goto out;
3431
3432                         if (sctx->is_dev_replace)
3433                                 sync_replace_for_zoned(sctx);
3434
3435                         if (extent_logical + extent_len <
3436                             key.objectid + bytes) {
3437                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3438                                         /*
3439                                          * loop until we find next data stripe
3440                                          * or we have finished all stripes.
3441                                          */
3442 loop:
3443                                         physical += map->stripe_len;
3444                                         ret = get_raid56_logic_offset(physical,
3445                                                         num, map, &logical,
3446                                                         &stripe_logical);
3447                                         logical += base;
3448
3449                                         if (ret && physical < physical_end) {
3450                                                 stripe_logical += base;
3451                                                 stripe_end = stripe_logical +
3452                                                                 increment;
3453                                                 ret = scrub_raid56_parity(sctx,
3454                                                         map, scrub_dev, ppath,
3455                                                         stripe_logical,
3456                                                         stripe_end);
3457                                                 if (ret)
3458                                                         goto out;
3459                                                 goto loop;
3460                                         }
3461                                 } else {
3462                                         physical += map->stripe_len;
3463                                         logical += increment;
3464                                 }
3465                                 if (logical < key.objectid + bytes) {
3466                                         cond_resched();
3467                                         goto again;
3468                                 }
3469
3470                                 if (physical >= physical_end) {
3471                                         stop_loop = 1;
3472                                         break;
3473                                 }
3474                         }
3475 next:
3476                         path->slots[0]++;
3477                 }
3478                 btrfs_release_path(path);
3479 skip:
3480                 logical += increment;
3481                 physical += map->stripe_len;
3482                 spin_lock(&sctx->stat_lock);
3483                 if (stop_loop)
3484                         sctx->stat.last_physical = map->stripes[num].physical +
3485                                                    length;
3486                 else
3487                         sctx->stat.last_physical = physical;
3488                 spin_unlock(&sctx->stat_lock);
3489                 if (stop_loop)
3490                         break;
3491         }
3492 out:
3493         /* push queued extents */
3494         scrub_submit(sctx);
3495         mutex_lock(&sctx->wr_lock);
3496         scrub_wr_submit(sctx);
3497         mutex_unlock(&sctx->wr_lock);
3498
3499         blk_finish_plug(&plug);
3500         btrfs_free_path(path);
3501         btrfs_free_path(ppath);
3502
3503         if (sctx->is_dev_replace && ret >= 0) {
3504                 int ret2;
3505
3506                 ret2 = sync_write_pointer_for_zoned(sctx, base + offset,
3507                                                     map->stripes[num].physical,
3508                                                     physical_end);
3509                 if (ret2)
3510                         ret = ret2;
3511         }
3512
3513         return ret < 0 ? ret : 0;
3514 }
3515
3516 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3517                                           struct btrfs_device *scrub_dev,
3518                                           u64 chunk_offset, u64 length,
3519                                           u64 dev_offset,
3520                                           struct btrfs_block_group *cache)
3521 {
3522         struct btrfs_fs_info *fs_info = sctx->fs_info;
3523         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3524         struct map_lookup *map;
3525         struct extent_map *em;
3526         int i;
3527         int ret = 0;
3528
3529         read_lock(&map_tree->lock);
3530         em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3531         read_unlock(&map_tree->lock);
3532
3533         if (!em) {
3534                 /*
3535                  * Might have been an unused block group deleted by the cleaner
3536                  * kthread or relocation.
3537                  */
3538                 spin_lock(&cache->lock);
3539                 if (!cache->removed)
3540                         ret = -EINVAL;
3541                 spin_unlock(&cache->lock);
3542
3543                 return ret;
3544         }
3545
3546         map = em->map_lookup;
3547         if (em->start != chunk_offset)
3548                 goto out;
3549
3550         if (em->len < length)
3551                 goto out;
3552
3553         for (i = 0; i < map->num_stripes; ++i) {
3554                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3555                     map->stripes[i].physical == dev_offset) {
3556                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3557                                            chunk_offset, length, cache);
3558                         if (ret)
3559                                 goto out;
3560                 }
3561         }
3562 out:
3563         free_extent_map(em);
3564
3565         return ret;
3566 }
3567
3568 static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3569                                           struct btrfs_block_group *cache)
3570 {
3571         struct btrfs_fs_info *fs_info = cache->fs_info;
3572         struct btrfs_trans_handle *trans;
3573
3574         if (!btrfs_is_zoned(fs_info))
3575                 return 0;
3576
3577         btrfs_wait_block_group_reservations(cache);
3578         btrfs_wait_nocow_writers(cache);
3579         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3580
3581         trans = btrfs_join_transaction(root);
3582         if (IS_ERR(trans))
3583                 return PTR_ERR(trans);
3584         return btrfs_commit_transaction(trans);
3585 }
3586
3587 static noinline_for_stack
3588 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3589                            struct btrfs_device *scrub_dev, u64 start, u64 end)
3590 {
3591         struct btrfs_dev_extent *dev_extent = NULL;
3592         struct btrfs_path *path;
3593         struct btrfs_fs_info *fs_info = sctx->fs_info;
3594         struct btrfs_root *root = fs_info->dev_root;
3595         u64 length;
3596         u64 chunk_offset;
3597         int ret = 0;
3598         int ro_set;
3599         int slot;
3600         struct extent_buffer *l;
3601         struct btrfs_key key;
3602         struct btrfs_key found_key;
3603         struct btrfs_block_group *cache;
3604         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3605
3606         path = btrfs_alloc_path();
3607         if (!path)
3608                 return -ENOMEM;
3609
3610         path->reada = READA_FORWARD;
3611         path->search_commit_root = 1;
3612         path->skip_locking = 1;
3613
3614         key.objectid = scrub_dev->devid;
3615         key.offset = 0ull;
3616         key.type = BTRFS_DEV_EXTENT_KEY;
3617
3618         while (1) {
3619                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3620                 if (ret < 0)
3621                         break;
3622                 if (ret > 0) {
3623                         if (path->slots[0] >=
3624                             btrfs_header_nritems(path->nodes[0])) {
3625                                 ret = btrfs_next_leaf(root, path);
3626                                 if (ret < 0)
3627                                         break;
3628                                 if (ret > 0) {
3629                                         ret = 0;
3630                                         break;
3631                                 }
3632                         } else {
3633                                 ret = 0;
3634                         }
3635                 }
3636
3637                 l = path->nodes[0];
3638                 slot = path->slots[0];
3639
3640                 btrfs_item_key_to_cpu(l, &found_key, slot);
3641
3642                 if (found_key.objectid != scrub_dev->devid)
3643                         break;
3644
3645                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3646                         break;
3647
3648                 if (found_key.offset >= end)
3649                         break;
3650
3651                 if (found_key.offset < key.offset)
3652                         break;
3653
3654                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3655                 length = btrfs_dev_extent_length(l, dev_extent);
3656
3657                 if (found_key.offset + length <= start)
3658                         goto skip;
3659
3660                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3661
3662                 /*
3663                  * get a reference on the corresponding block group to prevent
3664                  * the chunk from going away while we scrub it
3665                  */
3666                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3667
3668                 /* some chunks are removed but not committed to disk yet,
3669                  * continue scrubbing */
3670                 if (!cache)
3671                         goto skip;
3672
3673                 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3674                         spin_lock(&cache->lock);
3675                         if (!cache->to_copy) {
3676                                 spin_unlock(&cache->lock);
3677                                 btrfs_put_block_group(cache);
3678                                 goto skip;
3679                         }
3680                         spin_unlock(&cache->lock);
3681                 }
3682
3683                 /*
3684                  * Make sure that while we are scrubbing the corresponding block
3685                  * group doesn't get its logical address and its device extents
3686                  * reused for another block group, which can possibly be of a
3687                  * different type and different profile. We do this to prevent
3688                  * false error detections and crashes due to bogus attempts to
3689                  * repair extents.
3690                  */
3691                 spin_lock(&cache->lock);
3692                 if (cache->removed) {
3693                         spin_unlock(&cache->lock);
3694                         btrfs_put_block_group(cache);
3695                         goto skip;
3696                 }
3697                 btrfs_freeze_block_group(cache);
3698                 spin_unlock(&cache->lock);
3699
3700                 /*
3701                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3702                  * to avoid deadlock caused by:
3703                  * btrfs_inc_block_group_ro()
3704                  * -> btrfs_wait_for_commit()
3705                  * -> btrfs_commit_transaction()
3706                  * -> btrfs_scrub_pause()
3707                  */
3708                 scrub_pause_on(fs_info);
3709
3710                 /*
3711                  * Don't do chunk preallocation for scrub.
3712                  *
3713                  * This is especially important for SYSTEM bgs, or we can hit
3714                  * -EFBIG from btrfs_finish_chunk_alloc() like:
3715                  * 1. The only SYSTEM bg is marked RO.
3716                  *    Since SYSTEM bg is small, that's pretty common.
3717                  * 2. New SYSTEM bg will be allocated
3718                  *    Due to regular version will allocate new chunk.
3719                  * 3. New SYSTEM bg is empty and will get cleaned up
3720                  *    Before cleanup really happens, it's marked RO again.
3721                  * 4. Empty SYSTEM bg get scrubbed
3722                  *    We go back to 2.
3723                  *
3724                  * This can easily boost the amount of SYSTEM chunks if cleaner
3725                  * thread can't be triggered fast enough, and use up all space
3726                  * of btrfs_super_block::sys_chunk_array
3727                  *
3728                  * While for dev replace, we need to try our best to mark block
3729                  * group RO, to prevent race between:
3730                  * - Write duplication
3731                  *   Contains latest data
3732                  * - Scrub copy
3733                  *   Contains data from commit tree
3734                  *
3735                  * If target block group is not marked RO, nocow writes can
3736                  * be overwritten by scrub copy, causing data corruption.
3737                  * So for dev-replace, it's not allowed to continue if a block
3738                  * group is not RO.
3739                  */
3740                 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3741                 if (!ret && sctx->is_dev_replace) {
3742                         ret = finish_extent_writes_for_zoned(root, cache);
3743                         if (ret) {
3744                                 btrfs_dec_block_group_ro(cache);
3745                                 scrub_pause_off(fs_info);
3746                                 btrfs_put_block_group(cache);
3747                                 break;
3748                         }
3749                 }
3750
3751                 if (ret == 0) {
3752                         ro_set = 1;
3753                 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3754                         /*
3755                          * btrfs_inc_block_group_ro return -ENOSPC when it
3756                          * failed in creating new chunk for metadata.
3757                          * It is not a problem for scrub, because
3758                          * metadata are always cowed, and our scrub paused
3759                          * commit_transactions.
3760                          */
3761                         ro_set = 0;
3762                 } else if (ret == -ETXTBSY) {
3763                         btrfs_warn(fs_info,
3764                    "skipping scrub of block group %llu due to active swapfile",
3765                                    cache->start);
3766                         scrub_pause_off(fs_info);
3767                         ret = 0;
3768                         goto skip_unfreeze;
3769                 } else {
3770                         btrfs_warn(fs_info,
3771                                    "failed setting block group ro: %d", ret);
3772                         btrfs_unfreeze_block_group(cache);
3773                         btrfs_put_block_group(cache);
3774                         scrub_pause_off(fs_info);
3775                         break;
3776                 }
3777
3778                 /*
3779                  * Now the target block is marked RO, wait for nocow writes to
3780                  * finish before dev-replace.
3781                  * COW is fine, as COW never overwrites extents in commit tree.
3782                  */
3783                 if (sctx->is_dev_replace) {
3784                         btrfs_wait_nocow_writers(cache);
3785                         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3786                                         cache->length);
3787                 }
3788
3789                 scrub_pause_off(fs_info);
3790                 down_write(&dev_replace->rwsem);
3791                 dev_replace->cursor_right = found_key.offset + length;
3792                 dev_replace->cursor_left = found_key.offset;
3793                 dev_replace->item_needs_writeback = 1;
3794                 up_write(&dev_replace->rwsem);
3795
3796                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3797                                   found_key.offset, cache);
3798
3799                 /*
3800                  * flush, submit all pending read and write bios, afterwards
3801                  * wait for them.
3802                  * Note that in the dev replace case, a read request causes
3803                  * write requests that are submitted in the read completion
3804                  * worker. Therefore in the current situation, it is required
3805                  * that all write requests are flushed, so that all read and
3806                  * write requests are really completed when bios_in_flight
3807                  * changes to 0.
3808                  */
3809                 sctx->flush_all_writes = true;
3810                 scrub_submit(sctx);
3811                 mutex_lock(&sctx->wr_lock);
3812                 scrub_wr_submit(sctx);
3813                 mutex_unlock(&sctx->wr_lock);
3814
3815                 wait_event(sctx->list_wait,
3816                            atomic_read(&sctx->bios_in_flight) == 0);
3817
3818                 scrub_pause_on(fs_info);
3819
3820                 /*
3821                  * must be called before we decrease @scrub_paused.
3822                  * make sure we don't block transaction commit while
3823                  * we are waiting pending workers finished.
3824                  */
3825                 wait_event(sctx->list_wait,
3826                            atomic_read(&sctx->workers_pending) == 0);
3827                 sctx->flush_all_writes = false;
3828
3829                 scrub_pause_off(fs_info);
3830
3831                 if (sctx->is_dev_replace &&
3832                     !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
3833                                                       cache, found_key.offset))
3834                         ro_set = 0;
3835
3836                 down_write(&dev_replace->rwsem);
3837                 dev_replace->cursor_left = dev_replace->cursor_right;
3838                 dev_replace->item_needs_writeback = 1;
3839                 up_write(&dev_replace->rwsem);
3840
3841                 if (ro_set)
3842                         btrfs_dec_block_group_ro(cache);
3843
3844                 /*
3845                  * We might have prevented the cleaner kthread from deleting
3846                  * this block group if it was already unused because we raced
3847                  * and set it to RO mode first. So add it back to the unused
3848                  * list, otherwise it might not ever be deleted unless a manual
3849                  * balance is triggered or it becomes used and unused again.
3850                  */
3851                 spin_lock(&cache->lock);
3852                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3853                     cache->used == 0) {
3854                         spin_unlock(&cache->lock);
3855                         if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3856                                 btrfs_discard_queue_work(&fs_info->discard_ctl,
3857                                                          cache);
3858                         else
3859                                 btrfs_mark_bg_unused(cache);
3860                 } else {
3861                         spin_unlock(&cache->lock);
3862                 }
3863 skip_unfreeze:
3864                 btrfs_unfreeze_block_group(cache);
3865                 btrfs_put_block_group(cache);
3866                 if (ret)
3867                         break;
3868                 if (sctx->is_dev_replace &&
3869                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3870                         ret = -EIO;
3871                         break;
3872                 }
3873                 if (sctx->stat.malloc_errors > 0) {
3874                         ret = -ENOMEM;
3875                         break;
3876                 }
3877 skip:
3878                 key.offset = found_key.offset + length;
3879                 btrfs_release_path(path);
3880         }
3881
3882         btrfs_free_path(path);
3883
3884         return ret;
3885 }
3886
3887 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3888                                            struct btrfs_device *scrub_dev)
3889 {
3890         int     i;
3891         u64     bytenr;
3892         u64     gen;
3893         int     ret;
3894         struct btrfs_fs_info *fs_info = sctx->fs_info;
3895
3896         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3897                 return -EROFS;
3898
3899         /* Seed devices of a new filesystem has their own generation. */
3900         if (scrub_dev->fs_devices != fs_info->fs_devices)
3901                 gen = scrub_dev->generation;
3902         else
3903                 gen = fs_info->last_trans_committed;
3904
3905         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3906                 bytenr = btrfs_sb_offset(i);
3907                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3908                     scrub_dev->commit_total_bytes)
3909                         break;
3910                 if (!btrfs_check_super_location(scrub_dev, bytenr))
3911                         continue;
3912
3913                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3914                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3915                                   NULL, bytenr);
3916                 if (ret)
3917                         return ret;
3918         }
3919         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3920
3921         return 0;
3922 }
3923
3924 static void scrub_workers_put(struct btrfs_fs_info *fs_info)
3925 {
3926         if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
3927                                         &fs_info->scrub_lock)) {
3928                 struct btrfs_workqueue *scrub_workers = NULL;
3929                 struct btrfs_workqueue *scrub_wr_comp = NULL;
3930                 struct btrfs_workqueue *scrub_parity = NULL;
3931
3932                 scrub_workers = fs_info->scrub_workers;
3933                 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3934                 scrub_parity = fs_info->scrub_parity_workers;
3935
3936                 fs_info->scrub_workers = NULL;
3937                 fs_info->scrub_wr_completion_workers = NULL;
3938                 fs_info->scrub_parity_workers = NULL;
3939                 mutex_unlock(&fs_info->scrub_lock);
3940
3941                 btrfs_destroy_workqueue(scrub_workers);
3942                 btrfs_destroy_workqueue(scrub_wr_comp);
3943                 btrfs_destroy_workqueue(scrub_parity);
3944         }
3945 }
3946
3947 /*
3948  * get a reference count on fs_info->scrub_workers. start worker if necessary
3949  */
3950 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3951                                                 int is_dev_replace)
3952 {
3953         struct btrfs_workqueue *scrub_workers = NULL;
3954         struct btrfs_workqueue *scrub_wr_comp = NULL;
3955         struct btrfs_workqueue *scrub_parity = NULL;
3956         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3957         int max_active = fs_info->thread_pool_size;
3958         int ret = -ENOMEM;
3959
3960         if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
3961                 return 0;
3962
3963         scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub", flags,
3964                                               is_dev_replace ? 1 : max_active, 4);
3965         if (!scrub_workers)
3966                 goto fail_scrub_workers;
3967
3968         scrub_wr_comp = btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3969                                               max_active, 2);
3970         if (!scrub_wr_comp)
3971                 goto fail_scrub_wr_completion_workers;
3972
3973         scrub_parity = btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3974                                              max_active, 2);
3975         if (!scrub_parity)
3976                 goto fail_scrub_parity_workers;
3977
3978         mutex_lock(&fs_info->scrub_lock);
3979         if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3980                 ASSERT(fs_info->scrub_workers == NULL &&
3981                        fs_info->scrub_wr_completion_workers == NULL &&
3982                        fs_info->scrub_parity_workers == NULL);
3983                 fs_info->scrub_workers = scrub_workers;
3984                 fs_info->scrub_wr_completion_workers = scrub_wr_comp;
3985                 fs_info->scrub_parity_workers = scrub_parity;
3986                 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3987                 mutex_unlock(&fs_info->scrub_lock);
3988                 return 0;
3989         }
3990         /* Other thread raced in and created the workers for us */
3991         refcount_inc(&fs_info->scrub_workers_refcnt);
3992         mutex_unlock(&fs_info->scrub_lock);
3993
3994         ret = 0;
3995         btrfs_destroy_workqueue(scrub_parity);
3996 fail_scrub_parity_workers:
3997         btrfs_destroy_workqueue(scrub_wr_comp);
3998 fail_scrub_wr_completion_workers:
3999         btrfs_destroy_workqueue(scrub_workers);
4000 fail_scrub_workers:
4001         return ret;
4002 }
4003
4004 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4005                     u64 end, struct btrfs_scrub_progress *progress,
4006                     int readonly, int is_dev_replace)
4007 {
4008         struct scrub_ctx *sctx;
4009         int ret;
4010         struct btrfs_device *dev;
4011         unsigned int nofs_flag;
4012
4013         if (btrfs_fs_closing(fs_info))
4014                 return -EAGAIN;
4015
4016         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4017                 /*
4018                  * in this case scrub is unable to calculate the checksum
4019                  * the way scrub is implemented. Do not handle this
4020                  * situation at all because it won't ever happen.
4021                  */
4022                 btrfs_err(fs_info,
4023                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4024                        fs_info->nodesize,
4025                        BTRFS_STRIPE_LEN);
4026                 return -EINVAL;
4027         }
4028
4029         if (fs_info->nodesize >
4030             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4031             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4032                 /*
4033                  * would exhaust the array bounds of pagev member in
4034                  * struct scrub_block
4035                  */
4036                 btrfs_err(fs_info,
4037                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4038                        fs_info->nodesize,
4039                        SCRUB_MAX_PAGES_PER_BLOCK,
4040                        fs_info->sectorsize,
4041                        SCRUB_MAX_PAGES_PER_BLOCK);
4042                 return -EINVAL;
4043         }
4044
4045         /* Allocate outside of device_list_mutex */
4046         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4047         if (IS_ERR(sctx))
4048                 return PTR_ERR(sctx);
4049
4050         ret = scrub_workers_get(fs_info, is_dev_replace);
4051         if (ret)
4052                 goto out_free_ctx;
4053
4054         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4055         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4056         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4057                      !is_dev_replace)) {
4058                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4059                 ret = -ENODEV;
4060                 goto out;
4061         }
4062
4063         if (!is_dev_replace && !readonly &&
4064             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4065                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4066                 btrfs_err_in_rcu(fs_info,
4067                         "scrub on devid %llu: filesystem on %s is not writable",
4068                                  devid, rcu_str_deref(dev->name));
4069                 ret = -EROFS;
4070                 goto out;
4071         }
4072
4073         mutex_lock(&fs_info->scrub_lock);
4074         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4075             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4076                 mutex_unlock(&fs_info->scrub_lock);
4077                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4078                 ret = -EIO;
4079                 goto out;
4080         }
4081
4082         down_read(&fs_info->dev_replace.rwsem);
4083         if (dev->scrub_ctx ||
4084             (!is_dev_replace &&
4085              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4086                 up_read(&fs_info->dev_replace.rwsem);
4087                 mutex_unlock(&fs_info->scrub_lock);
4088                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4089                 ret = -EINPROGRESS;
4090                 goto out;
4091         }
4092         up_read(&fs_info->dev_replace.rwsem);
4093
4094         sctx->readonly = readonly;
4095         dev->scrub_ctx = sctx;
4096         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4097
4098         /*
4099          * checking @scrub_pause_req here, we can avoid
4100          * race between committing transaction and scrubbing.
4101          */
4102         __scrub_blocked_if_needed(fs_info);
4103         atomic_inc(&fs_info->scrubs_running);
4104         mutex_unlock(&fs_info->scrub_lock);
4105
4106         /*
4107          * In order to avoid deadlock with reclaim when there is a transaction
4108          * trying to pause scrub, make sure we use GFP_NOFS for all the
4109          * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
4110          * invoked by our callees. The pausing request is done when the
4111          * transaction commit starts, and it blocks the transaction until scrub
4112          * is paused (done at specific points at scrub_stripe() or right above
4113          * before incrementing fs_info->scrubs_running).
4114          */
4115         nofs_flag = memalloc_nofs_save();
4116         if (!is_dev_replace) {
4117                 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4118                 /*
4119                  * by holding device list mutex, we can
4120                  * kick off writing super in log tree sync.
4121                  */
4122                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4123                 ret = scrub_supers(sctx, dev);
4124                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4125         }
4126
4127         if (!ret)
4128                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
4129         memalloc_nofs_restore(nofs_flag);
4130
4131         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4132         atomic_dec(&fs_info->scrubs_running);
4133         wake_up(&fs_info->scrub_pause_wait);
4134
4135         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4136
4137         if (progress)
4138                 memcpy(progress, &sctx->stat, sizeof(*progress));
4139
4140         if (!is_dev_replace)
4141                 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4142                         ret ? "not finished" : "finished", devid, ret);
4143
4144         mutex_lock(&fs_info->scrub_lock);
4145         dev->scrub_ctx = NULL;
4146         mutex_unlock(&fs_info->scrub_lock);
4147
4148         scrub_workers_put(fs_info);
4149         scrub_put_ctx(sctx);
4150
4151         return ret;
4152 out:
4153         scrub_workers_put(fs_info);
4154 out_free_ctx:
4155         scrub_free_ctx(sctx);
4156
4157         return ret;
4158 }
4159
4160 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4161 {
4162         mutex_lock(&fs_info->scrub_lock);
4163         atomic_inc(&fs_info->scrub_pause_req);
4164         while (atomic_read(&fs_info->scrubs_paused) !=
4165                atomic_read(&fs_info->scrubs_running)) {
4166                 mutex_unlock(&fs_info->scrub_lock);
4167                 wait_event(fs_info->scrub_pause_wait,
4168                            atomic_read(&fs_info->scrubs_paused) ==
4169                            atomic_read(&fs_info->scrubs_running));
4170                 mutex_lock(&fs_info->scrub_lock);
4171         }
4172         mutex_unlock(&fs_info->scrub_lock);
4173 }
4174
4175 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4176 {
4177         atomic_dec(&fs_info->scrub_pause_req);
4178         wake_up(&fs_info->scrub_pause_wait);
4179 }
4180
4181 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4182 {
4183         mutex_lock(&fs_info->scrub_lock);
4184         if (!atomic_read(&fs_info->scrubs_running)) {
4185                 mutex_unlock(&fs_info->scrub_lock);
4186                 return -ENOTCONN;
4187         }
4188
4189         atomic_inc(&fs_info->scrub_cancel_req);
4190         while (atomic_read(&fs_info->scrubs_running)) {
4191                 mutex_unlock(&fs_info->scrub_lock);
4192                 wait_event(fs_info->scrub_pause_wait,
4193                            atomic_read(&fs_info->scrubs_running) == 0);
4194                 mutex_lock(&fs_info->scrub_lock);
4195         }
4196         atomic_dec(&fs_info->scrub_cancel_req);
4197         mutex_unlock(&fs_info->scrub_lock);
4198
4199         return 0;
4200 }
4201
4202 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4203 {
4204         struct btrfs_fs_info *fs_info = dev->fs_info;
4205         struct scrub_ctx *sctx;
4206
4207         mutex_lock(&fs_info->scrub_lock);
4208         sctx = dev->scrub_ctx;
4209         if (!sctx) {
4210                 mutex_unlock(&fs_info->scrub_lock);
4211                 return -ENOTCONN;
4212         }
4213         atomic_inc(&sctx->cancel_req);
4214         while (dev->scrub_ctx) {
4215                 mutex_unlock(&fs_info->scrub_lock);
4216                 wait_event(fs_info->scrub_pause_wait,
4217                            dev->scrub_ctx == NULL);
4218                 mutex_lock(&fs_info->scrub_lock);
4219         }
4220         mutex_unlock(&fs_info->scrub_lock);
4221
4222         return 0;
4223 }
4224
4225 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4226                          struct btrfs_scrub_progress *progress)
4227 {
4228         struct btrfs_device *dev;
4229         struct scrub_ctx *sctx = NULL;
4230
4231         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4232         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4233         if (dev)
4234                 sctx = dev->scrub_ctx;
4235         if (sctx)
4236                 memcpy(progress, &sctx->stat, sizeof(*progress));
4237         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4238
4239         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4240 }
4241
4242 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4243                                u64 extent_logical, u32 extent_len,
4244                                u64 *extent_physical,
4245                                struct btrfs_device **extent_dev,
4246                                int *extent_mirror_num)
4247 {
4248         u64 mapped_length;
4249         struct btrfs_bio *bbio = NULL;
4250         int ret;
4251
4252         mapped_length = extent_len;
4253         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4254                               &mapped_length, &bbio, 0);
4255         if (ret || !bbio || mapped_length < extent_len ||
4256             !bbio->stripes[0].dev->bdev) {
4257                 btrfs_put_bbio(bbio);
4258                 return;
4259         }
4260
4261         *extent_physical = bbio->stripes[0].physical;
4262         *extent_mirror_num = bbio->mirror_num;
4263         *extent_dev = bbio->stripes[0].dev;
4264         btrfs_put_bbio(bbio);
4265 }