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 "volumes.h"
  12 #include "disk-io.h"
  13 #include "ordered-data.h"
  14 #include "transaction.h"
  15 #include "backref.h"
  16 #include "extent_io.h"
  17 #include "dev-replace.h"
  18 #include "check-integrity.h"
  19 #include "rcu-string.h"
  20 #include "raid56.h"
  21 #include "block-group.h"
  22
  23 /*
  24  * This is only the first step towards a full-features scrub. It reads all
  25  * extent and super block and verifies the checksums. In case a bad checksum
  26  * is found or the extent cannot be read, good data will be written back if
  27  * any can be found.
  28  *
  29  * Future enhancements:
  30  *  - In case an unrepairable extent is encountered, track which files are
  31  *    affected and report them
  32  *  - track and record media errors, throw out bad devices
  33  *  - add a mode to also read unallocated space
  34  */
  35
  36 struct scrub_block;
  37 struct scrub_ctx;
  38
  39 /*
  40  * the following three values only influence the performance.
  41  * The last one configures the number of parallel and outstanding I/O
  42  * operations. The first two values configure an upper limit for the number
  43  * of (dynamically allocated) pages that are added to a bio.
  44  */
  45 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  46 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  47 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  48
  49 /*
  50  * the following value times PAGE_SIZE needs to be large enough to match the
  51  * largest node/leaf/sector size that shall be supported.
  52  * Values larger than BTRFS_STRIPE_LEN are not supported.
  53  */
  54 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  55
  56 struct scrub_recover {
  57         refcount_t              refs;
  58         struct btrfs_bio        *bbio;
  59         u64                     map_length;
  60 };
  61
  62 struct scrub_page {
  63         struct scrub_block      *sblock;
  64         struct page             *page;
  65         struct btrfs_device     *dev;
  66         struct list_head        list;
  67         u64                     flags;  /* extent flags */
  68         u64                     generation;
  69         u64                     logical;
  70         u64                     physical;
  71         u64                     physical_for_dev_replace;
  72         atomic_t                refs;
  73         struct {
  74                 unsigned int    mirror_num:8;
  75                 unsigned int    have_csum:1;
  76                 unsigned int    io_error:1;
  77         };
  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         u64                     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[0];
 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         u16                     csum_size;
 164         struct list_head        csum_list;
 165         atomic_t                cancel_req;
 166         int                     readonly;
 167         int                     pages_per_rd_bio;
 168
 169         int                     is_dev_replace;
 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 void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 210 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 211 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 212 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 213                                      struct scrub_block *sblocks_for_recheck);
 214 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 215                                 struct scrub_block *sblock,
 216                                 int retry_failed_mirror);
 217 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 218 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 219                                              struct scrub_block *sblock_good);
 220 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 221                                             struct scrub_block *sblock_good,
 222                                             int page_num, int force_write);
 223 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 224 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 225                                            int page_num);
 226 static int scrub_checksum_data(struct scrub_block *sblock);
 227 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 228 static int scrub_checksum_super(struct scrub_block *sblock);
 229 static void scrub_block_get(struct scrub_block *sblock);
 230 static void scrub_block_put(struct scrub_block *sblock);
 231 static void scrub_page_get(struct scrub_page *spage);
 232 static void scrub_page_put(struct scrub_page *spage);
 233 static void scrub_parity_get(struct scrub_parity *sparity);
 234 static void scrub_parity_put(struct scrub_parity *sparity);
 235 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 236                                     struct scrub_page *spage);
 237 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 238                        u64 physical, struct btrfs_device *dev, u64 flags,
 239                        u64 gen, int mirror_num, u8 *csum, int force,
 240                        u64 physical_for_dev_replace);
 241 static void scrub_bio_end_io(struct bio *bio);
 242 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 243 static void scrub_block_complete(struct scrub_block *sblock);
 244 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 245                                u64 extent_logical, u64 extent_len,
 246                                u64 *extent_physical,
 247                                struct btrfs_device **extent_dev,
 248                                int *extent_mirror_num);
 249 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 250                                     struct scrub_page *spage);
 251 static void scrub_wr_submit(struct scrub_ctx *sctx);
 252 static void scrub_wr_bio_end_io(struct bio *bio);
 253 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 254 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 255 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 256 static void scrub_put_ctx(struct scrub_ctx *sctx);
 257
 258 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
 259 {
 260         return page->recover &&
 261                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 262 }
 263
 264 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 265 {
 266         refcount_inc(&sctx->refs);
 267         atomic_inc(&sctx->bios_in_flight);
 268 }
 269
 270 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 271 {
 272         atomic_dec(&sctx->bios_in_flight);
 273         wake_up(&sctx->list_wait);
 274         scrub_put_ctx(sctx);
 275 }
 276
 277 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 278 {
 279         while (atomic_read(&fs_info->scrub_pause_req)) {
 280                 mutex_unlock(&fs_info->scrub_lock);
 281                 wait_event(fs_info->scrub_pause_wait,
 282                    atomic_read(&fs_info->scrub_pause_req) == 0);
 283                 mutex_lock(&fs_info->scrub_lock);
 284         }
 285 }
 286
 287 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 288 {
 289         atomic_inc(&fs_info->scrubs_paused);
 290         wake_up(&fs_info->scrub_pause_wait);
 291 }
 292
 293 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 294 {
 295         mutex_lock(&fs_info->scrub_lock);
 296         __scrub_blocked_if_needed(fs_info);
 297         atomic_dec(&fs_info->scrubs_paused);
 298         mutex_unlock(&fs_info->scrub_lock);
 299
 300         wake_up(&fs_info->scrub_pause_wait);
 301 }
 302
 303 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 304 {
 305         scrub_pause_on(fs_info);
 306         scrub_pause_off(fs_info);
 307 }
 308
 309 /*
 310  * Insert new full stripe lock into full stripe locks tree
 311  *
 312  * Return pointer to existing or newly inserted full_stripe_lock structure if
 313  * everything works well.
 314  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 315  *
 316  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 317  * function
 318  */
 319 static struct full_stripe_lock *insert_full_stripe_lock(
 320                 struct btrfs_full_stripe_locks_tree *locks_root,
 321                 u64 fstripe_logical)
 322 {
 323         struct rb_node **p;
 324         struct rb_node *parent = NULL;
 325         struct full_stripe_lock *entry;
 326         struct full_stripe_lock *ret;
 327
 328         lockdep_assert_held(&locks_root->lock);
 329
 330         p = &locks_root->root.rb_node;
 331         while (*p) {
 332                 parent = *p;
 333                 entry = rb_entry(parent, struct full_stripe_lock, node);
 334                 if (fstripe_logical < entry->logical) {
 335                         p = &(*p)->rb_left;
 336                 } else if (fstripe_logical > entry->logical) {
 337                         p = &(*p)->rb_right;
 338                 } else {
 339                         entry->refs++;
 340                         return entry;
 341                 }
 342         }
 343
 344         /*
 345          * Insert new lock.
 346          */
 347         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 348         if (!ret)
 349                 return ERR_PTR(-ENOMEM);
 350         ret->logical = fstripe_logical;
 351         ret->refs = 1;
 352         mutex_init(&ret->mutex);
 353
 354         rb_link_node(&ret->node, parent, p);
 355         rb_insert_color(&ret->node, &locks_root->root);
 356         return ret;
 357 }
 358
 359 /*
 360  * Search for a full stripe lock of a block group
 361  *
 362  * Return pointer to existing full stripe lock if found
 363  * Return NULL if not found
 364  */
 365 static struct full_stripe_lock *search_full_stripe_lock(
 366                 struct btrfs_full_stripe_locks_tree *locks_root,
 367                 u64 fstripe_logical)
 368 {
 369         struct rb_node *node;
 370         struct full_stripe_lock *entry;
 371
 372         lockdep_assert_held(&locks_root->lock);
 373
 374         node = locks_root->root.rb_node;
 375         while (node) {
 376                 entry = rb_entry(node, struct full_stripe_lock, node);
 377                 if (fstripe_logical < entry->logical)
 378                         node = node->rb_left;
 379                 else if (fstripe_logical > entry->logical)
 380                         node = node->rb_right;
 381                 else
 382                         return entry;
 383         }
 384         return NULL;
 385 }
 386
 387 /*
 388  * Helper to get full stripe logical from a normal bytenr.
 389  *
 390  * Caller must ensure @cache is a RAID56 block group.
 391  */
 392 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
 393                                    u64 bytenr)
 394 {
 395         u64 ret;
 396
 397         /*
 398          * Due to chunk item size limit, full stripe length should not be
 399          * larger than U32_MAX. Just a sanity check here.
 400          */
 401         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 402
 403         /*
 404          * round_down() can only handle power of 2, while RAID56 full
 405          * stripe length can be 64KiB * n, so we need to manually round down.
 406          */
 407         ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
 408                 cache->full_stripe_len + cache->key.objectid;
 409         return ret;
 410 }
 411
 412 /*
 413  * Lock a full stripe to avoid concurrency of recovery and read
 414  *
 415  * It's only used for profiles with parities (RAID5/6), for other profiles it
 416  * does nothing.
 417  *
 418  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 419  * So caller must call unlock_full_stripe() at the same context.
 420  *
 421  * Return <0 if encounters error.
 422  */
 423 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 424                             bool *locked_ret)
 425 {
 426         struct btrfs_block_group_cache *bg_cache;
 427         struct btrfs_full_stripe_locks_tree *locks_root;
 428         struct full_stripe_lock *existing;
 429         u64 fstripe_start;
 430         int ret = 0;
 431
 432         *locked_ret = false;
 433         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 434         if (!bg_cache) {
 435                 ASSERT(0);
 436                 return -ENOENT;
 437         }
 438
 439         /* Profiles not based on parity don't need full stripe lock */
 440         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 441                 goto out;
 442         locks_root = &bg_cache->full_stripe_locks_root;
 443
 444         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 445
 446         /* Now insert the full stripe lock */
 447         mutex_lock(&locks_root->lock);
 448         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 449         mutex_unlock(&locks_root->lock);
 450         if (IS_ERR(existing)) {
 451                 ret = PTR_ERR(existing);
 452                 goto out;
 453         }
 454         mutex_lock(&existing->mutex);
 455         *locked_ret = true;
 456 out:
 457         btrfs_put_block_group(bg_cache);
 458         return ret;
 459 }
 460
 461 /*
 462  * Unlock a full stripe.
 463  *
 464  * NOTE: Caller must ensure it's the same context calling corresponding
 465  * lock_full_stripe().
 466  *
 467  * Return 0 if we unlock full stripe without problem.
 468  * Return <0 for error
 469  */
 470 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 471                               bool locked)
 472 {
 473         struct btrfs_block_group_cache *bg_cache;
 474         struct btrfs_full_stripe_locks_tree *locks_root;
 475         struct full_stripe_lock *fstripe_lock;
 476         u64 fstripe_start;
 477         bool freeit = false;
 478         int ret = 0;
 479
 480         /* If we didn't acquire full stripe lock, no need to continue */
 481         if (!locked)
 482                 return 0;
 483
 484         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 485         if (!bg_cache) {
 486                 ASSERT(0);
 487                 return -ENOENT;
 488         }
 489         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 490                 goto out;
 491
 492         locks_root = &bg_cache->full_stripe_locks_root;
 493         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 494
 495         mutex_lock(&locks_root->lock);
 496         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 497         /* Unpaired unlock_full_stripe() detected */
 498         if (!fstripe_lock) {
 499                 WARN_ON(1);
 500                 ret = -ENOENT;
 501                 mutex_unlock(&locks_root->lock);
 502                 goto out;
 503         }
 504
 505         if (fstripe_lock->refs == 0) {
 506                 WARN_ON(1);
 507                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 508                         fstripe_lock->logical);
 509         } else {
 510                 fstripe_lock->refs--;
 511         }
 512
 513         if (fstripe_lock->refs == 0) {
 514                 rb_erase(&fstripe_lock->node, &locks_root->root);
 515                 freeit = true;
 516         }
 517         mutex_unlock(&locks_root->lock);
 518
 519         mutex_unlock(&fstripe_lock->mutex);
 520         if (freeit)
 521                 kfree(fstripe_lock);
 522 out:
 523         btrfs_put_block_group(bg_cache);
 524         return ret;
 525 }
 526
 527 static void scrub_free_csums(struct scrub_ctx *sctx)
 528 {
 529         while (!list_empty(&sctx->csum_list)) {
 530                 struct btrfs_ordered_sum *sum;
 531                 sum = list_first_entry(&sctx->csum_list,
 532                                        struct btrfs_ordered_sum, list);
 533                 list_del(&sum->list);
 534                 kfree(sum);
 535         }
 536 }
 537
 538 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 539 {
 540         int i;
 541
 542         if (!sctx)
 543                 return;
 544
 545         /* this can happen when scrub is cancelled */
 546         if (sctx->curr != -1) {
 547                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 548
 549                 for (i = 0; i < sbio->page_count; i++) {
 550                         WARN_ON(!sbio->pagev[i]->page);
 551                         scrub_block_put(sbio->pagev[i]->sblock);
 552                 }
 553                 bio_put(sbio->bio);
 554         }
 555
 556         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 557                 struct scrub_bio *sbio = sctx->bios[i];
 558
 559                 if (!sbio)
 560                         break;
 561                 kfree(sbio);
 562         }
 563
 564         kfree(sctx->wr_curr_bio);
 565         scrub_free_csums(sctx);
 566         kfree(sctx);
 567 }
 568
 569 static void scrub_put_ctx(struct scrub_ctx *sctx)
 570 {
 571         if (refcount_dec_and_test(&sctx->refs))
 572                 scrub_free_ctx(sctx);
 573 }
 574
 575 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 576                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 577 {
 578         struct scrub_ctx *sctx;
 579         int             i;
 580
 581         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 582         if (!sctx)
 583                 goto nomem;
 584         refcount_set(&sctx->refs, 1);
 585         sctx->is_dev_replace = is_dev_replace;
 586         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 587         sctx->curr = -1;
 588         sctx->fs_info = fs_info;
 589         INIT_LIST_HEAD(&sctx->csum_list);
 590         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 591                 struct scrub_bio *sbio;
 592
 593                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 594                 if (!sbio)
 595                         goto nomem;
 596                 sctx->bios[i] = sbio;
 597
 598                 sbio->index = i;
 599                 sbio->sctx = sctx;
 600                 sbio->page_count = 0;
 601                 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
 602                                 scrub_bio_end_io_worker, NULL, NULL);
 603
 604                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 605                         sctx->bios[i]->next_free = i + 1;
 606                 else
 607                         sctx->bios[i]->next_free = -1;
 608         }
 609         sctx->first_free = 0;
 610         atomic_set(&sctx->bios_in_flight, 0);
 611         atomic_set(&sctx->workers_pending, 0);
 612         atomic_set(&sctx->cancel_req, 0);
 613         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 614
 615         spin_lock_init(&sctx->list_lock);
 616         spin_lock_init(&sctx->stat_lock);
 617         init_waitqueue_head(&sctx->list_wait);
 618
 619         WARN_ON(sctx->wr_curr_bio != NULL);
 620         mutex_init(&sctx->wr_lock);
 621         sctx->wr_curr_bio = NULL;
 622         if (is_dev_replace) {
 623                 WARN_ON(!fs_info->dev_replace.tgtdev);
 624                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 625                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 626                 sctx->flush_all_writes = false;
 627         }
 628
 629         return sctx;
 630
 631 nomem:
 632         scrub_free_ctx(sctx);
 633         return ERR_PTR(-ENOMEM);
 634 }
 635
 636 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 637                                      void *warn_ctx)
 638 {
 639         u64 isize;
 640         u32 nlink;
 641         int ret;
 642         int i;
 643         unsigned nofs_flag;
 644         struct extent_buffer *eb;
 645         struct btrfs_inode_item *inode_item;
 646         struct scrub_warning *swarn = warn_ctx;
 647         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 648         struct inode_fs_paths *ipath = NULL;
 649         struct btrfs_root *local_root;
 650         struct btrfs_key root_key;
 651         struct btrfs_key key;
 652
 653         root_key.objectid = root;
 654         root_key.type = BTRFS_ROOT_ITEM_KEY;
 655         root_key.offset = (u64)-1;
 656         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 657         if (IS_ERR(local_root)) {
 658                 ret = PTR_ERR(local_root);
 659                 goto err;
 660         }
 661
 662         /*
 663          * this makes the path point to (inum INODE_ITEM ioff)
 664          */
 665         key.objectid = inum;
 666         key.type = BTRFS_INODE_ITEM_KEY;
 667         key.offset = 0;
 668
 669         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 670         if (ret) {
 671                 btrfs_release_path(swarn->path);
 672                 goto err;
 673         }
 674
 675         eb = swarn->path->nodes[0];
 676         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 677                                         struct btrfs_inode_item);
 678         isize = btrfs_inode_size(eb, inode_item);
 679         nlink = btrfs_inode_nlink(eb, inode_item);
 680         btrfs_release_path(swarn->path);
 681
 682         /*
 683          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 684          * uses GFP_NOFS in this context, so we keep it consistent but it does
 685          * not seem to be strictly necessary.
 686          */
 687         nofs_flag = memalloc_nofs_save();
 688         ipath = init_ipath(4096, local_root, swarn->path);
 689         memalloc_nofs_restore(nofs_flag);
 690         if (IS_ERR(ipath)) {
 691                 ret = PTR_ERR(ipath);
 692                 ipath = NULL;
 693                 goto err;
 694         }
 695         ret = paths_from_inode(inum, ipath);
 696
 697         if (ret < 0)
 698                 goto err;
 699
 700         /*
 701          * we deliberately ignore the bit ipath might have been too small to
 702          * hold all of the paths here
 703          */
 704         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 705                 btrfs_warn_in_rcu(fs_info,
 706 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 707                                   swarn->errstr, swarn->logical,
 708                                   rcu_str_deref(swarn->dev->name),
 709                                   swarn->physical,
 710                                   root, inum, offset,
 711                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
 712                                   (char *)(unsigned long)ipath->fspath->val[i]);
 713
 714         free_ipath(ipath);
 715         return 0;
 716
 717 err:
 718         btrfs_warn_in_rcu(fs_info,
 719                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 720                           swarn->errstr, swarn->logical,
 721                           rcu_str_deref(swarn->dev->name),
 722                           swarn->physical,
 723                           root, inum, offset, ret);
 724
 725         free_ipath(ipath);
 726         return 0;
 727 }
 728
 729 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 730 {
 731         struct btrfs_device *dev;
 732         struct btrfs_fs_info *fs_info;
 733         struct btrfs_path *path;
 734         struct btrfs_key found_key;
 735         struct extent_buffer *eb;
 736         struct btrfs_extent_item *ei;
 737         struct scrub_warning swarn;
 738         unsigned long ptr = 0;
 739         u64 extent_item_pos;
 740         u64 flags = 0;
 741         u64 ref_root;
 742         u32 item_size;
 743         u8 ref_level = 0;
 744         int ret;
 745
 746         WARN_ON(sblock->page_count < 1);
 747         dev = sblock->pagev[0]->dev;
 748         fs_info = sblock->sctx->fs_info;
 749
 750         path = btrfs_alloc_path();
 751         if (!path)
 752                 return;
 753
 754         swarn.physical = sblock->pagev[0]->physical;
 755         swarn.logical = sblock->pagev[0]->logical;
 756         swarn.errstr = errstr;
 757         swarn.dev = NULL;
 758
 759         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 760                                   &flags);
 761         if (ret < 0)
 762                 goto out;
 763
 764         extent_item_pos = swarn.logical - found_key.objectid;
 765         swarn.extent_item_size = found_key.offset;
 766
 767         eb = path->nodes[0];
 768         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 769         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 770
 771         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 772                 do {
 773                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 774                                                       item_size, &ref_root,
 775                                                       &ref_level);
 776                         btrfs_warn_in_rcu(fs_info,
 777 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 778                                 errstr, swarn.logical,
 779                                 rcu_str_deref(dev->name),
 780                                 swarn.physical,
 781                                 ref_level ? "node" : "leaf",
 782                                 ret < 0 ? -1 : ref_level,
 783                                 ret < 0 ? -1 : ref_root);
 784                 } while (ret != 1);
 785                 btrfs_release_path(path);
 786         } else {
 787                 btrfs_release_path(path);
 788                 swarn.path = path;
 789                 swarn.dev = dev;
 790                 iterate_extent_inodes(fs_info, found_key.objectid,
 791                                         extent_item_pos, 1,
 792                                         scrub_print_warning_inode, &swarn, false);
 793         }
 794
 795 out:
 796         btrfs_free_path(path);
 797 }
 798
 799 static inline void scrub_get_recover(struct scrub_recover *recover)
 800 {
 801         refcount_inc(&recover->refs);
 802 }
 803
 804 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 805                                      struct scrub_recover *recover)
 806 {
 807         if (refcount_dec_and_test(&recover->refs)) {
 808                 btrfs_bio_counter_dec(fs_info);
 809                 btrfs_put_bbio(recover->bbio);
 810                 kfree(recover);
 811         }
 812 }
 813
 814 /*
 815  * scrub_handle_errored_block gets called when either verification of the
 816  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 817  * case, this function handles all pages in the bio, even though only one
 818  * may be bad.
 819  * The goal of this function is to repair the errored block by using the
 820  * contents of one of the mirrors.
 821  */
 822 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 823 {
 824         struct scrub_ctx *sctx = sblock_to_check->sctx;
 825         struct btrfs_device *dev;
 826         struct btrfs_fs_info *fs_info;
 827         u64 logical;
 828         unsigned int failed_mirror_index;
 829         unsigned int is_metadata;
 830         unsigned int have_csum;
 831         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 832         struct scrub_block *sblock_bad;
 833         int ret;
 834         int mirror_index;
 835         int page_num;
 836         int success;
 837         bool full_stripe_locked;
 838         unsigned int nofs_flag;
 839         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 840                                       DEFAULT_RATELIMIT_BURST);
 841
 842         BUG_ON(sblock_to_check->page_count < 1);
 843         fs_info = sctx->fs_info;
 844         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 845                 /*
 846                  * if we find an error in a super block, we just report it.
 847                  * They will get written with the next transaction commit
 848                  * anyway
 849                  */
 850                 spin_lock(&sctx->stat_lock);
 851                 ++sctx->stat.super_errors;
 852                 spin_unlock(&sctx->stat_lock);
 853                 return 0;
 854         }
 855         logical = sblock_to_check->pagev[0]->logical;
 856         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 857         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 858         is_metadata = !(sblock_to_check->pagev[0]->flags &
 859                         BTRFS_EXTENT_FLAG_DATA);
 860         have_csum = sblock_to_check->pagev[0]->have_csum;
 861         dev = sblock_to_check->pagev[0]->dev;
 862
 863         /*
 864          * We must use GFP_NOFS because the scrub task might be waiting for a
 865          * worker task executing this function and in turn a transaction commit
 866          * might be waiting the scrub task to pause (which needs to wait for all
 867          * the worker tasks to complete before pausing).
 868          * We do allocations in the workers through insert_full_stripe_lock()
 869          * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 870          * this function.
 871          */
 872         nofs_flag = memalloc_nofs_save();
 873         /*
 874          * For RAID5/6, race can happen for a different device scrub thread.
 875          * For data corruption, Parity and Data threads will both try
 876          * to recovery the data.
 877          * Race can lead to doubly added csum error, or even unrecoverable
 878          * error.
 879          */
 880         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 881         if (ret < 0) {
 882                 memalloc_nofs_restore(nofs_flag);
 883                 spin_lock(&sctx->stat_lock);
 884                 if (ret == -ENOMEM)
 885                         sctx->stat.malloc_errors++;
 886                 sctx->stat.read_errors++;
 887                 sctx->stat.uncorrectable_errors++;
 888                 spin_unlock(&sctx->stat_lock);
 889                 return ret;
 890         }
 891
 892         /*
 893          * read all mirrors one after the other. This includes to
 894          * re-read the extent or metadata block that failed (that was
 895          * the cause that this fixup code is called) another time,
 896          * page by page this time in order to know which pages
 897          * caused I/O errors and which ones are good (for all mirrors).
 898          * It is the goal to handle the situation when more than one
 899          * mirror contains I/O errors, but the errors do not
 900          * overlap, i.e. the data can be repaired by selecting the
 901          * pages from those mirrors without I/O error on the
 902          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 903          * would be that mirror #1 has an I/O error on the first page,
 904          * the second page is good, and mirror #2 has an I/O error on
 905          * the second page, but the first page is good.
 906          * Then the first page of the first mirror can be repaired by
 907          * taking the first page of the second mirror, and the
 908          * second page of the second mirror can be repaired by
 909          * copying the contents of the 2nd page of the 1st mirror.
 910          * One more note: if the pages of one mirror contain I/O
 911          * errors, the checksum cannot be verified. In order to get
 912          * the best data for repairing, the first attempt is to find
 913          * a mirror without I/O errors and with a validated checksum.
 914          * Only if this is not possible, the pages are picked from
 915          * mirrors with I/O errors without considering the checksum.
 916          * If the latter is the case, at the end, the checksum of the
 917          * repaired area is verified in order to correctly maintain
 918          * the statistics.
 919          */
 920
 921         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 922                                       sizeof(*sblocks_for_recheck), GFP_KERNEL);
 923         if (!sblocks_for_recheck) {
 924                 spin_lock(&sctx->stat_lock);
 925                 sctx->stat.malloc_errors++;
 926                 sctx->stat.read_errors++;
 927                 sctx->stat.uncorrectable_errors++;
 928                 spin_unlock(&sctx->stat_lock);
 929                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 930                 goto out;
 931         }
 932
 933         /* setup the context, map the logical blocks and alloc the pages */
 934         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 935         if (ret) {
 936                 spin_lock(&sctx->stat_lock);
 937                 sctx->stat.read_errors++;
 938                 sctx->stat.uncorrectable_errors++;
 939                 spin_unlock(&sctx->stat_lock);
 940                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 941                 goto out;
 942         }
 943         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 944         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 945
 946         /* build and submit the bios for the failed mirror, check checksums */
 947         scrub_recheck_block(fs_info, sblock_bad, 1);
 948
 949         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 950             sblock_bad->no_io_error_seen) {
 951                 /*
 952                  * the error disappeared after reading page by page, or
 953                  * the area was part of a huge bio and other parts of the
 954                  * bio caused I/O errors, or the block layer merged several
 955                  * read requests into one and the error is caused by a
 956                  * different bio (usually one of the two latter cases is
 957                  * the cause)
 958                  */
 959                 spin_lock(&sctx->stat_lock);
 960                 sctx->stat.unverified_errors++;
 961                 sblock_to_check->data_corrected = 1;
 962                 spin_unlock(&sctx->stat_lock);
 963
 964                 if (sctx->is_dev_replace)
 965                         scrub_write_block_to_dev_replace(sblock_bad);
 966                 goto out;
 967         }
 968
 969         if (!sblock_bad->no_io_error_seen) {
 970                 spin_lock(&sctx->stat_lock);
 971                 sctx->stat.read_errors++;
 972                 spin_unlock(&sctx->stat_lock);
 973                 if (__ratelimit(&_rs))
 974                         scrub_print_warning("i/o error", sblock_to_check);
 975                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 976         } else if (sblock_bad->checksum_error) {
 977                 spin_lock(&sctx->stat_lock);
 978                 sctx->stat.csum_errors++;
 979                 spin_unlock(&sctx->stat_lock);
 980                 if (__ratelimit(&_rs))
 981                         scrub_print_warning("checksum error", sblock_to_check);
 982                 btrfs_dev_stat_inc_and_print(dev,
 983                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 984         } else if (sblock_bad->header_error) {
 985                 spin_lock(&sctx->stat_lock);
 986                 sctx->stat.verify_errors++;
 987                 spin_unlock(&sctx->stat_lock);
 988                 if (__ratelimit(&_rs))
 989                         scrub_print_warning("checksum/header error",
 990                                             sblock_to_check);
 991                 if (sblock_bad->generation_error)
 992                         btrfs_dev_stat_inc_and_print(dev,
 993                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 994                 else
 995                         btrfs_dev_stat_inc_and_print(dev,
 996                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 997         }
 998
 999         if (sctx->readonly) {
1000                 ASSERT(!sctx->is_dev_replace);
1001                 goto out;
1002         }
1003
1004         /*
1005          * now build and submit the bios for the other mirrors, check
1006          * checksums.
1007          * First try to pick the mirror which is completely without I/O
1008          * errors and also does not have a checksum error.
1009          * If one is found, and if a checksum is present, the full block
1010          * that is known to contain an error is rewritten. Afterwards
1011          * the block is known to be corrected.
1012          * If a mirror is found which is completely correct, and no
1013          * checksum is present, only those pages are rewritten that had
1014          * an I/O error in the block to be repaired, since it cannot be
1015          * determined, which copy of the other pages is better (and it
1016          * could happen otherwise that a correct page would be
1017          * overwritten by a bad one).
1018          */
1019         for (mirror_index = 0; ;mirror_index++) {
1020                 struct scrub_block *sblock_other;
1021
1022                 if (mirror_index == failed_mirror_index)
1023                         continue;
1024
1025                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1026                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1027                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1028                                 break;
1029                         if (!sblocks_for_recheck[mirror_index].page_count)
1030                                 break;
1031
1032                         sblock_other = sblocks_for_recheck + mirror_index;
1033                 } else {
1034                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1035                         int max_allowed = r->bbio->num_stripes -
1036                                                 r->bbio->num_tgtdevs;
1037
1038                         if (mirror_index >= max_allowed)
1039                                 break;
1040                         if (!sblocks_for_recheck[1].page_count)
1041                                 break;
1042
1043                         ASSERT(failed_mirror_index == 0);
1044                         sblock_other = sblocks_for_recheck + 1;
1045                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1046                 }
1047
1048                 /* build and submit the bios, check checksums */
1049                 scrub_recheck_block(fs_info, sblock_other, 0);
1050
1051                 if (!sblock_other->header_error &&
1052                     !sblock_other->checksum_error &&
1053                     sblock_other->no_io_error_seen) {
1054                         if (sctx->is_dev_replace) {
1055                                 scrub_write_block_to_dev_replace(sblock_other);
1056                                 goto corrected_error;
1057                         } else {
1058                                 ret = scrub_repair_block_from_good_copy(
1059                                                 sblock_bad, sblock_other);
1060                                 if (!ret)
1061                                         goto corrected_error;
1062                         }
1063                 }
1064         }
1065
1066         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1067                 goto did_not_correct_error;
1068
1069         /*
1070          * In case of I/O errors in the area that is supposed to be
1071          * repaired, continue by picking good copies of those pages.
1072          * Select the good pages from mirrors to rewrite bad pages from
1073          * the area to fix. Afterwards verify the checksum of the block
1074          * that is supposed to be repaired. This verification step is
1075          * only done for the purpose of statistic counting and for the
1076          * final scrub report, whether errors remain.
1077          * A perfect algorithm could make use of the checksum and try
1078          * all possible combinations of pages from the different mirrors
1079          * until the checksum verification succeeds. For example, when
1080          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1081          * of mirror #2 is readable but the final checksum test fails,
1082          * then the 2nd page of mirror #3 could be tried, whether now
1083          * the final checksum succeeds. But this would be a rare
1084          * exception and is therefore not implemented. At least it is
1085          * avoided that the good copy is overwritten.
1086          * A more useful improvement would be to pick the sectors
1087          * without I/O error based on sector sizes (512 bytes on legacy
1088          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1089          * mirror could be repaired by taking 512 byte of a different
1090          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1091          * area are unreadable.
1092          */
1093         success = 1;
1094         for (page_num = 0; page_num < sblock_bad->page_count;
1095              page_num++) {
1096                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1097                 struct scrub_block *sblock_other = NULL;
1098
1099                 /* skip no-io-error page in scrub */
1100                 if (!page_bad->io_error && !sctx->is_dev_replace)
1101                         continue;
1102
1103                 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1104                         /*
1105                          * In case of dev replace, if raid56 rebuild process
1106                          * didn't work out correct data, then copy the content
1107                          * in sblock_bad to make sure target device is identical
1108                          * to source device, instead of writing garbage data in
1109                          * sblock_for_recheck array to target device.
1110                          */
1111                         sblock_other = NULL;
1112                 } else if (page_bad->io_error) {
1113                         /* try to find no-io-error page in mirrors */
1114                         for (mirror_index = 0;
1115                              mirror_index < BTRFS_MAX_MIRRORS &&
1116                              sblocks_for_recheck[mirror_index].page_count > 0;
1117                              mirror_index++) {
1118                                 if (!sblocks_for_recheck[mirror_index].
1119                                     pagev[page_num]->io_error) {
1120                                         sblock_other = sblocks_for_recheck +
1121                                                        mirror_index;
1122                                         break;
1123                                 }
1124                         }
1125                         if (!sblock_other)
1126                                 success = 0;
1127                 }
1128
1129                 if (sctx->is_dev_replace) {
1130                         /*
1131                          * did not find a mirror to fetch the page
1132                          * from. scrub_write_page_to_dev_replace()
1133                          * handles this case (page->io_error), by
1134                          * filling the block with zeros before
1135                          * submitting the write request
1136                          */
1137                         if (!sblock_other)
1138                                 sblock_other = sblock_bad;
1139
1140                         if (scrub_write_page_to_dev_replace(sblock_other,
1141                                                             page_num) != 0) {
1142                                 atomic64_inc(
1143                                         &fs_info->dev_replace.num_write_errors);
1144                                 success = 0;
1145                         }
1146                 } else if (sblock_other) {
1147                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1148                                                                sblock_other,
1149                                                                page_num, 0);
1150                         if (0 == ret)
1151                                 page_bad->io_error = 0;
1152                         else
1153                                 success = 0;
1154                 }
1155         }
1156
1157         if (success && !sctx->is_dev_replace) {
1158                 if (is_metadata || have_csum) {
1159                         /*
1160                          * need to verify the checksum now that all
1161                          * sectors on disk are repaired (the write
1162                          * request for data to be repaired is on its way).
1163                          * Just be lazy and use scrub_recheck_block()
1164                          * which re-reads the data before the checksum
1165                          * is verified, but most likely the data comes out
1166                          * of the page cache.
1167                          */
1168                         scrub_recheck_block(fs_info, sblock_bad, 1);
1169                         if (!sblock_bad->header_error &&
1170                             !sblock_bad->checksum_error &&
1171                             sblock_bad->no_io_error_seen)
1172                                 goto corrected_error;
1173                         else
1174                                 goto did_not_correct_error;
1175                 } else {
1176 corrected_error:
1177                         spin_lock(&sctx->stat_lock);
1178                         sctx->stat.corrected_errors++;
1179                         sblock_to_check->data_corrected = 1;
1180                         spin_unlock(&sctx->stat_lock);
1181                         btrfs_err_rl_in_rcu(fs_info,
1182                                 "fixed up error at logical %llu on dev %s",
1183                                 logical, rcu_str_deref(dev->name));
1184                 }
1185         } else {
1186 did_not_correct_error:
1187                 spin_lock(&sctx->stat_lock);
1188                 sctx->stat.uncorrectable_errors++;
1189                 spin_unlock(&sctx->stat_lock);
1190                 btrfs_err_rl_in_rcu(fs_info,
1191                         "unable to fixup (regular) error at logical %llu on dev %s",
1192                         logical, rcu_str_deref(dev->name));
1193         }
1194
1195 out:
1196         if (sblocks_for_recheck) {
1197                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1198                      mirror_index++) {
1199                         struct scrub_block *sblock = sblocks_for_recheck +
1200                                                      mirror_index;
1201                         struct scrub_recover *recover;
1202                         int page_index;
1203
1204                         for (page_index = 0; page_index < sblock->page_count;
1205                              page_index++) {
1206                                 sblock->pagev[page_index]->sblock = NULL;
1207                                 recover = sblock->pagev[page_index]->recover;
1208                                 if (recover) {
1209                                         scrub_put_recover(fs_info, recover);
1210                                         sblock->pagev[page_index]->recover =
1211                                                                         NULL;
1212                                 }
1213                                 scrub_page_put(sblock->pagev[page_index]);
1214                         }
1215                 }
1216                 kfree(sblocks_for_recheck);
1217         }
1218
1219         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1220         memalloc_nofs_restore(nofs_flag);
1221         if (ret < 0)
1222                 return ret;
1223         return 0;
1224 }
1225
1226 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1227 {
1228         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1229                 return 2;
1230         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1231                 return 3;
1232         else
1233                 return (int)bbio->num_stripes;
1234 }
1235
1236 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1237                                                  u64 *raid_map,
1238                                                  u64 mapped_length,
1239                                                  int nstripes, int mirror,
1240                                                  int *stripe_index,
1241                                                  u64 *stripe_offset)
1242 {
1243         int i;
1244
1245         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1246                 /* RAID5/6 */
1247                 for (i = 0; i < nstripes; i++) {
1248                         if (raid_map[i] == RAID6_Q_STRIPE ||
1249                             raid_map[i] == RAID5_P_STRIPE)
1250                                 continue;
1251
1252                         if (logical >= raid_map[i] &&
1253                             logical < raid_map[i] + mapped_length)
1254                                 break;
1255                 }
1256
1257                 *stripe_index = i;
1258                 *stripe_offset = logical - raid_map[i];
1259         } else {
1260                 /* The other RAID type */
1261                 *stripe_index = mirror;
1262                 *stripe_offset = 0;
1263         }
1264 }
1265
1266 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1267                                      struct scrub_block *sblocks_for_recheck)
1268 {
1269         struct scrub_ctx *sctx = original_sblock->sctx;
1270         struct btrfs_fs_info *fs_info = sctx->fs_info;
1271         u64 length = original_sblock->page_count * PAGE_SIZE;
1272         u64 logical = original_sblock->pagev[0]->logical;
1273         u64 generation = original_sblock->pagev[0]->generation;
1274         u64 flags = original_sblock->pagev[0]->flags;
1275         u64 have_csum = original_sblock->pagev[0]->have_csum;
1276         struct scrub_recover *recover;
1277         struct btrfs_bio *bbio;
1278         u64 sublen;
1279         u64 mapped_length;
1280         u64 stripe_offset;
1281         int stripe_index;
1282         int page_index = 0;
1283         int mirror_index;
1284         int nmirrors;
1285         int ret;
1286
1287         /*
1288          * note: the two members refs and outstanding_pages
1289          * are not used (and not set) in the blocks that are used for
1290          * the recheck procedure
1291          */
1292
1293         while (length > 0) {
1294                 sublen = min_t(u64, length, PAGE_SIZE);
1295                 mapped_length = sublen;
1296                 bbio = NULL;
1297
1298                 /*
1299                  * with a length of PAGE_SIZE, each returned stripe
1300                  * represents one mirror
1301                  */
1302                 btrfs_bio_counter_inc_blocked(fs_info);
1303                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1304                                 logical, &mapped_length, &bbio);
1305                 if (ret || !bbio || mapped_length < sublen) {
1306                         btrfs_put_bbio(bbio);
1307                         btrfs_bio_counter_dec(fs_info);
1308                         return -EIO;
1309                 }
1310
1311                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1312                 if (!recover) {
1313                         btrfs_put_bbio(bbio);
1314                         btrfs_bio_counter_dec(fs_info);
1315                         return -ENOMEM;
1316                 }
1317
1318                 refcount_set(&recover->refs, 1);
1319                 recover->bbio = bbio;
1320                 recover->map_length = mapped_length;
1321
1322                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1323
1324                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1325
1326                 for (mirror_index = 0; mirror_index < nmirrors;
1327                      mirror_index++) {
1328                         struct scrub_block *sblock;
1329                         struct scrub_page *page;
1330
1331                         sblock = sblocks_for_recheck + mirror_index;
1332                         sblock->sctx = sctx;
1333
1334                         page = kzalloc(sizeof(*page), GFP_NOFS);
1335                         if (!page) {
1336 leave_nomem:
1337                                 spin_lock(&sctx->stat_lock);
1338                                 sctx->stat.malloc_errors++;
1339                                 spin_unlock(&sctx->stat_lock);
1340                                 scrub_put_recover(fs_info, recover);
1341                                 return -ENOMEM;
1342                         }
1343                         scrub_page_get(page);
1344                         sblock->pagev[page_index] = page;
1345                         page->sblock = sblock;
1346                         page->flags = flags;
1347                         page->generation = generation;
1348                         page->logical = logical;
1349                         page->have_csum = have_csum;
1350                         if (have_csum)
1351                                 memcpy(page->csum,
1352                                        original_sblock->pagev[0]->csum,
1353                                        sctx->csum_size);
1354
1355                         scrub_stripe_index_and_offset(logical,
1356                                                       bbio->map_type,
1357                                                       bbio->raid_map,
1358                                                       mapped_length,
1359                                                       bbio->num_stripes -
1360                                                       bbio->num_tgtdevs,
1361                                                       mirror_index,
1362                                                       &stripe_index,
1363                                                       &stripe_offset);
1364                         page->physical = bbio->stripes[stripe_index].physical +
1365                                          stripe_offset;
1366                         page->dev = bbio->stripes[stripe_index].dev;
1367
1368                         BUG_ON(page_index >= original_sblock->page_count);
1369                         page->physical_for_dev_replace =
1370                                 original_sblock->pagev[page_index]->
1371                                 physical_for_dev_replace;
1372                         /* for missing devices, dev->bdev is NULL */
1373                         page->mirror_num = mirror_index + 1;
1374                         sblock->page_count++;
1375                         page->page = alloc_page(GFP_NOFS);
1376                         if (!page->page)
1377                                 goto leave_nomem;
1378
1379                         scrub_get_recover(recover);
1380                         page->recover = recover;
1381                 }
1382                 scrub_put_recover(fs_info, recover);
1383                 length -= sublen;
1384                 logical += sublen;
1385                 page_index++;
1386         }
1387
1388         return 0;
1389 }
1390
1391 static void scrub_bio_wait_endio(struct bio *bio)
1392 {
1393         complete(bio->bi_private);
1394 }
1395
1396 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1397                                         struct bio *bio,
1398                                         struct scrub_page *page)
1399 {
1400         DECLARE_COMPLETION_ONSTACK(done);
1401         int ret;
1402         int mirror_num;
1403
1404         bio->bi_iter.bi_sector = page->logical >> 9;
1405         bio->bi_private = &done;
1406         bio->bi_end_io = scrub_bio_wait_endio;
1407
1408         mirror_num = page->sblock->pagev[0]->mirror_num;
1409         ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1410                                     page->recover->map_length,
1411                                     mirror_num, 0);
1412         if (ret)
1413                 return ret;
1414
1415         wait_for_completion_io(&done);
1416         return blk_status_to_errno(bio->bi_status);
1417 }
1418
1419 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1420                                           struct scrub_block *sblock)
1421 {
1422         struct scrub_page *first_page = sblock->pagev[0];
1423         struct bio *bio;
1424         int page_num;
1425
1426         /* All pages in sblock belong to the same stripe on the same device. */
1427         ASSERT(first_page->dev);
1428         if (!first_page->dev->bdev)
1429                 goto out;
1430
1431         bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1432         bio_set_dev(bio, first_page->dev->bdev);
1433
1434         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1435                 struct scrub_page *page = sblock->pagev[page_num];
1436
1437                 WARN_ON(!page->page);
1438                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1439         }
1440
1441         if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1442                 bio_put(bio);
1443                 goto out;
1444         }
1445
1446         bio_put(bio);
1447
1448         scrub_recheck_block_checksum(sblock);
1449
1450         return;
1451 out:
1452         for (page_num = 0; page_num < sblock->page_count; page_num++)
1453                 sblock->pagev[page_num]->io_error = 1;
1454
1455         sblock->no_io_error_seen = 0;
1456 }
1457
1458 /*
1459  * this function will check the on disk data for checksum errors, header
1460  * errors and read I/O errors. If any I/O errors happen, the exact pages
1461  * which are errored are marked as being bad. The goal is to enable scrub
1462  * to take those pages that are not errored from all the mirrors so that
1463  * the pages that are errored in the just handled mirror can be repaired.
1464  */
1465 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1466                                 struct scrub_block *sblock,
1467                                 int retry_failed_mirror)
1468 {
1469         int page_num;
1470
1471         sblock->no_io_error_seen = 1;
1472
1473         /* short cut for raid56 */
1474         if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1475                 return scrub_recheck_block_on_raid56(fs_info, sblock);
1476
1477         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1478                 struct bio *bio;
1479                 struct scrub_page *page = sblock->pagev[page_num];
1480
1481                 if (page->dev->bdev == NULL) {
1482                         page->io_error = 1;
1483                         sblock->no_io_error_seen = 0;
1484                         continue;
1485                 }
1486
1487                 WARN_ON(!page->page);
1488                 bio = btrfs_io_bio_alloc(1);
1489                 bio_set_dev(bio, page->dev->bdev);
1490
1491                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1492                 bio->bi_iter.bi_sector = page->physical >> 9;
1493                 bio->bi_opf = REQ_OP_READ;
1494
1495                 if (btrfsic_submit_bio_wait(bio)) {
1496                         page->io_error = 1;
1497                         sblock->no_io_error_seen = 0;
1498                 }
1499
1500                 bio_put(bio);
1501         }
1502
1503         if (sblock->no_io_error_seen)
1504                 scrub_recheck_block_checksum(sblock);
1505 }
1506
1507 static inline int scrub_check_fsid(u8 fsid[],
1508                                    struct scrub_page *spage)
1509 {
1510         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1511         int ret;
1512
1513         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1514         return !ret;
1515 }
1516
1517 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1518 {
1519         sblock->header_error = 0;
1520         sblock->checksum_error = 0;
1521         sblock->generation_error = 0;
1522
1523         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1524                 scrub_checksum_data(sblock);
1525         else
1526                 scrub_checksum_tree_block(sblock);
1527 }
1528
1529 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1530                                              struct scrub_block *sblock_good)
1531 {
1532         int page_num;
1533         int ret = 0;
1534
1535         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1536                 int ret_sub;
1537
1538                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1539                                                            sblock_good,
1540                                                            page_num, 1);
1541                 if (ret_sub)
1542                         ret = ret_sub;
1543         }
1544
1545         return ret;
1546 }
1547
1548 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1549                                             struct scrub_block *sblock_good,
1550                                             int page_num, int force_write)
1551 {
1552         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1553         struct scrub_page *page_good = sblock_good->pagev[page_num];
1554         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1555
1556         BUG_ON(page_bad->page == NULL);
1557         BUG_ON(page_good->page == NULL);
1558         if (force_write || sblock_bad->header_error ||
1559             sblock_bad->checksum_error || page_bad->io_error) {
1560                 struct bio *bio;
1561                 int ret;
1562
1563                 if (!page_bad->dev->bdev) {
1564                         btrfs_warn_rl(fs_info,
1565                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1566                         return -EIO;
1567                 }
1568
1569                 bio = btrfs_io_bio_alloc(1);
1570                 bio_set_dev(bio, page_bad->dev->bdev);
1571                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1572                 bio->bi_opf = REQ_OP_WRITE;
1573
1574                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1575                 if (PAGE_SIZE != ret) {
1576                         bio_put(bio);
1577                         return -EIO;
1578                 }
1579
1580                 if (btrfsic_submit_bio_wait(bio)) {
1581                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1582                                 BTRFS_DEV_STAT_WRITE_ERRS);
1583                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1584                         bio_put(bio);
1585                         return -EIO;
1586                 }
1587                 bio_put(bio);
1588         }
1589
1590         return 0;
1591 }
1592
1593 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1594 {
1595         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1596         int page_num;
1597
1598         /*
1599          * This block is used for the check of the parity on the source device,
1600          * so the data needn't be written into the destination device.
1601          */
1602         if (sblock->sparity)
1603                 return;
1604
1605         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1606                 int ret;
1607
1608                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1609                 if (ret)
1610                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1611         }
1612 }
1613
1614 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1615                                            int page_num)
1616 {
1617         struct scrub_page *spage = sblock->pagev[page_num];
1618
1619         BUG_ON(spage->page == NULL);
1620         if (spage->io_error) {
1621                 void *mapped_buffer = kmap_atomic(spage->page);
1622
1623                 clear_page(mapped_buffer);
1624                 flush_dcache_page(spage->page);
1625                 kunmap_atomic(mapped_buffer);
1626         }
1627         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1628 }
1629
1630 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1631                                     struct scrub_page *spage)
1632 {
1633         struct scrub_bio *sbio;
1634         int ret;
1635
1636         mutex_lock(&sctx->wr_lock);
1637 again:
1638         if (!sctx->wr_curr_bio) {
1639                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1640                                               GFP_KERNEL);
1641                 if (!sctx->wr_curr_bio) {
1642                         mutex_unlock(&sctx->wr_lock);
1643                         return -ENOMEM;
1644                 }
1645                 sctx->wr_curr_bio->sctx = sctx;
1646                 sctx->wr_curr_bio->page_count = 0;
1647         }
1648         sbio = sctx->wr_curr_bio;
1649         if (sbio->page_count == 0) {
1650                 struct bio *bio;
1651
1652                 sbio->physical = spage->physical_for_dev_replace;
1653                 sbio->logical = spage->logical;
1654                 sbio->dev = sctx->wr_tgtdev;
1655                 bio = sbio->bio;
1656                 if (!bio) {
1657                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1658                         sbio->bio = bio;
1659                 }
1660
1661                 bio->bi_private = sbio;
1662                 bio->bi_end_io = scrub_wr_bio_end_io;
1663                 bio_set_dev(bio, sbio->dev->bdev);
1664                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1665                 bio->bi_opf = REQ_OP_WRITE;
1666                 sbio->status = 0;
1667         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1668                    spage->physical_for_dev_replace ||
1669                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1670                    spage->logical) {
1671                 scrub_wr_submit(sctx);
1672                 goto again;
1673         }
1674
1675         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1676         if (ret != PAGE_SIZE) {
1677                 if (sbio->page_count < 1) {
1678                         bio_put(sbio->bio);
1679                         sbio->bio = NULL;
1680                         mutex_unlock(&sctx->wr_lock);
1681                         return -EIO;
1682                 }
1683                 scrub_wr_submit(sctx);
1684                 goto again;
1685         }
1686
1687         sbio->pagev[sbio->page_count] = spage;
1688         scrub_page_get(spage);
1689         sbio->page_count++;
1690         if (sbio->page_count == sctx->pages_per_wr_bio)
1691                 scrub_wr_submit(sctx);
1692         mutex_unlock(&sctx->wr_lock);
1693
1694         return 0;
1695 }
1696
1697 static void scrub_wr_submit(struct scrub_ctx *sctx)
1698 {
1699         struct scrub_bio *sbio;
1700
1701         if (!sctx->wr_curr_bio)
1702                 return;
1703
1704         sbio = sctx->wr_curr_bio;
1705         sctx->wr_curr_bio = NULL;
1706         WARN_ON(!sbio->bio->bi_disk);
1707         scrub_pending_bio_inc(sctx);
1708         /* process all writes in a single worker thread. Then the block layer
1709          * orders the requests before sending them to the driver which
1710          * doubled the write performance on spinning disks when measured
1711          * with Linux 3.5 */
1712         btrfsic_submit_bio(sbio->bio);
1713 }
1714
1715 static void scrub_wr_bio_end_io(struct bio *bio)
1716 {
1717         struct scrub_bio *sbio = bio->bi_private;
1718         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1719
1720         sbio->status = bio->bi_status;
1721         sbio->bio = bio;
1722
1723         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1724                          scrub_wr_bio_end_io_worker, NULL, NULL);
1725         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1726 }
1727
1728 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1729 {
1730         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1731         struct scrub_ctx *sctx = sbio->sctx;
1732         int i;
1733
1734         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1735         if (sbio->status) {
1736                 struct btrfs_dev_replace *dev_replace =
1737                         &sbio->sctx->fs_info->dev_replace;
1738
1739                 for (i = 0; i < sbio->page_count; i++) {
1740                         struct scrub_page *spage = sbio->pagev[i];
1741
1742                         spage->io_error = 1;
1743                         atomic64_inc(&dev_replace->num_write_errors);
1744                 }
1745         }
1746
1747         for (i = 0; i < sbio->page_count; i++)
1748                 scrub_page_put(sbio->pagev[i]);
1749
1750         bio_put(sbio->bio);
1751         kfree(sbio);
1752         scrub_pending_bio_dec(sctx);
1753 }
1754
1755 static int scrub_checksum(struct scrub_block *sblock)
1756 {
1757         u64 flags;
1758         int ret;
1759
1760         /*
1761          * No need to initialize these stats currently,
1762          * because this function only use return value
1763          * instead of these stats value.
1764          *
1765          * Todo:
1766          * always use stats
1767          */
1768         sblock->header_error = 0;
1769         sblock->generation_error = 0;
1770         sblock->checksum_error = 0;
1771
1772         WARN_ON(sblock->page_count < 1);
1773         flags = sblock->pagev[0]->flags;
1774         ret = 0;
1775         if (flags & BTRFS_EXTENT_FLAG_DATA)
1776                 ret = scrub_checksum_data(sblock);
1777         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1778                 ret = scrub_checksum_tree_block(sblock);
1779         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1780                 (void)scrub_checksum_super(sblock);
1781         else
1782                 WARN_ON(1);
1783         if (ret)
1784                 scrub_handle_errored_block(sblock);
1785
1786         return ret;
1787 }
1788
1789 static int scrub_checksum_data(struct scrub_block *sblock)
1790 {
1791         struct scrub_ctx *sctx = sblock->sctx;
1792         struct btrfs_fs_info *fs_info = sctx->fs_info;
1793         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1794         u8 csum[BTRFS_CSUM_SIZE];
1795         u8 *on_disk_csum;
1796         struct page *page;
1797         void *buffer;
1798         u64 len;
1799         int index;
1800
1801         BUG_ON(sblock->page_count < 1);
1802         if (!sblock->pagev[0]->have_csum)
1803                 return 0;
1804
1805         shash->tfm = fs_info->csum_shash;
1806         crypto_shash_init(shash);
1807
1808         on_disk_csum = sblock->pagev[0]->csum;
1809         page = sblock->pagev[0]->page;
1810         buffer = kmap_atomic(page);
1811
1812         len = sctx->fs_info->sectorsize;
1813         index = 0;
1814         for (;;) {
1815                 u64 l = min_t(u64, len, PAGE_SIZE);
1816
1817                 crypto_shash_update(shash, buffer, l);
1818                 kunmap_atomic(buffer);
1819                 len -= l;
1820                 if (len == 0)
1821                         break;
1822                 index++;
1823                 BUG_ON(index >= sblock->page_count);
1824                 BUG_ON(!sblock->pagev[index]->page);
1825                 page = sblock->pagev[index]->page;
1826                 buffer = kmap_atomic(page);
1827         }
1828
1829         crypto_shash_final(shash, csum);
1830         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1831                 sblock->checksum_error = 1;
1832
1833         return sblock->checksum_error;
1834 }
1835
1836 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1837 {
1838         struct scrub_ctx *sctx = sblock->sctx;
1839         struct btrfs_header *h;
1840         struct btrfs_fs_info *fs_info = sctx->fs_info;
1841         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1842         u8 calculated_csum[BTRFS_CSUM_SIZE];
1843         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1844         struct page *page;
1845         void *mapped_buffer;
1846         u64 mapped_size;
1847         void *p;
1848         u64 len;
1849         int index;
1850
1851         shash->tfm = fs_info->csum_shash;
1852         crypto_shash_init(shash);
1853
1854         BUG_ON(sblock->page_count < 1);
1855         page = sblock->pagev[0]->page;
1856         mapped_buffer = kmap_atomic(page);
1857         h = (struct btrfs_header *)mapped_buffer;
1858         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1859
1860         /*
1861          * we don't use the getter functions here, as we
1862          * a) don't have an extent buffer and
1863          * b) the page is already kmapped
1864          */
1865         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1866                 sblock->header_error = 1;
1867
1868         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1869                 sblock->header_error = 1;
1870                 sblock->generation_error = 1;
1871         }
1872
1873         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1874                 sblock->header_error = 1;
1875
1876         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1877                    BTRFS_UUID_SIZE))
1878                 sblock->header_error = 1;
1879
1880         len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1881         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1882         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1883         index = 0;
1884         for (;;) {
1885                 u64 l = min_t(u64, len, mapped_size);
1886
1887                 crypto_shash_update(shash, p, l);
1888                 kunmap_atomic(mapped_buffer);
1889                 len -= l;
1890                 if (len == 0)
1891                         break;
1892                 index++;
1893                 BUG_ON(index >= sblock->page_count);
1894                 BUG_ON(!sblock->pagev[index]->page);
1895                 page = sblock->pagev[index]->page;
1896                 mapped_buffer = kmap_atomic(page);
1897                 mapped_size = PAGE_SIZE;
1898                 p = mapped_buffer;
1899         }
1900
1901         crypto_shash_final(shash, calculated_csum);
1902         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1903                 sblock->checksum_error = 1;
1904
1905         return sblock->header_error || sblock->checksum_error;
1906 }
1907
1908 static int scrub_checksum_super(struct scrub_block *sblock)
1909 {
1910         struct btrfs_super_block *s;
1911         struct scrub_ctx *sctx = sblock->sctx;
1912         struct btrfs_fs_info *fs_info = sctx->fs_info;
1913         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1914         u8 calculated_csum[BTRFS_CSUM_SIZE];
1915         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1916         struct page *page;
1917         void *mapped_buffer;
1918         u64 mapped_size;
1919         void *p;
1920         int fail_gen = 0;
1921         int fail_cor = 0;
1922         u64 len;
1923         int index;
1924
1925         shash->tfm = fs_info->csum_shash;
1926         crypto_shash_init(shash);
1927
1928         BUG_ON(sblock->page_count < 1);
1929         page = sblock->pagev[0]->page;
1930         mapped_buffer = kmap_atomic(page);
1931         s = (struct btrfs_super_block *)mapped_buffer;
1932         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1933
1934         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1935                 ++fail_cor;
1936
1937         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1938                 ++fail_gen;
1939
1940         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1941                 ++fail_cor;
1942
1943         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1944         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1945         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1946         index = 0;
1947         for (;;) {
1948                 u64 l = min_t(u64, len, mapped_size);
1949
1950                 crypto_shash_update(shash, p, l);
1951                 kunmap_atomic(mapped_buffer);
1952                 len -= l;
1953                 if (len == 0)
1954                         break;
1955                 index++;
1956                 BUG_ON(index >= sblock->page_count);
1957                 BUG_ON(!sblock->pagev[index]->page);
1958                 page = sblock->pagev[index]->page;
1959                 mapped_buffer = kmap_atomic(page);
1960                 mapped_size = PAGE_SIZE;
1961                 p = mapped_buffer;
1962         }
1963
1964         crypto_shash_final(shash, calculated_csum);
1965         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1966                 ++fail_cor;
1967
1968         if (fail_cor + fail_gen) {
1969                 /*
1970                  * if we find an error in a super block, we just report it.
1971                  * They will get written with the next transaction commit
1972                  * anyway
1973                  */
1974                 spin_lock(&sctx->stat_lock);
1975                 ++sctx->stat.super_errors;
1976                 spin_unlock(&sctx->stat_lock);
1977                 if (fail_cor)
1978                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1979                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1980                 else
1981                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1982                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1983         }
1984
1985         return fail_cor + fail_gen;
1986 }
1987
1988 static void scrub_block_get(struct scrub_block *sblock)
1989 {
1990         refcount_inc(&sblock->refs);
1991 }
1992
1993 static void scrub_block_put(struct scrub_block *sblock)
1994 {
1995         if (refcount_dec_and_test(&sblock->refs)) {
1996                 int i;
1997
1998                 if (sblock->sparity)
1999                         scrub_parity_put(sblock->sparity);
2000
2001                 for (i = 0; i < sblock->page_count; i++)
2002                         scrub_page_put(sblock->pagev[i]);
2003                 kfree(sblock);
2004         }
2005 }
2006
2007 static void scrub_page_get(struct scrub_page *spage)
2008 {
2009         atomic_inc(&spage->refs);
2010 }
2011
2012 static void scrub_page_put(struct scrub_page *spage)
2013 {
2014         if (atomic_dec_and_test(&spage->refs)) {
2015                 if (spage->page)
2016                         __free_page(spage->page);
2017                 kfree(spage);
2018         }
2019 }
2020
2021 static void scrub_submit(struct scrub_ctx *sctx)
2022 {
2023         struct scrub_bio *sbio;
2024
2025         if (sctx->curr == -1)
2026                 return;
2027
2028         sbio = sctx->bios[sctx->curr];
2029         sctx->curr = -1;
2030         scrub_pending_bio_inc(sctx);
2031         btrfsic_submit_bio(sbio->bio);
2032 }
2033
2034 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2035                                     struct scrub_page *spage)
2036 {
2037         struct scrub_block *sblock = spage->sblock;
2038         struct scrub_bio *sbio;
2039         int ret;
2040
2041 again:
2042         /*
2043          * grab a fresh bio or wait for one to become available
2044          */
2045         while (sctx->curr == -1) {
2046                 spin_lock(&sctx->list_lock);
2047                 sctx->curr = sctx->first_free;
2048                 if (sctx->curr != -1) {
2049                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2050                         sctx->bios[sctx->curr]->next_free = -1;
2051                         sctx->bios[sctx->curr]->page_count = 0;
2052                         spin_unlock(&sctx->list_lock);
2053                 } else {
2054                         spin_unlock(&sctx->list_lock);
2055                         wait_event(sctx->list_wait, sctx->first_free != -1);
2056                 }
2057         }
2058         sbio = sctx->bios[sctx->curr];
2059         if (sbio->page_count == 0) {
2060                 struct bio *bio;
2061
2062                 sbio->physical = spage->physical;
2063                 sbio->logical = spage->logical;
2064                 sbio->dev = spage->dev;
2065                 bio = sbio->bio;
2066                 if (!bio) {
2067                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2068                         sbio->bio = bio;
2069                 }
2070
2071                 bio->bi_private = sbio;
2072                 bio->bi_end_io = scrub_bio_end_io;
2073                 bio_set_dev(bio, sbio->dev->bdev);
2074                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2075                 bio->bi_opf = REQ_OP_READ;
2076                 sbio->status = 0;
2077         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2078                    spage->physical ||
2079                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2080                    spage->logical ||
2081                    sbio->dev != spage->dev) {
2082                 scrub_submit(sctx);
2083                 goto again;
2084         }
2085
2086         sbio->pagev[sbio->page_count] = spage;
2087         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2088         if (ret != PAGE_SIZE) {
2089                 if (sbio->page_count < 1) {
2090                         bio_put(sbio->bio);
2091                         sbio->bio = NULL;
2092                         return -EIO;
2093                 }
2094                 scrub_submit(sctx);
2095                 goto again;
2096         }
2097
2098         scrub_block_get(sblock); /* one for the page added to the bio */
2099         atomic_inc(&sblock->outstanding_pages);
2100         sbio->page_count++;
2101         if (sbio->page_count == sctx->pages_per_rd_bio)
2102                 scrub_submit(sctx);
2103
2104         return 0;
2105 }
2106
2107 static void scrub_missing_raid56_end_io(struct bio *bio)
2108 {
2109         struct scrub_block *sblock = bio->bi_private;
2110         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2111
2112         if (bio->bi_status)
2113                 sblock->no_io_error_seen = 0;
2114
2115         bio_put(bio);
2116
2117         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2118 }
2119
2120 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2121 {
2122         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2123         struct scrub_ctx *sctx = sblock->sctx;
2124         struct btrfs_fs_info *fs_info = sctx->fs_info;
2125         u64 logical;
2126         struct btrfs_device *dev;
2127
2128         logical = sblock->pagev[0]->logical;
2129         dev = sblock->pagev[0]->dev;
2130
2131         if (sblock->no_io_error_seen)
2132                 scrub_recheck_block_checksum(sblock);
2133
2134         if (!sblock->no_io_error_seen) {
2135                 spin_lock(&sctx->stat_lock);
2136                 sctx->stat.read_errors++;
2137                 spin_unlock(&sctx->stat_lock);
2138                 btrfs_err_rl_in_rcu(fs_info,
2139                         "IO error rebuilding logical %llu for dev %s",
2140                         logical, rcu_str_deref(dev->name));
2141         } else if (sblock->header_error || sblock->checksum_error) {
2142                 spin_lock(&sctx->stat_lock);
2143                 sctx->stat.uncorrectable_errors++;
2144                 spin_unlock(&sctx->stat_lock);
2145                 btrfs_err_rl_in_rcu(fs_info,
2146                         "failed to rebuild valid logical %llu for dev %s",
2147                         logical, rcu_str_deref(dev->name));
2148         } else {
2149                 scrub_write_block_to_dev_replace(sblock);
2150         }
2151
2152         scrub_block_put(sblock);
2153
2154         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2155                 mutex_lock(&sctx->wr_lock);
2156                 scrub_wr_submit(sctx);
2157                 mutex_unlock(&sctx->wr_lock);
2158         }
2159
2160         scrub_pending_bio_dec(sctx);
2161 }
2162
2163 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2164 {
2165         struct scrub_ctx *sctx = sblock->sctx;
2166         struct btrfs_fs_info *fs_info = sctx->fs_info;
2167         u64 length = sblock->page_count * PAGE_SIZE;
2168         u64 logical = sblock->pagev[0]->logical;
2169         struct btrfs_bio *bbio = NULL;
2170         struct bio *bio;
2171         struct btrfs_raid_bio *rbio;
2172         int ret;
2173         int i;
2174
2175         btrfs_bio_counter_inc_blocked(fs_info);
2176         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2177                         &length, &bbio);
2178         if (ret || !bbio || !bbio->raid_map)
2179                 goto bbio_out;
2180
2181         if (WARN_ON(!sctx->is_dev_replace ||
2182                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2183                 /*
2184                  * We shouldn't be scrubbing a missing device. Even for dev
2185                  * replace, we should only get here for RAID 5/6. We either
2186                  * managed to mount something with no mirrors remaining or
2187                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2188                  */
2189                 goto bbio_out;
2190         }
2191
2192         bio = btrfs_io_bio_alloc(0);
2193         bio->bi_iter.bi_sector = logical >> 9;
2194         bio->bi_private = sblock;
2195         bio->bi_end_io = scrub_missing_raid56_end_io;
2196
2197         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2198         if (!rbio)
2199                 goto rbio_out;
2200
2201         for (i = 0; i < sblock->page_count; i++) {
2202                 struct scrub_page *spage = sblock->pagev[i];
2203
2204                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2205         }
2206
2207         btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2208                         scrub_missing_raid56_worker, NULL, NULL);
2209         scrub_block_get(sblock);
2210         scrub_pending_bio_inc(sctx);
2211         raid56_submit_missing_rbio(rbio);
2212         return;
2213
2214 rbio_out:
2215         bio_put(bio);
2216 bbio_out:
2217         btrfs_bio_counter_dec(fs_info);
2218         btrfs_put_bbio(bbio);
2219         spin_lock(&sctx->stat_lock);
2220         sctx->stat.malloc_errors++;
2221         spin_unlock(&sctx->stat_lock);
2222 }
2223
2224 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2225                        u64 physical, struct btrfs_device *dev, u64 flags,
2226                        u64 gen, int mirror_num, u8 *csum, int force,
2227                        u64 physical_for_dev_replace)
2228 {
2229         struct scrub_block *sblock;
2230         int index;
2231
2232         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2233         if (!sblock) {
2234                 spin_lock(&sctx->stat_lock);
2235                 sctx->stat.malloc_errors++;
2236                 spin_unlock(&sctx->stat_lock);
2237                 return -ENOMEM;
2238         }
2239
2240         /* one ref inside this function, plus one for each page added to
2241          * a bio later on */
2242         refcount_set(&sblock->refs, 1);
2243         sblock->sctx = sctx;
2244         sblock->no_io_error_seen = 1;
2245
2246         for (index = 0; len > 0; index++) {
2247                 struct scrub_page *spage;
2248                 u64 l = min_t(u64, len, PAGE_SIZE);
2249
2250                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2251                 if (!spage) {
2252 leave_nomem:
2253                         spin_lock(&sctx->stat_lock);
2254                         sctx->stat.malloc_errors++;
2255                         spin_unlock(&sctx->stat_lock);
2256                         scrub_block_put(sblock);
2257                         return -ENOMEM;
2258                 }
2259                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2260                 scrub_page_get(spage);
2261                 sblock->pagev[index] = spage;
2262                 spage->sblock = sblock;
2263                 spage->dev = dev;
2264                 spage->flags = flags;
2265                 spage->generation = gen;
2266                 spage->logical = logical;
2267                 spage->physical = physical;
2268                 spage->physical_for_dev_replace = physical_for_dev_replace;
2269                 spage->mirror_num = mirror_num;
2270                 if (csum) {
2271                         spage->have_csum = 1;
2272                         memcpy(spage->csum, csum, sctx->csum_size);
2273                 } else {
2274                         spage->have_csum = 0;
2275                 }
2276                 sblock->page_count++;
2277                 spage->page = alloc_page(GFP_KERNEL);
2278                 if (!spage->page)
2279                         goto leave_nomem;
2280                 len -= l;
2281                 logical += l;
2282                 physical += l;
2283                 physical_for_dev_replace += l;
2284         }
2285
2286         WARN_ON(sblock->page_count == 0);
2287         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2288                 /*
2289                  * This case should only be hit for RAID 5/6 device replace. See
2290                  * the comment in scrub_missing_raid56_pages() for details.
2291                  */
2292                 scrub_missing_raid56_pages(sblock);
2293         } else {
2294                 for (index = 0; index < sblock->page_count; index++) {
2295                         struct scrub_page *spage = sblock->pagev[index];
2296                         int ret;
2297
2298                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2299                         if (ret) {
2300                                 scrub_block_put(sblock);
2301                                 return ret;
2302                         }
2303                 }
2304
2305                 if (force)
2306                         scrub_submit(sctx);
2307         }
2308
2309         /* last one frees, either here or in bio completion for last page */
2310         scrub_block_put(sblock);
2311         return 0;
2312 }
2313
2314 static void scrub_bio_end_io(struct bio *bio)
2315 {
2316         struct scrub_bio *sbio = bio->bi_private;
2317         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2318
2319         sbio->status = bio->bi_status;
2320         sbio->bio = bio;
2321
2322         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2323 }
2324
2325 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2326 {
2327         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2328         struct scrub_ctx *sctx = sbio->sctx;
2329         int i;
2330
2331         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2332         if (sbio->status) {
2333                 for (i = 0; i < sbio->page_count; i++) {
2334                         struct scrub_page *spage = sbio->pagev[i];
2335
2336                         spage->io_error = 1;
2337                         spage->sblock->no_io_error_seen = 0;
2338                 }
2339         }
2340
2341         /* now complete the scrub_block items that have all pages completed */
2342         for (i = 0; i < sbio->page_count; i++) {
2343                 struct scrub_page *spage = sbio->pagev[i];
2344                 struct scrub_block *sblock = spage->sblock;
2345
2346                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2347                         scrub_block_complete(sblock);
2348                 scrub_block_put(sblock);
2349         }
2350
2351         bio_put(sbio->bio);
2352         sbio->bio = NULL;
2353         spin_lock(&sctx->list_lock);
2354         sbio->next_free = sctx->first_free;
2355         sctx->first_free = sbio->index;
2356         spin_unlock(&sctx->list_lock);
2357
2358         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2359                 mutex_lock(&sctx->wr_lock);
2360                 scrub_wr_submit(sctx);
2361                 mutex_unlock(&sctx->wr_lock);
2362         }
2363
2364         scrub_pending_bio_dec(sctx);
2365 }
2366
2367 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2368                                        unsigned long *bitmap,
2369                                        u64 start, u64 len)
2370 {
2371         u64 offset;
2372         u64 nsectors64;
2373         u32 nsectors;
2374         int sectorsize = sparity->sctx->fs_info->sectorsize;
2375
2376         if (len >= sparity->stripe_len) {
2377                 bitmap_set(bitmap, 0, sparity->nsectors);
2378                 return;
2379         }
2380
2381         start -= sparity->logic_start;
2382         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2383         offset = div_u64(offset, sectorsize);
2384         nsectors64 = div_u64(len, sectorsize);
2385
2386         ASSERT(nsectors64 < UINT_MAX);
2387         nsectors = (u32)nsectors64;
2388
2389         if (offset + nsectors <= sparity->nsectors) {
2390                 bitmap_set(bitmap, offset, nsectors);
2391                 return;
2392         }
2393
2394         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2395         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2396 }
2397
2398 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2399                                                    u64 start, u64 len)
2400 {
2401         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2402 }
2403
2404 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2405                                                   u64 start, u64 len)
2406 {
2407         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2408 }
2409
2410 static void scrub_block_complete(struct scrub_block *sblock)
2411 {
2412         int corrupted = 0;
2413
2414         if (!sblock->no_io_error_seen) {
2415                 corrupted = 1;
2416                 scrub_handle_errored_block(sblock);
2417         } else {
2418                 /*
2419                  * if has checksum error, write via repair mechanism in
2420                  * dev replace case, otherwise write here in dev replace
2421                  * case.
2422                  */
2423                 corrupted = scrub_checksum(sblock);
2424                 if (!corrupted && sblock->sctx->is_dev_replace)
2425                         scrub_write_block_to_dev_replace(sblock);
2426         }
2427
2428         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2429                 u64 start = sblock->pagev[0]->logical;
2430                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2431                           PAGE_SIZE;
2432
2433                 scrub_parity_mark_sectors_error(sblock->sparity,
2434                                                 start, end - start);
2435         }
2436 }
2437
2438 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2439 {
2440         struct btrfs_ordered_sum *sum = NULL;
2441         unsigned long index;
2442         unsigned long num_sectors;
2443
2444         while (!list_empty(&sctx->csum_list)) {
2445                 sum = list_first_entry(&sctx->csum_list,
2446                                        struct btrfs_ordered_sum, list);
2447                 if (sum->bytenr > logical)
2448                         return 0;
2449                 if (sum->bytenr + sum->len > logical)
2450                         break;
2451
2452                 ++sctx->stat.csum_discards;
2453                 list_del(&sum->list);
2454                 kfree(sum);
2455                 sum = NULL;
2456         }
2457         if (!sum)
2458                 return 0;
2459
2460         index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2461         ASSERT(index < UINT_MAX);
2462
2463         num_sectors = sum->len / sctx->fs_info->sectorsize;
2464         memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2465         if (index == num_sectors - 1) {
2466                 list_del(&sum->list);
2467                 kfree(sum);
2468         }
2469         return 1;
2470 }
2471
2472 /* scrub extent tries to collect up to 64 kB for each bio */
2473 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2474                         u64 logical, u64 len,
2475                         u64 physical, struct btrfs_device *dev, u64 flags,
2476                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2477 {
2478         int ret;
2479         u8 csum[BTRFS_CSUM_SIZE];
2480         u32 blocksize;
2481
2482         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2483                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2484                         blocksize = map->stripe_len;
2485                 else
2486                         blocksize = sctx->fs_info->sectorsize;
2487                 spin_lock(&sctx->stat_lock);
2488                 sctx->stat.data_extents_scrubbed++;
2489                 sctx->stat.data_bytes_scrubbed += len;
2490                 spin_unlock(&sctx->stat_lock);
2491         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2492                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2493                         blocksize = map->stripe_len;
2494                 else
2495                         blocksize = sctx->fs_info->nodesize;
2496                 spin_lock(&sctx->stat_lock);
2497                 sctx->stat.tree_extents_scrubbed++;
2498                 sctx->stat.tree_bytes_scrubbed += len;
2499                 spin_unlock(&sctx->stat_lock);
2500         } else {
2501                 blocksize = sctx->fs_info->sectorsize;
2502                 WARN_ON(1);
2503         }
2504
2505         while (len) {
2506                 u64 l = min_t(u64, len, blocksize);
2507                 int have_csum = 0;
2508
2509                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2510                         /* push csums to sbio */
2511                         have_csum = scrub_find_csum(sctx, logical, csum);
2512                         if (have_csum == 0)
2513                                 ++sctx->stat.no_csum;
2514                 }
2515                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2516                                   mirror_num, have_csum ? csum : NULL, 0,
2517                                   physical_for_dev_replace);
2518                 if (ret)
2519                         return ret;
2520                 len -= l;
2521                 logical += l;
2522                 physical += l;
2523                 physical_for_dev_replace += l;
2524         }
2525         return 0;
2526 }
2527
2528 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2529                                   u64 logical, u64 len,
2530                                   u64 physical, struct btrfs_device *dev,
2531                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2532 {
2533         struct scrub_ctx *sctx = sparity->sctx;
2534         struct scrub_block *sblock;
2535         int index;
2536
2537         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2538         if (!sblock) {
2539                 spin_lock(&sctx->stat_lock);
2540                 sctx->stat.malloc_errors++;
2541                 spin_unlock(&sctx->stat_lock);
2542                 return -ENOMEM;
2543         }
2544
2545         /* one ref inside this function, plus one for each page added to
2546          * a bio later on */
2547         refcount_set(&sblock->refs, 1);
2548         sblock->sctx = sctx;
2549         sblock->no_io_error_seen = 1;
2550         sblock->sparity = sparity;
2551         scrub_parity_get(sparity);
2552
2553         for (index = 0; len > 0; index++) {
2554                 struct scrub_page *spage;
2555                 u64 l = min_t(u64, len, PAGE_SIZE);
2556
2557                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2558                 if (!spage) {
2559 leave_nomem:
2560                         spin_lock(&sctx->stat_lock);
2561                         sctx->stat.malloc_errors++;
2562                         spin_unlock(&sctx->stat_lock);
2563                         scrub_block_put(sblock);
2564                         return -ENOMEM;
2565                 }
2566                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2567                 /* For scrub block */
2568                 scrub_page_get(spage);
2569                 sblock->pagev[index] = spage;
2570                 /* For scrub parity */
2571                 scrub_page_get(spage);
2572                 list_add_tail(&spage->list, &sparity->spages);
2573                 spage->sblock = sblock;
2574                 spage->dev = dev;
2575                 spage->flags = flags;
2576                 spage->generation = gen;
2577                 spage->logical = logical;
2578                 spage->physical = physical;
2579                 spage->mirror_num = mirror_num;
2580                 if (csum) {
2581                         spage->have_csum = 1;
2582                         memcpy(spage->csum, csum, sctx->csum_size);
2583                 } else {
2584                         spage->have_csum = 0;
2585                 }
2586                 sblock->page_count++;
2587                 spage->page = alloc_page(GFP_KERNEL);
2588                 if (!spage->page)
2589                         goto leave_nomem;
2590                 len -= l;
2591                 logical += l;
2592                 physical += l;
2593         }
2594
2595         WARN_ON(sblock->page_count == 0);
2596         for (index = 0; index < sblock->page_count; index++) {
2597                 struct scrub_page *spage = sblock->pagev[index];
2598                 int ret;
2599
2600                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2601                 if (ret) {
2602                         scrub_block_put(sblock);
2603                         return ret;
2604                 }
2605         }
2606
2607         /* last one frees, either here or in bio completion for last page */
2608         scrub_block_put(sblock);
2609         return 0;
2610 }
2611
2612 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2613                                    u64 logical, u64 len,
2614                                    u64 physical, struct btrfs_device *dev,
2615                                    u64 flags, u64 gen, int mirror_num)
2616 {
2617         struct scrub_ctx *sctx = sparity->sctx;
2618         int ret;
2619         u8 csum[BTRFS_CSUM_SIZE];
2620         u32 blocksize;
2621
2622         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2623                 scrub_parity_mark_sectors_error(sparity, logical, len);
2624                 return 0;
2625         }
2626
2627         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2628                 blocksize = sparity->stripe_len;
2629         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2630                 blocksize = sparity->stripe_len;
2631         } else {
2632                 blocksize = sctx->fs_info->sectorsize;
2633                 WARN_ON(1);
2634         }
2635
2636         while (len) {
2637                 u64 l = min_t(u64, len, blocksize);
2638                 int have_csum = 0;
2639
2640                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2641                         /* push csums to sbio */
2642                         have_csum = scrub_find_csum(sctx, logical, csum);
2643                         if (have_csum == 0)
2644                                 goto skip;
2645                 }
2646                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2647                                              flags, gen, mirror_num,
2648                                              have_csum ? csum : NULL);
2649                 if (ret)
2650                         return ret;
2651 skip:
2652                 len -= l;
2653                 logical += l;
2654                 physical += l;
2655         }
2656         return 0;
2657 }
2658
2659 /*
2660  * Given a physical address, this will calculate it's
2661  * logical offset. if this is a parity stripe, it will return
2662  * the most left data stripe's logical offset.
2663  *
2664  * return 0 if it is a data stripe, 1 means parity stripe.
2665  */
2666 static int get_raid56_logic_offset(u64 physical, int num,
2667                                    struct map_lookup *map, u64 *offset,
2668                                    u64 *stripe_start)
2669 {
2670         int i;
2671         int j = 0;
2672         u64 stripe_nr;
2673         u64 last_offset;
2674         u32 stripe_index;
2675         u32 rot;
2676         const int data_stripes = nr_data_stripes(map);
2677
2678         last_offset = (physical - map->stripes[num].physical) * data_stripes;
2679         if (stripe_start)
2680                 *stripe_start = last_offset;
2681
2682         *offset = last_offset;
2683         for (i = 0; i < data_stripes; i++) {
2684                 *offset = last_offset + i * map->stripe_len;
2685
2686                 stripe_nr = div64_u64(*offset, map->stripe_len);
2687                 stripe_nr = div_u64(stripe_nr, data_stripes);
2688
2689                 /* Work out the disk rotation on this stripe-set */
2690                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2691                 /* calculate which stripe this data locates */
2692                 rot += i;
2693                 stripe_index = rot % map->num_stripes;
2694                 if (stripe_index == num)
2695                         return 0;
2696                 if (stripe_index < num)
2697                         j++;
2698         }
2699         *offset = last_offset + j * map->stripe_len;
2700         return 1;
2701 }
2702
2703 static void scrub_free_parity(struct scrub_parity *sparity)
2704 {
2705         struct scrub_ctx *sctx = sparity->sctx;
2706         struct scrub_page *curr, *next;
2707         int nbits;
2708
2709         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2710         if (nbits) {
2711                 spin_lock(&sctx->stat_lock);
2712                 sctx->stat.read_errors += nbits;
2713                 sctx->stat.uncorrectable_errors += nbits;
2714                 spin_unlock(&sctx->stat_lock);
2715         }
2716
2717         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2718                 list_del_init(&curr->list);
2719                 scrub_page_put(curr);
2720         }
2721
2722         kfree(sparity);
2723 }
2724
2725 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2726 {
2727         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2728                                                     work);
2729         struct scrub_ctx *sctx = sparity->sctx;
2730
2731         scrub_free_parity(sparity);
2732         scrub_pending_bio_dec(sctx);
2733 }
2734
2735 static void scrub_parity_bio_endio(struct bio *bio)
2736 {
2737         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2738         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2739
2740         if (bio->bi_status)
2741                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2742                           sparity->nsectors);
2743
2744         bio_put(bio);
2745
2746         btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2747                         scrub_parity_bio_endio_worker, NULL, NULL);
2748         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2749 }
2750
2751 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2752 {
2753         struct scrub_ctx *sctx = sparity->sctx;
2754         struct btrfs_fs_info *fs_info = sctx->fs_info;
2755         struct bio *bio;
2756         struct btrfs_raid_bio *rbio;
2757         struct btrfs_bio *bbio = NULL;
2758         u64 length;
2759         int ret;
2760
2761         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2762                            sparity->nsectors))
2763                 goto out;
2764
2765         length = sparity->logic_end - sparity->logic_start;
2766
2767         btrfs_bio_counter_inc_blocked(fs_info);
2768         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2769                                &length, &bbio);
2770         if (ret || !bbio || !bbio->raid_map)
2771                 goto bbio_out;
2772
2773         bio = btrfs_io_bio_alloc(0);
2774         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2775         bio->bi_private = sparity;
2776         bio->bi_end_io = scrub_parity_bio_endio;
2777
2778         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2779                                               length, sparity->scrub_dev,
2780                                               sparity->dbitmap,
2781                                               sparity->nsectors);
2782         if (!rbio)
2783                 goto rbio_out;
2784
2785         scrub_pending_bio_inc(sctx);
2786         raid56_parity_submit_scrub_rbio(rbio);
2787         return;
2788
2789 rbio_out:
2790         bio_put(bio);
2791 bbio_out:
2792         btrfs_bio_counter_dec(fs_info);
2793         btrfs_put_bbio(bbio);
2794         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2795                   sparity->nsectors);
2796         spin_lock(&sctx->stat_lock);
2797         sctx->stat.malloc_errors++;
2798         spin_unlock(&sctx->stat_lock);
2799 out:
2800         scrub_free_parity(sparity);
2801 }
2802
2803 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2804 {
2805         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2806 }
2807
2808 static void scrub_parity_get(struct scrub_parity *sparity)
2809 {
2810         refcount_inc(&sparity->refs);
2811 }
2812
2813 static void scrub_parity_put(struct scrub_parity *sparity)
2814 {
2815         if (!refcount_dec_and_test(&sparity->refs))
2816                 return;
2817
2818         scrub_parity_check_and_repair(sparity);
2819 }
2820
2821 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2822                                                   struct map_lookup *map,
2823                                                   struct btrfs_device *sdev,
2824                                                   struct btrfs_path *path,
2825                                                   u64 logic_start,
2826                                                   u64 logic_end)
2827 {
2828         struct btrfs_fs_info *fs_info = sctx->fs_info;
2829         struct btrfs_root *root = fs_info->extent_root;
2830         struct btrfs_root *csum_root = fs_info->csum_root;
2831         struct btrfs_extent_item *extent;
2832         struct btrfs_bio *bbio = NULL;
2833         u64 flags;
2834         int ret;
2835         int slot;
2836         struct extent_buffer *l;
2837         struct btrfs_key key;
2838         u64 generation;
2839         u64 extent_logical;
2840         u64 extent_physical;
2841         u64 extent_len;
2842         u64 mapped_length;
2843         struct btrfs_device *extent_dev;
2844         struct scrub_parity *sparity;
2845         int nsectors;
2846         int bitmap_len;
2847         int extent_mirror_num;
2848         int stop_loop = 0;
2849
2850         nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2851         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2852         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2853                           GFP_NOFS);
2854         if (!sparity) {
2855                 spin_lock(&sctx->stat_lock);
2856                 sctx->stat.malloc_errors++;
2857                 spin_unlock(&sctx->stat_lock);
2858                 return -ENOMEM;
2859         }
2860
2861         sparity->stripe_len = map->stripe_len;
2862         sparity->nsectors = nsectors;
2863         sparity->sctx = sctx;
2864         sparity->scrub_dev = sdev;
2865         sparity->logic_start = logic_start;
2866         sparity->logic_end = logic_end;
2867         refcount_set(&sparity->refs, 1);
2868         INIT_LIST_HEAD(&sparity->spages);
2869         sparity->dbitmap = sparity->bitmap;
2870         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2871
2872         ret = 0;
2873         while (logic_start < logic_end) {
2874                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2875                         key.type = BTRFS_METADATA_ITEM_KEY;
2876                 else
2877                         key.type = BTRFS_EXTENT_ITEM_KEY;
2878                 key.objectid = logic_start;
2879                 key.offset = (u64)-1;
2880
2881                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2882                 if (ret < 0)
2883                         goto out;
2884
2885                 if (ret > 0) {
2886                         ret = btrfs_previous_extent_item(root, path, 0);
2887                         if (ret < 0)
2888                                 goto out;
2889                         if (ret > 0) {
2890                                 btrfs_release_path(path);
2891                                 ret = btrfs_search_slot(NULL, root, &key,
2892                                                         path, 0, 0);
2893                                 if (ret < 0)
2894                                         goto out;
2895                         }
2896                 }
2897
2898                 stop_loop = 0;
2899                 while (1) {
2900                         u64 bytes;
2901
2902                         l = path->nodes[0];
2903                         slot = path->slots[0];
2904                         if (slot >= btrfs_header_nritems(l)) {
2905                                 ret = btrfs_next_leaf(root, path);
2906                                 if (ret == 0)
2907                                         continue;
2908                                 if (ret < 0)
2909                                         goto out;
2910
2911                                 stop_loop = 1;
2912                                 break;
2913                         }
2914                         btrfs_item_key_to_cpu(l, &key, slot);
2915
2916                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2917                             key.type != BTRFS_METADATA_ITEM_KEY)
2918                                 goto next;
2919
2920                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2921                                 bytes = fs_info->nodesize;
2922                         else
2923                                 bytes = key.offset;
2924
2925                         if (key.objectid + bytes <= logic_start)
2926                                 goto next;
2927
2928                         if (key.objectid >= logic_end) {
2929                                 stop_loop = 1;
2930                                 break;
2931                         }
2932
2933                         while (key.objectid >= logic_start + map->stripe_len)
2934                                 logic_start += map->stripe_len;
2935
2936                         extent = btrfs_item_ptr(l, slot,
2937                                                 struct btrfs_extent_item);
2938                         flags = btrfs_extent_flags(l, extent);
2939                         generation = btrfs_extent_generation(l, extent);
2940
2941                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2942                             (key.objectid < logic_start ||
2943                              key.objectid + bytes >
2944                              logic_start + map->stripe_len)) {
2945                                 btrfs_err(fs_info,
2946                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2947                                           key.objectid, logic_start);
2948                                 spin_lock(&sctx->stat_lock);
2949                                 sctx->stat.uncorrectable_errors++;
2950                                 spin_unlock(&sctx->stat_lock);
2951                                 goto next;
2952                         }
2953 again:
2954                         extent_logical = key.objectid;
2955                         extent_len = bytes;
2956
2957                         if (extent_logical < logic_start) {
2958                                 extent_len -= logic_start - extent_logical;
2959                                 extent_logical = logic_start;
2960                         }
2961
2962                         if (extent_logical + extent_len >
2963                             logic_start + map->stripe_len)
2964                                 extent_len = logic_start + map->stripe_len -
2965                                              extent_logical;
2966
2967                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2968                                                        extent_len);
2969
2970                         mapped_length = extent_len;
2971                         bbio = NULL;
2972                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2973                                         extent_logical, &mapped_length, &bbio,
2974                                         0);
2975                         if (!ret) {
2976                                 if (!bbio || mapped_length < extent_len)
2977                                         ret = -EIO;
2978                         }
2979                         if (ret) {
2980                                 btrfs_put_bbio(bbio);
2981                                 goto out;
2982                         }
2983                         extent_physical = bbio->stripes[0].physical;
2984                         extent_mirror_num = bbio->mirror_num;
2985                         extent_dev = bbio->stripes[0].dev;
2986                         btrfs_put_bbio(bbio);
2987
2988                         ret = btrfs_lookup_csums_range(csum_root,
2989                                                 extent_logical,
2990                                                 extent_logical + extent_len - 1,
2991                                                 &sctx->csum_list, 1);
2992                         if (ret)
2993                                 goto out;
2994
2995                         ret = scrub_extent_for_parity(sparity, extent_logical,
2996                                                       extent_len,
2997                                                       extent_physical,
2998                                                       extent_dev, flags,
2999                                                       generation,
3000                                                       extent_mirror_num);
3001
3002                         scrub_free_csums(sctx);
3003
3004                         if (ret)
3005                                 goto out;
3006
3007                         if (extent_logical + extent_len <
3008                             key.objectid + bytes) {
3009                                 logic_start += map->stripe_len;
3010
3011                                 if (logic_start >= logic_end) {
3012                                         stop_loop = 1;
3013                                         break;
3014                                 }
3015
3016                                 if (logic_start < key.objectid + bytes) {
3017                                         cond_resched();
3018                                         goto again;
3019                                 }
3020                         }
3021 next:
3022                         path->slots[0]++;
3023                 }
3024
3025                 btrfs_release_path(path);
3026
3027                 if (stop_loop)
3028                         break;
3029
3030                 logic_start += map->stripe_len;
3031         }
3032 out:
3033         if (ret < 0)
3034                 scrub_parity_mark_sectors_error(sparity, logic_start,
3035                                                 logic_end - logic_start);
3036         scrub_parity_put(sparity);
3037         scrub_submit(sctx);
3038         mutex_lock(&sctx->wr_lock);
3039         scrub_wr_submit(sctx);
3040         mutex_unlock(&sctx->wr_lock);
3041
3042         btrfs_release_path(path);
3043         return ret < 0 ? ret : 0;
3044 }
3045
3046 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3047                                            struct map_lookup *map,
3048                                            struct btrfs_device *scrub_dev,
3049                                            int num, u64 base, u64 length)
3050 {
3051         struct btrfs_path *path, *ppath;
3052         struct btrfs_fs_info *fs_info = sctx->fs_info;
3053         struct btrfs_root *root = fs_info->extent_root;
3054         struct btrfs_root *csum_root = fs_info->csum_root;
3055         struct btrfs_extent_item *extent;
3056         struct blk_plug plug;
3057         u64 flags;
3058         int ret;
3059         int slot;
3060         u64 nstripes;
3061         struct extent_buffer *l;
3062         u64 physical;
3063         u64 logical;
3064         u64 logic_end;
3065         u64 physical_end;
3066         u64 generation;
3067         int mirror_num;
3068         struct reada_control *reada1;
3069         struct reada_control *reada2;
3070         struct btrfs_key key;
3071         struct btrfs_key key_end;
3072         u64 increment = map->stripe_len;
3073         u64 offset;
3074         u64 extent_logical;
3075         u64 extent_physical;
3076         u64 extent_len;
3077         u64 stripe_logical;
3078         u64 stripe_end;
3079         struct btrfs_device *extent_dev;
3080         int extent_mirror_num;
3081         int stop_loop = 0;
3082
3083         physical = map->stripes[num].physical;
3084         offset = 0;
3085         nstripes = div64_u64(length, map->stripe_len);
3086         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3087                 offset = map->stripe_len * num;
3088                 increment = map->stripe_len * map->num_stripes;
3089                 mirror_num = 1;
3090         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3091                 int factor = map->num_stripes / map->sub_stripes;
3092                 offset = map->stripe_len * (num / map->sub_stripes);
3093                 increment = map->stripe_len * factor;
3094                 mirror_num = num % map->sub_stripes + 1;
3095         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3096                 increment = map->stripe_len;
3097                 mirror_num = num % map->num_stripes + 1;
3098         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3099                 increment = map->stripe_len;
3100                 mirror_num = num % map->num_stripes + 1;
3101         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3102                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3103                 increment = map->stripe_len * nr_data_stripes(map);
3104                 mirror_num = 1;
3105         } else {
3106                 increment = map->stripe_len;
3107                 mirror_num = 1;
3108         }
3109
3110         path = btrfs_alloc_path();
3111         if (!path)
3112                 return -ENOMEM;
3113
3114         ppath = btrfs_alloc_path();
3115         if (!ppath) {
3116                 btrfs_free_path(path);
3117                 return -ENOMEM;
3118         }
3119
3120         /*
3121          * work on commit root. The related disk blocks are static as
3122          * long as COW is applied. This means, it is save to rewrite
3123          * them to repair disk errors without any race conditions
3124          */
3125         path->search_commit_root = 1;
3126         path->skip_locking = 1;
3127
3128         ppath->search_commit_root = 1;
3129         ppath->skip_locking = 1;
3130         /*
3131          * trigger the readahead for extent tree csum tree and wait for
3132          * completion. During readahead, the scrub is officially paused
3133          * to not hold off transaction commits
3134          */
3135         logical = base + offset;
3136         physical_end = physical + nstripes * map->stripe_len;
3137         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3138                 get_raid56_logic_offset(physical_end, num,
3139                                         map, &logic_end, NULL);
3140                 logic_end += base;
3141         } else {
3142                 logic_end = logical + increment * nstripes;
3143         }
3144         wait_event(sctx->list_wait,
3145                    atomic_read(&sctx->bios_in_flight) == 0);
3146         scrub_blocked_if_needed(fs_info);
3147
3148         /* FIXME it might be better to start readahead at commit root */
3149         key.objectid = logical;
3150         key.type = BTRFS_EXTENT_ITEM_KEY;
3151         key.offset = (u64)0;
3152         key_end.objectid = logic_end;
3153         key_end.type = BTRFS_METADATA_ITEM_KEY;
3154         key_end.offset = (u64)-1;
3155         reada1 = btrfs_reada_add(root, &key, &key_end);
3156
3157         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3158         key.type = BTRFS_EXTENT_CSUM_KEY;
3159         key.offset = logical;
3160         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3161         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3162         key_end.offset = logic_end;
3163         reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3164
3165         if (!IS_ERR(reada1))
3166                 btrfs_reada_wait(reada1);
3167         if (!IS_ERR(reada2))
3168                 btrfs_reada_wait(reada2);
3169
3170
3171         /*
3172          * collect all data csums for the stripe to avoid seeking during
3173          * the scrub. This might currently (crc32) end up to be about 1MB
3174          */
3175         blk_start_plug(&plug);
3176
3177         /*
3178          * now find all extents for each stripe and scrub them
3179          */
3180         ret = 0;
3181         while (physical < physical_end) {
3182                 /*
3183                  * canceled?
3184                  */
3185                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3186                     atomic_read(&sctx->cancel_req)) {
3187                         ret = -ECANCELED;
3188                         goto out;
3189                 }
3190                 /*
3191                  * check to see if we have to pause
3192                  */
3193                 if (atomic_read(&fs_info->scrub_pause_req)) {
3194                         /* push queued extents */
3195                         sctx->flush_all_writes = true;
3196                         scrub_submit(sctx);
3197                         mutex_lock(&sctx->wr_lock);
3198                         scrub_wr_submit(sctx);
3199                         mutex_unlock(&sctx->wr_lock);
3200                         wait_event(sctx->list_wait,
3201                                    atomic_read(&sctx->bios_in_flight) == 0);
3202                         sctx->flush_all_writes = false;
3203                         scrub_blocked_if_needed(fs_info);
3204                 }
3205
3206                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3207                         ret = get_raid56_logic_offset(physical, num, map,
3208                                                       &logical,
3209                                                       &stripe_logical);
3210                         logical += base;
3211                         if (ret) {
3212                                 /* it is parity strip */
3213                                 stripe_logical += base;
3214                                 stripe_end = stripe_logical + increment;
3215                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3216                                                           ppath, stripe_logical,
3217                                                           stripe_end);
3218                                 if (ret)
3219                                         goto out;
3220                                 goto skip;
3221                         }
3222                 }
3223
3224                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3225                         key.type = BTRFS_METADATA_ITEM_KEY;
3226                 else
3227                         key.type = BTRFS_EXTENT_ITEM_KEY;
3228                 key.objectid = logical;
3229                 key.offset = (u64)-1;
3230
3231                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3232                 if (ret < 0)
3233                         goto out;
3234
3235                 if (ret > 0) {
3236                         ret = btrfs_previous_extent_item(root, path, 0);
3237                         if (ret < 0)
3238                                 goto out;
3239                         if (ret > 0) {
3240                                 /* there's no smaller item, so stick with the
3241                                  * larger one */
3242                                 btrfs_release_path(path);
3243                                 ret = btrfs_search_slot(NULL, root, &key,
3244                                                         path, 0, 0);
3245                                 if (ret < 0)
3246                                         goto out;
3247                         }
3248                 }
3249
3250                 stop_loop = 0;
3251                 while (1) {
3252                         u64 bytes;
3253
3254                         l = path->nodes[0];
3255                         slot = path->slots[0];
3256                         if (slot >= btrfs_header_nritems(l)) {
3257                                 ret = btrfs_next_leaf(root, path);
3258                                 if (ret == 0)
3259                                         continue;
3260                                 if (ret < 0)
3261                                         goto out;
3262
3263                                 stop_loop = 1;
3264                                 break;
3265                         }
3266                         btrfs_item_key_to_cpu(l, &key, slot);
3267
3268                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3269                             key.type != BTRFS_METADATA_ITEM_KEY)
3270                                 goto next;
3271
3272                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3273                                 bytes = fs_info->nodesize;
3274                         else
3275                                 bytes = key.offset;
3276
3277                         if (key.objectid + bytes <= logical)
3278                                 goto next;
3279
3280                         if (key.objectid >= logical + map->stripe_len) {
3281                                 /* out of this device extent */
3282                                 if (key.objectid >= logic_end)
3283                                         stop_loop = 1;
3284                                 break;
3285                         }
3286
3287                         extent = btrfs_item_ptr(l, slot,
3288                                                 struct btrfs_extent_item);
3289                         flags = btrfs_extent_flags(l, extent);
3290                         generation = btrfs_extent_generation(l, extent);
3291
3292                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3293                             (key.objectid < logical ||
3294                              key.objectid + bytes >
3295                              logical + map->stripe_len)) {
3296                                 btrfs_err(fs_info,
3297                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3298                                        key.objectid, logical);
3299                                 spin_lock(&sctx->stat_lock);
3300                                 sctx->stat.uncorrectable_errors++;
3301                                 spin_unlock(&sctx->stat_lock);
3302                                 goto next;
3303                         }
3304
3305 again:
3306                         extent_logical = key.objectid;
3307                         extent_len = bytes;
3308
3309                         /*
3310                          * trim extent to this stripe
3311                          */
3312                         if (extent_logical < logical) {
3313                                 extent_len -= logical - extent_logical;
3314                                 extent_logical = logical;
3315                         }
3316                         if (extent_logical + extent_len >
3317                             logical + map->stripe_len) {
3318                                 extent_len = logical + map->stripe_len -
3319                                              extent_logical;
3320                         }
3321
3322                         extent_physical = extent_logical - logical + physical;
3323                         extent_dev = scrub_dev;
3324                         extent_mirror_num = mirror_num;
3325                         if (sctx->is_dev_replace)
3326                                 scrub_remap_extent(fs_info, extent_logical,
3327                                                    extent_len, &extent_physical,
3328                                                    &extent_dev,
3329                                                    &extent_mirror_num);
3330
3331                         ret = btrfs_lookup_csums_range(csum_root,
3332                                                        extent_logical,
3333                                                        extent_logical +
3334                                                        extent_len - 1,
3335                                                        &sctx->csum_list, 1);
3336                         if (ret)
3337                                 goto out;
3338
3339                         ret = scrub_extent(sctx, map, extent_logical, extent_len,
3340                                            extent_physical, extent_dev, flags,
3341                                            generation, extent_mirror_num,
3342                                            extent_logical - logical + physical);
3343
3344                         scrub_free_csums(sctx);
3345
3346                         if (ret)
3347                                 goto out;
3348
3349                         if (extent_logical + extent_len <
3350                             key.objectid + bytes) {
3351                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3352                                         /*
3353                                          * loop until we find next data stripe
3354                                          * or we have finished all stripes.
3355                                          */
3356 loop:
3357                                         physical += map->stripe_len;
3358                                         ret = get_raid56_logic_offset(physical,
3359                                                         num, map, &logical,
3360                                                         &stripe_logical);
3361                                         logical += base;
3362
3363                                         if (ret && physical < physical_end) {
3364                                                 stripe_logical += base;
3365                                                 stripe_end = stripe_logical +
3366                                                                 increment;
3367                                                 ret = scrub_raid56_parity(sctx,
3368                                                         map, scrub_dev, ppath,
3369                                                         stripe_logical,
3370                                                         stripe_end);
3371                                                 if (ret)
3372                                                         goto out;
3373                                                 goto loop;
3374                                         }
3375                                 } else {
3376                                         physical += map->stripe_len;
3377                                         logical += increment;
3378                                 }
3379                                 if (logical < key.objectid + bytes) {
3380                                         cond_resched();
3381                                         goto again;
3382                                 }
3383
3384                                 if (physical >= physical_end) {
3385                                         stop_loop = 1;
3386                                         break;
3387                                 }
3388                         }
3389 next:
3390                         path->slots[0]++;
3391                 }
3392                 btrfs_release_path(path);
3393 skip:
3394                 logical += increment;
3395                 physical += map->stripe_len;
3396                 spin_lock(&sctx->stat_lock);
3397                 if (stop_loop)
3398                         sctx->stat.last_physical = map->stripes[num].physical +
3399                                                    length;
3400                 else
3401                         sctx->stat.last_physical = physical;
3402                 spin_unlock(&sctx->stat_lock);
3403                 if (stop_loop)
3404                         break;
3405         }
3406 out:
3407         /* push queued extents */
3408         scrub_submit(sctx);
3409         mutex_lock(&sctx->wr_lock);
3410         scrub_wr_submit(sctx);
3411         mutex_unlock(&sctx->wr_lock);
3412
3413         blk_finish_plug(&plug);
3414         btrfs_free_path(path);
3415         btrfs_free_path(ppath);
3416         return ret < 0 ? ret : 0;
3417 }
3418
3419 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3420                                           struct btrfs_device *scrub_dev,
3421                                           u64 chunk_offset, u64 length,
3422                                           u64 dev_offset,
3423                                           struct btrfs_block_group_cache *cache)
3424 {
3425         struct btrfs_fs_info *fs_info = sctx->fs_info;
3426         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3427         struct map_lookup *map;
3428         struct extent_map *em;
3429         int i;
3430         int ret = 0;
3431
3432         read_lock(&map_tree->lock);
3433         em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3434         read_unlock(&map_tree->lock);
3435
3436         if (!em) {
3437                 /*
3438                  * Might have been an unused block group deleted by the cleaner
3439                  * kthread or relocation.
3440                  */
3441                 spin_lock(&cache->lock);
3442                 if (!cache->removed)
3443                         ret = -EINVAL;
3444                 spin_unlock(&cache->lock);
3445
3446                 return ret;
3447         }
3448
3449         map = em->map_lookup;
3450         if (em->start != chunk_offset)
3451                 goto out;
3452
3453         if (em->len < length)
3454                 goto out;
3455
3456         for (i = 0; i < map->num_stripes; ++i) {
3457                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3458                     map->stripes[i].physical == dev_offset) {
3459                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3460                                            chunk_offset, length);
3461                         if (ret)
3462                                 goto out;
3463                 }
3464         }
3465 out:
3466         free_extent_map(em);
3467
3468         return ret;
3469 }
3470
3471 static noinline_for_stack
3472 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3473                            struct btrfs_device *scrub_dev, u64 start, u64 end)
3474 {
3475         struct btrfs_dev_extent *dev_extent = NULL;
3476         struct btrfs_path *path;
3477         struct btrfs_fs_info *fs_info = sctx->fs_info;
3478         struct btrfs_root *root = fs_info->dev_root;
3479         u64 length;
3480         u64 chunk_offset;
3481         int ret = 0;
3482         int ro_set;
3483         int slot;
3484         struct extent_buffer *l;
3485         struct btrfs_key key;
3486         struct btrfs_key found_key;
3487         struct btrfs_block_group_cache *cache;
3488         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3489
3490         path = btrfs_alloc_path();
3491         if (!path)
3492                 return -ENOMEM;
3493
3494         path->reada = READA_FORWARD;
3495         path->search_commit_root = 1;
3496         path->skip_locking = 1;
3497
3498         key.objectid = scrub_dev->devid;
3499         key.offset = 0ull;
3500         key.type = BTRFS_DEV_EXTENT_KEY;
3501
3502         while (1) {
3503                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3504                 if (ret < 0)
3505                         break;
3506                 if (ret > 0) {
3507                         if (path->slots[0] >=
3508                             btrfs_header_nritems(path->nodes[0])) {
3509                                 ret = btrfs_next_leaf(root, path);
3510                                 if (ret < 0)
3511                                         break;
3512                                 if (ret > 0) {
3513                                         ret = 0;
3514                                         break;
3515                                 }
3516                         } else {
3517                                 ret = 0;
3518                         }
3519                 }
3520
3521                 l = path->nodes[0];
3522                 slot = path->slots[0];
3523
3524                 btrfs_item_key_to_cpu(l, &found_key, slot);
3525
3526                 if (found_key.objectid != scrub_dev->devid)
3527                         break;
3528
3529                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3530                         break;
3531
3532                 if (found_key.offset >= end)
3533                         break;
3534
3535                 if (found_key.offset < key.offset)
3536                         break;
3537
3538                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3539                 length = btrfs_dev_extent_length(l, dev_extent);
3540
3541                 if (found_key.offset + length <= start)
3542                         goto skip;
3543
3544                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3545
3546                 /*
3547                  * get a reference on the corresponding block group to prevent
3548                  * the chunk from going away while we scrub it
3549                  */
3550                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3551
3552                 /* some chunks are removed but not committed to disk yet,
3553                  * continue scrubbing */
3554                 if (!cache)
3555                         goto skip;
3556
3557                 /*
3558                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3559                  * to avoid deadlock caused by:
3560                  * btrfs_inc_block_group_ro()
3561                  * -> btrfs_wait_for_commit()
3562                  * -> btrfs_commit_transaction()
3563                  * -> btrfs_scrub_pause()
3564                  */
3565                 scrub_pause_on(fs_info);
3566                 ret = btrfs_inc_block_group_ro(cache);
3567                 if (!ret && sctx->is_dev_replace) {
3568                         /*
3569                          * If we are doing a device replace wait for any tasks
3570                          * that started delalloc right before we set the block
3571                          * group to RO mode, as they might have just allocated
3572                          * an extent from it or decided they could do a nocow
3573                          * write. And if any such tasks did that, wait for their
3574                          * ordered extents to complete and then commit the
3575                          * current transaction, so that we can later see the new
3576                          * extent items in the extent tree - the ordered extents
3577                          * create delayed data references (for cow writes) when
3578                          * they complete, which will be run and insert the
3579                          * corresponding extent items into the extent tree when
3580                          * we commit the transaction they used when running
3581                          * inode.c:btrfs_finish_ordered_io(). We later use
3582                          * the commit root of the extent tree to find extents
3583                          * to copy from the srcdev into the tgtdev, and we don't
3584                          * want to miss any new extents.
3585                          */
3586                         btrfs_wait_block_group_reservations(cache);
3587                         btrfs_wait_nocow_writers(cache);
3588                         ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3589                                                        cache->key.objectid,
3590                                                        cache->key.offset);
3591                         if (ret > 0) {
3592                                 struct btrfs_trans_handle *trans;
3593
3594                                 trans = btrfs_join_transaction(root);
3595                                 if (IS_ERR(trans))
3596                                         ret = PTR_ERR(trans);
3597                                 else
3598                                         ret = btrfs_commit_transaction(trans);
3599                                 if (ret) {
3600                                         scrub_pause_off(fs_info);
3601                                         btrfs_put_block_group(cache);
3602                                         break;
3603                                 }
3604                         }
3605                 }
3606                 scrub_pause_off(fs_info);
3607
3608                 if (ret == 0) {
3609                         ro_set = 1;
3610                 } else if (ret == -ENOSPC) {
3611                         /*
3612                          * btrfs_inc_block_group_ro return -ENOSPC when it
3613                          * failed in creating new chunk for metadata.
3614                          * It is not a problem for scrub/replace, because
3615                          * metadata are always cowed, and our scrub paused
3616                          * commit_transactions.
3617                          */
3618                         ro_set = 0;
3619                 } else {
3620                         btrfs_warn(fs_info,
3621                                    "failed setting block group ro: %d", ret);
3622                         btrfs_put_block_group(cache);
3623                         break;
3624                 }
3625
3626                 down_write(&fs_info->dev_replace.rwsem);
3627                 dev_replace->cursor_right = found_key.offset + length;
3628                 dev_replace->cursor_left = found_key.offset;
3629                 dev_replace->item_needs_writeback = 1;
3630                 up_write(&dev_replace->rwsem);
3631
3632                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3633                                   found_key.offset, cache);
3634
3635                 /*
3636                  * flush, submit all pending read and write bios, afterwards
3637                  * wait for them.
3638                  * Note that in the dev replace case, a read request causes
3639                  * write requests that are submitted in the read completion
3640                  * worker. Therefore in the current situation, it is required
3641                  * that all write requests are flushed, so that all read and
3642                  * write requests are really completed when bios_in_flight
3643                  * changes to 0.
3644                  */
3645                 sctx->flush_all_writes = true;
3646                 scrub_submit(sctx);
3647                 mutex_lock(&sctx->wr_lock);
3648                 scrub_wr_submit(sctx);
3649                 mutex_unlock(&sctx->wr_lock);
3650
3651                 wait_event(sctx->list_wait,
3652                            atomic_read(&sctx->bios_in_flight) == 0);
3653
3654                 scrub_pause_on(fs_info);
3655
3656                 /*
3657                  * must be called before we decrease @scrub_paused.
3658                  * make sure we don't block transaction commit while
3659                  * we are waiting pending workers finished.
3660                  */
3661                 wait_event(sctx->list_wait,
3662                            atomic_read(&sctx->workers_pending) == 0);
3663                 sctx->flush_all_writes = false;
3664
3665                 scrub_pause_off(fs_info);
3666
3667                 down_write(&fs_info->dev_replace.rwsem);
3668                 dev_replace->cursor_left = dev_replace->cursor_right;
3669                 dev_replace->item_needs_writeback = 1;
3670                 up_write(&fs_info->dev_replace.rwsem);
3671
3672                 if (ro_set)
3673                         btrfs_dec_block_group_ro(cache);
3674
3675                 /*
3676                  * We might have prevented the cleaner kthread from deleting
3677                  * this block group if it was already unused because we raced
3678                  * and set it to RO mode first. So add it back to the unused
3679                  * list, otherwise it might not ever be deleted unless a manual
3680                  * balance is triggered or it becomes used and unused again.
3681                  */
3682                 spin_lock(&cache->lock);
3683                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3684                     btrfs_block_group_used(&cache->item) == 0) {
3685                         spin_unlock(&cache->lock);
3686                         btrfs_mark_bg_unused(cache);
3687                 } else {
3688                         spin_unlock(&cache->lock);
3689                 }
3690
3691                 btrfs_put_block_group(cache);
3692                 if (ret)
3693                         break;
3694                 if (sctx->is_dev_replace &&
3695                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3696                         ret = -EIO;
3697                         break;
3698                 }
3699                 if (sctx->stat.malloc_errors > 0) {
3700                         ret = -ENOMEM;
3701                         break;
3702                 }
3703 skip:
3704                 key.offset = found_key.offset + length;
3705                 btrfs_release_path(path);
3706         }
3707
3708         btrfs_free_path(path);
3709
3710         return ret;
3711 }
3712
3713 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3714                                            struct btrfs_device *scrub_dev)
3715 {
3716         int     i;
3717         u64     bytenr;
3718         u64     gen;
3719         int     ret;
3720         struct btrfs_fs_info *fs_info = sctx->fs_info;
3721
3722         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3723                 return -EIO;
3724
3725         /* Seed devices of a new filesystem has their own generation. */
3726         if (scrub_dev->fs_devices != fs_info->fs_devices)
3727                 gen = scrub_dev->generation;
3728         else
3729                 gen = fs_info->last_trans_committed;
3730
3731         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3732                 bytenr = btrfs_sb_offset(i);
3733                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3734                     scrub_dev->commit_total_bytes)
3735                         break;
3736
3737                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3738                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3739                                   NULL, 1, bytenr);
3740                 if (ret)
3741                         return ret;
3742         }
3743         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3744
3745         return 0;
3746 }
3747
3748 /*
3749  * get a reference count on fs_info->scrub_workers. start worker if necessary
3750  */
3751 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3752                                                 int is_dev_replace)
3753 {
3754         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3755         int max_active = fs_info->thread_pool_size;
3756
3757         lockdep_assert_held(&fs_info->scrub_lock);
3758
3759         if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3760                 ASSERT(fs_info->scrub_workers == NULL);
3761                 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3762                                 flags, is_dev_replace ? 1 : max_active, 4);
3763                 if (!fs_info->scrub_workers)
3764                         goto fail_scrub_workers;
3765
3766                 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3767                 fs_info->scrub_wr_completion_workers =
3768                         btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3769                                               max_active, 2);
3770                 if (!fs_info->scrub_wr_completion_workers)
3771                         goto fail_scrub_wr_completion_workers;
3772
3773                 ASSERT(fs_info->scrub_parity_workers == NULL);
3774                 fs_info->scrub_parity_workers =
3775                         btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3776                                               max_active, 2);
3777                 if (!fs_info->scrub_parity_workers)
3778                         goto fail_scrub_parity_workers;
3779
3780                 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3781         } else {
3782                 refcount_inc(&fs_info->scrub_workers_refcnt);
3783         }
3784         return 0;
3785
3786 fail_scrub_parity_workers:
3787         btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3788 fail_scrub_wr_completion_workers:
3789         btrfs_destroy_workqueue(fs_info->scrub_workers);
3790 fail_scrub_workers:
3791         return -ENOMEM;
3792 }
3793
3794 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3795                     u64 end, struct btrfs_scrub_progress *progress,
3796                     int readonly, int is_dev_replace)
3797 {
3798         struct scrub_ctx *sctx;
3799         int ret;
3800         struct btrfs_device *dev;
3801         unsigned int nofs_flag;
3802         struct btrfs_workqueue *scrub_workers = NULL;
3803         struct btrfs_workqueue *scrub_wr_comp = NULL;
3804         struct btrfs_workqueue *scrub_parity = NULL;
3805
3806         if (btrfs_fs_closing(fs_info))
3807                 return -EAGAIN;
3808
3809         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3810                 /*
3811                  * in this case scrub is unable to calculate the checksum
3812                  * the way scrub is implemented. Do not handle this
3813                  * situation at all because it won't ever happen.
3814                  */
3815                 btrfs_err(fs_info,
3816                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3817                        fs_info->nodesize,
3818                        BTRFS_STRIPE_LEN);
3819                 return -EINVAL;
3820         }
3821
3822         if (fs_info->sectorsize != PAGE_SIZE) {
3823                 /* not supported for data w/o checksums */
3824                 btrfs_err_rl(fs_info,
3825                            "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3826                        fs_info->sectorsize, PAGE_SIZE);
3827                 return -EINVAL;
3828         }
3829
3830         if (fs_info->nodesize >
3831             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3832             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3833                 /*
3834                  * would exhaust the array bounds of pagev member in
3835                  * struct scrub_block
3836                  */
3837                 btrfs_err(fs_info,
3838                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3839                        fs_info->nodesize,
3840                        SCRUB_MAX_PAGES_PER_BLOCK,
3841                        fs_info->sectorsize,
3842                        SCRUB_MAX_PAGES_PER_BLOCK);
3843                 return -EINVAL;
3844         }
3845
3846         /* Allocate outside of device_list_mutex */
3847         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3848         if (IS_ERR(sctx))
3849                 return PTR_ERR(sctx);
3850
3851         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3852         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3853         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3854                      !is_dev_replace)) {
3855                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3856                 ret = -ENODEV;
3857                 goto out_free_ctx;
3858         }
3859
3860         if (!is_dev_replace && !readonly &&
3861             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3862                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3863                 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3864                                 rcu_str_deref(dev->name));
3865                 ret = -EROFS;
3866                 goto out_free_ctx;
3867         }
3868
3869         mutex_lock(&fs_info->scrub_lock);
3870         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3871             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3872                 mutex_unlock(&fs_info->scrub_lock);
3873                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3874                 ret = -EIO;
3875                 goto out_free_ctx;
3876         }
3877
3878         down_read(&fs_info->dev_replace.rwsem);
3879         if (dev->scrub_ctx ||
3880             (!is_dev_replace &&
3881              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3882                 up_read(&fs_info->dev_replace.rwsem);
3883                 mutex_unlock(&fs_info->scrub_lock);
3884                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3885                 ret = -EINPROGRESS;
3886                 goto out_free_ctx;
3887         }
3888         up_read(&fs_info->dev_replace.rwsem);
3889
3890         ret = scrub_workers_get(fs_info, is_dev_replace);
3891         if (ret) {
3892                 mutex_unlock(&fs_info->scrub_lock);
3893                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3894                 goto out_free_ctx;
3895         }
3896
3897         sctx->readonly = readonly;
3898         dev->scrub_ctx = sctx;
3899         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3900
3901         /*
3902          * checking @scrub_pause_req here, we can avoid
3903          * race between committing transaction and scrubbing.
3904          */
3905         __scrub_blocked_if_needed(fs_info);
3906         atomic_inc(&fs_info->scrubs_running);
3907         mutex_unlock(&fs_info->scrub_lock);
3908
3909         /*
3910          * In order to avoid deadlock with reclaim when there is a transaction
3911          * trying to pause scrub, make sure we use GFP_NOFS for all the
3912          * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3913          * invoked by our callees. The pausing request is done when the
3914          * transaction commit starts, and it blocks the transaction until scrub
3915          * is paused (done at specific points at scrub_stripe() or right above
3916          * before incrementing fs_info->scrubs_running).
3917          */
3918         nofs_flag = memalloc_nofs_save();
3919         if (!is_dev_replace) {
3920                 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3921                 /*
3922                  * by holding device list mutex, we can
3923                  * kick off writing super in log tree sync.
3924                  */
3925                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3926                 ret = scrub_supers(sctx, dev);
3927                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3928         }
3929
3930         if (!ret)
3931                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3932         memalloc_nofs_restore(nofs_flag);
3933
3934         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3935         atomic_dec(&fs_info->scrubs_running);
3936         wake_up(&fs_info->scrub_pause_wait);
3937
3938         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3939
3940         if (progress)
3941                 memcpy(progress, &sctx->stat, sizeof(*progress));
3942
3943         if (!is_dev_replace)
3944                 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3945                         ret ? "not finished" : "finished", devid, ret);
3946
3947         mutex_lock(&fs_info->scrub_lock);
3948         dev->scrub_ctx = NULL;
3949         if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3950                 scrub_workers = fs_info->scrub_workers;
3951                 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3952                 scrub_parity = fs_info->scrub_parity_workers;
3953
3954                 fs_info->scrub_workers = NULL;
3955                 fs_info->scrub_wr_completion_workers = NULL;
3956                 fs_info->scrub_parity_workers = NULL;
3957         }
3958         mutex_unlock(&fs_info->scrub_lock);
3959
3960         btrfs_destroy_workqueue(scrub_workers);
3961         btrfs_destroy_workqueue(scrub_wr_comp);
3962         btrfs_destroy_workqueue(scrub_parity);
3963         scrub_put_ctx(sctx);
3964
3965         return ret;
3966
3967 out_free_ctx:
3968         scrub_free_ctx(sctx);
3969
3970         return ret;
3971 }
3972
3973 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3974 {
3975         mutex_lock(&fs_info->scrub_lock);
3976         atomic_inc(&fs_info->scrub_pause_req);
3977         while (atomic_read(&fs_info->scrubs_paused) !=
3978                atomic_read(&fs_info->scrubs_running)) {
3979                 mutex_unlock(&fs_info->scrub_lock);
3980                 wait_event(fs_info->scrub_pause_wait,
3981                            atomic_read(&fs_info->scrubs_paused) ==
3982                            atomic_read(&fs_info->scrubs_running));
3983                 mutex_lock(&fs_info->scrub_lock);
3984         }
3985         mutex_unlock(&fs_info->scrub_lock);
3986 }
3987
3988 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3989 {
3990         atomic_dec(&fs_info->scrub_pause_req);
3991         wake_up(&fs_info->scrub_pause_wait);
3992 }
3993
3994 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3995 {
3996         mutex_lock(&fs_info->scrub_lock);
3997         if (!atomic_read(&fs_info->scrubs_running)) {
3998                 mutex_unlock(&fs_info->scrub_lock);
3999                 return -ENOTCONN;
4000         }
4001
4002         atomic_inc(&fs_info->scrub_cancel_req);
4003         while (atomic_read(&fs_info->scrubs_running)) {
4004                 mutex_unlock(&fs_info->scrub_lock);
4005                 wait_event(fs_info->scrub_pause_wait,
4006                            atomic_read(&fs_info->scrubs_running) == 0);
4007                 mutex_lock(&fs_info->scrub_lock);
4008         }
4009         atomic_dec(&fs_info->scrub_cancel_req);
4010         mutex_unlock(&fs_info->scrub_lock);
4011
4012         return 0;
4013 }
4014
4015 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4016 {
4017         struct btrfs_fs_info *fs_info = dev->fs_info;
4018         struct scrub_ctx *sctx;
4019
4020         mutex_lock(&fs_info->scrub_lock);
4021         sctx = dev->scrub_ctx;
4022         if (!sctx) {
4023                 mutex_unlock(&fs_info->scrub_lock);
4024                 return -ENOTCONN;
4025         }
4026         atomic_inc(&sctx->cancel_req);
4027         while (dev->scrub_ctx) {
4028                 mutex_unlock(&fs_info->scrub_lock);
4029                 wait_event(fs_info->scrub_pause_wait,
4030                            dev->scrub_ctx == NULL);
4031                 mutex_lock(&fs_info->scrub_lock);
4032         }
4033         mutex_unlock(&fs_info->scrub_lock);
4034
4035         return 0;
4036 }
4037
4038 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4039                          struct btrfs_scrub_progress *progress)
4040 {
4041         struct btrfs_device *dev;
4042         struct scrub_ctx *sctx = NULL;
4043
4044         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4045         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4046         if (dev)
4047                 sctx = dev->scrub_ctx;
4048         if (sctx)
4049                 memcpy(progress, &sctx->stat, sizeof(*progress));
4050         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4051
4052         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4053 }
4054
4055 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4056                                u64 extent_logical, u64 extent_len,
4057                                u64 *extent_physical,
4058                                struct btrfs_device **extent_dev,
4059                                int *extent_mirror_num)
4060 {
4061         u64 mapped_length;
4062         struct btrfs_bio *bbio = NULL;
4063         int ret;
4064
4065         mapped_length = extent_len;
4066         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4067                               &mapped_length, &bbio, 0);
4068         if (ret || !bbio || mapped_length < extent_len ||
4069             !bbio->stripes[0].dev->bdev) {
4070                 btrfs_put_bbio(bbio);
4071                 return;
4072         }
4073
4074         *extent_physical = bbio->stripes[0].physical;
4075         *extent_mirror_num = bbio->mirror_num;
4076         *extent_dev = bbio->stripes[0].dev;
4077         btrfs_put_bbio(bbio);
4078 }