fs/btrfs/scrub.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   4  */
   5
   6 #include <linux/blkdev.h>
   7 #include <linux/ratelimit.h>
   8 #include <linux/sched/mm.h>
   9 #include <crypto/hash.h>
  10 #include "ctree.h"
  11 #include "discard.h"
  12 #include "volumes.h"
  13 #include "disk-io.h"
  14 #include "ordered-data.h"
  15 #include "transaction.h"
  16 #include "backref.h"
  17 #include "extent_io.h"
  18 #include "dev-replace.h"
  19 #include "check-integrity.h"
  20 #include "rcu-string.h"
  21 #include "raid56.h"
  22 #include "block-group.h"
  23 #include "zoned.h"
  24
  25 /*
  26  * This is only the first step towards a full-features scrub. It reads all
  27  * extent and super block and verifies the checksums. In case a bad checksum
  28  * is found or the extent cannot be read, good data will be written back if
  29  * any can be found.
  30  *
  31  * Future enhancements:
  32  *  - In case an unrepairable extent is encountered, track which files are
  33  *    affected and report them
  34  *  - track and record media errors, throw out bad devices
  35  *  - add a mode to also read unallocated space
  36  */
  37
  38 struct scrub_block;
  39 struct scrub_ctx;
  40
  41 /*
  42  * the following three values only influence the performance.
  43  * The last one configures the number of parallel and outstanding I/O
  44  * operations. The first two values configure an upper limit for the number
  45  * of (dynamically allocated) pages that are added to a bio.
  46  */
  47 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  48 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  49 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  50
  51 /*
  52  * the following value times PAGE_SIZE needs to be large enough to match the
  53  * largest node/leaf/sector size that shall be supported.
  54  * Values larger than BTRFS_STRIPE_LEN are not supported.
  55  */
  56 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  57
  58 struct scrub_recover {
  59         refcount_t              refs;
  60         struct btrfs_bio        *bbio;
  61         u64                     map_length;
  62 };
  63
  64 struct scrub_page {
  65         struct scrub_block      *sblock;
  66         struct page             *page;
  67         struct btrfs_device     *dev;
  68         struct list_head        list;
  69         u64                     flags;  /* extent flags */
  70         u64                     generation;
  71         u64                     logical;
  72         u64                     physical;
  73         u64                     physical_for_dev_replace;
  74         atomic_t                refs;
  75         u8                      mirror_num;
  76         int                     have_csum:1;
  77         int                     io_error:1;
  78         u8                      csum[BTRFS_CSUM_SIZE];
  79
  80         struct scrub_recover    *recover;
  81 };
  82
  83 struct scrub_bio {
  84         int                     index;
  85         struct scrub_ctx        *sctx;
  86         struct btrfs_device     *dev;
  87         struct bio              *bio;
  88         blk_status_t            status;
  89         u64                     logical;
  90         u64                     physical;
  91 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  92         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  93 #else
  94         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  95 #endif
  96         int                     page_count;
  97         int                     next_free;
  98         struct btrfs_work       work;
  99 };
 100
 101 struct scrub_block {
 102         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 103         int                     page_count;
 104         atomic_t                outstanding_pages;
 105         refcount_t              refs; /* free mem on transition to zero */
 106         struct scrub_ctx        *sctx;
 107         struct scrub_parity     *sparity;
 108         struct {
 109                 unsigned int    header_error:1;
 110                 unsigned int    checksum_error:1;
 111                 unsigned int    no_io_error_seen:1;
 112                 unsigned int    generation_error:1; /* also sets header_error */
 113
 114                 /* The following is for the data used to check parity */
 115                 /* It is for the data with checksum */
 116                 unsigned int    data_corrected:1;
 117         };
 118         struct btrfs_work       work;
 119 };
 120
 121 /* Used for the chunks with parity stripe such RAID5/6 */
 122 struct scrub_parity {
 123         struct scrub_ctx        *sctx;
 124
 125         struct btrfs_device     *scrub_dev;
 126
 127         u64                     logic_start;
 128
 129         u64                     logic_end;
 130
 131         int                     nsectors;
 132
 133         u32                     stripe_len;
 134
 135         refcount_t              refs;
 136
 137         struct list_head        spages;
 138
 139         /* Work of parity check and repair */
 140         struct btrfs_work       work;
 141
 142         /* Mark the parity blocks which have data */
 143         unsigned long           *dbitmap;
 144
 145         /*
 146          * Mark the parity blocks which have data, but errors happen when
 147          * read data or check data
 148          */
 149         unsigned long           *ebitmap;
 150
 151         unsigned long           bitmap[];
 152 };
 153
 154 struct scrub_ctx {
 155         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 156         struct btrfs_fs_info    *fs_info;
 157         int                     first_free;
 158         int                     curr;
 159         atomic_t                bios_in_flight;
 160         atomic_t                workers_pending;
 161         spinlock_t              list_lock;
 162         wait_queue_head_t       list_wait;
 163         struct list_head        csum_list;
 164         atomic_t                cancel_req;
 165         int                     readonly;
 166         int                     pages_per_rd_bio;
 167
 168         int                     is_dev_replace;
 169         u64                     write_pointer;
 170
 171         struct scrub_bio        *wr_curr_bio;
 172         struct mutex            wr_lock;
 173         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 174         struct btrfs_device     *wr_tgtdev;
 175         bool                    flush_all_writes;
 176
 177         /*
 178          * statistics
 179          */
 180         struct btrfs_scrub_progress stat;
 181         spinlock_t              stat_lock;
 182
 183         /*
 184          * Use a ref counter to avoid use-after-free issues. Scrub workers
 185          * decrement bios_in_flight and workers_pending and then do a wakeup
 186          * on the list_wait wait queue. We must ensure the main scrub task
 187          * doesn't free the scrub context before or while the workers are
 188          * doing the wakeup() call.
 189          */
 190         refcount_t              refs;
 191 };
 192
 193 struct scrub_warning {
 194         struct btrfs_path       *path;
 195         u64                     extent_item_size;
 196         const char              *errstr;
 197         u64                     physical;
 198         u64                     logical;
 199         struct btrfs_device     *dev;
 200 };
 201
 202 struct full_stripe_lock {
 203         struct rb_node node;
 204         u64 logical;
 205         u64 refs;
 206         struct mutex mutex;
 207 };
 208
 209 static 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, u32 len,
 238                        u64 physical, struct btrfs_device *dev, u64 flags,
 239                        u64 gen, int mirror_num, u8 *csum,
 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, u32 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 *spage)
 259 {
 260         return spage->recover &&
 261                (spage->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, u64 bytenr)
 393 {
 394         u64 ret;
 395
 396         /*
 397          * Due to chunk item size limit, full stripe length should not be
 398          * larger than U32_MAX. Just a sanity check here.
 399          */
 400         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 401
 402         /*
 403          * round_down() can only handle power of 2, while RAID56 full
 404          * stripe length can be 64KiB * n, so we need to manually round down.
 405          */
 406         ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 407                         cache->full_stripe_len + cache->start;
 408         return ret;
 409 }
 410
 411 /*
 412  * Lock a full stripe to avoid concurrency of recovery and read
 413  *
 414  * It's only used for profiles with parities (RAID5/6), for other profiles it
 415  * does nothing.
 416  *
 417  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 418  * So caller must call unlock_full_stripe() at the same context.
 419  *
 420  * Return <0 if encounters error.
 421  */
 422 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 423                             bool *locked_ret)
 424 {
 425         struct btrfs_block_group *bg_cache;
 426         struct btrfs_full_stripe_locks_tree *locks_root;
 427         struct full_stripe_lock *existing;
 428         u64 fstripe_start;
 429         int ret = 0;
 430
 431         *locked_ret = false;
 432         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 433         if (!bg_cache) {
 434                 ASSERT(0);
 435                 return -ENOENT;
 436         }
 437
 438         /* Profiles not based on parity don't need full stripe lock */
 439         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 440                 goto out;
 441         locks_root = &bg_cache->full_stripe_locks_root;
 442
 443         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 444
 445         /* Now insert the full stripe lock */
 446         mutex_lock(&locks_root->lock);
 447         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 448         mutex_unlock(&locks_root->lock);
 449         if (IS_ERR(existing)) {
 450                 ret = PTR_ERR(existing);
 451                 goto out;
 452         }
 453         mutex_lock(&existing->mutex);
 454         *locked_ret = true;
 455 out:
 456         btrfs_put_block_group(bg_cache);
 457         return ret;
 458 }
 459
 460 /*
 461  * Unlock a full stripe.
 462  *
 463  * NOTE: Caller must ensure it's the same context calling corresponding
 464  * lock_full_stripe().
 465  *
 466  * Return 0 if we unlock full stripe without problem.
 467  * Return <0 for error
 468  */
 469 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 470                               bool locked)
 471 {
 472         struct btrfs_block_group *bg_cache;
 473         struct btrfs_full_stripe_locks_tree *locks_root;
 474         struct full_stripe_lock *fstripe_lock;
 475         u64 fstripe_start;
 476         bool freeit = false;
 477         int ret = 0;
 478
 479         /* If we didn't acquire full stripe lock, no need to continue */
 480         if (!locked)
 481                 return 0;
 482
 483         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 484         if (!bg_cache) {
 485                 ASSERT(0);
 486                 return -ENOENT;
 487         }
 488         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 489                 goto out;
 490
 491         locks_root = &bg_cache->full_stripe_locks_root;
 492         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 493
 494         mutex_lock(&locks_root->lock);
 495         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 496         /* Unpaired unlock_full_stripe() detected */
 497         if (!fstripe_lock) {
 498                 WARN_ON(1);
 499                 ret = -ENOENT;
 500                 mutex_unlock(&locks_root->lock);
 501                 goto out;
 502         }
 503
 504         if (fstripe_lock->refs == 0) {
 505                 WARN_ON(1);
 506                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 507                         fstripe_lock->logical);
 508         } else {
 509                 fstripe_lock->refs--;
 510         }
 511
 512         if (fstripe_lock->refs == 0) {
 513                 rb_erase(&fstripe_lock->node, &locks_root->root);
 514                 freeit = true;
 515         }
 516         mutex_unlock(&locks_root->lock);
 517
 518         mutex_unlock(&fstripe_lock->mutex);
 519         if (freeit)
 520                 kfree(fstripe_lock);
 521 out:
 522         btrfs_put_block_group(bg_cache);
 523         return ret;
 524 }
 525
 526 static void scrub_free_csums(struct scrub_ctx *sctx)
 527 {
 528         while (!list_empty(&sctx->csum_list)) {
 529                 struct btrfs_ordered_sum *sum;
 530                 sum = list_first_entry(&sctx->csum_list,
 531                                        struct btrfs_ordered_sum, list);
 532                 list_del(&sum->list);
 533                 kfree(sum);
 534         }
 535 }
 536
 537 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 538 {
 539         int i;
 540
 541         if (!sctx)
 542                 return;
 543
 544         /* this can happen when scrub is cancelled */
 545         if (sctx->curr != -1) {
 546                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 547
 548                 for (i = 0; i < sbio->page_count; i++) {
 549                         WARN_ON(!sbio->pagev[i]->page);
 550                         scrub_block_put(sbio->pagev[i]->sblock);
 551                 }
 552                 bio_put(sbio->bio);
 553         }
 554
 555         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 556                 struct scrub_bio *sbio = sctx->bios[i];
 557
 558                 if (!sbio)
 559                         break;
 560                 kfree(sbio);
 561         }
 562
 563         kfree(sctx->wr_curr_bio);
 564         scrub_free_csums(sctx);
 565         kfree(sctx);
 566 }
 567
 568 static void scrub_put_ctx(struct scrub_ctx *sctx)
 569 {
 570         if (refcount_dec_and_test(&sctx->refs))
 571                 scrub_free_ctx(sctx);
 572 }
 573
 574 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 575                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 576 {
 577         struct scrub_ctx *sctx;
 578         int             i;
 579
 580         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 581         if (!sctx)
 582                 goto nomem;
 583         refcount_set(&sctx->refs, 1);
 584         sctx->is_dev_replace = is_dev_replace;
 585         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 586         sctx->curr = -1;
 587         sctx->fs_info = fs_info;
 588         INIT_LIST_HEAD(&sctx->csum_list);
 589         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 590                 struct scrub_bio *sbio;
 591
 592                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 593                 if (!sbio)
 594                         goto nomem;
 595                 sctx->bios[i] = sbio;
 596
 597                 sbio->index = i;
 598                 sbio->sctx = sctx;
 599                 sbio->page_count = 0;
 600                 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
 601                                 NULL);
 602
 603                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 604                         sctx->bios[i]->next_free = i + 1;
 605                 else
 606                         sctx->bios[i]->next_free = -1;
 607         }
 608         sctx->first_free = 0;
 609         atomic_set(&sctx->bios_in_flight, 0);
 610         atomic_set(&sctx->workers_pending, 0);
 611         atomic_set(&sctx->cancel_req, 0);
 612
 613         spin_lock_init(&sctx->list_lock);
 614         spin_lock_init(&sctx->stat_lock);
 615         init_waitqueue_head(&sctx->list_wait);
 616
 617         WARN_ON(sctx->wr_curr_bio != NULL);
 618         mutex_init(&sctx->wr_lock);
 619         sctx->wr_curr_bio = NULL;
 620         if (is_dev_replace) {
 621                 WARN_ON(!fs_info->dev_replace.tgtdev);
 622                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 623                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 624                 sctx->flush_all_writes = false;
 625         }
 626
 627         return sctx;
 628
 629 nomem:
 630         scrub_free_ctx(sctx);
 631         return ERR_PTR(-ENOMEM);
 632 }
 633
 634 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 635                                      void *warn_ctx)
 636 {
 637         u64 isize;
 638         u32 nlink;
 639         int ret;
 640         int i;
 641         unsigned nofs_flag;
 642         struct extent_buffer *eb;
 643         struct btrfs_inode_item *inode_item;
 644         struct scrub_warning *swarn = warn_ctx;
 645         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 646         struct inode_fs_paths *ipath = NULL;
 647         struct btrfs_root *local_root;
 648         struct btrfs_key key;
 649
 650         local_root = btrfs_get_fs_root(fs_info, root, true);
 651         if (IS_ERR(local_root)) {
 652                 ret = PTR_ERR(local_root);
 653                 goto err;
 654         }
 655
 656         /*
 657          * this makes the path point to (inum INODE_ITEM ioff)
 658          */
 659         key.objectid = inum;
 660         key.type = BTRFS_INODE_ITEM_KEY;
 661         key.offset = 0;
 662
 663         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 664         if (ret) {
 665                 btrfs_put_root(local_root);
 666                 btrfs_release_path(swarn->path);
 667                 goto err;
 668         }
 669
 670         eb = swarn->path->nodes[0];
 671         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 672                                         struct btrfs_inode_item);
 673         isize = btrfs_inode_size(eb, inode_item);
 674         nlink = btrfs_inode_nlink(eb, inode_item);
 675         btrfs_release_path(swarn->path);
 676
 677         /*
 678          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 679          * uses GFP_NOFS in this context, so we keep it consistent but it does
 680          * not seem to be strictly necessary.
 681          */
 682         nofs_flag = memalloc_nofs_save();
 683         ipath = init_ipath(4096, local_root, swarn->path);
 684         memalloc_nofs_restore(nofs_flag);
 685         if (IS_ERR(ipath)) {
 686                 btrfs_put_root(local_root);
 687                 ret = PTR_ERR(ipath);
 688                 ipath = NULL;
 689                 goto err;
 690         }
 691         ret = paths_from_inode(inum, ipath);
 692
 693         if (ret < 0)
 694                 goto err;
 695
 696         /*
 697          * we deliberately ignore the bit ipath might have been too small to
 698          * hold all of the paths here
 699          */
 700         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 701                 btrfs_warn_in_rcu(fs_info,
 702 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 703                                   swarn->errstr, swarn->logical,
 704                                   rcu_str_deref(swarn->dev->name),
 705                                   swarn->physical,
 706                                   root, inum, offset,
 707                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
 708                                   (char *)(unsigned long)ipath->fspath->val[i]);
 709
 710         btrfs_put_root(local_root);
 711         free_ipath(ipath);
 712         return 0;
 713
 714 err:
 715         btrfs_warn_in_rcu(fs_info,
 716                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 717                           swarn->errstr, swarn->logical,
 718                           rcu_str_deref(swarn->dev->name),
 719                           swarn->physical,
 720                           root, inum, offset, ret);
 721
 722         free_ipath(ipath);
 723         return 0;
 724 }
 725
 726 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 727 {
 728         struct btrfs_device *dev;
 729         struct btrfs_fs_info *fs_info;
 730         struct btrfs_path *path;
 731         struct btrfs_key found_key;
 732         struct extent_buffer *eb;
 733         struct btrfs_extent_item *ei;
 734         struct scrub_warning swarn;
 735         unsigned long ptr = 0;
 736         u64 extent_item_pos;
 737         u64 flags = 0;
 738         u64 ref_root;
 739         u32 item_size;
 740         u8 ref_level = 0;
 741         int ret;
 742
 743         WARN_ON(sblock->page_count < 1);
 744         dev = sblock->pagev[0]->dev;
 745         fs_info = sblock->sctx->fs_info;
 746
 747         path = btrfs_alloc_path();
 748         if (!path)
 749                 return;
 750
 751         swarn.physical = sblock->pagev[0]->physical;
 752         swarn.logical = sblock->pagev[0]->logical;
 753         swarn.errstr = errstr;
 754         swarn.dev = NULL;
 755
 756         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 757                                   &flags);
 758         if (ret < 0)
 759                 goto out;
 760
 761         extent_item_pos = swarn.logical - found_key.objectid;
 762         swarn.extent_item_size = found_key.offset;
 763
 764         eb = path->nodes[0];
 765         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 766         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 767
 768         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 769                 do {
 770                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 771                                                       item_size, &ref_root,
 772                                                       &ref_level);
 773                         btrfs_warn_in_rcu(fs_info,
 774 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 775                                 errstr, swarn.logical,
 776                                 rcu_str_deref(dev->name),
 777                                 swarn.physical,
 778                                 ref_level ? "node" : "leaf",
 779                                 ret < 0 ? -1 : ref_level,
 780                                 ret < 0 ? -1 : ref_root);
 781                 } while (ret != 1);
 782                 btrfs_release_path(path);
 783         } else {
 784                 btrfs_release_path(path);
 785                 swarn.path = path;
 786                 swarn.dev = dev;
 787                 iterate_extent_inodes(fs_info, found_key.objectid,
 788                                         extent_item_pos, 1,
 789                                         scrub_print_warning_inode, &swarn, false);
 790         }
 791
 792 out:
 793         btrfs_free_path(path);
 794 }
 795
 796 static inline void scrub_get_recover(struct scrub_recover *recover)
 797 {
 798         refcount_inc(&recover->refs);
 799 }
 800
 801 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 802                                      struct scrub_recover *recover)
 803 {
 804         if (refcount_dec_and_test(&recover->refs)) {
 805                 btrfs_bio_counter_dec(fs_info);
 806                 btrfs_put_bbio(recover->bbio);
 807                 kfree(recover);
 808         }
 809 }
 810
 811 /*
 812  * scrub_handle_errored_block gets called when either verification of the
 813  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 814  * case, this function handles all pages in the bio, even though only one
 815  * may be bad.
 816  * The goal of this function is to repair the errored block by using the
 817  * contents of one of the mirrors.
 818  */
 819 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 820 {
 821         struct scrub_ctx *sctx = sblock_to_check->sctx;
 822         struct btrfs_device *dev;
 823         struct btrfs_fs_info *fs_info;
 824         u64 logical;
 825         unsigned int failed_mirror_index;
 826         unsigned int is_metadata;
 827         unsigned int have_csum;
 828         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 829         struct scrub_block *sblock_bad;
 830         int ret;
 831         int mirror_index;
 832         int page_num;
 833         int success;
 834         bool full_stripe_locked;
 835         unsigned int nofs_flag;
 836         static DEFINE_RATELIMIT_STATE(rs, DEFAULT_RATELIMIT_INTERVAL,
 837                                       DEFAULT_RATELIMIT_BURST);
 838
 839         BUG_ON(sblock_to_check->page_count < 1);
 840         fs_info = sctx->fs_info;
 841         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 842                 /*
 843                  * if we find an error in a super block, we just report it.
 844                  * They will get written with the next transaction commit
 845                  * anyway
 846                  */
 847                 spin_lock(&sctx->stat_lock);
 848                 ++sctx->stat.super_errors;
 849                 spin_unlock(&sctx->stat_lock);
 850                 return 0;
 851         }
 852         logical = sblock_to_check->pagev[0]->logical;
 853         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 854         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 855         is_metadata = !(sblock_to_check->pagev[0]->flags &
 856                         BTRFS_EXTENT_FLAG_DATA);
 857         have_csum = sblock_to_check->pagev[0]->have_csum;
 858         dev = sblock_to_check->pagev[0]->dev;
 859
 860         if (btrfs_is_zoned(fs_info) && !sctx->is_dev_replace)
 861                 return btrfs_repair_one_zone(fs_info, logical);
 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 *spage_bad = sblock_bad->pagev[page_num];
1097                 struct scrub_block *sblock_other = NULL;
1098
1099                 /* skip no-io-error page in scrub */
1100                 if (!spage_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 (spage_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                                 spage_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 *spage;
1330
1331                         sblock = sblocks_for_recheck + mirror_index;
1332                         sblock->sctx = sctx;
1333
1334                         spage = kzalloc(sizeof(*spage), GFP_NOFS);
1335                         if (!spage) {
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(spage);
1344                         sblock->pagev[page_index] = spage;
1345                         spage->sblock = sblock;
1346                         spage->flags = flags;
1347                         spage->generation = generation;
1348                         spage->logical = logical;
1349                         spage->have_csum = have_csum;
1350                         if (have_csum)
1351                                 memcpy(spage->csum,
1352                                        original_sblock->pagev[0]->csum,
1353                                        sctx->fs_info->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                         spage->physical = bbio->stripes[stripe_index].physical +
1365                                          stripe_offset;
1366                         spage->dev = bbio->stripes[stripe_index].dev;
1367
1368                         BUG_ON(page_index >= original_sblock->page_count);
1369                         spage->physical_for_dev_replace =
1370                                 original_sblock->pagev[page_index]->
1371                                 physical_for_dev_replace;
1372                         /* for missing devices, dev->bdev is NULL */
1373                         spage->mirror_num = mirror_index + 1;
1374                         sblock->page_count++;
1375                         spage->page = alloc_page(GFP_NOFS);
1376                         if (!spage->page)
1377                                 goto leave_nomem;
1378
1379                         scrub_get_recover(recover);
1380                         spage->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 *spage)
1399 {
1400         DECLARE_COMPLETION_ONSTACK(done);
1401         int ret;
1402         int mirror_num;
1403
1404         bio->bi_iter.bi_sector = spage->logical >> 9;
1405         bio->bi_private = &done;
1406         bio->bi_end_io = scrub_bio_wait_endio;
1407
1408         mirror_num = spage->sblock->pagev[0]->mirror_num;
1409         ret = raid56_parity_recover(fs_info, bio, spage->recover->bbio,
1410                                     spage->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 *spage = sblock->pagev[page_num];
1436
1437                 WARN_ON(!spage->page);
1438                 bio_add_page(bio, spage->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 *spage = sblock->pagev[page_num];
1480
1481                 if (spage->dev->bdev == NULL) {
1482                         spage->io_error = 1;
1483                         sblock->no_io_error_seen = 0;
1484                         continue;
1485                 }
1486
1487                 WARN_ON(!spage->page);
1488                 bio = btrfs_io_bio_alloc(1);
1489                 bio_set_dev(bio, spage->dev->bdev);
1490
1491                 bio_add_page(bio, spage->page, PAGE_SIZE, 0);
1492                 bio->bi_iter.bi_sector = spage->physical >> 9;
1493                 bio->bi_opf = REQ_OP_READ;
1494
1495                 if (btrfsic_submit_bio_wait(bio)) {
1496                         spage->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 *spage_bad = sblock_bad->pagev[page_num];
1553         struct scrub_page *spage_good = sblock_good->pagev[page_num];
1554         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1555
1556         BUG_ON(spage_bad->page == NULL);
1557         BUG_ON(spage_good->page == NULL);
1558         if (force_write || sblock_bad->header_error ||
1559             sblock_bad->checksum_error || spage_bad->io_error) {
1560                 struct bio *bio;
1561                 int ret;
1562
1563                 if (!spage_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, spage_bad->dev->bdev);
1571                 bio->bi_iter.bi_sector = spage_bad->physical >> 9;
1572                 bio->bi_opf = REQ_OP_WRITE;
1573
1574                 ret = bio_add_page(bio, spage_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(spage_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                 clear_page(page_address(spage->page));
1622
1623         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1624 }
1625
1626 static int fill_writer_pointer_gap(struct scrub_ctx *sctx, u64 physical)
1627 {
1628         int ret = 0;
1629         u64 length;
1630
1631         if (!btrfs_is_zoned(sctx->fs_info))
1632                 return 0;
1633
1634         if (!btrfs_dev_is_sequential(sctx->wr_tgtdev, physical))
1635                 return 0;
1636
1637         if (sctx->write_pointer < physical) {
1638                 length = physical - sctx->write_pointer;
1639
1640                 ret = btrfs_zoned_issue_zeroout(sctx->wr_tgtdev,
1641                                                 sctx->write_pointer, length);
1642                 if (!ret)
1643                         sctx->write_pointer = physical;
1644         }
1645         return ret;
1646 }
1647
1648 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1649                                     struct scrub_page *spage)
1650 {
1651         struct scrub_bio *sbio;
1652         int ret;
1653
1654         mutex_lock(&sctx->wr_lock);
1655 again:
1656         if (!sctx->wr_curr_bio) {
1657                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1658                                               GFP_KERNEL);
1659                 if (!sctx->wr_curr_bio) {
1660                         mutex_unlock(&sctx->wr_lock);
1661                         return -ENOMEM;
1662                 }
1663                 sctx->wr_curr_bio->sctx = sctx;
1664                 sctx->wr_curr_bio->page_count = 0;
1665         }
1666         sbio = sctx->wr_curr_bio;
1667         if (sbio->page_count == 0) {
1668                 struct bio *bio;
1669
1670                 ret = fill_writer_pointer_gap(sctx,
1671                                               spage->physical_for_dev_replace);
1672                 if (ret) {
1673                         mutex_unlock(&sctx->wr_lock);
1674                         return ret;
1675                 }
1676
1677                 sbio->physical = spage->physical_for_dev_replace;
1678                 sbio->logical = spage->logical;
1679                 sbio->dev = sctx->wr_tgtdev;
1680                 bio = sbio->bio;
1681                 if (!bio) {
1682                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1683                         sbio->bio = bio;
1684                 }
1685
1686                 bio->bi_private = sbio;
1687                 bio->bi_end_io = scrub_wr_bio_end_io;
1688                 bio_set_dev(bio, sbio->dev->bdev);
1689                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1690                 bio->bi_opf = REQ_OP_WRITE;
1691                 sbio->status = 0;
1692         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1693                    spage->physical_for_dev_replace ||
1694                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1695                    spage->logical) {
1696                 scrub_wr_submit(sctx);
1697                 goto again;
1698         }
1699
1700         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1701         if (ret != PAGE_SIZE) {
1702                 if (sbio->page_count < 1) {
1703                         bio_put(sbio->bio);
1704                         sbio->bio = NULL;
1705                         mutex_unlock(&sctx->wr_lock);
1706                         return -EIO;
1707                 }
1708                 scrub_wr_submit(sctx);
1709                 goto again;
1710         }
1711
1712         sbio->pagev[sbio->page_count] = spage;
1713         scrub_page_get(spage);
1714         sbio->page_count++;
1715         if (sbio->page_count == sctx->pages_per_wr_bio)
1716                 scrub_wr_submit(sctx);
1717         mutex_unlock(&sctx->wr_lock);
1718
1719         return 0;
1720 }
1721
1722 static void scrub_wr_submit(struct scrub_ctx *sctx)
1723 {
1724         struct scrub_bio *sbio;
1725
1726         if (!sctx->wr_curr_bio)
1727                 return;
1728
1729         sbio = sctx->wr_curr_bio;
1730         sctx->wr_curr_bio = NULL;
1731         WARN_ON(!sbio->bio->bi_bdev);
1732         scrub_pending_bio_inc(sctx);
1733         /* process all writes in a single worker thread. Then the block layer
1734          * orders the requests before sending them to the driver which
1735          * doubled the write performance on spinning disks when measured
1736          * with Linux 3.5 */
1737         btrfsic_submit_bio(sbio->bio);
1738
1739         if (btrfs_is_zoned(sctx->fs_info))
1740                 sctx->write_pointer = sbio->physical + sbio->page_count * PAGE_SIZE;
1741 }
1742
1743 static void scrub_wr_bio_end_io(struct bio *bio)
1744 {
1745         struct scrub_bio *sbio = bio->bi_private;
1746         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1747
1748         sbio->status = bio->bi_status;
1749         sbio->bio = bio;
1750
1751         btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1752         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1753 }
1754
1755 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1756 {
1757         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1758         struct scrub_ctx *sctx = sbio->sctx;
1759         int i;
1760
1761         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1762         if (sbio->status) {
1763                 struct btrfs_dev_replace *dev_replace =
1764                         &sbio->sctx->fs_info->dev_replace;
1765
1766                 for (i = 0; i < sbio->page_count; i++) {
1767                         struct scrub_page *spage = sbio->pagev[i];
1768
1769                         spage->io_error = 1;
1770                         atomic64_inc(&dev_replace->num_write_errors);
1771                 }
1772         }
1773
1774         for (i = 0; i < sbio->page_count; i++)
1775                 scrub_page_put(sbio->pagev[i]);
1776
1777         bio_put(sbio->bio);
1778         kfree(sbio);
1779         scrub_pending_bio_dec(sctx);
1780 }
1781
1782 static int scrub_checksum(struct scrub_block *sblock)
1783 {
1784         u64 flags;
1785         int ret;
1786
1787         /*
1788          * No need to initialize these stats currently,
1789          * because this function only use return value
1790          * instead of these stats value.
1791          *
1792          * Todo:
1793          * always use stats
1794          */
1795         sblock->header_error = 0;
1796         sblock->generation_error = 0;
1797         sblock->checksum_error = 0;
1798
1799         WARN_ON(sblock->page_count < 1);
1800         flags = sblock->pagev[0]->flags;
1801         ret = 0;
1802         if (flags & BTRFS_EXTENT_FLAG_DATA)
1803                 ret = scrub_checksum_data(sblock);
1804         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1805                 ret = scrub_checksum_tree_block(sblock);
1806         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1807                 (void)scrub_checksum_super(sblock);
1808         else
1809                 WARN_ON(1);
1810         if (ret)
1811                 scrub_handle_errored_block(sblock);
1812
1813         return ret;
1814 }
1815
1816 static int scrub_checksum_data(struct scrub_block *sblock)
1817 {
1818         struct scrub_ctx *sctx = sblock->sctx;
1819         struct btrfs_fs_info *fs_info = sctx->fs_info;
1820         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1821         u8 csum[BTRFS_CSUM_SIZE];
1822         struct scrub_page *spage;
1823         char *kaddr;
1824
1825         BUG_ON(sblock->page_count < 1);
1826         spage = sblock->pagev[0];
1827         if (!spage->have_csum)
1828                 return 0;
1829
1830         kaddr = page_address(spage->page);
1831
1832         shash->tfm = fs_info->csum_shash;
1833         crypto_shash_init(shash);
1834
1835         /*
1836          * In scrub_pages() and scrub_pages_for_parity() we ensure each spage
1837          * only contains one sector of data.
1838          */
1839         crypto_shash_digest(shash, kaddr, fs_info->sectorsize, csum);
1840
1841         if (memcmp(csum, spage->csum, fs_info->csum_size))
1842                 sblock->checksum_error = 1;
1843         return sblock->checksum_error;
1844 }
1845
1846 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1847 {
1848         struct scrub_ctx *sctx = sblock->sctx;
1849         struct btrfs_header *h;
1850         struct btrfs_fs_info *fs_info = sctx->fs_info;
1851         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1852         u8 calculated_csum[BTRFS_CSUM_SIZE];
1853         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1854         /*
1855          * This is done in sectorsize steps even for metadata as there's a
1856          * constraint for nodesize to be aligned to sectorsize. This will need
1857          * to change so we don't misuse data and metadata units like that.
1858          */
1859         const u32 sectorsize = sctx->fs_info->sectorsize;
1860         const int num_sectors = fs_info->nodesize >> fs_info->sectorsize_bits;
1861         int i;
1862         struct scrub_page *spage;
1863         char *kaddr;
1864
1865         BUG_ON(sblock->page_count < 1);
1866
1867         /* Each member in pagev is just one block, not a full page */
1868         ASSERT(sblock->page_count == num_sectors);
1869
1870         spage = sblock->pagev[0];
1871         kaddr = page_address(spage->page);
1872         h = (struct btrfs_header *)kaddr;
1873         memcpy(on_disk_csum, h->csum, sctx->fs_info->csum_size);
1874
1875         /*
1876          * we don't use the getter functions here, as we
1877          * a) don't have an extent buffer and
1878          * b) the page is already kmapped
1879          */
1880         if (spage->logical != btrfs_stack_header_bytenr(h))
1881                 sblock->header_error = 1;
1882
1883         if (spage->generation != btrfs_stack_header_generation(h)) {
1884                 sblock->header_error = 1;
1885                 sblock->generation_error = 1;
1886         }
1887
1888         if (!scrub_check_fsid(h->fsid, spage))
1889                 sblock->header_error = 1;
1890
1891         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1892                    BTRFS_UUID_SIZE))
1893                 sblock->header_error = 1;
1894
1895         shash->tfm = fs_info->csum_shash;
1896         crypto_shash_init(shash);
1897         crypto_shash_update(shash, kaddr + BTRFS_CSUM_SIZE,
1898                             sectorsize - BTRFS_CSUM_SIZE);
1899
1900         for (i = 1; i < num_sectors; i++) {
1901                 kaddr = page_address(sblock->pagev[i]->page);
1902                 crypto_shash_update(shash, kaddr, sectorsize);
1903         }
1904
1905         crypto_shash_final(shash, calculated_csum);
1906         if (memcmp(calculated_csum, on_disk_csum, sctx->fs_info->csum_size))
1907                 sblock->checksum_error = 1;
1908
1909         return sblock->header_error || sblock->checksum_error;
1910 }
1911
1912 static int scrub_checksum_super(struct scrub_block *sblock)
1913 {
1914         struct btrfs_super_block *s;
1915         struct scrub_ctx *sctx = sblock->sctx;
1916         struct btrfs_fs_info *fs_info = sctx->fs_info;
1917         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1918         u8 calculated_csum[BTRFS_CSUM_SIZE];
1919         struct scrub_page *spage;
1920         char *kaddr;
1921         int fail_gen = 0;
1922         int fail_cor = 0;
1923
1924         BUG_ON(sblock->page_count < 1);
1925         spage = sblock->pagev[0];
1926         kaddr = page_address(spage->page);
1927         s = (struct btrfs_super_block *)kaddr;
1928
1929         if (spage->logical != btrfs_super_bytenr(s))
1930                 ++fail_cor;
1931
1932         if (spage->generation != btrfs_super_generation(s))
1933                 ++fail_gen;
1934
1935         if (!scrub_check_fsid(s->fsid, spage))
1936                 ++fail_cor;
1937
1938         shash->tfm = fs_info->csum_shash;
1939         crypto_shash_init(shash);
1940         crypto_shash_digest(shash, kaddr + BTRFS_CSUM_SIZE,
1941                         BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE, calculated_csum);
1942
1943         if (memcmp(calculated_csum, s->csum, sctx->fs_info->csum_size))
1944                 ++fail_cor;
1945
1946         if (fail_cor + fail_gen) {
1947                 /*
1948                  * if we find an error in a super block, we just report it.
1949                  * They will get written with the next transaction commit
1950                  * anyway
1951                  */
1952                 spin_lock(&sctx->stat_lock);
1953                 ++sctx->stat.super_errors;
1954                 spin_unlock(&sctx->stat_lock);
1955                 if (fail_cor)
1956                         btrfs_dev_stat_inc_and_print(spage->dev,
1957                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1958                 else
1959                         btrfs_dev_stat_inc_and_print(spage->dev,
1960                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1961         }
1962
1963         return fail_cor + fail_gen;
1964 }
1965
1966 static void scrub_block_get(struct scrub_block *sblock)
1967 {
1968         refcount_inc(&sblock->refs);
1969 }
1970
1971 static void scrub_block_put(struct scrub_block *sblock)
1972 {
1973         if (refcount_dec_and_test(&sblock->refs)) {
1974                 int i;
1975
1976                 if (sblock->sparity)
1977                         scrub_parity_put(sblock->sparity);
1978
1979                 for (i = 0; i < sblock->page_count; i++)
1980                         scrub_page_put(sblock->pagev[i]);
1981                 kfree(sblock);
1982         }
1983 }
1984
1985 static void scrub_page_get(struct scrub_page *spage)
1986 {
1987         atomic_inc(&spage->refs);
1988 }
1989
1990 static void scrub_page_put(struct scrub_page *spage)
1991 {
1992         if (atomic_dec_and_test(&spage->refs)) {
1993                 if (spage->page)
1994                         __free_page(spage->page);
1995                 kfree(spage);
1996         }
1997 }
1998
1999 static void scrub_submit(struct scrub_ctx *sctx)
2000 {
2001         struct scrub_bio *sbio;
2002
2003         if (sctx->curr == -1)
2004                 return;
2005
2006         sbio = sctx->bios[sctx->curr];
2007         sctx->curr = -1;
2008         scrub_pending_bio_inc(sctx);
2009         btrfsic_submit_bio(sbio->bio);
2010 }
2011
2012 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2013                                     struct scrub_page *spage)
2014 {
2015         struct scrub_block *sblock = spage->sblock;
2016         struct scrub_bio *sbio;
2017         int ret;
2018
2019 again:
2020         /*
2021          * grab a fresh bio or wait for one to become available
2022          */
2023         while (sctx->curr == -1) {
2024                 spin_lock(&sctx->list_lock);
2025                 sctx->curr = sctx->first_free;
2026                 if (sctx->curr != -1) {
2027                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2028                         sctx->bios[sctx->curr]->next_free = -1;
2029                         sctx->bios[sctx->curr]->page_count = 0;
2030                         spin_unlock(&sctx->list_lock);
2031                 } else {
2032                         spin_unlock(&sctx->list_lock);
2033                         wait_event(sctx->list_wait, sctx->first_free != -1);
2034                 }
2035         }
2036         sbio = sctx->bios[sctx->curr];
2037         if (sbio->page_count == 0) {
2038                 struct bio *bio;
2039
2040                 sbio->physical = spage->physical;
2041                 sbio->logical = spage->logical;
2042                 sbio->dev = spage->dev;
2043                 bio = sbio->bio;
2044                 if (!bio) {
2045                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2046                         sbio->bio = bio;
2047                 }
2048
2049                 bio->bi_private = sbio;
2050                 bio->bi_end_io = scrub_bio_end_io;
2051                 bio_set_dev(bio, sbio->dev->bdev);
2052                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2053                 bio->bi_opf = REQ_OP_READ;
2054                 sbio->status = 0;
2055         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2056                    spage->physical ||
2057                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2058                    spage->logical ||
2059                    sbio->dev != spage->dev) {
2060                 scrub_submit(sctx);
2061                 goto again;
2062         }
2063
2064         sbio->pagev[sbio->page_count] = spage;
2065         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2066         if (ret != PAGE_SIZE) {
2067                 if (sbio->page_count < 1) {
2068                         bio_put(sbio->bio);
2069                         sbio->bio = NULL;
2070                         return -EIO;
2071                 }
2072                 scrub_submit(sctx);
2073                 goto again;
2074         }
2075
2076         scrub_block_get(sblock); /* one for the page added to the bio */
2077         atomic_inc(&sblock->outstanding_pages);
2078         sbio->page_count++;
2079         if (sbio->page_count == sctx->pages_per_rd_bio)
2080                 scrub_submit(sctx);
2081
2082         return 0;
2083 }
2084
2085 static void scrub_missing_raid56_end_io(struct bio *bio)
2086 {
2087         struct scrub_block *sblock = bio->bi_private;
2088         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2089
2090         if (bio->bi_status)
2091                 sblock->no_io_error_seen = 0;
2092
2093         bio_put(bio);
2094
2095         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2096 }
2097
2098 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2099 {
2100         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2101         struct scrub_ctx *sctx = sblock->sctx;
2102         struct btrfs_fs_info *fs_info = sctx->fs_info;
2103         u64 logical;
2104         struct btrfs_device *dev;
2105
2106         logical = sblock->pagev[0]->logical;
2107         dev = sblock->pagev[0]->dev;
2108
2109         if (sblock->no_io_error_seen)
2110                 scrub_recheck_block_checksum(sblock);
2111
2112         if (!sblock->no_io_error_seen) {
2113                 spin_lock(&sctx->stat_lock);
2114                 sctx->stat.read_errors++;
2115                 spin_unlock(&sctx->stat_lock);
2116                 btrfs_err_rl_in_rcu(fs_info,
2117                         "IO error rebuilding logical %llu for dev %s",
2118                         logical, rcu_str_deref(dev->name));
2119         } else if (sblock->header_error || sblock->checksum_error) {
2120                 spin_lock(&sctx->stat_lock);
2121                 sctx->stat.uncorrectable_errors++;
2122                 spin_unlock(&sctx->stat_lock);
2123                 btrfs_err_rl_in_rcu(fs_info,
2124                         "failed to rebuild valid logical %llu for dev %s",
2125                         logical, rcu_str_deref(dev->name));
2126         } else {
2127                 scrub_write_block_to_dev_replace(sblock);
2128         }
2129
2130         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2131                 mutex_lock(&sctx->wr_lock);
2132                 scrub_wr_submit(sctx);
2133                 mutex_unlock(&sctx->wr_lock);
2134         }
2135
2136         scrub_block_put(sblock);
2137         scrub_pending_bio_dec(sctx);
2138 }
2139
2140 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2141 {
2142         struct scrub_ctx *sctx = sblock->sctx;
2143         struct btrfs_fs_info *fs_info = sctx->fs_info;
2144         u64 length = sblock->page_count * PAGE_SIZE;
2145         u64 logical = sblock->pagev[0]->logical;
2146         struct btrfs_bio *bbio = NULL;
2147         struct bio *bio;
2148         struct btrfs_raid_bio *rbio;
2149         int ret;
2150         int i;
2151
2152         btrfs_bio_counter_inc_blocked(fs_info);
2153         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2154                         &length, &bbio);
2155         if (ret || !bbio || !bbio->raid_map)
2156                 goto bbio_out;
2157
2158         if (WARN_ON(!sctx->is_dev_replace ||
2159                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2160                 /*
2161                  * We shouldn't be scrubbing a missing device. Even for dev
2162                  * replace, we should only get here for RAID 5/6. We either
2163                  * managed to mount something with no mirrors remaining or
2164                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2165                  */
2166                 goto bbio_out;
2167         }
2168
2169         bio = btrfs_io_bio_alloc(0);
2170         bio->bi_iter.bi_sector = logical >> 9;
2171         bio->bi_private = sblock;
2172         bio->bi_end_io = scrub_missing_raid56_end_io;
2173
2174         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2175         if (!rbio)
2176                 goto rbio_out;
2177
2178         for (i = 0; i < sblock->page_count; i++) {
2179                 struct scrub_page *spage = sblock->pagev[i];
2180
2181                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2182         }
2183
2184         btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2185         scrub_block_get(sblock);
2186         scrub_pending_bio_inc(sctx);
2187         raid56_submit_missing_rbio(rbio);
2188         return;
2189
2190 rbio_out:
2191         bio_put(bio);
2192 bbio_out:
2193         btrfs_bio_counter_dec(fs_info);
2194         btrfs_put_bbio(bbio);
2195         spin_lock(&sctx->stat_lock);
2196         sctx->stat.malloc_errors++;
2197         spin_unlock(&sctx->stat_lock);
2198 }
2199
2200 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u32 len,
2201                        u64 physical, struct btrfs_device *dev, u64 flags,
2202                        u64 gen, int mirror_num, u8 *csum,
2203                        u64 physical_for_dev_replace)
2204 {
2205         struct scrub_block *sblock;
2206         const u32 sectorsize = sctx->fs_info->sectorsize;
2207         int index;
2208
2209         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2210         if (!sblock) {
2211                 spin_lock(&sctx->stat_lock);
2212                 sctx->stat.malloc_errors++;
2213                 spin_unlock(&sctx->stat_lock);
2214                 return -ENOMEM;
2215         }
2216
2217         /* one ref inside this function, plus one for each page added to
2218          * a bio later on */
2219         refcount_set(&sblock->refs, 1);
2220         sblock->sctx = sctx;
2221         sblock->no_io_error_seen = 1;
2222
2223         for (index = 0; len > 0; index++) {
2224                 struct scrub_page *spage;
2225                 /*
2226                  * Here we will allocate one page for one sector to scrub.
2227                  * This is fine if PAGE_SIZE == sectorsize, but will cost
2228                  * more memory for PAGE_SIZE > sectorsize case.
2229                  */
2230                 u32 l = min(sectorsize, len);
2231
2232                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2233                 if (!spage) {
2234 leave_nomem:
2235                         spin_lock(&sctx->stat_lock);
2236                         sctx->stat.malloc_errors++;
2237                         spin_unlock(&sctx->stat_lock);
2238                         scrub_block_put(sblock);
2239                         return -ENOMEM;
2240                 }
2241                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2242                 scrub_page_get(spage);
2243                 sblock->pagev[index] = spage;
2244                 spage->sblock = sblock;
2245                 spage->dev = dev;
2246                 spage->flags = flags;
2247                 spage->generation = gen;
2248                 spage->logical = logical;
2249                 spage->physical = physical;
2250                 spage->physical_for_dev_replace = physical_for_dev_replace;
2251                 spage->mirror_num = mirror_num;
2252                 if (csum) {
2253                         spage->have_csum = 1;
2254                         memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2255                 } else {
2256                         spage->have_csum = 0;
2257                 }
2258                 sblock->page_count++;
2259                 spage->page = alloc_page(GFP_KERNEL);
2260                 if (!spage->page)
2261                         goto leave_nomem;
2262                 len -= l;
2263                 logical += l;
2264                 physical += l;
2265                 physical_for_dev_replace += l;
2266         }
2267
2268         WARN_ON(sblock->page_count == 0);
2269         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2270                 /*
2271                  * This case should only be hit for RAID 5/6 device replace. See
2272                  * the comment in scrub_missing_raid56_pages() for details.
2273                  */
2274                 scrub_missing_raid56_pages(sblock);
2275         } else {
2276                 for (index = 0; index < sblock->page_count; index++) {
2277                         struct scrub_page *spage = sblock->pagev[index];
2278                         int ret;
2279
2280                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2281                         if (ret) {
2282                                 scrub_block_put(sblock);
2283                                 return ret;
2284                         }
2285                 }
2286
2287                 if (flags & BTRFS_EXTENT_FLAG_SUPER)
2288                         scrub_submit(sctx);
2289         }
2290
2291         /* last one frees, either here or in bio completion for last page */
2292         scrub_block_put(sblock);
2293         return 0;
2294 }
2295
2296 static void scrub_bio_end_io(struct bio *bio)
2297 {
2298         struct scrub_bio *sbio = bio->bi_private;
2299         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2300
2301         sbio->status = bio->bi_status;
2302         sbio->bio = bio;
2303
2304         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2305 }
2306
2307 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2308 {
2309         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2310         struct scrub_ctx *sctx = sbio->sctx;
2311         int i;
2312
2313         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2314         if (sbio->status) {
2315                 for (i = 0; i < sbio->page_count; i++) {
2316                         struct scrub_page *spage = sbio->pagev[i];
2317
2318                         spage->io_error = 1;
2319                         spage->sblock->no_io_error_seen = 0;
2320                 }
2321         }
2322
2323         /* now complete the scrub_block items that have all pages completed */
2324         for (i = 0; i < sbio->page_count; i++) {
2325                 struct scrub_page *spage = sbio->pagev[i];
2326                 struct scrub_block *sblock = spage->sblock;
2327
2328                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2329                         scrub_block_complete(sblock);
2330                 scrub_block_put(sblock);
2331         }
2332
2333         bio_put(sbio->bio);
2334         sbio->bio = NULL;
2335         spin_lock(&sctx->list_lock);
2336         sbio->next_free = sctx->first_free;
2337         sctx->first_free = sbio->index;
2338         spin_unlock(&sctx->list_lock);
2339
2340         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2341                 mutex_lock(&sctx->wr_lock);
2342                 scrub_wr_submit(sctx);
2343                 mutex_unlock(&sctx->wr_lock);
2344         }
2345
2346         scrub_pending_bio_dec(sctx);
2347 }
2348
2349 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2350                                        unsigned long *bitmap,
2351                                        u64 start, u32 len)
2352 {
2353         u64 offset;
2354         u32 nsectors;
2355         u32 sectorsize_bits = sparity->sctx->fs_info->sectorsize_bits;
2356
2357         if (len >= sparity->stripe_len) {
2358                 bitmap_set(bitmap, 0, sparity->nsectors);
2359                 return;
2360         }
2361
2362         start -= sparity->logic_start;
2363         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2364         offset = offset >> sectorsize_bits;
2365         nsectors = len >> sectorsize_bits;
2366
2367         if (offset + nsectors <= sparity->nsectors) {
2368                 bitmap_set(bitmap, offset, nsectors);
2369                 return;
2370         }
2371
2372         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2373         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2374 }
2375
2376 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2377                                                    u64 start, u32 len)
2378 {
2379         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2380 }
2381
2382 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2383                                                   u64 start, u32 len)
2384 {
2385         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2386 }
2387
2388 static void scrub_block_complete(struct scrub_block *sblock)
2389 {
2390         int corrupted = 0;
2391
2392         if (!sblock->no_io_error_seen) {
2393                 corrupted = 1;
2394                 scrub_handle_errored_block(sblock);
2395         } else {
2396                 /*
2397                  * if has checksum error, write via repair mechanism in
2398                  * dev replace case, otherwise write here in dev replace
2399                  * case.
2400                  */
2401                 corrupted = scrub_checksum(sblock);
2402                 if (!corrupted && sblock->sctx->is_dev_replace)
2403                         scrub_write_block_to_dev_replace(sblock);
2404         }
2405
2406         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2407                 u64 start = sblock->pagev[0]->logical;
2408                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2409                           PAGE_SIZE;
2410
2411                 ASSERT(end - start <= U32_MAX);
2412                 scrub_parity_mark_sectors_error(sblock->sparity,
2413                                                 start, end - start);
2414         }
2415 }
2416
2417 static void drop_csum_range(struct scrub_ctx *sctx, struct btrfs_ordered_sum *sum)
2418 {
2419         sctx->stat.csum_discards += sum->len >> sctx->fs_info->sectorsize_bits;
2420         list_del(&sum->list);
2421         kfree(sum);
2422 }
2423
2424 /*
2425  * Find the desired csum for range [logical, logical + sectorsize), and store
2426  * the csum into @csum.
2427  *
2428  * The search source is sctx->csum_list, which is a pre-populated list
2429  * storing bytenr ordered csum ranges.  We're reponsible to cleanup any range
2430  * that is before @logical.
2431  *
2432  * Return 0 if there is no csum for the range.
2433  * Return 1 if there is csum for the range and copied to @csum.
2434  */
2435 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2436 {
2437         bool found = false;
2438
2439         while (!list_empty(&sctx->csum_list)) {
2440                 struct btrfs_ordered_sum *sum = NULL;
2441                 unsigned long index;
2442                 unsigned long num_sectors;
2443
2444                 sum = list_first_entry(&sctx->csum_list,
2445                                        struct btrfs_ordered_sum, list);
2446                 /* The current csum range is beyond our range, no csum found */
2447                 if (sum->bytenr > logical)
2448                         break;
2449
2450                 /*
2451                  * The current sum is before our bytenr, since scrub is always
2452                  * done in bytenr order, the csum will never be used anymore,
2453                  * clean it up so that later calls won't bother with the range,
2454                  * and continue search the next range.
2455                  */
2456                 if (sum->bytenr + sum->len <= logical) {
2457                         drop_csum_range(sctx, sum);
2458                         continue;
2459                 }
2460
2461                 /* Now the csum range covers our bytenr, copy the csum */
2462                 found = true;
2463                 index = (logical - sum->bytenr) >> sctx->fs_info->sectorsize_bits;
2464                 num_sectors = sum->len >> sctx->fs_info->sectorsize_bits;
2465
2466                 memcpy(csum, sum->sums + index * sctx->fs_info->csum_size,
2467                        sctx->fs_info->csum_size);
2468
2469                 /* Cleanup the range if we're at the end of the csum range */
2470                 if (index == num_sectors - 1)
2471                         drop_csum_range(sctx, sum);
2472                 break;
2473         }
2474         if (!found)
2475                 return 0;
2476         return 1;
2477 }
2478
2479 /* scrub extent tries to collect up to 64 kB for each bio */
2480 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2481                         u64 logical, u32 len,
2482                         u64 physical, struct btrfs_device *dev, u64 flags,
2483                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2484 {
2485         int ret;
2486         u8 csum[BTRFS_CSUM_SIZE];
2487         u32 blocksize;
2488
2489         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2490                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2491                         blocksize = map->stripe_len;
2492                 else
2493                         blocksize = sctx->fs_info->sectorsize;
2494                 spin_lock(&sctx->stat_lock);
2495                 sctx->stat.data_extents_scrubbed++;
2496                 sctx->stat.data_bytes_scrubbed += len;
2497                 spin_unlock(&sctx->stat_lock);
2498         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2499                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2500                         blocksize = map->stripe_len;
2501                 else
2502                         blocksize = sctx->fs_info->nodesize;
2503                 spin_lock(&sctx->stat_lock);
2504                 sctx->stat.tree_extents_scrubbed++;
2505                 sctx->stat.tree_bytes_scrubbed += len;
2506                 spin_unlock(&sctx->stat_lock);
2507         } else {
2508                 blocksize = sctx->fs_info->sectorsize;
2509                 WARN_ON(1);
2510         }
2511
2512         while (len) {
2513                 u32 l = min(len, blocksize);
2514                 int have_csum = 0;
2515
2516                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2517                         /* push csums to sbio */
2518                         have_csum = scrub_find_csum(sctx, logical, csum);
2519                         if (have_csum == 0)
2520                                 ++sctx->stat.no_csum;
2521                 }
2522                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2523                                   mirror_num, have_csum ? csum : NULL,
2524                                   physical_for_dev_replace);
2525                 if (ret)
2526                         return ret;
2527                 len -= l;
2528                 logical += l;
2529                 physical += l;
2530                 physical_for_dev_replace += l;
2531         }
2532         return 0;
2533 }
2534
2535 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2536                                   u64 logical, u32 len,
2537                                   u64 physical, struct btrfs_device *dev,
2538                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2539 {
2540         struct scrub_ctx *sctx = sparity->sctx;
2541         struct scrub_block *sblock;
2542         const u32 sectorsize = sctx->fs_info->sectorsize;
2543         int index;
2544
2545         ASSERT(IS_ALIGNED(len, sectorsize));
2546
2547         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2548         if (!sblock) {
2549                 spin_lock(&sctx->stat_lock);
2550                 sctx->stat.malloc_errors++;
2551                 spin_unlock(&sctx->stat_lock);
2552                 return -ENOMEM;
2553         }
2554
2555         /* one ref inside this function, plus one for each page added to
2556          * a bio later on */
2557         refcount_set(&sblock->refs, 1);
2558         sblock->sctx = sctx;
2559         sblock->no_io_error_seen = 1;
2560         sblock->sparity = sparity;
2561         scrub_parity_get(sparity);
2562
2563         for (index = 0; len > 0; index++) {
2564                 struct scrub_page *spage;
2565
2566                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2567                 if (!spage) {
2568 leave_nomem:
2569                         spin_lock(&sctx->stat_lock);
2570                         sctx->stat.malloc_errors++;
2571                         spin_unlock(&sctx->stat_lock);
2572                         scrub_block_put(sblock);
2573                         return -ENOMEM;
2574                 }
2575                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2576                 /* For scrub block */
2577                 scrub_page_get(spage);
2578                 sblock->pagev[index] = spage;
2579                 /* For scrub parity */
2580                 scrub_page_get(spage);
2581                 list_add_tail(&spage->list, &sparity->spages);
2582                 spage->sblock = sblock;
2583                 spage->dev = dev;
2584                 spage->flags = flags;
2585                 spage->generation = gen;
2586                 spage->logical = logical;
2587                 spage->physical = physical;
2588                 spage->mirror_num = mirror_num;
2589                 if (csum) {
2590                         spage->have_csum = 1;
2591                         memcpy(spage->csum, csum, sctx->fs_info->csum_size);
2592                 } else {
2593                         spage->have_csum = 0;
2594                 }
2595                 sblock->page_count++;
2596                 spage->page = alloc_page(GFP_KERNEL);
2597                 if (!spage->page)
2598                         goto leave_nomem;
2599
2600
2601                 /* Iterate over the stripe range in sectorsize steps */
2602                 len -= sectorsize;
2603                 logical += sectorsize;
2604                 physical += sectorsize;
2605         }
2606
2607         WARN_ON(sblock->page_count == 0);
2608         for (index = 0; index < sblock->page_count; index++) {
2609                 struct scrub_page *spage = sblock->pagev[index];
2610                 int ret;
2611
2612                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2613                 if (ret) {
2614                         scrub_block_put(sblock);
2615                         return ret;
2616                 }
2617         }
2618
2619         /* last one frees, either here or in bio completion for last page */
2620         scrub_block_put(sblock);
2621         return 0;
2622 }
2623
2624 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2625                                    u64 logical, u32 len,
2626                                    u64 physical, struct btrfs_device *dev,
2627                                    u64 flags, u64 gen, int mirror_num)
2628 {
2629         struct scrub_ctx *sctx = sparity->sctx;
2630         int ret;
2631         u8 csum[BTRFS_CSUM_SIZE];
2632         u32 blocksize;
2633
2634         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2635                 scrub_parity_mark_sectors_error(sparity, logical, len);
2636                 return 0;
2637         }
2638
2639         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2640                 blocksize = sparity->stripe_len;
2641         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2642                 blocksize = sparity->stripe_len;
2643         } else {
2644                 blocksize = sctx->fs_info->sectorsize;
2645                 WARN_ON(1);
2646         }
2647
2648         while (len) {
2649                 u32 l = min(len, blocksize);
2650                 int have_csum = 0;
2651
2652                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2653                         /* push csums to sbio */
2654                         have_csum = scrub_find_csum(sctx, logical, csum);
2655                         if (have_csum == 0)
2656                                 goto skip;
2657                 }
2658                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2659                                              flags, gen, mirror_num,
2660                                              have_csum ? csum : NULL);
2661                 if (ret)
2662                         return ret;
2663 skip:
2664                 len -= l;
2665                 logical += l;
2666                 physical += l;
2667         }
2668         return 0;
2669 }
2670
2671 /*
2672  * Given a physical address, this will calculate it's
2673  * logical offset. if this is a parity stripe, it will return
2674  * the most left data stripe's logical offset.
2675  *
2676  * return 0 if it is a data stripe, 1 means parity stripe.
2677  */
2678 static int get_raid56_logic_offset(u64 physical, int num,
2679                                    struct map_lookup *map, u64 *offset,
2680                                    u64 *stripe_start)
2681 {
2682         int i;
2683         int j = 0;
2684         u64 stripe_nr;
2685         u64 last_offset;
2686         u32 stripe_index;
2687         u32 rot;
2688         const int data_stripes = nr_data_stripes(map);
2689
2690         last_offset = (physical - map->stripes[num].physical) * data_stripes;
2691         if (stripe_start)
2692                 *stripe_start = last_offset;
2693
2694         *offset = last_offset;
2695         for (i = 0; i < data_stripes; i++) {
2696                 *offset = last_offset + i * map->stripe_len;
2697
2698                 stripe_nr = div64_u64(*offset, map->stripe_len);
2699                 stripe_nr = div_u64(stripe_nr, data_stripes);
2700
2701                 /* Work out the disk rotation on this stripe-set */
2702                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2703                 /* calculate which stripe this data locates */
2704                 rot += i;
2705                 stripe_index = rot % map->num_stripes;
2706                 if (stripe_index == num)
2707                         return 0;
2708                 if (stripe_index < num)
2709                         j++;
2710         }
2711         *offset = last_offset + j * map->stripe_len;
2712         return 1;
2713 }
2714
2715 static void scrub_free_parity(struct scrub_parity *sparity)
2716 {
2717         struct scrub_ctx *sctx = sparity->sctx;
2718         struct scrub_page *curr, *next;
2719         int nbits;
2720
2721         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2722         if (nbits) {
2723                 spin_lock(&sctx->stat_lock);
2724                 sctx->stat.read_errors += nbits;
2725                 sctx->stat.uncorrectable_errors += nbits;
2726                 spin_unlock(&sctx->stat_lock);
2727         }
2728
2729         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2730                 list_del_init(&curr->list);
2731                 scrub_page_put(curr);
2732         }
2733
2734         kfree(sparity);
2735 }
2736
2737 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2738 {
2739         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2740                                                     work);
2741         struct scrub_ctx *sctx = sparity->sctx;
2742
2743         scrub_free_parity(sparity);
2744         scrub_pending_bio_dec(sctx);
2745 }
2746
2747 static void scrub_parity_bio_endio(struct bio *bio)
2748 {
2749         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2750         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2751
2752         if (bio->bi_status)
2753                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2754                           sparity->nsectors);
2755
2756         bio_put(bio);
2757
2758         btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2759                         NULL);
2760         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2761 }
2762
2763 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2764 {
2765         struct scrub_ctx *sctx = sparity->sctx;
2766         struct btrfs_fs_info *fs_info = sctx->fs_info;
2767         struct bio *bio;
2768         struct btrfs_raid_bio *rbio;
2769         struct btrfs_bio *bbio = NULL;
2770         u64 length;
2771         int ret;
2772
2773         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2774                            sparity->nsectors))
2775                 goto out;
2776
2777         length = sparity->logic_end - sparity->logic_start;
2778
2779         btrfs_bio_counter_inc_blocked(fs_info);
2780         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2781                                &length, &bbio);
2782         if (ret || !bbio || !bbio->raid_map)
2783                 goto bbio_out;
2784
2785         bio = btrfs_io_bio_alloc(0);
2786         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2787         bio->bi_private = sparity;
2788         bio->bi_end_io = scrub_parity_bio_endio;
2789
2790         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2791                                               length, sparity->scrub_dev,
2792                                               sparity->dbitmap,
2793                                               sparity->nsectors);
2794         if (!rbio)
2795                 goto rbio_out;
2796
2797         scrub_pending_bio_inc(sctx);
2798         raid56_parity_submit_scrub_rbio(rbio);
2799         return;
2800
2801 rbio_out:
2802         bio_put(bio);
2803 bbio_out:
2804         btrfs_bio_counter_dec(fs_info);
2805         btrfs_put_bbio(bbio);
2806         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2807                   sparity->nsectors);
2808         spin_lock(&sctx->stat_lock);
2809         sctx->stat.malloc_errors++;
2810         spin_unlock(&sctx->stat_lock);
2811 out:
2812         scrub_free_parity(sparity);
2813 }
2814
2815 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2816 {
2817         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2818 }
2819
2820 static void scrub_parity_get(struct scrub_parity *sparity)
2821 {
2822         refcount_inc(&sparity->refs);
2823 }
2824
2825 static void scrub_parity_put(struct scrub_parity *sparity)
2826 {
2827         if (!refcount_dec_and_test(&sparity->refs))
2828                 return;
2829
2830         scrub_parity_check_and_repair(sparity);
2831 }
2832
2833 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2834                                                   struct map_lookup *map,
2835                                                   struct btrfs_device *sdev,
2836                                                   struct btrfs_path *path,
2837                                                   u64 logic_start,
2838                                                   u64 logic_end)
2839 {
2840         struct btrfs_fs_info *fs_info = sctx->fs_info;
2841         struct btrfs_root *root = fs_info->extent_root;
2842         struct btrfs_root *csum_root = fs_info->csum_root;
2843         struct btrfs_extent_item *extent;
2844         struct btrfs_bio *bbio = NULL;
2845         u64 flags;
2846         int ret;
2847         int slot;
2848         struct extent_buffer *l;
2849         struct btrfs_key key;
2850         u64 generation;
2851         u64 extent_logical;
2852         u64 extent_physical;
2853         /* Check the comment in scrub_stripe() for why u32 is enough here */
2854         u32 extent_len;
2855         u64 mapped_length;
2856         struct btrfs_device *extent_dev;
2857         struct scrub_parity *sparity;
2858         int nsectors;
2859         int bitmap_len;
2860         int extent_mirror_num;
2861         int stop_loop = 0;
2862
2863         ASSERT(map->stripe_len <= U32_MAX);
2864         nsectors = map->stripe_len >> fs_info->sectorsize_bits;
2865         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2866         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2867                           GFP_NOFS);
2868         if (!sparity) {
2869                 spin_lock(&sctx->stat_lock);
2870                 sctx->stat.malloc_errors++;
2871                 spin_unlock(&sctx->stat_lock);
2872                 return -ENOMEM;
2873         }
2874
2875         ASSERT(map->stripe_len <= U32_MAX);
2876         sparity->stripe_len = map->stripe_len;
2877         sparity->nsectors = nsectors;
2878         sparity->sctx = sctx;
2879         sparity->scrub_dev = sdev;
2880         sparity->logic_start = logic_start;
2881         sparity->logic_end = logic_end;
2882         refcount_set(&sparity->refs, 1);
2883         INIT_LIST_HEAD(&sparity->spages);
2884         sparity->dbitmap = sparity->bitmap;
2885         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2886
2887         ret = 0;
2888         while (logic_start < logic_end) {
2889                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2890                         key.type = BTRFS_METADATA_ITEM_KEY;
2891                 else
2892                         key.type = BTRFS_EXTENT_ITEM_KEY;
2893                 key.objectid = logic_start;
2894                 key.offset = (u64)-1;
2895
2896                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2897                 if (ret < 0)
2898                         goto out;
2899
2900                 if (ret > 0) {
2901                         ret = btrfs_previous_extent_item(root, path, 0);
2902                         if (ret < 0)
2903                                 goto out;
2904                         if (ret > 0) {
2905                                 btrfs_release_path(path);
2906                                 ret = btrfs_search_slot(NULL, root, &key,
2907                                                         path, 0, 0);
2908                                 if (ret < 0)
2909                                         goto out;
2910                         }
2911                 }
2912
2913                 stop_loop = 0;
2914                 while (1) {
2915                         u64 bytes;
2916
2917                         l = path->nodes[0];
2918                         slot = path->slots[0];
2919                         if (slot >= btrfs_header_nritems(l)) {
2920                                 ret = btrfs_next_leaf(root, path);
2921                                 if (ret == 0)
2922                                         continue;
2923                                 if (ret < 0)
2924                                         goto out;
2925
2926                                 stop_loop = 1;
2927                                 break;
2928                         }
2929                         btrfs_item_key_to_cpu(l, &key, slot);
2930
2931                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2932                             key.type != BTRFS_METADATA_ITEM_KEY)
2933                                 goto next;
2934
2935                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2936                                 bytes = fs_info->nodesize;
2937                         else
2938                                 bytes = key.offset;
2939
2940                         if (key.objectid + bytes <= logic_start)
2941                                 goto next;
2942
2943                         if (key.objectid >= logic_end) {
2944                                 stop_loop = 1;
2945                                 break;
2946                         }
2947
2948                         while (key.objectid >= logic_start + map->stripe_len)
2949                                 logic_start += map->stripe_len;
2950
2951                         extent = btrfs_item_ptr(l, slot,
2952                                                 struct btrfs_extent_item);
2953                         flags = btrfs_extent_flags(l, extent);
2954                         generation = btrfs_extent_generation(l, extent);
2955
2956                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2957                             (key.objectid < logic_start ||
2958                              key.objectid + bytes >
2959                              logic_start + map->stripe_len)) {
2960                                 btrfs_err(fs_info,
2961                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2962                                           key.objectid, logic_start);
2963                                 spin_lock(&sctx->stat_lock);
2964                                 sctx->stat.uncorrectable_errors++;
2965                                 spin_unlock(&sctx->stat_lock);
2966                                 goto next;
2967                         }
2968 again:
2969                         extent_logical = key.objectid;
2970                         ASSERT(bytes <= U32_MAX);
2971                         extent_len = bytes;
2972
2973                         if (extent_logical < logic_start) {
2974                                 extent_len -= logic_start - extent_logical;
2975                                 extent_logical = logic_start;
2976                         }
2977
2978                         if (extent_logical + extent_len >
2979                             logic_start + map->stripe_len)
2980                                 extent_len = logic_start + map->stripe_len -
2981                                              extent_logical;
2982
2983                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2984                                                        extent_len);
2985
2986                         mapped_length = extent_len;
2987                         bbio = NULL;
2988                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2989                                         extent_logical, &mapped_length, &bbio,
2990                                         0);
2991                         if (!ret) {
2992                                 if (!bbio || mapped_length < extent_len)
2993                                         ret = -EIO;
2994                         }
2995                         if (ret) {
2996                                 btrfs_put_bbio(bbio);
2997                                 goto out;
2998                         }
2999                         extent_physical = bbio->stripes[0].physical;
3000                         extent_mirror_num = bbio->mirror_num;
3001                         extent_dev = bbio->stripes[0].dev;
3002                         btrfs_put_bbio(bbio);
3003
3004                         ret = btrfs_lookup_csums_range(csum_root,
3005                                                 extent_logical,
3006                                                 extent_logical + extent_len - 1,
3007                                                 &sctx->csum_list, 1);
3008                         if (ret)
3009                                 goto out;
3010
3011                         ret = scrub_extent_for_parity(sparity, extent_logical,
3012                                                       extent_len,
3013                                                       extent_physical,
3014                                                       extent_dev, flags,
3015                                                       generation,
3016                                                       extent_mirror_num);
3017
3018                         scrub_free_csums(sctx);
3019
3020                         if (ret)
3021                                 goto out;
3022
3023                         if (extent_logical + extent_len <
3024                             key.objectid + bytes) {
3025                                 logic_start += map->stripe_len;
3026
3027                                 if (logic_start >= logic_end) {
3028                                         stop_loop = 1;
3029                                         break;
3030                                 }
3031
3032                                 if (logic_start < key.objectid + bytes) {
3033                                         cond_resched();
3034                                         goto again;
3035                                 }
3036                         }
3037 next:
3038                         path->slots[0]++;
3039                 }
3040
3041                 btrfs_release_path(path);
3042
3043                 if (stop_loop)
3044                         break;
3045
3046                 logic_start += map->stripe_len;
3047         }
3048 out:
3049         if (ret < 0) {
3050                 ASSERT(logic_end - logic_start <= U32_MAX);
3051                 scrub_parity_mark_sectors_error(sparity, logic_start,
3052                                                 logic_end - logic_start);
3053         }
3054         scrub_parity_put(sparity);
3055         scrub_submit(sctx);
3056         mutex_lock(&sctx->wr_lock);
3057         scrub_wr_submit(sctx);
3058         mutex_unlock(&sctx->wr_lock);
3059
3060         btrfs_release_path(path);
3061         return ret < 0 ? ret : 0;
3062 }
3063
3064 static void sync_replace_for_zoned(struct scrub_ctx *sctx)
3065 {
3066         if (!btrfs_is_zoned(sctx->fs_info))
3067                 return;
3068
3069         sctx->flush_all_writes = true;
3070         scrub_submit(sctx);
3071         mutex_lock(&sctx->wr_lock);
3072         scrub_wr_submit(sctx);
3073         mutex_unlock(&sctx->wr_lock);
3074
3075         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3076 }
3077
3078 static int sync_write_pointer_for_zoned(struct scrub_ctx *sctx, u64 logical,
3079                                         u64 physical, u64 physical_end)
3080 {
3081         struct btrfs_fs_info *fs_info = sctx->fs_info;
3082         int ret = 0;
3083
3084         if (!btrfs_is_zoned(fs_info))
3085                 return 0;
3086
3087         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3088
3089         mutex_lock(&sctx->wr_lock);
3090         if (sctx->write_pointer < physical_end) {
3091                 ret = btrfs_sync_zone_write_pointer(sctx->wr_tgtdev, logical,
3092                                                     physical,
3093                                                     sctx->write_pointer);
3094                 if (ret)
3095                         btrfs_err(fs_info,
3096                                   "zoned: failed to recover write pointer");
3097         }
3098         mutex_unlock(&sctx->wr_lock);
3099         btrfs_dev_clear_zone_empty(sctx->wr_tgtdev, physical);
3100
3101         return ret;
3102 }
3103
3104 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3105                                            struct map_lookup *map,
3106                                            struct btrfs_device *scrub_dev,
3107                                            int num, u64 base, u64 length,
3108                                            struct btrfs_block_group *cache)
3109 {
3110         struct btrfs_path *path, *ppath;
3111         struct btrfs_fs_info *fs_info = sctx->fs_info;
3112         struct btrfs_root *root = fs_info->extent_root;
3113         struct btrfs_root *csum_root = fs_info->csum_root;
3114         struct btrfs_extent_item *extent;
3115         struct blk_plug plug;
3116         u64 flags;
3117         int ret;
3118         int slot;
3119         u64 nstripes;
3120         struct extent_buffer *l;
3121         u64 physical;
3122         u64 logical;
3123         u64 logic_end;
3124         u64 physical_end;
3125         u64 generation;
3126         int mirror_num;
3127         struct reada_control *reada1;
3128         struct reada_control *reada2;
3129         struct btrfs_key key;
3130         struct btrfs_key key_end;
3131         u64 increment = map->stripe_len;
3132         u64 offset;
3133         u64 extent_logical;
3134         u64 extent_physical;
3135         /*
3136          * Unlike chunk length, extent length should never go beyond
3137          * BTRFS_MAX_EXTENT_SIZE, thus u32 is enough here.
3138          */
3139         u32 extent_len;
3140         u64 stripe_logical;
3141         u64 stripe_end;
3142         struct btrfs_device *extent_dev;
3143         int extent_mirror_num;
3144         int stop_loop = 0;
3145
3146         physical = map->stripes[num].physical;
3147         offset = 0;
3148         nstripes = div64_u64(length, map->stripe_len);
3149         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3150                 offset = map->stripe_len * num;
3151                 increment = map->stripe_len * map->num_stripes;
3152                 mirror_num = 1;
3153         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3154                 int factor = map->num_stripes / map->sub_stripes;
3155                 offset = map->stripe_len * (num / map->sub_stripes);
3156                 increment = map->stripe_len * factor;
3157                 mirror_num = num % map->sub_stripes + 1;
3158         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3159                 increment = map->stripe_len;
3160                 mirror_num = num % map->num_stripes + 1;
3161         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3162                 increment = map->stripe_len;
3163                 mirror_num = num % map->num_stripes + 1;
3164         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3165                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3166                 increment = map->stripe_len * nr_data_stripes(map);
3167                 mirror_num = 1;
3168         } else {
3169                 increment = map->stripe_len;
3170                 mirror_num = 1;
3171         }
3172
3173         path = btrfs_alloc_path();
3174         if (!path)
3175                 return -ENOMEM;
3176
3177         ppath = btrfs_alloc_path();
3178         if (!ppath) {
3179                 btrfs_free_path(path);
3180                 return -ENOMEM;
3181         }
3182
3183         /*
3184          * work on commit root. The related disk blocks are static as
3185          * long as COW is applied. This means, it is save to rewrite
3186          * them to repair disk errors without any race conditions
3187          */
3188         path->search_commit_root = 1;
3189         path->skip_locking = 1;
3190
3191         ppath->search_commit_root = 1;
3192         ppath->skip_locking = 1;
3193         /*
3194          * trigger the readahead for extent tree csum tree and wait for
3195          * completion. During readahead, the scrub is officially paused
3196          * to not hold off transaction commits
3197          */
3198         logical = base + offset;
3199         physical_end = physical + nstripes * map->stripe_len;
3200         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3201                 get_raid56_logic_offset(physical_end, num,
3202                                         map, &logic_end, NULL);
3203                 logic_end += base;
3204         } else {
3205                 logic_end = logical + increment * nstripes;
3206         }
3207         wait_event(sctx->list_wait,
3208                    atomic_read(&sctx->bios_in_flight) == 0);
3209         scrub_blocked_if_needed(fs_info);
3210
3211         /* FIXME it might be better to start readahead at commit root */
3212         key.objectid = logical;
3213         key.type = BTRFS_EXTENT_ITEM_KEY;
3214         key.offset = (u64)0;
3215         key_end.objectid = logic_end;
3216         key_end.type = BTRFS_METADATA_ITEM_KEY;
3217         key_end.offset = (u64)-1;
3218         reada1 = btrfs_reada_add(root, &key, &key_end);
3219
3220         if (cache->flags & BTRFS_BLOCK_GROUP_DATA) {
3221                 key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3222                 key.type = BTRFS_EXTENT_CSUM_KEY;
3223                 key.offset = logical;
3224                 key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3225                 key_end.type = BTRFS_EXTENT_CSUM_KEY;
3226                 key_end.offset = logic_end;
3227                 reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3228         } else {
3229                 reada2 = NULL;
3230         }
3231
3232         if (!IS_ERR(reada1))
3233                 btrfs_reada_wait(reada1);
3234         if (!IS_ERR_OR_NULL(reada2))
3235                 btrfs_reada_wait(reada2);
3236
3237
3238         /*
3239          * collect all data csums for the stripe to avoid seeking during
3240          * the scrub. This might currently (crc32) end up to be about 1MB
3241          */
3242         blk_start_plug(&plug);
3243
3244         if (sctx->is_dev_replace &&
3245             btrfs_dev_is_sequential(sctx->wr_tgtdev, physical)) {
3246                 mutex_lock(&sctx->wr_lock);
3247                 sctx->write_pointer = physical;
3248                 mutex_unlock(&sctx->wr_lock);
3249                 sctx->flush_all_writes = true;
3250         }
3251
3252         /*
3253          * now find all extents for each stripe and scrub them
3254          */
3255         ret = 0;
3256         while (physical < physical_end) {
3257                 /*
3258                  * canceled?
3259                  */
3260                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3261                     atomic_read(&sctx->cancel_req)) {
3262                         ret = -ECANCELED;
3263                         goto out;
3264                 }
3265                 /*
3266                  * check to see if we have to pause
3267                  */
3268                 if (atomic_read(&fs_info->scrub_pause_req)) {
3269                         /* push queued extents */
3270                         sctx->flush_all_writes = true;
3271                         scrub_submit(sctx);
3272                         mutex_lock(&sctx->wr_lock);
3273                         scrub_wr_submit(sctx);
3274                         mutex_unlock(&sctx->wr_lock);
3275                         wait_event(sctx->list_wait,
3276                                    atomic_read(&sctx->bios_in_flight) == 0);
3277                         sctx->flush_all_writes = false;
3278                         scrub_blocked_if_needed(fs_info);
3279                 }
3280
3281                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3282                         ret = get_raid56_logic_offset(physical, num, map,
3283                                                       &logical,
3284                                                       &stripe_logical);
3285                         logical += base;
3286                         if (ret) {
3287                                 /* it is parity strip */
3288                                 stripe_logical += base;
3289                                 stripe_end = stripe_logical + increment;
3290                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3291                                                           ppath, stripe_logical,
3292                                                           stripe_end);
3293                                 if (ret)
3294                                         goto out;
3295                                 goto skip;
3296                         }
3297                 }
3298
3299                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3300                         key.type = BTRFS_METADATA_ITEM_KEY;
3301                 else
3302                         key.type = BTRFS_EXTENT_ITEM_KEY;
3303                 key.objectid = logical;
3304                 key.offset = (u64)-1;
3305
3306                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3307                 if (ret < 0)
3308                         goto out;
3309
3310                 if (ret > 0) {
3311                         ret = btrfs_previous_extent_item(root, path, 0);
3312                         if (ret < 0)
3313                                 goto out;
3314                         if (ret > 0) {
3315                                 /* there's no smaller item, so stick with the
3316                                  * larger one */
3317                                 btrfs_release_path(path);
3318                                 ret = btrfs_search_slot(NULL, root, &key,
3319                                                         path, 0, 0);
3320                                 if (ret < 0)
3321                                         goto out;
3322                         }
3323                 }
3324
3325                 stop_loop = 0;
3326                 while (1) {
3327                         u64 bytes;
3328
3329                         l = path->nodes[0];
3330                         slot = path->slots[0];
3331                         if (slot >= btrfs_header_nritems(l)) {
3332                                 ret = btrfs_next_leaf(root, path);
3333                                 if (ret == 0)
3334                                         continue;
3335                                 if (ret < 0)
3336                                         goto out;
3337
3338                                 stop_loop = 1;
3339                                 break;
3340                         }
3341                         btrfs_item_key_to_cpu(l, &key, slot);
3342
3343                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3344                             key.type != BTRFS_METADATA_ITEM_KEY)
3345                                 goto next;
3346
3347                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3348                                 bytes = fs_info->nodesize;
3349                         else
3350                                 bytes = key.offset;
3351
3352                         if (key.objectid + bytes <= logical)
3353                                 goto next;
3354
3355                         if (key.objectid >= logical + map->stripe_len) {
3356                                 /* out of this device extent */
3357                                 if (key.objectid >= logic_end)
3358                                         stop_loop = 1;
3359                                 break;
3360                         }
3361
3362                         /*
3363                          * If our block group was removed in the meanwhile, just
3364                          * stop scrubbing since there is no point in continuing.
3365                          * Continuing would prevent reusing its device extents
3366                          * for new block groups for a long time.
3367                          */
3368                         spin_lock(&cache->lock);
3369                         if (cache->removed) {
3370                                 spin_unlock(&cache->lock);
3371                                 ret = 0;
3372                                 goto out;
3373                         }
3374                         spin_unlock(&cache->lock);
3375
3376                         extent = btrfs_item_ptr(l, slot,
3377                                                 struct btrfs_extent_item);
3378                         flags = btrfs_extent_flags(l, extent);
3379                         generation = btrfs_extent_generation(l, extent);
3380
3381                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3382                             (key.objectid < logical ||
3383                              key.objectid + bytes >
3384                              logical + map->stripe_len)) {
3385                                 btrfs_err(fs_info,
3386                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3387                                        key.objectid, logical);
3388                                 spin_lock(&sctx->stat_lock);
3389                                 sctx->stat.uncorrectable_errors++;
3390                                 spin_unlock(&sctx->stat_lock);
3391                                 goto next;
3392                         }
3393
3394 again:
3395                         extent_logical = key.objectid;
3396                         ASSERT(bytes <= U32_MAX);
3397                         extent_len = bytes;
3398
3399                         /*
3400                          * trim extent to this stripe
3401                          */
3402                         if (extent_logical < logical) {
3403                                 extent_len -= logical - extent_logical;
3404                                 extent_logical = logical;
3405                         }
3406                         if (extent_logical + extent_len >
3407                             logical + map->stripe_len) {
3408                                 extent_len = logical + map->stripe_len -
3409                                              extent_logical;
3410                         }
3411
3412                         extent_physical = extent_logical - logical + physical;
3413                         extent_dev = scrub_dev;
3414                         extent_mirror_num = mirror_num;
3415                         if (sctx->is_dev_replace)
3416                                 scrub_remap_extent(fs_info, extent_logical,
3417                                                    extent_len, &extent_physical,
3418                                                    &extent_dev,
3419                                                    &extent_mirror_num);
3420
3421                         if (flags & BTRFS_EXTENT_FLAG_DATA) {
3422                                 ret = btrfs_lookup_csums_range(csum_root,
3423                                                 extent_logical,
3424                                                 extent_logical + extent_len - 1,
3425                                                 &sctx->csum_list, 1);
3426                                 if (ret)
3427                                         goto out;
3428                         }
3429
3430                         ret = scrub_extent(sctx, map, extent_logical, extent_len,
3431                                            extent_physical, extent_dev, flags,
3432                                            generation, extent_mirror_num,
3433                                            extent_logical - logical + physical);
3434
3435                         scrub_free_csums(sctx);
3436
3437                         if (ret)
3438                                 goto out;
3439
3440                         if (sctx->is_dev_replace)
3441                                 sync_replace_for_zoned(sctx);
3442
3443                         if (extent_logical + extent_len <
3444                             key.objectid + bytes) {
3445                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3446                                         /*
3447                                          * loop until we find next data stripe
3448                                          * or we have finished all stripes.
3449                                          */
3450 loop:
3451                                         physical += map->stripe_len;
3452                                         ret = get_raid56_logic_offset(physical,
3453                                                         num, map, &logical,
3454                                                         &stripe_logical);
3455                                         logical += base;
3456
3457                                         if (ret && physical < physical_end) {
3458                                                 stripe_logical += base;
3459                                                 stripe_end = stripe_logical +
3460                                                                 increment;
3461                                                 ret = scrub_raid56_parity(sctx,
3462                                                         map, scrub_dev, ppath,
3463                                                         stripe_logical,
3464                                                         stripe_end);
3465                                                 if (ret)
3466                                                         goto out;
3467                                                 goto loop;
3468                                         }
3469                                 } else {
3470                                         physical += map->stripe_len;
3471                                         logical += increment;
3472                                 }
3473                                 if (logical < key.objectid + bytes) {
3474                                         cond_resched();
3475                                         goto again;
3476                                 }
3477
3478                                 if (physical >= physical_end) {
3479                                         stop_loop = 1;
3480                                         break;
3481                                 }
3482                         }
3483 next:
3484                         path->slots[0]++;
3485                 }
3486                 btrfs_release_path(path);
3487 skip:
3488                 logical += increment;
3489                 physical += map->stripe_len;
3490                 spin_lock(&sctx->stat_lock);
3491                 if (stop_loop)
3492                         sctx->stat.last_physical = map->stripes[num].physical +
3493                                                    length;
3494                 else
3495                         sctx->stat.last_physical = physical;
3496                 spin_unlock(&sctx->stat_lock);
3497                 if (stop_loop)
3498                         break;
3499         }
3500 out:
3501         /* push queued extents */
3502         scrub_submit(sctx);
3503         mutex_lock(&sctx->wr_lock);
3504         scrub_wr_submit(sctx);
3505         mutex_unlock(&sctx->wr_lock);
3506
3507         blk_finish_plug(&plug);
3508         btrfs_free_path(path);
3509         btrfs_free_path(ppath);
3510
3511         if (sctx->is_dev_replace && ret >= 0) {
3512                 int ret2;
3513
3514                 ret2 = sync_write_pointer_for_zoned(sctx, base + offset,
3515                                                     map->stripes[num].physical,
3516                                                     physical_end);
3517                 if (ret2)
3518                         ret = ret2;
3519         }
3520
3521         return ret < 0 ? ret : 0;
3522 }
3523
3524 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3525                                           struct btrfs_device *scrub_dev,
3526                                           u64 chunk_offset, u64 length,
3527                                           u64 dev_offset,
3528                                           struct btrfs_block_group *cache)
3529 {
3530         struct btrfs_fs_info *fs_info = sctx->fs_info;
3531         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3532         struct map_lookup *map;
3533         struct extent_map *em;
3534         int i;
3535         int ret = 0;
3536
3537         read_lock(&map_tree->lock);
3538         em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3539         read_unlock(&map_tree->lock);
3540
3541         if (!em) {
3542                 /*
3543                  * Might have been an unused block group deleted by the cleaner
3544                  * kthread or relocation.
3545                  */
3546                 spin_lock(&cache->lock);
3547                 if (!cache->removed)
3548                         ret = -EINVAL;
3549                 spin_unlock(&cache->lock);
3550
3551                 return ret;
3552         }
3553
3554         map = em->map_lookup;
3555         if (em->start != chunk_offset)
3556                 goto out;
3557
3558         if (em->len < length)
3559                 goto out;
3560
3561         for (i = 0; i < map->num_stripes; ++i) {
3562                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3563                     map->stripes[i].physical == dev_offset) {
3564                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3565                                            chunk_offset, length, cache);
3566                         if (ret)
3567                                 goto out;
3568                 }
3569         }
3570 out:
3571         free_extent_map(em);
3572
3573         return ret;
3574 }
3575
3576 static int finish_extent_writes_for_zoned(struct btrfs_root *root,
3577                                           struct btrfs_block_group *cache)
3578 {
3579         struct btrfs_fs_info *fs_info = cache->fs_info;
3580         struct btrfs_trans_handle *trans;
3581
3582         if (!btrfs_is_zoned(fs_info))
3583                 return 0;
3584
3585         btrfs_wait_block_group_reservations(cache);
3586         btrfs_wait_nocow_writers(cache);
3587         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start, cache->length);
3588
3589         trans = btrfs_join_transaction(root);
3590         if (IS_ERR(trans))
3591                 return PTR_ERR(trans);
3592         return btrfs_commit_transaction(trans);
3593 }
3594
3595 static noinline_for_stack
3596 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3597                            struct btrfs_device *scrub_dev, u64 start, u64 end)
3598 {
3599         struct btrfs_dev_extent *dev_extent = NULL;
3600         struct btrfs_path *path;
3601         struct btrfs_fs_info *fs_info = sctx->fs_info;
3602         struct btrfs_root *root = fs_info->dev_root;
3603         u64 length;
3604         u64 chunk_offset;
3605         int ret = 0;
3606         int ro_set;
3607         int slot;
3608         struct extent_buffer *l;
3609         struct btrfs_key key;
3610         struct btrfs_key found_key;
3611         struct btrfs_block_group *cache;
3612         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3613
3614         path = btrfs_alloc_path();
3615         if (!path)
3616                 return -ENOMEM;
3617
3618         path->reada = READA_FORWARD;
3619         path->search_commit_root = 1;
3620         path->skip_locking = 1;
3621
3622         key.objectid = scrub_dev->devid;
3623         key.offset = 0ull;
3624         key.type = BTRFS_DEV_EXTENT_KEY;
3625
3626         while (1) {
3627                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3628                 if (ret < 0)
3629                         break;
3630                 if (ret > 0) {
3631                         if (path->slots[0] >=
3632                             btrfs_header_nritems(path->nodes[0])) {
3633                                 ret = btrfs_next_leaf(root, path);
3634                                 if (ret < 0)
3635                                         break;
3636                                 if (ret > 0) {
3637                                         ret = 0;
3638                                         break;
3639                                 }
3640                         } else {
3641                                 ret = 0;
3642                         }
3643                 }
3644
3645                 l = path->nodes[0];
3646                 slot = path->slots[0];
3647
3648                 btrfs_item_key_to_cpu(l, &found_key, slot);
3649
3650                 if (found_key.objectid != scrub_dev->devid)
3651                         break;
3652
3653                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3654                         break;
3655
3656                 if (found_key.offset >= end)
3657                         break;
3658
3659                 if (found_key.offset < key.offset)
3660                         break;
3661
3662                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3663                 length = btrfs_dev_extent_length(l, dev_extent);
3664
3665                 if (found_key.offset + length <= start)
3666                         goto skip;
3667
3668                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3669
3670                 /*
3671                  * get a reference on the corresponding block group to prevent
3672                  * the chunk from going away while we scrub it
3673                  */
3674                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3675
3676                 /* some chunks are removed but not committed to disk yet,
3677                  * continue scrubbing */
3678                 if (!cache)
3679                         goto skip;
3680
3681                 if (sctx->is_dev_replace && btrfs_is_zoned(fs_info)) {
3682                         spin_lock(&cache->lock);
3683                         if (!cache->to_copy) {
3684                                 spin_unlock(&cache->lock);
3685                                 ro_set = 0;
3686                                 goto done;
3687                         }
3688                         spin_unlock(&cache->lock);
3689                 }
3690
3691                 /*
3692                  * Make sure that while we are scrubbing the corresponding block
3693                  * group doesn't get its logical address and its device extents
3694                  * reused for another block group, which can possibly be of a
3695                  * different type and different profile. We do this to prevent
3696                  * false error detections and crashes due to bogus attempts to
3697                  * repair extents.
3698                  */
3699                 spin_lock(&cache->lock);
3700                 if (cache->removed) {
3701                         spin_unlock(&cache->lock);
3702                         btrfs_put_block_group(cache);
3703                         goto skip;
3704                 }
3705                 btrfs_freeze_block_group(cache);
3706                 spin_unlock(&cache->lock);
3707
3708                 /*
3709                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3710                  * to avoid deadlock caused by:
3711                  * btrfs_inc_block_group_ro()
3712                  * -> btrfs_wait_for_commit()
3713                  * -> btrfs_commit_transaction()
3714                  * -> btrfs_scrub_pause()
3715                  */
3716                 scrub_pause_on(fs_info);
3717
3718                 /*
3719                  * Don't do chunk preallocation for scrub.
3720                  *
3721                  * This is especially important for SYSTEM bgs, or we can hit
3722                  * -EFBIG from btrfs_finish_chunk_alloc() like:
3723                  * 1. The only SYSTEM bg is marked RO.
3724                  *    Since SYSTEM bg is small, that's pretty common.
3725                  * 2. New SYSTEM bg will be allocated
3726                  *    Due to regular version will allocate new chunk.
3727                  * 3. New SYSTEM bg is empty and will get cleaned up
3728                  *    Before cleanup really happens, it's marked RO again.
3729                  * 4. Empty SYSTEM bg get scrubbed
3730                  *    We go back to 2.
3731                  *
3732                  * This can easily boost the amount of SYSTEM chunks if cleaner
3733                  * thread can't be triggered fast enough, and use up all space
3734                  * of btrfs_super_block::sys_chunk_array
3735                  *
3736                  * While for dev replace, we need to try our best to mark block
3737                  * group RO, to prevent race between:
3738                  * - Write duplication
3739                  *   Contains latest data
3740                  * - Scrub copy
3741                  *   Contains data from commit tree
3742                  *
3743                  * If target block group is not marked RO, nocow writes can
3744                  * be overwritten by scrub copy, causing data corruption.
3745                  * So for dev-replace, it's not allowed to continue if a block
3746                  * group is not RO.
3747                  */
3748                 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3749                 if (!ret && sctx->is_dev_replace) {
3750                         ret = finish_extent_writes_for_zoned(root, cache);
3751                         if (ret) {
3752                                 btrfs_dec_block_group_ro(cache);
3753                                 scrub_pause_off(fs_info);
3754                                 btrfs_put_block_group(cache);
3755                                 break;
3756                         }
3757                 }
3758
3759                 if (ret == 0) {
3760                         ro_set = 1;
3761                 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3762                         /*
3763                          * btrfs_inc_block_group_ro return -ENOSPC when it
3764                          * failed in creating new chunk for metadata.
3765                          * It is not a problem for scrub, because
3766                          * metadata are always cowed, and our scrub paused
3767                          * commit_transactions.
3768                          */
3769                         ro_set = 0;
3770                 } else {
3771                         btrfs_warn(fs_info,
3772                                    "failed setting block group ro: %d", ret);
3773                         btrfs_unfreeze_block_group(cache);
3774                         btrfs_put_block_group(cache);
3775                         scrub_pause_off(fs_info);
3776                         break;
3777                 }
3778
3779                 /*
3780                  * Now the target block is marked RO, wait for nocow writes to
3781                  * finish before dev-replace.
3782                  * COW is fine, as COW never overwrites extents in commit tree.
3783                  */
3784                 if (sctx->is_dev_replace) {
3785                         btrfs_wait_nocow_writers(cache);
3786                         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3787                                         cache->length);
3788                 }
3789
3790                 scrub_pause_off(fs_info);
3791                 down_write(&dev_replace->rwsem);
3792                 dev_replace->cursor_right = found_key.offset + length;
3793                 dev_replace->cursor_left = found_key.offset;
3794                 dev_replace->item_needs_writeback = 1;
3795                 up_write(&dev_replace->rwsem);
3796
3797                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3798                                   found_key.offset, cache);
3799
3800                 /*
3801                  * flush, submit all pending read and write bios, afterwards
3802                  * wait for them.
3803                  * Note that in the dev replace case, a read request causes
3804                  * write requests that are submitted in the read completion
3805                  * worker. Therefore in the current situation, it is required
3806                  * that all write requests are flushed, so that all read and
3807                  * write requests are really completed when bios_in_flight
3808                  * changes to 0.
3809                  */
3810                 sctx->flush_all_writes = true;
3811                 scrub_submit(sctx);
3812                 mutex_lock(&sctx->wr_lock);
3813                 scrub_wr_submit(sctx);
3814                 mutex_unlock(&sctx->wr_lock);
3815
3816                 wait_event(sctx->list_wait,
3817                            atomic_read(&sctx->bios_in_flight) == 0);
3818
3819                 scrub_pause_on(fs_info);
3820
3821                 /*
3822                  * must be called before we decrease @scrub_paused.
3823                  * make sure we don't block transaction commit while
3824                  * we are waiting pending workers finished.
3825                  */
3826                 wait_event(sctx->list_wait,
3827                            atomic_read(&sctx->workers_pending) == 0);
3828                 sctx->flush_all_writes = false;
3829
3830                 scrub_pause_off(fs_info);
3831
3832                 if (sctx->is_dev_replace &&
3833                     !btrfs_finish_block_group_to_copy(dev_replace->srcdev,
3834                                                       cache, found_key.offset))
3835                         ro_set = 0;
3836
3837 done:
3838                 down_write(&dev_replace->rwsem);
3839                 dev_replace->cursor_left = dev_replace->cursor_right;
3840                 dev_replace->item_needs_writeback = 1;
3841                 up_write(&dev_replace->rwsem);
3842
3843                 if (ro_set)
3844                         btrfs_dec_block_group_ro(cache);
3845
3846                 /*
3847                  * We might have prevented the cleaner kthread from deleting
3848                  * this block group if it was already unused because we raced
3849                  * and set it to RO mode first. So add it back to the unused
3850                  * list, otherwise it might not ever be deleted unless a manual
3851                  * balance is triggered or it becomes used and unused again.
3852                  */
3853                 spin_lock(&cache->lock);
3854                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3855                     cache->used == 0) {
3856                         spin_unlock(&cache->lock);
3857                         if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3858                                 btrfs_discard_queue_work(&fs_info->discard_ctl,
3859                                                          cache);
3860                         else
3861                                 btrfs_mark_bg_unused(cache);
3862                 } else {
3863                         spin_unlock(&cache->lock);
3864                 }
3865
3866                 btrfs_unfreeze_block_group(cache);
3867                 btrfs_put_block_group(cache);
3868                 if (ret)
3869                         break;
3870                 if (sctx->is_dev_replace &&
3871                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3872                         ret = -EIO;
3873                         break;
3874                 }
3875                 if (sctx->stat.malloc_errors > 0) {
3876                         ret = -ENOMEM;
3877                         break;
3878                 }
3879 skip:
3880                 key.offset = found_key.offset + length;
3881                 btrfs_release_path(path);
3882         }
3883
3884         btrfs_free_path(path);
3885
3886         return ret;
3887 }
3888
3889 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3890                                            struct btrfs_device *scrub_dev)
3891 {
3892         int     i;
3893         u64     bytenr;
3894         u64     gen;
3895         int     ret;
3896         struct btrfs_fs_info *fs_info = sctx->fs_info;
3897
3898         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3899                 return -EROFS;
3900
3901         /* Seed devices of a new filesystem has their own generation. */
3902         if (scrub_dev->fs_devices != fs_info->fs_devices)
3903                 gen = scrub_dev->generation;
3904         else
3905                 gen = fs_info->last_trans_committed;
3906
3907         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3908                 bytenr = btrfs_sb_offset(i);
3909                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3910                     scrub_dev->commit_total_bytes)
3911                         break;
3912                 if (!btrfs_check_super_location(scrub_dev, bytenr))
3913                         continue;
3914
3915                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3916                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3917                                   NULL, bytenr);
3918                 if (ret)
3919                         return ret;
3920         }
3921         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3922
3923         return 0;
3924 }
3925
3926 static void scrub_workers_put(struct btrfs_fs_info *fs_info)
3927 {
3928         if (refcount_dec_and_mutex_lock(&fs_info->scrub_workers_refcnt,
3929                                         &fs_info->scrub_lock)) {
3930                 struct btrfs_workqueue *scrub_workers = NULL;
3931                 struct btrfs_workqueue *scrub_wr_comp = NULL;
3932                 struct btrfs_workqueue *scrub_parity = NULL;
3933
3934                 scrub_workers = fs_info->scrub_workers;
3935                 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3936                 scrub_parity = fs_info->scrub_parity_workers;
3937
3938                 fs_info->scrub_workers = NULL;
3939                 fs_info->scrub_wr_completion_workers = NULL;
3940                 fs_info->scrub_parity_workers = NULL;
3941                 mutex_unlock(&fs_info->scrub_lock);
3942
3943                 btrfs_destroy_workqueue(scrub_workers);
3944                 btrfs_destroy_workqueue(scrub_wr_comp);
3945                 btrfs_destroy_workqueue(scrub_parity);
3946         }
3947 }
3948
3949 /*
3950  * get a reference count on fs_info->scrub_workers. start worker if necessary
3951  */
3952 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3953                                                 int is_dev_replace)
3954 {
3955         struct btrfs_workqueue *scrub_workers = NULL;
3956         struct btrfs_workqueue *scrub_wr_comp = NULL;
3957         struct btrfs_workqueue *scrub_parity = NULL;
3958         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3959         int max_active = fs_info->thread_pool_size;
3960         int ret = -ENOMEM;
3961
3962         if (refcount_inc_not_zero(&fs_info->scrub_workers_refcnt))
3963                 return 0;
3964
3965         scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub", flags,
3966                                               is_dev_replace ? 1 : max_active, 4);
3967         if (!scrub_workers)
3968                 goto fail_scrub_workers;
3969
3970         scrub_wr_comp = btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3971                                               max_active, 2);
3972         if (!scrub_wr_comp)
3973                 goto fail_scrub_wr_completion_workers;
3974
3975         scrub_parity = btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3976                                              max_active, 2);
3977         if (!scrub_parity)
3978                 goto fail_scrub_parity_workers;
3979
3980         mutex_lock(&fs_info->scrub_lock);
3981         if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3982                 ASSERT(fs_info->scrub_workers == NULL &&
3983                        fs_info->scrub_wr_completion_workers == NULL &&
3984                        fs_info->scrub_parity_workers == NULL);
3985                 fs_info->scrub_workers = scrub_workers;
3986                 fs_info->scrub_wr_completion_workers = scrub_wr_comp;
3987                 fs_info->scrub_parity_workers = scrub_parity;
3988                 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3989                 mutex_unlock(&fs_info->scrub_lock);
3990                 return 0;
3991         }
3992         /* Other thread raced in and created the workers for us */
3993         refcount_inc(&fs_info->scrub_workers_refcnt);
3994         mutex_unlock(&fs_info->scrub_lock);
3995
3996         ret = 0;
3997         btrfs_destroy_workqueue(scrub_parity);
3998 fail_scrub_parity_workers:
3999         btrfs_destroy_workqueue(scrub_wr_comp);
4000 fail_scrub_wr_completion_workers:
4001         btrfs_destroy_workqueue(scrub_workers);
4002 fail_scrub_workers:
4003         return ret;
4004 }
4005
4006 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4007                     u64 end, struct btrfs_scrub_progress *progress,
4008                     int readonly, int is_dev_replace)
4009 {
4010         struct scrub_ctx *sctx;
4011         int ret;
4012         struct btrfs_device *dev;
4013         unsigned int nofs_flag;
4014
4015         if (btrfs_fs_closing(fs_info))
4016                 return -EAGAIN;
4017
4018         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4019                 /*
4020                  * in this case scrub is unable to calculate the checksum
4021                  * the way scrub is implemented. Do not handle this
4022                  * situation at all because it won't ever happen.
4023                  */
4024                 btrfs_err(fs_info,
4025                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4026                        fs_info->nodesize,
4027                        BTRFS_STRIPE_LEN);
4028                 return -EINVAL;
4029         }
4030
4031         if (fs_info->nodesize >
4032             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4033             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4034                 /*
4035                  * would exhaust the array bounds of pagev member in
4036                  * struct scrub_block
4037                  */
4038                 btrfs_err(fs_info,
4039                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4040                        fs_info->nodesize,
4041                        SCRUB_MAX_PAGES_PER_BLOCK,
4042                        fs_info->sectorsize,
4043                        SCRUB_MAX_PAGES_PER_BLOCK);
4044                 return -EINVAL;
4045         }
4046
4047         /* Allocate outside of device_list_mutex */
4048         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
4049         if (IS_ERR(sctx))
4050                 return PTR_ERR(sctx);
4051
4052         ret = scrub_workers_get(fs_info, is_dev_replace);
4053         if (ret)
4054                 goto out_free_ctx;
4055
4056         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4057         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4058         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
4059                      !is_dev_replace)) {
4060                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4061                 ret = -ENODEV;
4062                 goto out;
4063         }
4064
4065         if (!is_dev_replace && !readonly &&
4066             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
4067                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4068                 btrfs_err_in_rcu(fs_info,
4069                         "scrub on devid %llu: filesystem on %s is not writable",
4070                                  devid, rcu_str_deref(dev->name));
4071                 ret = -EROFS;
4072                 goto out;
4073         }
4074
4075         mutex_lock(&fs_info->scrub_lock);
4076         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
4077             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
4078                 mutex_unlock(&fs_info->scrub_lock);
4079                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4080                 ret = -EIO;
4081                 goto out;
4082         }
4083
4084         down_read(&fs_info->dev_replace.rwsem);
4085         if (dev->scrub_ctx ||
4086             (!is_dev_replace &&
4087              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4088                 up_read(&fs_info->dev_replace.rwsem);
4089                 mutex_unlock(&fs_info->scrub_lock);
4090                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4091                 ret = -EINPROGRESS;
4092                 goto out;
4093         }
4094         up_read(&fs_info->dev_replace.rwsem);
4095
4096         sctx->readonly = readonly;
4097         dev->scrub_ctx = sctx;
4098         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4099
4100         /*
4101          * checking @scrub_pause_req here, we can avoid
4102          * race between committing transaction and scrubbing.
4103          */
4104         __scrub_blocked_if_needed(fs_info);
4105         atomic_inc(&fs_info->scrubs_running);
4106         mutex_unlock(&fs_info->scrub_lock);
4107
4108         /*
4109          * In order to avoid deadlock with reclaim when there is a transaction
4110          * trying to pause scrub, make sure we use GFP_NOFS for all the
4111          * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
4112          * invoked by our callees. The pausing request is done when the
4113          * transaction commit starts, and it blocks the transaction until scrub
4114          * is paused (done at specific points at scrub_stripe() or right above
4115          * before incrementing fs_info->scrubs_running).
4116          */
4117         nofs_flag = memalloc_nofs_save();
4118         if (!is_dev_replace) {
4119                 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
4120                 /*
4121                  * by holding device list mutex, we can
4122                  * kick off writing super in log tree sync.
4123                  */
4124                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4125                 ret = scrub_supers(sctx, dev);
4126                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4127         }
4128
4129         if (!ret)
4130                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
4131         memalloc_nofs_restore(nofs_flag);
4132
4133         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4134         atomic_dec(&fs_info->scrubs_running);
4135         wake_up(&fs_info->scrub_pause_wait);
4136
4137         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4138
4139         if (progress)
4140                 memcpy(progress, &sctx->stat, sizeof(*progress));
4141
4142         if (!is_dev_replace)
4143                 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
4144                         ret ? "not finished" : "finished", devid, ret);
4145
4146         mutex_lock(&fs_info->scrub_lock);
4147         dev->scrub_ctx = NULL;
4148         mutex_unlock(&fs_info->scrub_lock);
4149
4150         scrub_workers_put(fs_info);
4151         scrub_put_ctx(sctx);
4152
4153         return ret;
4154 out:
4155         scrub_workers_put(fs_info);
4156 out_free_ctx:
4157         scrub_free_ctx(sctx);
4158
4159         return ret;
4160 }
4161
4162 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4163 {
4164         mutex_lock(&fs_info->scrub_lock);
4165         atomic_inc(&fs_info->scrub_pause_req);
4166         while (atomic_read(&fs_info->scrubs_paused) !=
4167                atomic_read(&fs_info->scrubs_running)) {
4168                 mutex_unlock(&fs_info->scrub_lock);
4169                 wait_event(fs_info->scrub_pause_wait,
4170                            atomic_read(&fs_info->scrubs_paused) ==
4171                            atomic_read(&fs_info->scrubs_running));
4172                 mutex_lock(&fs_info->scrub_lock);
4173         }
4174         mutex_unlock(&fs_info->scrub_lock);
4175 }
4176
4177 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4178 {
4179         atomic_dec(&fs_info->scrub_pause_req);
4180         wake_up(&fs_info->scrub_pause_wait);
4181 }
4182
4183 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4184 {
4185         mutex_lock(&fs_info->scrub_lock);
4186         if (!atomic_read(&fs_info->scrubs_running)) {
4187                 mutex_unlock(&fs_info->scrub_lock);
4188                 return -ENOTCONN;
4189         }
4190
4191         atomic_inc(&fs_info->scrub_cancel_req);
4192         while (atomic_read(&fs_info->scrubs_running)) {
4193                 mutex_unlock(&fs_info->scrub_lock);
4194                 wait_event(fs_info->scrub_pause_wait,
4195                            atomic_read(&fs_info->scrubs_running) == 0);
4196                 mutex_lock(&fs_info->scrub_lock);
4197         }
4198         atomic_dec(&fs_info->scrub_cancel_req);
4199         mutex_unlock(&fs_info->scrub_lock);
4200
4201         return 0;
4202 }
4203
4204 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4205 {
4206         struct btrfs_fs_info *fs_info = dev->fs_info;
4207         struct scrub_ctx *sctx;
4208
4209         mutex_lock(&fs_info->scrub_lock);
4210         sctx = dev->scrub_ctx;
4211         if (!sctx) {
4212                 mutex_unlock(&fs_info->scrub_lock);
4213                 return -ENOTCONN;
4214         }
4215         atomic_inc(&sctx->cancel_req);
4216         while (dev->scrub_ctx) {
4217                 mutex_unlock(&fs_info->scrub_lock);
4218                 wait_event(fs_info->scrub_pause_wait,
4219                            dev->scrub_ctx == NULL);
4220                 mutex_lock(&fs_info->scrub_lock);
4221         }
4222         mutex_unlock(&fs_info->scrub_lock);
4223
4224         return 0;
4225 }
4226
4227 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4228                          struct btrfs_scrub_progress *progress)
4229 {
4230         struct btrfs_device *dev;
4231         struct scrub_ctx *sctx = NULL;
4232
4233         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4234         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL);
4235         if (dev)
4236                 sctx = dev->scrub_ctx;
4237         if (sctx)
4238                 memcpy(progress, &sctx->stat, sizeof(*progress));
4239         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4240
4241         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4242 }
4243
4244 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4245                                u64 extent_logical, u32 extent_len,
4246                                u64 *extent_physical,
4247                                struct btrfs_device **extent_dev,
4248                                int *extent_mirror_num)
4249 {
4250         u64 mapped_length;
4251         struct btrfs_bio *bbio = NULL;
4252         int ret;
4253
4254         mapped_length = extent_len;
4255         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4256                               &mapped_length, &bbio, 0);
4257         if (ret || !bbio || mapped_length < extent_len ||
4258             !bbio->stripes[0].dev->bdev) {
4259                 btrfs_put_bbio(bbio);
4260                 return;
4261         }
4262
4263         *extent_physical = bbio->stripes[0].physical;
4264         *extent_mirror_num = bbio->mirror_num;
4265         *extent_dev = bbio->stripes[0].dev;
4266         btrfs_put_bbio(bbio);
4267 }