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
  24 /*
  25  * This is only the first step towards a full-features scrub. It reads all
  26  * extent and super block and verifies the checksums. In case a bad checksum
  27  * is found or the extent cannot be read, good data will be written back if
  28  * any can be found.
  29  *
  30  * Future enhancements:
  31  *  - In case an unrepairable extent is encountered, track which files are
  32  *    affected and report them
  33  *  - track and record media errors, throw out bad devices
  34  *  - add a mode to also read unallocated space
  35  */
  36
  37 struct scrub_block;
  38 struct scrub_ctx;
  39
  40 /*
  41  * the following three values only influence the performance.
  42  * The last one configures the number of parallel and outstanding I/O
  43  * operations. The first two values configure an upper limit for the number
  44  * of (dynamically allocated) pages that are added to a bio.
  45  */
  46 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  47 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  48 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  49
  50 /*
  51  * the following value times PAGE_SIZE needs to be large enough to match the
  52  * largest node/leaf/sector size that shall be supported.
  53  * Values larger than BTRFS_STRIPE_LEN are not supported.
  54  */
  55 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  56
  57 struct scrub_recover {
  58         refcount_t              refs;
  59         struct btrfs_bio        *bbio;
  60         u64                     map_length;
  61 };
  62
  63 struct scrub_page {
  64         struct scrub_block      *sblock;
  65         struct page             *page;
  66         struct btrfs_device     *dev;
  67         struct list_head        list;
  68         u64                     flags;  /* extent flags */
  69         u64                     generation;
  70         u64                     logical;
  71         u64                     physical;
  72         u64                     physical_for_dev_replace;
  73         atomic_t                refs;
  74         struct {
  75                 unsigned int    mirror_num:8;
  76                 unsigned int    have_csum:1;
  77                 unsigned int    io_error:1;
  78         };
  79         u8                      csum[BTRFS_CSUM_SIZE];
  80
  81         struct scrub_recover    *recover;
  82 };
  83
  84 struct scrub_bio {
  85         int                     index;
  86         struct scrub_ctx        *sctx;
  87         struct btrfs_device     *dev;
  88         struct bio              *bio;
  89         blk_status_t            status;
  90         u64                     logical;
  91         u64                     physical;
  92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  93         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  94 #else
  95         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  96 #endif
  97         int                     page_count;
  98         int                     next_free;
  99         struct btrfs_work       work;
 100 };
 101
 102 struct scrub_block {
 103         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 104         int                     page_count;
 105         atomic_t                outstanding_pages;
 106         refcount_t              refs; /* free mem on transition to zero */
 107         struct scrub_ctx        *sctx;
 108         struct scrub_parity     *sparity;
 109         struct {
 110                 unsigned int    header_error:1;
 111                 unsigned int    checksum_error:1;
 112                 unsigned int    no_io_error_seen:1;
 113                 unsigned int    generation_error:1; /* also sets header_error */
 114
 115                 /* The following is for the data used to check parity */
 116                 /* It is for the data with checksum */
 117                 unsigned int    data_corrected:1;
 118         };
 119         struct btrfs_work       work;
 120 };
 121
 122 /* Used for the chunks with parity stripe such RAID5/6 */
 123 struct scrub_parity {
 124         struct scrub_ctx        *sctx;
 125
 126         struct btrfs_device     *scrub_dev;
 127
 128         u64                     logic_start;
 129
 130         u64                     logic_end;
 131
 132         int                     nsectors;
 133
 134         u64                     stripe_len;
 135
 136         refcount_t              refs;
 137
 138         struct list_head        spages;
 139
 140         /* Work of parity check and repair */
 141         struct btrfs_work       work;
 142
 143         /* Mark the parity blocks which have data */
 144         unsigned long           *dbitmap;
 145
 146         /*
 147          * Mark the parity blocks which have data, but errors happen when
 148          * read data or check data
 149          */
 150         unsigned long           *ebitmap;
 151
 152         unsigned long           bitmap[];
 153 };
 154
 155 struct scrub_ctx {
 156         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 157         struct btrfs_fs_info    *fs_info;
 158         int                     first_free;
 159         int                     curr;
 160         atomic_t                bios_in_flight;
 161         atomic_t                workers_pending;
 162         spinlock_t              list_lock;
 163         wait_queue_head_t       list_wait;
 164         u16                     csum_size;
 165         struct list_head        csum_list;
 166         atomic_t                cancel_req;
 167         int                     readonly;
 168         int                     pages_per_rd_bio;
 169
 170         int                     is_dev_replace;
 171
 172         struct scrub_bio        *wr_curr_bio;
 173         struct mutex            wr_lock;
 174         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 175         struct btrfs_device     *wr_tgtdev;
 176         bool                    flush_all_writes;
 177
 178         /*
 179          * statistics
 180          */
 181         struct btrfs_scrub_progress stat;
 182         spinlock_t              stat_lock;
 183
 184         /*
 185          * Use a ref counter to avoid use-after-free issues. Scrub workers
 186          * decrement bios_in_flight and workers_pending and then do a wakeup
 187          * on the list_wait wait queue. We must ensure the main scrub task
 188          * doesn't free the scrub context before or while the workers are
 189          * doing the wakeup() call.
 190          */
 191         refcount_t              refs;
 192 };
 193
 194 struct scrub_warning {
 195         struct btrfs_path       *path;
 196         u64                     extent_item_size;
 197         const char              *errstr;
 198         u64                     physical;
 199         u64                     logical;
 200         struct btrfs_device     *dev;
 201 };
 202
 203 struct full_stripe_lock {
 204         struct rb_node node;
 205         u64 logical;
 206         u64 refs;
 207         struct mutex mutex;
 208 };
 209
 210 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 211 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 212 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 213 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 214                                      struct scrub_block *sblocks_for_recheck);
 215 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 216                                 struct scrub_block *sblock,
 217                                 int retry_failed_mirror);
 218 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 219 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 220                                              struct scrub_block *sblock_good);
 221 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 222                                             struct scrub_block *sblock_good,
 223                                             int page_num, int force_write);
 224 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 225 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 226                                            int page_num);
 227 static int scrub_checksum_data(struct scrub_block *sblock);
 228 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 229 static int scrub_checksum_super(struct scrub_block *sblock);
 230 static void scrub_block_get(struct scrub_block *sblock);
 231 static void scrub_block_put(struct scrub_block *sblock);
 232 static void scrub_page_get(struct scrub_page *spage);
 233 static void scrub_page_put(struct scrub_page *spage);
 234 static void scrub_parity_get(struct scrub_parity *sparity);
 235 static void scrub_parity_put(struct scrub_parity *sparity);
 236 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 237                                     struct scrub_page *spage);
 238 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 239                        u64 physical, struct btrfs_device *dev, u64 flags,
 240                        u64 gen, int mirror_num, u8 *csum, int force,
 241                        u64 physical_for_dev_replace);
 242 static void scrub_bio_end_io(struct bio *bio);
 243 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 244 static void scrub_block_complete(struct scrub_block *sblock);
 245 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 246                                u64 extent_logical, u64 extent_len,
 247                                u64 *extent_physical,
 248                                struct btrfs_device **extent_dev,
 249                                int *extent_mirror_num);
 250 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 251                                     struct scrub_page *spage);
 252 static void scrub_wr_submit(struct scrub_ctx *sctx);
 253 static void scrub_wr_bio_end_io(struct bio *bio);
 254 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 255 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 256 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 257 static void scrub_put_ctx(struct scrub_ctx *sctx);
 258
 259 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
 260 {
 261         return page->recover &&
 262                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 263 }
 264
 265 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 266 {
 267         refcount_inc(&sctx->refs);
 268         atomic_inc(&sctx->bios_in_flight);
 269 }
 270
 271 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 272 {
 273         atomic_dec(&sctx->bios_in_flight);
 274         wake_up(&sctx->list_wait);
 275         scrub_put_ctx(sctx);
 276 }
 277
 278 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 279 {
 280         while (atomic_read(&fs_info->scrub_pause_req)) {
 281                 mutex_unlock(&fs_info->scrub_lock);
 282                 wait_event(fs_info->scrub_pause_wait,
 283                    atomic_read(&fs_info->scrub_pause_req) == 0);
 284                 mutex_lock(&fs_info->scrub_lock);
 285         }
 286 }
 287
 288 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 289 {
 290         atomic_inc(&fs_info->scrubs_paused);
 291         wake_up(&fs_info->scrub_pause_wait);
 292 }
 293
 294 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 295 {
 296         mutex_lock(&fs_info->scrub_lock);
 297         __scrub_blocked_if_needed(fs_info);
 298         atomic_dec(&fs_info->scrubs_paused);
 299         mutex_unlock(&fs_info->scrub_lock);
 300
 301         wake_up(&fs_info->scrub_pause_wait);
 302 }
 303
 304 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 305 {
 306         scrub_pause_on(fs_info);
 307         scrub_pause_off(fs_info);
 308 }
 309
 310 /*
 311  * Insert new full stripe lock into full stripe locks tree
 312  *
 313  * Return pointer to existing or newly inserted full_stripe_lock structure if
 314  * everything works well.
 315  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 316  *
 317  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 318  * function
 319  */
 320 static struct full_stripe_lock *insert_full_stripe_lock(
 321                 struct btrfs_full_stripe_locks_tree *locks_root,
 322                 u64 fstripe_logical)
 323 {
 324         struct rb_node **p;
 325         struct rb_node *parent = NULL;
 326         struct full_stripe_lock *entry;
 327         struct full_stripe_lock *ret;
 328
 329         lockdep_assert_held(&locks_root->lock);
 330
 331         p = &locks_root->root.rb_node;
 332         while (*p) {
 333                 parent = *p;
 334                 entry = rb_entry(parent, struct full_stripe_lock, node);
 335                 if (fstripe_logical < entry->logical) {
 336                         p = &(*p)->rb_left;
 337                 } else if (fstripe_logical > entry->logical) {
 338                         p = &(*p)->rb_right;
 339                 } else {
 340                         entry->refs++;
 341                         return entry;
 342                 }
 343         }
 344
 345         /*
 346          * Insert new lock.
 347          */
 348         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 349         if (!ret)
 350                 return ERR_PTR(-ENOMEM);
 351         ret->logical = fstripe_logical;
 352         ret->refs = 1;
 353         mutex_init(&ret->mutex);
 354
 355         rb_link_node(&ret->node, parent, p);
 356         rb_insert_color(&ret->node, &locks_root->root);
 357         return ret;
 358 }
 359
 360 /*
 361  * Search for a full stripe lock of a block group
 362  *
 363  * Return pointer to existing full stripe lock if found
 364  * Return NULL if not found
 365  */
 366 static struct full_stripe_lock *search_full_stripe_lock(
 367                 struct btrfs_full_stripe_locks_tree *locks_root,
 368                 u64 fstripe_logical)
 369 {
 370         struct rb_node *node;
 371         struct full_stripe_lock *entry;
 372
 373         lockdep_assert_held(&locks_root->lock);
 374
 375         node = locks_root->root.rb_node;
 376         while (node) {
 377                 entry = rb_entry(node, struct full_stripe_lock, node);
 378                 if (fstripe_logical < entry->logical)
 379                         node = node->rb_left;
 380                 else if (fstripe_logical > entry->logical)
 381                         node = node->rb_right;
 382                 else
 383                         return entry;
 384         }
 385         return NULL;
 386 }
 387
 388 /*
 389  * Helper to get full stripe logical from a normal bytenr.
 390  *
 391  * Caller must ensure @cache is a RAID56 block group.
 392  */
 393 static u64 get_full_stripe_logical(struct btrfs_block_group *cache, u64 bytenr)
 394 {
 395         u64 ret;
 396
 397         /*
 398          * Due to chunk item size limit, full stripe length should not be
 399          * larger than U32_MAX. Just a sanity check here.
 400          */
 401         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 402
 403         /*
 404          * round_down() can only handle power of 2, while RAID56 full
 405          * stripe length can be 64KiB * n, so we need to manually round down.
 406          */
 407         ret = div64_u64(bytenr - cache->start, cache->full_stripe_len) *
 408                         cache->full_stripe_len + cache->start;
 409         return ret;
 410 }
 411
 412 /*
 413  * Lock a full stripe to avoid concurrency of recovery and read
 414  *
 415  * It's only used for profiles with parities (RAID5/6), for other profiles it
 416  * does nothing.
 417  *
 418  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 419  * So caller must call unlock_full_stripe() at the same context.
 420  *
 421  * Return <0 if encounters error.
 422  */
 423 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 424                             bool *locked_ret)
 425 {
 426         struct btrfs_block_group *bg_cache;
 427         struct btrfs_full_stripe_locks_tree *locks_root;
 428         struct full_stripe_lock *existing;
 429         u64 fstripe_start;
 430         int ret = 0;
 431
 432         *locked_ret = false;
 433         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 434         if (!bg_cache) {
 435                 ASSERT(0);
 436                 return -ENOENT;
 437         }
 438
 439         /* Profiles not based on parity don't need full stripe lock */
 440         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 441                 goto out;
 442         locks_root = &bg_cache->full_stripe_locks_root;
 443
 444         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 445
 446         /* Now insert the full stripe lock */
 447         mutex_lock(&locks_root->lock);
 448         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 449         mutex_unlock(&locks_root->lock);
 450         if (IS_ERR(existing)) {
 451                 ret = PTR_ERR(existing);
 452                 goto out;
 453         }
 454         mutex_lock(&existing->mutex);
 455         *locked_ret = true;
 456 out:
 457         btrfs_put_block_group(bg_cache);
 458         return ret;
 459 }
 460
 461 /*
 462  * Unlock a full stripe.
 463  *
 464  * NOTE: Caller must ensure it's the same context calling corresponding
 465  * lock_full_stripe().
 466  *
 467  * Return 0 if we unlock full stripe without problem.
 468  * Return <0 for error
 469  */
 470 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 471                               bool locked)
 472 {
 473         struct btrfs_block_group *bg_cache;
 474         struct btrfs_full_stripe_locks_tree *locks_root;
 475         struct full_stripe_lock *fstripe_lock;
 476         u64 fstripe_start;
 477         bool freeit = false;
 478         int ret = 0;
 479
 480         /* If we didn't acquire full stripe lock, no need to continue */
 481         if (!locked)
 482                 return 0;
 483
 484         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 485         if (!bg_cache) {
 486                 ASSERT(0);
 487                 return -ENOENT;
 488         }
 489         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 490                 goto out;
 491
 492         locks_root = &bg_cache->full_stripe_locks_root;
 493         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 494
 495         mutex_lock(&locks_root->lock);
 496         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 497         /* Unpaired unlock_full_stripe() detected */
 498         if (!fstripe_lock) {
 499                 WARN_ON(1);
 500                 ret = -ENOENT;
 501                 mutex_unlock(&locks_root->lock);
 502                 goto out;
 503         }
 504
 505         if (fstripe_lock->refs == 0) {
 506                 WARN_ON(1);
 507                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 508                         fstripe_lock->logical);
 509         } else {
 510                 fstripe_lock->refs--;
 511         }
 512
 513         if (fstripe_lock->refs == 0) {
 514                 rb_erase(&fstripe_lock->node, &locks_root->root);
 515                 freeit = true;
 516         }
 517         mutex_unlock(&locks_root->lock);
 518
 519         mutex_unlock(&fstripe_lock->mutex);
 520         if (freeit)
 521                 kfree(fstripe_lock);
 522 out:
 523         btrfs_put_block_group(bg_cache);
 524         return ret;
 525 }
 526
 527 static void scrub_free_csums(struct scrub_ctx *sctx)
 528 {
 529         while (!list_empty(&sctx->csum_list)) {
 530                 struct btrfs_ordered_sum *sum;
 531                 sum = list_first_entry(&sctx->csum_list,
 532                                        struct btrfs_ordered_sum, list);
 533                 list_del(&sum->list);
 534                 kfree(sum);
 535         }
 536 }
 537
 538 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 539 {
 540         int i;
 541
 542         if (!sctx)
 543                 return;
 544
 545         /* this can happen when scrub is cancelled */
 546         if (sctx->curr != -1) {
 547                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 548
 549                 for (i = 0; i < sbio->page_count; i++) {
 550                         WARN_ON(!sbio->pagev[i]->page);
 551                         scrub_block_put(sbio->pagev[i]->sblock);
 552                 }
 553                 bio_put(sbio->bio);
 554         }
 555
 556         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 557                 struct scrub_bio *sbio = sctx->bios[i];
 558
 559                 if (!sbio)
 560                         break;
 561                 kfree(sbio);
 562         }
 563
 564         kfree(sctx->wr_curr_bio);
 565         scrub_free_csums(sctx);
 566         kfree(sctx);
 567 }
 568
 569 static void scrub_put_ctx(struct scrub_ctx *sctx)
 570 {
 571         if (refcount_dec_and_test(&sctx->refs))
 572                 scrub_free_ctx(sctx);
 573 }
 574
 575 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 576                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 577 {
 578         struct scrub_ctx *sctx;
 579         int             i;
 580
 581         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 582         if (!sctx)
 583                 goto nomem;
 584         refcount_set(&sctx->refs, 1);
 585         sctx->is_dev_replace = is_dev_replace;
 586         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 587         sctx->curr = -1;
 588         sctx->fs_info = fs_info;
 589         INIT_LIST_HEAD(&sctx->csum_list);
 590         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 591                 struct scrub_bio *sbio;
 592
 593                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 594                 if (!sbio)
 595                         goto nomem;
 596                 sctx->bios[i] = sbio;
 597
 598                 sbio->index = i;
 599                 sbio->sctx = sctx;
 600                 sbio->page_count = 0;
 601                 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker, NULL,
 602                                 NULL);
 603
 604                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 605                         sctx->bios[i]->next_free = i + 1;
 606                 else
 607                         sctx->bios[i]->next_free = -1;
 608         }
 609         sctx->first_free = 0;
 610         atomic_set(&sctx->bios_in_flight, 0);
 611         atomic_set(&sctx->workers_pending, 0);
 612         atomic_set(&sctx->cancel_req, 0);
 613         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 614
 615         spin_lock_init(&sctx->list_lock);
 616         spin_lock_init(&sctx->stat_lock);
 617         init_waitqueue_head(&sctx->list_wait);
 618
 619         WARN_ON(sctx->wr_curr_bio != NULL);
 620         mutex_init(&sctx->wr_lock);
 621         sctx->wr_curr_bio = NULL;
 622         if (is_dev_replace) {
 623                 WARN_ON(!fs_info->dev_replace.tgtdev);
 624                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 625                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 626                 sctx->flush_all_writes = false;
 627         }
 628
 629         return sctx;
 630
 631 nomem:
 632         scrub_free_ctx(sctx);
 633         return ERR_PTR(-ENOMEM);
 634 }
 635
 636 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 637                                      void *warn_ctx)
 638 {
 639         u64 isize;
 640         u32 nlink;
 641         int ret;
 642         int i;
 643         unsigned nofs_flag;
 644         struct extent_buffer *eb;
 645         struct btrfs_inode_item *inode_item;
 646         struct scrub_warning *swarn = warn_ctx;
 647         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 648         struct inode_fs_paths *ipath = NULL;
 649         struct btrfs_root *local_root;
 650         struct btrfs_key key;
 651
 652         local_root = btrfs_get_fs_root(fs_info, root, true);
 653         if (IS_ERR(local_root)) {
 654                 ret = PTR_ERR(local_root);
 655                 goto err;
 656         }
 657
 658         /*
 659          * this makes the path point to (inum INODE_ITEM ioff)
 660          */
 661         key.objectid = inum;
 662         key.type = BTRFS_INODE_ITEM_KEY;
 663         key.offset = 0;
 664
 665         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 666         if (ret) {
 667                 btrfs_put_root(local_root);
 668                 btrfs_release_path(swarn->path);
 669                 goto err;
 670         }
 671
 672         eb = swarn->path->nodes[0];
 673         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 674                                         struct btrfs_inode_item);
 675         isize = btrfs_inode_size(eb, inode_item);
 676         nlink = btrfs_inode_nlink(eb, inode_item);
 677         btrfs_release_path(swarn->path);
 678
 679         /*
 680          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 681          * uses GFP_NOFS in this context, so we keep it consistent but it does
 682          * not seem to be strictly necessary.
 683          */
 684         nofs_flag = memalloc_nofs_save();
 685         ipath = init_ipath(4096, local_root, swarn->path);
 686         memalloc_nofs_restore(nofs_flag);
 687         if (IS_ERR(ipath)) {
 688                 btrfs_put_root(local_root);
 689                 ret = PTR_ERR(ipath);
 690                 ipath = NULL;
 691                 goto err;
 692         }
 693         ret = paths_from_inode(inum, ipath);
 694
 695         if (ret < 0)
 696                 goto err;
 697
 698         /*
 699          * we deliberately ignore the bit ipath might have been too small to
 700          * hold all of the paths here
 701          */
 702         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 703                 btrfs_warn_in_rcu(fs_info,
 704 "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 705                                   swarn->errstr, swarn->logical,
 706                                   rcu_str_deref(swarn->dev->name),
 707                                   swarn->physical,
 708                                   root, inum, offset,
 709                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
 710                                   (char *)(unsigned long)ipath->fspath->val[i]);
 711
 712         btrfs_put_root(local_root);
 713         free_ipath(ipath);
 714         return 0;
 715
 716 err:
 717         btrfs_warn_in_rcu(fs_info,
 718                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 719                           swarn->errstr, swarn->logical,
 720                           rcu_str_deref(swarn->dev->name),
 721                           swarn->physical,
 722                           root, inum, offset, ret);
 723
 724         free_ipath(ipath);
 725         return 0;
 726 }
 727
 728 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 729 {
 730         struct btrfs_device *dev;
 731         struct btrfs_fs_info *fs_info;
 732         struct btrfs_path *path;
 733         struct btrfs_key found_key;
 734         struct extent_buffer *eb;
 735         struct btrfs_extent_item *ei;
 736         struct scrub_warning swarn;
 737         unsigned long ptr = 0;
 738         u64 extent_item_pos;
 739         u64 flags = 0;
 740         u64 ref_root;
 741         u32 item_size;
 742         u8 ref_level = 0;
 743         int ret;
 744
 745         WARN_ON(sblock->page_count < 1);
 746         dev = sblock->pagev[0]->dev;
 747         fs_info = sblock->sctx->fs_info;
 748
 749         path = btrfs_alloc_path();
 750         if (!path)
 751                 return;
 752
 753         swarn.physical = sblock->pagev[0]->physical;
 754         swarn.logical = sblock->pagev[0]->logical;
 755         swarn.errstr = errstr;
 756         swarn.dev = NULL;
 757
 758         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 759                                   &flags);
 760         if (ret < 0)
 761                 goto out;
 762
 763         extent_item_pos = swarn.logical - found_key.objectid;
 764         swarn.extent_item_size = found_key.offset;
 765
 766         eb = path->nodes[0];
 767         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 768         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 769
 770         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 771                 do {
 772                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 773                                                       item_size, &ref_root,
 774                                                       &ref_level);
 775                         btrfs_warn_in_rcu(fs_info,
 776 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 777                                 errstr, swarn.logical,
 778                                 rcu_str_deref(dev->name),
 779                                 swarn.physical,
 780                                 ref_level ? "node" : "leaf",
 781                                 ret < 0 ? -1 : ref_level,
 782                                 ret < 0 ? -1 : ref_root);
 783                 } while (ret != 1);
 784                 btrfs_release_path(path);
 785         } else {
 786                 btrfs_release_path(path);
 787                 swarn.path = path;
 788                 swarn.dev = dev;
 789                 iterate_extent_inodes(fs_info, found_key.objectid,
 790                                         extent_item_pos, 1,
 791                                         scrub_print_warning_inode, &swarn, false);
 792         }
 793
 794 out:
 795         btrfs_free_path(path);
 796 }
 797
 798 static inline void scrub_get_recover(struct scrub_recover *recover)
 799 {
 800         refcount_inc(&recover->refs);
 801 }
 802
 803 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 804                                      struct scrub_recover *recover)
 805 {
 806         if (refcount_dec_and_test(&recover->refs)) {
 807                 btrfs_bio_counter_dec(fs_info);
 808                 btrfs_put_bbio(recover->bbio);
 809                 kfree(recover);
 810         }
 811 }
 812
 813 /*
 814  * scrub_handle_errored_block gets called when either verification of the
 815  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 816  * case, this function handles all pages in the bio, even though only one
 817  * may be bad.
 818  * The goal of this function is to repair the errored block by using the
 819  * contents of one of the mirrors.
 820  */
 821 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 822 {
 823         struct scrub_ctx *sctx = sblock_to_check->sctx;
 824         struct btrfs_device *dev;
 825         struct btrfs_fs_info *fs_info;
 826         u64 logical;
 827         unsigned int failed_mirror_index;
 828         unsigned int is_metadata;
 829         unsigned int have_csum;
 830         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 831         struct scrub_block *sblock_bad;
 832         int ret;
 833         int mirror_index;
 834         int page_num;
 835         int success;
 836         bool full_stripe_locked;
 837         unsigned int nofs_flag;
 838         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 839                                       DEFAULT_RATELIMIT_BURST);
 840
 841         BUG_ON(sblock_to_check->page_count < 1);
 842         fs_info = sctx->fs_info;
 843         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 844                 /*
 845                  * if we find an error in a super block, we just report it.
 846                  * They will get written with the next transaction commit
 847                  * anyway
 848                  */
 849                 spin_lock(&sctx->stat_lock);
 850                 ++sctx->stat.super_errors;
 851                 spin_unlock(&sctx->stat_lock);
 852                 return 0;
 853         }
 854         logical = sblock_to_check->pagev[0]->logical;
 855         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 856         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 857         is_metadata = !(sblock_to_check->pagev[0]->flags &
 858                         BTRFS_EXTENT_FLAG_DATA);
 859         have_csum = sblock_to_check->pagev[0]->have_csum;
 860         dev = sblock_to_check->pagev[0]->dev;
 861
 862         /*
 863          * We must use GFP_NOFS because the scrub task might be waiting for a
 864          * worker task executing this function and in turn a transaction commit
 865          * might be waiting the scrub task to pause (which needs to wait for all
 866          * the worker tasks to complete before pausing).
 867          * We do allocations in the workers through insert_full_stripe_lock()
 868          * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 869          * this function.
 870          */
 871         nofs_flag = memalloc_nofs_save();
 872         /*
 873          * For RAID5/6, race can happen for a different device scrub thread.
 874          * For data corruption, Parity and Data threads will both try
 875          * to recovery the data.
 876          * Race can lead to doubly added csum error, or even unrecoverable
 877          * error.
 878          */
 879         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 880         if (ret < 0) {
 881                 memalloc_nofs_restore(nofs_flag);
 882                 spin_lock(&sctx->stat_lock);
 883                 if (ret == -ENOMEM)
 884                         sctx->stat.malloc_errors++;
 885                 sctx->stat.read_errors++;
 886                 sctx->stat.uncorrectable_errors++;
 887                 spin_unlock(&sctx->stat_lock);
 888                 return ret;
 889         }
 890
 891         /*
 892          * read all mirrors one after the other. This includes to
 893          * re-read the extent or metadata block that failed (that was
 894          * the cause that this fixup code is called) another time,
 895          * page by page this time in order to know which pages
 896          * caused I/O errors and which ones are good (for all mirrors).
 897          * It is the goal to handle the situation when more than one
 898          * mirror contains I/O errors, but the errors do not
 899          * overlap, i.e. the data can be repaired by selecting the
 900          * pages from those mirrors without I/O error on the
 901          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 902          * would be that mirror #1 has an I/O error on the first page,
 903          * the second page is good, and mirror #2 has an I/O error on
 904          * the second page, but the first page is good.
 905          * Then the first page of the first mirror can be repaired by
 906          * taking the first page of the second mirror, and the
 907          * second page of the second mirror can be repaired by
 908          * copying the contents of the 2nd page of the 1st mirror.
 909          * One more note: if the pages of one mirror contain I/O
 910          * errors, the checksum cannot be verified. In order to get
 911          * the best data for repairing, the first attempt is to find
 912          * a mirror without I/O errors and with a validated checksum.
 913          * Only if this is not possible, the pages are picked from
 914          * mirrors with I/O errors without considering the checksum.
 915          * If the latter is the case, at the end, the checksum of the
 916          * repaired area is verified in order to correctly maintain
 917          * the statistics.
 918          */
 919
 920         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 921                                       sizeof(*sblocks_for_recheck), GFP_KERNEL);
 922         if (!sblocks_for_recheck) {
 923                 spin_lock(&sctx->stat_lock);
 924                 sctx->stat.malloc_errors++;
 925                 sctx->stat.read_errors++;
 926                 sctx->stat.uncorrectable_errors++;
 927                 spin_unlock(&sctx->stat_lock);
 928                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 929                 goto out;
 930         }
 931
 932         /* setup the context, map the logical blocks and alloc the pages */
 933         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 934         if (ret) {
 935                 spin_lock(&sctx->stat_lock);
 936                 sctx->stat.read_errors++;
 937                 sctx->stat.uncorrectable_errors++;
 938                 spin_unlock(&sctx->stat_lock);
 939                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 940                 goto out;
 941         }
 942         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 943         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 944
 945         /* build and submit the bios for the failed mirror, check checksums */
 946         scrub_recheck_block(fs_info, sblock_bad, 1);
 947
 948         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 949             sblock_bad->no_io_error_seen) {
 950                 /*
 951                  * the error disappeared after reading page by page, or
 952                  * the area was part of a huge bio and other parts of the
 953                  * bio caused I/O errors, or the block layer merged several
 954                  * read requests into one and the error is caused by a
 955                  * different bio (usually one of the two latter cases is
 956                  * the cause)
 957                  */
 958                 spin_lock(&sctx->stat_lock);
 959                 sctx->stat.unverified_errors++;
 960                 sblock_to_check->data_corrected = 1;
 961                 spin_unlock(&sctx->stat_lock);
 962
 963                 if (sctx->is_dev_replace)
 964                         scrub_write_block_to_dev_replace(sblock_bad);
 965                 goto out;
 966         }
 967
 968         if (!sblock_bad->no_io_error_seen) {
 969                 spin_lock(&sctx->stat_lock);
 970                 sctx->stat.read_errors++;
 971                 spin_unlock(&sctx->stat_lock);
 972                 if (__ratelimit(&_rs))
 973                         scrub_print_warning("i/o error", sblock_to_check);
 974                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 975         } else if (sblock_bad->checksum_error) {
 976                 spin_lock(&sctx->stat_lock);
 977                 sctx->stat.csum_errors++;
 978                 spin_unlock(&sctx->stat_lock);
 979                 if (__ratelimit(&_rs))
 980                         scrub_print_warning("checksum error", sblock_to_check);
 981                 btrfs_dev_stat_inc_and_print(dev,
 982                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 983         } else if (sblock_bad->header_error) {
 984                 spin_lock(&sctx->stat_lock);
 985                 sctx->stat.verify_errors++;
 986                 spin_unlock(&sctx->stat_lock);
 987                 if (__ratelimit(&_rs))
 988                         scrub_print_warning("checksum/header error",
 989                                             sblock_to_check);
 990                 if (sblock_bad->generation_error)
 991                         btrfs_dev_stat_inc_and_print(dev,
 992                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 993                 else
 994                         btrfs_dev_stat_inc_and_print(dev,
 995                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 996         }
 997
 998         if (sctx->readonly) {
 999                 ASSERT(!sctx->is_dev_replace);
1000                 goto out;
1001         }
1002
1003         /*
1004          * now build and submit the bios for the other mirrors, check
1005          * checksums.
1006          * First try to pick the mirror which is completely without I/O
1007          * errors and also does not have a checksum error.
1008          * If one is found, and if a checksum is present, the full block
1009          * that is known to contain an error is rewritten. Afterwards
1010          * the block is known to be corrected.
1011          * If a mirror is found which is completely correct, and no
1012          * checksum is present, only those pages are rewritten that had
1013          * an I/O error in the block to be repaired, since it cannot be
1014          * determined, which copy of the other pages is better (and it
1015          * could happen otherwise that a correct page would be
1016          * overwritten by a bad one).
1017          */
1018         for (mirror_index = 0; ;mirror_index++) {
1019                 struct scrub_block *sblock_other;
1020
1021                 if (mirror_index == failed_mirror_index)
1022                         continue;
1023
1024                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1025                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1026                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1027                                 break;
1028                         if (!sblocks_for_recheck[mirror_index].page_count)
1029                                 break;
1030
1031                         sblock_other = sblocks_for_recheck + mirror_index;
1032                 } else {
1033                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1034                         int max_allowed = r->bbio->num_stripes -
1035                                                 r->bbio->num_tgtdevs;
1036
1037                         if (mirror_index >= max_allowed)
1038                                 break;
1039                         if (!sblocks_for_recheck[1].page_count)
1040                                 break;
1041
1042                         ASSERT(failed_mirror_index == 0);
1043                         sblock_other = sblocks_for_recheck + 1;
1044                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1045                 }
1046
1047                 /* build and submit the bios, check checksums */
1048                 scrub_recheck_block(fs_info, sblock_other, 0);
1049
1050                 if (!sblock_other->header_error &&
1051                     !sblock_other->checksum_error &&
1052                     sblock_other->no_io_error_seen) {
1053                         if (sctx->is_dev_replace) {
1054                                 scrub_write_block_to_dev_replace(sblock_other);
1055                                 goto corrected_error;
1056                         } else {
1057                                 ret = scrub_repair_block_from_good_copy(
1058                                                 sblock_bad, sblock_other);
1059                                 if (!ret)
1060                                         goto corrected_error;
1061                         }
1062                 }
1063         }
1064
1065         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1066                 goto did_not_correct_error;
1067
1068         /*
1069          * In case of I/O errors in the area that is supposed to be
1070          * repaired, continue by picking good copies of those pages.
1071          * Select the good pages from mirrors to rewrite bad pages from
1072          * the area to fix. Afterwards verify the checksum of the block
1073          * that is supposed to be repaired. This verification step is
1074          * only done for the purpose of statistic counting and for the
1075          * final scrub report, whether errors remain.
1076          * A perfect algorithm could make use of the checksum and try
1077          * all possible combinations of pages from the different mirrors
1078          * until the checksum verification succeeds. For example, when
1079          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1080          * of mirror #2 is readable but the final checksum test fails,
1081          * then the 2nd page of mirror #3 could be tried, whether now
1082          * the final checksum succeeds. But this would be a rare
1083          * exception and is therefore not implemented. At least it is
1084          * avoided that the good copy is overwritten.
1085          * A more useful improvement would be to pick the sectors
1086          * without I/O error based on sector sizes (512 bytes on legacy
1087          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1088          * mirror could be repaired by taking 512 byte of a different
1089          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1090          * area are unreadable.
1091          */
1092         success = 1;
1093         for (page_num = 0; page_num < sblock_bad->page_count;
1094              page_num++) {
1095                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1096                 struct scrub_block *sblock_other = NULL;
1097
1098                 /* skip no-io-error page in scrub */
1099                 if (!page_bad->io_error && !sctx->is_dev_replace)
1100                         continue;
1101
1102                 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1103                         /*
1104                          * In case of dev replace, if raid56 rebuild process
1105                          * didn't work out correct data, then copy the content
1106                          * in sblock_bad to make sure target device is identical
1107                          * to source device, instead of writing garbage data in
1108                          * sblock_for_recheck array to target device.
1109                          */
1110                         sblock_other = NULL;
1111                 } else if (page_bad->io_error) {
1112                         /* try to find no-io-error page in mirrors */
1113                         for (mirror_index = 0;
1114                              mirror_index < BTRFS_MAX_MIRRORS &&
1115                              sblocks_for_recheck[mirror_index].page_count > 0;
1116                              mirror_index++) {
1117                                 if (!sblocks_for_recheck[mirror_index].
1118                                     pagev[page_num]->io_error) {
1119                                         sblock_other = sblocks_for_recheck +
1120                                                        mirror_index;
1121                                         break;
1122                                 }
1123                         }
1124                         if (!sblock_other)
1125                                 success = 0;
1126                 }
1127
1128                 if (sctx->is_dev_replace) {
1129                         /*
1130                          * did not find a mirror to fetch the page
1131                          * from. scrub_write_page_to_dev_replace()
1132                          * handles this case (page->io_error), by
1133                          * filling the block with zeros before
1134                          * submitting the write request
1135                          */
1136                         if (!sblock_other)
1137                                 sblock_other = sblock_bad;
1138
1139                         if (scrub_write_page_to_dev_replace(sblock_other,
1140                                                             page_num) != 0) {
1141                                 atomic64_inc(
1142                                         &fs_info->dev_replace.num_write_errors);
1143                                 success = 0;
1144                         }
1145                 } else if (sblock_other) {
1146                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1147                                                                sblock_other,
1148                                                                page_num, 0);
1149                         if (0 == ret)
1150                                 page_bad->io_error = 0;
1151                         else
1152                                 success = 0;
1153                 }
1154         }
1155
1156         if (success && !sctx->is_dev_replace) {
1157                 if (is_metadata || have_csum) {
1158                         /*
1159                          * need to verify the checksum now that all
1160                          * sectors on disk are repaired (the write
1161                          * request for data to be repaired is on its way).
1162                          * Just be lazy and use scrub_recheck_block()
1163                          * which re-reads the data before the checksum
1164                          * is verified, but most likely the data comes out
1165                          * of the page cache.
1166                          */
1167                         scrub_recheck_block(fs_info, sblock_bad, 1);
1168                         if (!sblock_bad->header_error &&
1169                             !sblock_bad->checksum_error &&
1170                             sblock_bad->no_io_error_seen)
1171                                 goto corrected_error;
1172                         else
1173                                 goto did_not_correct_error;
1174                 } else {
1175 corrected_error:
1176                         spin_lock(&sctx->stat_lock);
1177                         sctx->stat.corrected_errors++;
1178                         sblock_to_check->data_corrected = 1;
1179                         spin_unlock(&sctx->stat_lock);
1180                         btrfs_err_rl_in_rcu(fs_info,
1181                                 "fixed up error at logical %llu on dev %s",
1182                                 logical, rcu_str_deref(dev->name));
1183                 }
1184         } else {
1185 did_not_correct_error:
1186                 spin_lock(&sctx->stat_lock);
1187                 sctx->stat.uncorrectable_errors++;
1188                 spin_unlock(&sctx->stat_lock);
1189                 btrfs_err_rl_in_rcu(fs_info,
1190                         "unable to fixup (regular) error at logical %llu on dev %s",
1191                         logical, rcu_str_deref(dev->name));
1192         }
1193
1194 out:
1195         if (sblocks_for_recheck) {
1196                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1197                      mirror_index++) {
1198                         struct scrub_block *sblock = sblocks_for_recheck +
1199                                                      mirror_index;
1200                         struct scrub_recover *recover;
1201                         int page_index;
1202
1203                         for (page_index = 0; page_index < sblock->page_count;
1204                              page_index++) {
1205                                 sblock->pagev[page_index]->sblock = NULL;
1206                                 recover = sblock->pagev[page_index]->recover;
1207                                 if (recover) {
1208                                         scrub_put_recover(fs_info, recover);
1209                                         sblock->pagev[page_index]->recover =
1210                                                                         NULL;
1211                                 }
1212                                 scrub_page_put(sblock->pagev[page_index]);
1213                         }
1214                 }
1215                 kfree(sblocks_for_recheck);
1216         }
1217
1218         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1219         memalloc_nofs_restore(nofs_flag);
1220         if (ret < 0)
1221                 return ret;
1222         return 0;
1223 }
1224
1225 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1226 {
1227         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1228                 return 2;
1229         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1230                 return 3;
1231         else
1232                 return (int)bbio->num_stripes;
1233 }
1234
1235 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1236                                                  u64 *raid_map,
1237                                                  u64 mapped_length,
1238                                                  int nstripes, int mirror,
1239                                                  int *stripe_index,
1240                                                  u64 *stripe_offset)
1241 {
1242         int i;
1243
1244         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1245                 /* RAID5/6 */
1246                 for (i = 0; i < nstripes; i++) {
1247                         if (raid_map[i] == RAID6_Q_STRIPE ||
1248                             raid_map[i] == RAID5_P_STRIPE)
1249                                 continue;
1250
1251                         if (logical >= raid_map[i] &&
1252                             logical < raid_map[i] + mapped_length)
1253                                 break;
1254                 }
1255
1256                 *stripe_index = i;
1257                 *stripe_offset = logical - raid_map[i];
1258         } else {
1259                 /* The other RAID type */
1260                 *stripe_index = mirror;
1261                 *stripe_offset = 0;
1262         }
1263 }
1264
1265 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1266                                      struct scrub_block *sblocks_for_recheck)
1267 {
1268         struct scrub_ctx *sctx = original_sblock->sctx;
1269         struct btrfs_fs_info *fs_info = sctx->fs_info;
1270         u64 length = original_sblock->page_count * PAGE_SIZE;
1271         u64 logical = original_sblock->pagev[0]->logical;
1272         u64 generation = original_sblock->pagev[0]->generation;
1273         u64 flags = original_sblock->pagev[0]->flags;
1274         u64 have_csum = original_sblock->pagev[0]->have_csum;
1275         struct scrub_recover *recover;
1276         struct btrfs_bio *bbio;
1277         u64 sublen;
1278         u64 mapped_length;
1279         u64 stripe_offset;
1280         int stripe_index;
1281         int page_index = 0;
1282         int mirror_index;
1283         int nmirrors;
1284         int ret;
1285
1286         /*
1287          * note: the two members refs and outstanding_pages
1288          * are not used (and not set) in the blocks that are used for
1289          * the recheck procedure
1290          */
1291
1292         while (length > 0) {
1293                 sublen = min_t(u64, length, PAGE_SIZE);
1294                 mapped_length = sublen;
1295                 bbio = NULL;
1296
1297                 /*
1298                  * with a length of PAGE_SIZE, each returned stripe
1299                  * represents one mirror
1300                  */
1301                 btrfs_bio_counter_inc_blocked(fs_info);
1302                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1303                                 logical, &mapped_length, &bbio);
1304                 if (ret || !bbio || mapped_length < sublen) {
1305                         btrfs_put_bbio(bbio);
1306                         btrfs_bio_counter_dec(fs_info);
1307                         return -EIO;
1308                 }
1309
1310                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1311                 if (!recover) {
1312                         btrfs_put_bbio(bbio);
1313                         btrfs_bio_counter_dec(fs_info);
1314                         return -ENOMEM;
1315                 }
1316
1317                 refcount_set(&recover->refs, 1);
1318                 recover->bbio = bbio;
1319                 recover->map_length = mapped_length;
1320
1321                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1322
1323                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1324
1325                 for (mirror_index = 0; mirror_index < nmirrors;
1326                      mirror_index++) {
1327                         struct scrub_block *sblock;
1328                         struct scrub_page *page;
1329
1330                         sblock = sblocks_for_recheck + mirror_index;
1331                         sblock->sctx = sctx;
1332
1333                         page = kzalloc(sizeof(*page), GFP_NOFS);
1334                         if (!page) {
1335 leave_nomem:
1336                                 spin_lock(&sctx->stat_lock);
1337                                 sctx->stat.malloc_errors++;
1338                                 spin_unlock(&sctx->stat_lock);
1339                                 scrub_put_recover(fs_info, recover);
1340                                 return -ENOMEM;
1341                         }
1342                         scrub_page_get(page);
1343                         sblock->pagev[page_index] = page;
1344                         page->sblock = sblock;
1345                         page->flags = flags;
1346                         page->generation = generation;
1347                         page->logical = logical;
1348                         page->have_csum = have_csum;
1349                         if (have_csum)
1350                                 memcpy(page->csum,
1351                                        original_sblock->pagev[0]->csum,
1352                                        sctx->csum_size);
1353
1354                         scrub_stripe_index_and_offset(logical,
1355                                                       bbio->map_type,
1356                                                       bbio->raid_map,
1357                                                       mapped_length,
1358                                                       bbio->num_stripes -
1359                                                       bbio->num_tgtdevs,
1360                                                       mirror_index,
1361                                                       &stripe_index,
1362                                                       &stripe_offset);
1363                         page->physical = bbio->stripes[stripe_index].physical +
1364                                          stripe_offset;
1365                         page->dev = bbio->stripes[stripe_index].dev;
1366
1367                         BUG_ON(page_index >= original_sblock->page_count);
1368                         page->physical_for_dev_replace =
1369                                 original_sblock->pagev[page_index]->
1370                                 physical_for_dev_replace;
1371                         /* for missing devices, dev->bdev is NULL */
1372                         page->mirror_num = mirror_index + 1;
1373                         sblock->page_count++;
1374                         page->page = alloc_page(GFP_NOFS);
1375                         if (!page->page)
1376                                 goto leave_nomem;
1377
1378                         scrub_get_recover(recover);
1379                         page->recover = recover;
1380                 }
1381                 scrub_put_recover(fs_info, recover);
1382                 length -= sublen;
1383                 logical += sublen;
1384                 page_index++;
1385         }
1386
1387         return 0;
1388 }
1389
1390 static void scrub_bio_wait_endio(struct bio *bio)
1391 {
1392         complete(bio->bi_private);
1393 }
1394
1395 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1396                                         struct bio *bio,
1397                                         struct scrub_page *page)
1398 {
1399         DECLARE_COMPLETION_ONSTACK(done);
1400         int ret;
1401         int mirror_num;
1402
1403         bio->bi_iter.bi_sector = page->logical >> 9;
1404         bio->bi_private = &done;
1405         bio->bi_end_io = scrub_bio_wait_endio;
1406
1407         mirror_num = page->sblock->pagev[0]->mirror_num;
1408         ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1409                                     page->recover->map_length,
1410                                     mirror_num, 0);
1411         if (ret)
1412                 return ret;
1413
1414         wait_for_completion_io(&done);
1415         return blk_status_to_errno(bio->bi_status);
1416 }
1417
1418 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1419                                           struct scrub_block *sblock)
1420 {
1421         struct scrub_page *first_page = sblock->pagev[0];
1422         struct bio *bio;
1423         int page_num;
1424
1425         /* All pages in sblock belong to the same stripe on the same device. */
1426         ASSERT(first_page->dev);
1427         if (!first_page->dev->bdev)
1428                 goto out;
1429
1430         bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1431         bio_set_dev(bio, first_page->dev->bdev);
1432
1433         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1434                 struct scrub_page *page = sblock->pagev[page_num];
1435
1436                 WARN_ON(!page->page);
1437                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1438         }
1439
1440         if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1441                 bio_put(bio);
1442                 goto out;
1443         }
1444
1445         bio_put(bio);
1446
1447         scrub_recheck_block_checksum(sblock);
1448
1449         return;
1450 out:
1451         for (page_num = 0; page_num < sblock->page_count; page_num++)
1452                 sblock->pagev[page_num]->io_error = 1;
1453
1454         sblock->no_io_error_seen = 0;
1455 }
1456
1457 /*
1458  * this function will check the on disk data for checksum errors, header
1459  * errors and read I/O errors. If any I/O errors happen, the exact pages
1460  * which are errored are marked as being bad. The goal is to enable scrub
1461  * to take those pages that are not errored from all the mirrors so that
1462  * the pages that are errored in the just handled mirror can be repaired.
1463  */
1464 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1465                                 struct scrub_block *sblock,
1466                                 int retry_failed_mirror)
1467 {
1468         int page_num;
1469
1470         sblock->no_io_error_seen = 1;
1471
1472         /* short cut for raid56 */
1473         if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1474                 return scrub_recheck_block_on_raid56(fs_info, sblock);
1475
1476         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1477                 struct bio *bio;
1478                 struct scrub_page *page = sblock->pagev[page_num];
1479
1480                 if (page->dev->bdev == NULL) {
1481                         page->io_error = 1;
1482                         sblock->no_io_error_seen = 0;
1483                         continue;
1484                 }
1485
1486                 WARN_ON(!page->page);
1487                 bio = btrfs_io_bio_alloc(1);
1488                 bio_set_dev(bio, page->dev->bdev);
1489
1490                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1491                 bio->bi_iter.bi_sector = page->physical >> 9;
1492                 bio->bi_opf = REQ_OP_READ;
1493
1494                 if (btrfsic_submit_bio_wait(bio)) {
1495                         page->io_error = 1;
1496                         sblock->no_io_error_seen = 0;
1497                 }
1498
1499                 bio_put(bio);
1500         }
1501
1502         if (sblock->no_io_error_seen)
1503                 scrub_recheck_block_checksum(sblock);
1504 }
1505
1506 static inline int scrub_check_fsid(u8 fsid[],
1507                                    struct scrub_page *spage)
1508 {
1509         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1510         int ret;
1511
1512         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1513         return !ret;
1514 }
1515
1516 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1517 {
1518         sblock->header_error = 0;
1519         sblock->checksum_error = 0;
1520         sblock->generation_error = 0;
1521
1522         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1523                 scrub_checksum_data(sblock);
1524         else
1525                 scrub_checksum_tree_block(sblock);
1526 }
1527
1528 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1529                                              struct scrub_block *sblock_good)
1530 {
1531         int page_num;
1532         int ret = 0;
1533
1534         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1535                 int ret_sub;
1536
1537                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1538                                                            sblock_good,
1539                                                            page_num, 1);
1540                 if (ret_sub)
1541                         ret = ret_sub;
1542         }
1543
1544         return ret;
1545 }
1546
1547 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1548                                             struct scrub_block *sblock_good,
1549                                             int page_num, int force_write)
1550 {
1551         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1552         struct scrub_page *page_good = sblock_good->pagev[page_num];
1553         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1554
1555         BUG_ON(page_bad->page == NULL);
1556         BUG_ON(page_good->page == NULL);
1557         if (force_write || sblock_bad->header_error ||
1558             sblock_bad->checksum_error || page_bad->io_error) {
1559                 struct bio *bio;
1560                 int ret;
1561
1562                 if (!page_bad->dev->bdev) {
1563                         btrfs_warn_rl(fs_info,
1564                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1565                         return -EIO;
1566                 }
1567
1568                 bio = btrfs_io_bio_alloc(1);
1569                 bio_set_dev(bio, page_bad->dev->bdev);
1570                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1571                 bio->bi_opf = REQ_OP_WRITE;
1572
1573                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1574                 if (PAGE_SIZE != ret) {
1575                         bio_put(bio);
1576                         return -EIO;
1577                 }
1578
1579                 if (btrfsic_submit_bio_wait(bio)) {
1580                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1581                                 BTRFS_DEV_STAT_WRITE_ERRS);
1582                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1583                         bio_put(bio);
1584                         return -EIO;
1585                 }
1586                 bio_put(bio);
1587         }
1588
1589         return 0;
1590 }
1591
1592 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1593 {
1594         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1595         int page_num;
1596
1597         /*
1598          * This block is used for the check of the parity on the source device,
1599          * so the data needn't be written into the destination device.
1600          */
1601         if (sblock->sparity)
1602                 return;
1603
1604         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1605                 int ret;
1606
1607                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1608                 if (ret)
1609                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1610         }
1611 }
1612
1613 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1614                                            int page_num)
1615 {
1616         struct scrub_page *spage = sblock->pagev[page_num];
1617
1618         BUG_ON(spage->page == NULL);
1619         if (spage->io_error) {
1620                 void *mapped_buffer = kmap_atomic(spage->page);
1621
1622                 clear_page(mapped_buffer);
1623                 flush_dcache_page(spage->page);
1624                 kunmap_atomic(mapped_buffer);
1625         }
1626         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1627 }
1628
1629 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1630                                     struct scrub_page *spage)
1631 {
1632         struct scrub_bio *sbio;
1633         int ret;
1634
1635         mutex_lock(&sctx->wr_lock);
1636 again:
1637         if (!sctx->wr_curr_bio) {
1638                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1639                                               GFP_KERNEL);
1640                 if (!sctx->wr_curr_bio) {
1641                         mutex_unlock(&sctx->wr_lock);
1642                         return -ENOMEM;
1643                 }
1644                 sctx->wr_curr_bio->sctx = sctx;
1645                 sctx->wr_curr_bio->page_count = 0;
1646         }
1647         sbio = sctx->wr_curr_bio;
1648         if (sbio->page_count == 0) {
1649                 struct bio *bio;
1650
1651                 sbio->physical = spage->physical_for_dev_replace;
1652                 sbio->logical = spage->logical;
1653                 sbio->dev = sctx->wr_tgtdev;
1654                 bio = sbio->bio;
1655                 if (!bio) {
1656                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1657                         sbio->bio = bio;
1658                 }
1659
1660                 bio->bi_private = sbio;
1661                 bio->bi_end_io = scrub_wr_bio_end_io;
1662                 bio_set_dev(bio, sbio->dev->bdev);
1663                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1664                 bio->bi_opf = REQ_OP_WRITE;
1665                 sbio->status = 0;
1666         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1667                    spage->physical_for_dev_replace ||
1668                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1669                    spage->logical) {
1670                 scrub_wr_submit(sctx);
1671                 goto again;
1672         }
1673
1674         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1675         if (ret != PAGE_SIZE) {
1676                 if (sbio->page_count < 1) {
1677                         bio_put(sbio->bio);
1678                         sbio->bio = NULL;
1679                         mutex_unlock(&sctx->wr_lock);
1680                         return -EIO;
1681                 }
1682                 scrub_wr_submit(sctx);
1683                 goto again;
1684         }
1685
1686         sbio->pagev[sbio->page_count] = spage;
1687         scrub_page_get(spage);
1688         sbio->page_count++;
1689         if (sbio->page_count == sctx->pages_per_wr_bio)
1690                 scrub_wr_submit(sctx);
1691         mutex_unlock(&sctx->wr_lock);
1692
1693         return 0;
1694 }
1695
1696 static void scrub_wr_submit(struct scrub_ctx *sctx)
1697 {
1698         struct scrub_bio *sbio;
1699
1700         if (!sctx->wr_curr_bio)
1701                 return;
1702
1703         sbio = sctx->wr_curr_bio;
1704         sctx->wr_curr_bio = NULL;
1705         WARN_ON(!sbio->bio->bi_disk);
1706         scrub_pending_bio_inc(sctx);
1707         /* process all writes in a single worker thread. Then the block layer
1708          * orders the requests before sending them to the driver which
1709          * doubled the write performance on spinning disks when measured
1710          * with Linux 3.5 */
1711         btrfsic_submit_bio(sbio->bio);
1712 }
1713
1714 static void scrub_wr_bio_end_io(struct bio *bio)
1715 {
1716         struct scrub_bio *sbio = bio->bi_private;
1717         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1718
1719         sbio->status = bio->bi_status;
1720         sbio->bio = bio;
1721
1722         btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1723         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1724 }
1725
1726 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1727 {
1728         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1729         struct scrub_ctx *sctx = sbio->sctx;
1730         int i;
1731
1732         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1733         if (sbio->status) {
1734                 struct btrfs_dev_replace *dev_replace =
1735                         &sbio->sctx->fs_info->dev_replace;
1736
1737                 for (i = 0; i < sbio->page_count; i++) {
1738                         struct scrub_page *spage = sbio->pagev[i];
1739
1740                         spage->io_error = 1;
1741                         atomic64_inc(&dev_replace->num_write_errors);
1742                 }
1743         }
1744
1745         for (i = 0; i < sbio->page_count; i++)
1746                 scrub_page_put(sbio->pagev[i]);
1747
1748         bio_put(sbio->bio);
1749         kfree(sbio);
1750         scrub_pending_bio_dec(sctx);
1751 }
1752
1753 static int scrub_checksum(struct scrub_block *sblock)
1754 {
1755         u64 flags;
1756         int ret;
1757
1758         /*
1759          * No need to initialize these stats currently,
1760          * because this function only use return value
1761          * instead of these stats value.
1762          *
1763          * Todo:
1764          * always use stats
1765          */
1766         sblock->header_error = 0;
1767         sblock->generation_error = 0;
1768         sblock->checksum_error = 0;
1769
1770         WARN_ON(sblock->page_count < 1);
1771         flags = sblock->pagev[0]->flags;
1772         ret = 0;
1773         if (flags & BTRFS_EXTENT_FLAG_DATA)
1774                 ret = scrub_checksum_data(sblock);
1775         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1776                 ret = scrub_checksum_tree_block(sblock);
1777         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1778                 (void)scrub_checksum_super(sblock);
1779         else
1780                 WARN_ON(1);
1781         if (ret)
1782                 scrub_handle_errored_block(sblock);
1783
1784         return ret;
1785 }
1786
1787 static int scrub_checksum_data(struct scrub_block *sblock)
1788 {
1789         struct scrub_ctx *sctx = sblock->sctx;
1790         struct btrfs_fs_info *fs_info = sctx->fs_info;
1791         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1792         u8 csum[BTRFS_CSUM_SIZE];
1793         u8 *on_disk_csum;
1794         struct page *page;
1795         void *buffer;
1796         u64 len;
1797         int index;
1798
1799         BUG_ON(sblock->page_count < 1);
1800         if (!sblock->pagev[0]->have_csum)
1801                 return 0;
1802
1803         shash->tfm = fs_info->csum_shash;
1804         crypto_shash_init(shash);
1805
1806         on_disk_csum = sblock->pagev[0]->csum;
1807         page = sblock->pagev[0]->page;
1808         buffer = kmap_atomic(page);
1809
1810         len = sctx->fs_info->sectorsize;
1811         index = 0;
1812         for (;;) {
1813                 u64 l = min_t(u64, len, PAGE_SIZE);
1814
1815                 crypto_shash_update(shash, buffer, l);
1816                 kunmap_atomic(buffer);
1817                 len -= l;
1818                 if (len == 0)
1819                         break;
1820                 index++;
1821                 BUG_ON(index >= sblock->page_count);
1822                 BUG_ON(!sblock->pagev[index]->page);
1823                 page = sblock->pagev[index]->page;
1824                 buffer = kmap_atomic(page);
1825         }
1826
1827         crypto_shash_final(shash, csum);
1828         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1829                 sblock->checksum_error = 1;
1830
1831         return sblock->checksum_error;
1832 }
1833
1834 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1835 {
1836         struct scrub_ctx *sctx = sblock->sctx;
1837         struct btrfs_header *h;
1838         struct btrfs_fs_info *fs_info = sctx->fs_info;
1839         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1840         u8 calculated_csum[BTRFS_CSUM_SIZE];
1841         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1842         struct page *page;
1843         void *mapped_buffer;
1844         u64 mapped_size;
1845         void *p;
1846         u64 len;
1847         int index;
1848
1849         shash->tfm = fs_info->csum_shash;
1850         crypto_shash_init(shash);
1851
1852         BUG_ON(sblock->page_count < 1);
1853         page = sblock->pagev[0]->page;
1854         mapped_buffer = kmap_atomic(page);
1855         h = (struct btrfs_header *)mapped_buffer;
1856         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1857
1858         /*
1859          * we don't use the getter functions here, as we
1860          * a) don't have an extent buffer and
1861          * b) the page is already kmapped
1862          */
1863         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1864                 sblock->header_error = 1;
1865
1866         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1867                 sblock->header_error = 1;
1868                 sblock->generation_error = 1;
1869         }
1870
1871         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1872                 sblock->header_error = 1;
1873
1874         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1875                    BTRFS_UUID_SIZE))
1876                 sblock->header_error = 1;
1877
1878         len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1879         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1880         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1881         index = 0;
1882         for (;;) {
1883                 u64 l = min_t(u64, len, mapped_size);
1884
1885                 crypto_shash_update(shash, p, l);
1886                 kunmap_atomic(mapped_buffer);
1887                 len -= l;
1888                 if (len == 0)
1889                         break;
1890                 index++;
1891                 BUG_ON(index >= sblock->page_count);
1892                 BUG_ON(!sblock->pagev[index]->page);
1893                 page = sblock->pagev[index]->page;
1894                 mapped_buffer = kmap_atomic(page);
1895                 mapped_size = PAGE_SIZE;
1896                 p = mapped_buffer;
1897         }
1898
1899         crypto_shash_final(shash, calculated_csum);
1900         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1901                 sblock->checksum_error = 1;
1902
1903         return sblock->header_error || sblock->checksum_error;
1904 }
1905
1906 static int scrub_checksum_super(struct scrub_block *sblock)
1907 {
1908         struct btrfs_super_block *s;
1909         struct scrub_ctx *sctx = sblock->sctx;
1910         struct btrfs_fs_info *fs_info = sctx->fs_info;
1911         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
1912         u8 calculated_csum[BTRFS_CSUM_SIZE];
1913         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1914         struct page *page;
1915         void *mapped_buffer;
1916         u64 mapped_size;
1917         void *p;
1918         int fail_gen = 0;
1919         int fail_cor = 0;
1920         u64 len;
1921         int index;
1922
1923         shash->tfm = fs_info->csum_shash;
1924         crypto_shash_init(shash);
1925
1926         BUG_ON(sblock->page_count < 1);
1927         page = sblock->pagev[0]->page;
1928         mapped_buffer = kmap_atomic(page);
1929         s = (struct btrfs_super_block *)mapped_buffer;
1930         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1931
1932         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1933                 ++fail_cor;
1934
1935         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1936                 ++fail_gen;
1937
1938         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1939                 ++fail_cor;
1940
1941         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1942         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1943         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1944         index = 0;
1945         for (;;) {
1946                 u64 l = min_t(u64, len, mapped_size);
1947
1948                 crypto_shash_update(shash, p, l);
1949                 kunmap_atomic(mapped_buffer);
1950                 len -= l;
1951                 if (len == 0)
1952                         break;
1953                 index++;
1954                 BUG_ON(index >= sblock->page_count);
1955                 BUG_ON(!sblock->pagev[index]->page);
1956                 page = sblock->pagev[index]->page;
1957                 mapped_buffer = kmap_atomic(page);
1958                 mapped_size = PAGE_SIZE;
1959                 p = mapped_buffer;
1960         }
1961
1962         crypto_shash_final(shash, calculated_csum);
1963         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1964                 ++fail_cor;
1965
1966         if (fail_cor + fail_gen) {
1967                 /*
1968                  * if we find an error in a super block, we just report it.
1969                  * They will get written with the next transaction commit
1970                  * anyway
1971                  */
1972                 spin_lock(&sctx->stat_lock);
1973                 ++sctx->stat.super_errors;
1974                 spin_unlock(&sctx->stat_lock);
1975                 if (fail_cor)
1976                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1977                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1978                 else
1979                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1980                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1981         }
1982
1983         return fail_cor + fail_gen;
1984 }
1985
1986 static void scrub_block_get(struct scrub_block *sblock)
1987 {
1988         refcount_inc(&sblock->refs);
1989 }
1990
1991 static void scrub_block_put(struct scrub_block *sblock)
1992 {
1993         if (refcount_dec_and_test(&sblock->refs)) {
1994                 int i;
1995
1996                 if (sblock->sparity)
1997                         scrub_parity_put(sblock->sparity);
1998
1999                 for (i = 0; i < sblock->page_count; i++)
2000                         scrub_page_put(sblock->pagev[i]);
2001                 kfree(sblock);
2002         }
2003 }
2004
2005 static void scrub_page_get(struct scrub_page *spage)
2006 {
2007         atomic_inc(&spage->refs);
2008 }
2009
2010 static void scrub_page_put(struct scrub_page *spage)
2011 {
2012         if (atomic_dec_and_test(&spage->refs)) {
2013                 if (spage->page)
2014                         __free_page(spage->page);
2015                 kfree(spage);
2016         }
2017 }
2018
2019 static void scrub_submit(struct scrub_ctx *sctx)
2020 {
2021         struct scrub_bio *sbio;
2022
2023         if (sctx->curr == -1)
2024                 return;
2025
2026         sbio = sctx->bios[sctx->curr];
2027         sctx->curr = -1;
2028         scrub_pending_bio_inc(sctx);
2029         btrfsic_submit_bio(sbio->bio);
2030 }
2031
2032 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2033                                     struct scrub_page *spage)
2034 {
2035         struct scrub_block *sblock = spage->sblock;
2036         struct scrub_bio *sbio;
2037         int ret;
2038
2039 again:
2040         /*
2041          * grab a fresh bio or wait for one to become available
2042          */
2043         while (sctx->curr == -1) {
2044                 spin_lock(&sctx->list_lock);
2045                 sctx->curr = sctx->first_free;
2046                 if (sctx->curr != -1) {
2047                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2048                         sctx->bios[sctx->curr]->next_free = -1;
2049                         sctx->bios[sctx->curr]->page_count = 0;
2050                         spin_unlock(&sctx->list_lock);
2051                 } else {
2052                         spin_unlock(&sctx->list_lock);
2053                         wait_event(sctx->list_wait, sctx->first_free != -1);
2054                 }
2055         }
2056         sbio = sctx->bios[sctx->curr];
2057         if (sbio->page_count == 0) {
2058                 struct bio *bio;
2059
2060                 sbio->physical = spage->physical;
2061                 sbio->logical = spage->logical;
2062                 sbio->dev = spage->dev;
2063                 bio = sbio->bio;
2064                 if (!bio) {
2065                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2066                         sbio->bio = bio;
2067                 }
2068
2069                 bio->bi_private = sbio;
2070                 bio->bi_end_io = scrub_bio_end_io;
2071                 bio_set_dev(bio, sbio->dev->bdev);
2072                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2073                 bio->bi_opf = REQ_OP_READ;
2074                 sbio->status = 0;
2075         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2076                    spage->physical ||
2077                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2078                    spage->logical ||
2079                    sbio->dev != spage->dev) {
2080                 scrub_submit(sctx);
2081                 goto again;
2082         }
2083
2084         sbio->pagev[sbio->page_count] = spage;
2085         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2086         if (ret != PAGE_SIZE) {
2087                 if (sbio->page_count < 1) {
2088                         bio_put(sbio->bio);
2089                         sbio->bio = NULL;
2090                         return -EIO;
2091                 }
2092                 scrub_submit(sctx);
2093                 goto again;
2094         }
2095
2096         scrub_block_get(sblock); /* one for the page added to the bio */
2097         atomic_inc(&sblock->outstanding_pages);
2098         sbio->page_count++;
2099         if (sbio->page_count == sctx->pages_per_rd_bio)
2100                 scrub_submit(sctx);
2101
2102         return 0;
2103 }
2104
2105 static void scrub_missing_raid56_end_io(struct bio *bio)
2106 {
2107         struct scrub_block *sblock = bio->bi_private;
2108         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2109
2110         if (bio->bi_status)
2111                 sblock->no_io_error_seen = 0;
2112
2113         bio_put(bio);
2114
2115         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2116 }
2117
2118 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2119 {
2120         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2121         struct scrub_ctx *sctx = sblock->sctx;
2122         struct btrfs_fs_info *fs_info = sctx->fs_info;
2123         u64 logical;
2124         struct btrfs_device *dev;
2125
2126         logical = sblock->pagev[0]->logical;
2127         dev = sblock->pagev[0]->dev;
2128
2129         if (sblock->no_io_error_seen)
2130                 scrub_recheck_block_checksum(sblock);
2131
2132         if (!sblock->no_io_error_seen) {
2133                 spin_lock(&sctx->stat_lock);
2134                 sctx->stat.read_errors++;
2135                 spin_unlock(&sctx->stat_lock);
2136                 btrfs_err_rl_in_rcu(fs_info,
2137                         "IO error rebuilding logical %llu for dev %s",
2138                         logical, rcu_str_deref(dev->name));
2139         } else if (sblock->header_error || sblock->checksum_error) {
2140                 spin_lock(&sctx->stat_lock);
2141                 sctx->stat.uncorrectable_errors++;
2142                 spin_unlock(&sctx->stat_lock);
2143                 btrfs_err_rl_in_rcu(fs_info,
2144                         "failed to rebuild valid logical %llu for dev %s",
2145                         logical, rcu_str_deref(dev->name));
2146         } else {
2147                 scrub_write_block_to_dev_replace(sblock);
2148         }
2149
2150         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2151                 mutex_lock(&sctx->wr_lock);
2152                 scrub_wr_submit(sctx);
2153                 mutex_unlock(&sctx->wr_lock);
2154         }
2155
2156         scrub_block_put(sblock);
2157         scrub_pending_bio_dec(sctx);
2158 }
2159
2160 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2161 {
2162         struct scrub_ctx *sctx = sblock->sctx;
2163         struct btrfs_fs_info *fs_info = sctx->fs_info;
2164         u64 length = sblock->page_count * PAGE_SIZE;
2165         u64 logical = sblock->pagev[0]->logical;
2166         struct btrfs_bio *bbio = NULL;
2167         struct bio *bio;
2168         struct btrfs_raid_bio *rbio;
2169         int ret;
2170         int i;
2171
2172         btrfs_bio_counter_inc_blocked(fs_info);
2173         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2174                         &length, &bbio);
2175         if (ret || !bbio || !bbio->raid_map)
2176                 goto bbio_out;
2177
2178         if (WARN_ON(!sctx->is_dev_replace ||
2179                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2180                 /*
2181                  * We shouldn't be scrubbing a missing device. Even for dev
2182                  * replace, we should only get here for RAID 5/6. We either
2183                  * managed to mount something with no mirrors remaining or
2184                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2185                  */
2186                 goto bbio_out;
2187         }
2188
2189         bio = btrfs_io_bio_alloc(0);
2190         bio->bi_iter.bi_sector = logical >> 9;
2191         bio->bi_private = sblock;
2192         bio->bi_end_io = scrub_missing_raid56_end_io;
2193
2194         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2195         if (!rbio)
2196                 goto rbio_out;
2197
2198         for (i = 0; i < sblock->page_count; i++) {
2199                 struct scrub_page *spage = sblock->pagev[i];
2200
2201                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2202         }
2203
2204         btrfs_init_work(&sblock->work, scrub_missing_raid56_worker, NULL, NULL);
2205         scrub_block_get(sblock);
2206         scrub_pending_bio_inc(sctx);
2207         raid56_submit_missing_rbio(rbio);
2208         return;
2209
2210 rbio_out:
2211         bio_put(bio);
2212 bbio_out:
2213         btrfs_bio_counter_dec(fs_info);
2214         btrfs_put_bbio(bbio);
2215         spin_lock(&sctx->stat_lock);
2216         sctx->stat.malloc_errors++;
2217         spin_unlock(&sctx->stat_lock);
2218 }
2219
2220 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2221                        u64 physical, struct btrfs_device *dev, u64 flags,
2222                        u64 gen, int mirror_num, u8 *csum, int force,
2223                        u64 physical_for_dev_replace)
2224 {
2225         struct scrub_block *sblock;
2226         int index;
2227
2228         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2229         if (!sblock) {
2230                 spin_lock(&sctx->stat_lock);
2231                 sctx->stat.malloc_errors++;
2232                 spin_unlock(&sctx->stat_lock);
2233                 return -ENOMEM;
2234         }
2235
2236         /* one ref inside this function, plus one for each page added to
2237          * a bio later on */
2238         refcount_set(&sblock->refs, 1);
2239         sblock->sctx = sctx;
2240         sblock->no_io_error_seen = 1;
2241
2242         for (index = 0; len > 0; index++) {
2243                 struct scrub_page *spage;
2244                 u64 l = min_t(u64, len, PAGE_SIZE);
2245
2246                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2247                 if (!spage) {
2248 leave_nomem:
2249                         spin_lock(&sctx->stat_lock);
2250                         sctx->stat.malloc_errors++;
2251                         spin_unlock(&sctx->stat_lock);
2252                         scrub_block_put(sblock);
2253                         return -ENOMEM;
2254                 }
2255                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2256                 scrub_page_get(spage);
2257                 sblock->pagev[index] = spage;
2258                 spage->sblock = sblock;
2259                 spage->dev = dev;
2260                 spage->flags = flags;
2261                 spage->generation = gen;
2262                 spage->logical = logical;
2263                 spage->physical = physical;
2264                 spage->physical_for_dev_replace = physical_for_dev_replace;
2265                 spage->mirror_num = mirror_num;
2266                 if (csum) {
2267                         spage->have_csum = 1;
2268                         memcpy(spage->csum, csum, sctx->csum_size);
2269                 } else {
2270                         spage->have_csum = 0;
2271                 }
2272                 sblock->page_count++;
2273                 spage->page = alloc_page(GFP_KERNEL);
2274                 if (!spage->page)
2275                         goto leave_nomem;
2276                 len -= l;
2277                 logical += l;
2278                 physical += l;
2279                 physical_for_dev_replace += l;
2280         }
2281
2282         WARN_ON(sblock->page_count == 0);
2283         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2284                 /*
2285                  * This case should only be hit for RAID 5/6 device replace. See
2286                  * the comment in scrub_missing_raid56_pages() for details.
2287                  */
2288                 scrub_missing_raid56_pages(sblock);
2289         } else {
2290                 for (index = 0; index < sblock->page_count; index++) {
2291                         struct scrub_page *spage = sblock->pagev[index];
2292                         int ret;
2293
2294                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2295                         if (ret) {
2296                                 scrub_block_put(sblock);
2297                                 return ret;
2298                         }
2299                 }
2300
2301                 if (force)
2302                         scrub_submit(sctx);
2303         }
2304
2305         /* last one frees, either here or in bio completion for last page */
2306         scrub_block_put(sblock);
2307         return 0;
2308 }
2309
2310 static void scrub_bio_end_io(struct bio *bio)
2311 {
2312         struct scrub_bio *sbio = bio->bi_private;
2313         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2314
2315         sbio->status = bio->bi_status;
2316         sbio->bio = bio;
2317
2318         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2319 }
2320
2321 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2322 {
2323         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2324         struct scrub_ctx *sctx = sbio->sctx;
2325         int i;
2326
2327         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2328         if (sbio->status) {
2329                 for (i = 0; i < sbio->page_count; i++) {
2330                         struct scrub_page *spage = sbio->pagev[i];
2331
2332                         spage->io_error = 1;
2333                         spage->sblock->no_io_error_seen = 0;
2334                 }
2335         }
2336
2337         /* now complete the scrub_block items that have all pages completed */
2338         for (i = 0; i < sbio->page_count; i++) {
2339                 struct scrub_page *spage = sbio->pagev[i];
2340                 struct scrub_block *sblock = spage->sblock;
2341
2342                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2343                         scrub_block_complete(sblock);
2344                 scrub_block_put(sblock);
2345         }
2346
2347         bio_put(sbio->bio);
2348         sbio->bio = NULL;
2349         spin_lock(&sctx->list_lock);
2350         sbio->next_free = sctx->first_free;
2351         sctx->first_free = sbio->index;
2352         spin_unlock(&sctx->list_lock);
2353
2354         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2355                 mutex_lock(&sctx->wr_lock);
2356                 scrub_wr_submit(sctx);
2357                 mutex_unlock(&sctx->wr_lock);
2358         }
2359
2360         scrub_pending_bio_dec(sctx);
2361 }
2362
2363 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2364                                        unsigned long *bitmap,
2365                                        u64 start, u64 len)
2366 {
2367         u64 offset;
2368         u64 nsectors64;
2369         u32 nsectors;
2370         int sectorsize = sparity->sctx->fs_info->sectorsize;
2371
2372         if (len >= sparity->stripe_len) {
2373                 bitmap_set(bitmap, 0, sparity->nsectors);
2374                 return;
2375         }
2376
2377         start -= sparity->logic_start;
2378         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2379         offset = div_u64(offset, sectorsize);
2380         nsectors64 = div_u64(len, sectorsize);
2381
2382         ASSERT(nsectors64 < UINT_MAX);
2383         nsectors = (u32)nsectors64;
2384
2385         if (offset + nsectors <= sparity->nsectors) {
2386                 bitmap_set(bitmap, offset, nsectors);
2387                 return;
2388         }
2389
2390         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2391         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2392 }
2393
2394 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2395                                                    u64 start, u64 len)
2396 {
2397         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2398 }
2399
2400 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2401                                                   u64 start, u64 len)
2402 {
2403         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2404 }
2405
2406 static void scrub_block_complete(struct scrub_block *sblock)
2407 {
2408         int corrupted = 0;
2409
2410         if (!sblock->no_io_error_seen) {
2411                 corrupted = 1;
2412                 scrub_handle_errored_block(sblock);
2413         } else {
2414                 /*
2415                  * if has checksum error, write via repair mechanism in
2416                  * dev replace case, otherwise write here in dev replace
2417                  * case.
2418                  */
2419                 corrupted = scrub_checksum(sblock);
2420                 if (!corrupted && sblock->sctx->is_dev_replace)
2421                         scrub_write_block_to_dev_replace(sblock);
2422         }
2423
2424         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2425                 u64 start = sblock->pagev[0]->logical;
2426                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2427                           PAGE_SIZE;
2428
2429                 scrub_parity_mark_sectors_error(sblock->sparity,
2430                                                 start, end - start);
2431         }
2432 }
2433
2434 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2435 {
2436         struct btrfs_ordered_sum *sum = NULL;
2437         unsigned long index;
2438         unsigned long num_sectors;
2439
2440         while (!list_empty(&sctx->csum_list)) {
2441                 sum = list_first_entry(&sctx->csum_list,
2442                                        struct btrfs_ordered_sum, list);
2443                 if (sum->bytenr > logical)
2444                         return 0;
2445                 if (sum->bytenr + sum->len > logical)
2446                         break;
2447
2448                 ++sctx->stat.csum_discards;
2449                 list_del(&sum->list);
2450                 kfree(sum);
2451                 sum = NULL;
2452         }
2453         if (!sum)
2454                 return 0;
2455
2456         index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2457         ASSERT(index < UINT_MAX);
2458
2459         num_sectors = sum->len / sctx->fs_info->sectorsize;
2460         memcpy(csum, sum->sums + index * sctx->csum_size, sctx->csum_size);
2461         if (index == num_sectors - 1) {
2462                 list_del(&sum->list);
2463                 kfree(sum);
2464         }
2465         return 1;
2466 }
2467
2468 /* scrub extent tries to collect up to 64 kB for each bio */
2469 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2470                         u64 logical, u64 len,
2471                         u64 physical, struct btrfs_device *dev, u64 flags,
2472                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2473 {
2474         int ret;
2475         u8 csum[BTRFS_CSUM_SIZE];
2476         u32 blocksize;
2477
2478         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2479                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2480                         blocksize = map->stripe_len;
2481                 else
2482                         blocksize = sctx->fs_info->sectorsize;
2483                 spin_lock(&sctx->stat_lock);
2484                 sctx->stat.data_extents_scrubbed++;
2485                 sctx->stat.data_bytes_scrubbed += len;
2486                 spin_unlock(&sctx->stat_lock);
2487         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2488                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2489                         blocksize = map->stripe_len;
2490                 else
2491                         blocksize = sctx->fs_info->nodesize;
2492                 spin_lock(&sctx->stat_lock);
2493                 sctx->stat.tree_extents_scrubbed++;
2494                 sctx->stat.tree_bytes_scrubbed += len;
2495                 spin_unlock(&sctx->stat_lock);
2496         } else {
2497                 blocksize = sctx->fs_info->sectorsize;
2498                 WARN_ON(1);
2499         }
2500
2501         while (len) {
2502                 u64 l = min_t(u64, len, blocksize);
2503                 int have_csum = 0;
2504
2505                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2506                         /* push csums to sbio */
2507                         have_csum = scrub_find_csum(sctx, logical, csum);
2508                         if (have_csum == 0)
2509                                 ++sctx->stat.no_csum;
2510                 }
2511                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2512                                   mirror_num, have_csum ? csum : NULL, 0,
2513                                   physical_for_dev_replace);
2514                 if (ret)
2515                         return ret;
2516                 len -= l;
2517                 logical += l;
2518                 physical += l;
2519                 physical_for_dev_replace += l;
2520         }
2521         return 0;
2522 }
2523
2524 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2525                                   u64 logical, u64 len,
2526                                   u64 physical, struct btrfs_device *dev,
2527                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2528 {
2529         struct scrub_ctx *sctx = sparity->sctx;
2530         struct scrub_block *sblock;
2531         int index;
2532
2533         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2534         if (!sblock) {
2535                 spin_lock(&sctx->stat_lock);
2536                 sctx->stat.malloc_errors++;
2537                 spin_unlock(&sctx->stat_lock);
2538                 return -ENOMEM;
2539         }
2540
2541         /* one ref inside this function, plus one for each page added to
2542          * a bio later on */
2543         refcount_set(&sblock->refs, 1);
2544         sblock->sctx = sctx;
2545         sblock->no_io_error_seen = 1;
2546         sblock->sparity = sparity;
2547         scrub_parity_get(sparity);
2548
2549         for (index = 0; len > 0; index++) {
2550                 struct scrub_page *spage;
2551                 u64 l = min_t(u64, len, PAGE_SIZE);
2552
2553                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2554                 if (!spage) {
2555 leave_nomem:
2556                         spin_lock(&sctx->stat_lock);
2557                         sctx->stat.malloc_errors++;
2558                         spin_unlock(&sctx->stat_lock);
2559                         scrub_block_put(sblock);
2560                         return -ENOMEM;
2561                 }
2562                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2563                 /* For scrub block */
2564                 scrub_page_get(spage);
2565                 sblock->pagev[index] = spage;
2566                 /* For scrub parity */
2567                 scrub_page_get(spage);
2568                 list_add_tail(&spage->list, &sparity->spages);
2569                 spage->sblock = sblock;
2570                 spage->dev = dev;
2571                 spage->flags = flags;
2572                 spage->generation = gen;
2573                 spage->logical = logical;
2574                 spage->physical = physical;
2575                 spage->mirror_num = mirror_num;
2576                 if (csum) {
2577                         spage->have_csum = 1;
2578                         memcpy(spage->csum, csum, sctx->csum_size);
2579                 } else {
2580                         spage->have_csum = 0;
2581                 }
2582                 sblock->page_count++;
2583                 spage->page = alloc_page(GFP_KERNEL);
2584                 if (!spage->page)
2585                         goto leave_nomem;
2586                 len -= l;
2587                 logical += l;
2588                 physical += l;
2589         }
2590
2591         WARN_ON(sblock->page_count == 0);
2592         for (index = 0; index < sblock->page_count; index++) {
2593                 struct scrub_page *spage = sblock->pagev[index];
2594                 int ret;
2595
2596                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2597                 if (ret) {
2598                         scrub_block_put(sblock);
2599                         return ret;
2600                 }
2601         }
2602
2603         /* last one frees, either here or in bio completion for last page */
2604         scrub_block_put(sblock);
2605         return 0;
2606 }
2607
2608 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2609                                    u64 logical, u64 len,
2610                                    u64 physical, struct btrfs_device *dev,
2611                                    u64 flags, u64 gen, int mirror_num)
2612 {
2613         struct scrub_ctx *sctx = sparity->sctx;
2614         int ret;
2615         u8 csum[BTRFS_CSUM_SIZE];
2616         u32 blocksize;
2617
2618         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2619                 scrub_parity_mark_sectors_error(sparity, logical, len);
2620                 return 0;
2621         }
2622
2623         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2624                 blocksize = sparity->stripe_len;
2625         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2626                 blocksize = sparity->stripe_len;
2627         } else {
2628                 blocksize = sctx->fs_info->sectorsize;
2629                 WARN_ON(1);
2630         }
2631
2632         while (len) {
2633                 u64 l = min_t(u64, len, blocksize);
2634                 int have_csum = 0;
2635
2636                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2637                         /* push csums to sbio */
2638                         have_csum = scrub_find_csum(sctx, logical, csum);
2639                         if (have_csum == 0)
2640                                 goto skip;
2641                 }
2642                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2643                                              flags, gen, mirror_num,
2644                                              have_csum ? csum : NULL);
2645                 if (ret)
2646                         return ret;
2647 skip:
2648                 len -= l;
2649                 logical += l;
2650                 physical += l;
2651         }
2652         return 0;
2653 }
2654
2655 /*
2656  * Given a physical address, this will calculate it's
2657  * logical offset. if this is a parity stripe, it will return
2658  * the most left data stripe's logical offset.
2659  *
2660  * return 0 if it is a data stripe, 1 means parity stripe.
2661  */
2662 static int get_raid56_logic_offset(u64 physical, int num,
2663                                    struct map_lookup *map, u64 *offset,
2664                                    u64 *stripe_start)
2665 {
2666         int i;
2667         int j = 0;
2668         u64 stripe_nr;
2669         u64 last_offset;
2670         u32 stripe_index;
2671         u32 rot;
2672         const int data_stripes = nr_data_stripes(map);
2673
2674         last_offset = (physical - map->stripes[num].physical) * data_stripes;
2675         if (stripe_start)
2676                 *stripe_start = last_offset;
2677
2678         *offset = last_offset;
2679         for (i = 0; i < data_stripes; i++) {
2680                 *offset = last_offset + i * map->stripe_len;
2681
2682                 stripe_nr = div64_u64(*offset, map->stripe_len);
2683                 stripe_nr = div_u64(stripe_nr, data_stripes);
2684
2685                 /* Work out the disk rotation on this stripe-set */
2686                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2687                 /* calculate which stripe this data locates */
2688                 rot += i;
2689                 stripe_index = rot % map->num_stripes;
2690                 if (stripe_index == num)
2691                         return 0;
2692                 if (stripe_index < num)
2693                         j++;
2694         }
2695         *offset = last_offset + j * map->stripe_len;
2696         return 1;
2697 }
2698
2699 static void scrub_free_parity(struct scrub_parity *sparity)
2700 {
2701         struct scrub_ctx *sctx = sparity->sctx;
2702         struct scrub_page *curr, *next;
2703         int nbits;
2704
2705         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2706         if (nbits) {
2707                 spin_lock(&sctx->stat_lock);
2708                 sctx->stat.read_errors += nbits;
2709                 sctx->stat.uncorrectable_errors += nbits;
2710                 spin_unlock(&sctx->stat_lock);
2711         }
2712
2713         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2714                 list_del_init(&curr->list);
2715                 scrub_page_put(curr);
2716         }
2717
2718         kfree(sparity);
2719 }
2720
2721 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2722 {
2723         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2724                                                     work);
2725         struct scrub_ctx *sctx = sparity->sctx;
2726
2727         scrub_free_parity(sparity);
2728         scrub_pending_bio_dec(sctx);
2729 }
2730
2731 static void scrub_parity_bio_endio(struct bio *bio)
2732 {
2733         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2734         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2735
2736         if (bio->bi_status)
2737                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2738                           sparity->nsectors);
2739
2740         bio_put(bio);
2741
2742         btrfs_init_work(&sparity->work, scrub_parity_bio_endio_worker, NULL,
2743                         NULL);
2744         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2745 }
2746
2747 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2748 {
2749         struct scrub_ctx *sctx = sparity->sctx;
2750         struct btrfs_fs_info *fs_info = sctx->fs_info;
2751         struct bio *bio;
2752         struct btrfs_raid_bio *rbio;
2753         struct btrfs_bio *bbio = NULL;
2754         u64 length;
2755         int ret;
2756
2757         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2758                            sparity->nsectors))
2759                 goto out;
2760
2761         length = sparity->logic_end - sparity->logic_start;
2762
2763         btrfs_bio_counter_inc_blocked(fs_info);
2764         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2765                                &length, &bbio);
2766         if (ret || !bbio || !bbio->raid_map)
2767                 goto bbio_out;
2768
2769         bio = btrfs_io_bio_alloc(0);
2770         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2771         bio->bi_private = sparity;
2772         bio->bi_end_io = scrub_parity_bio_endio;
2773
2774         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2775                                               length, sparity->scrub_dev,
2776                                               sparity->dbitmap,
2777                                               sparity->nsectors);
2778         if (!rbio)
2779                 goto rbio_out;
2780
2781         scrub_pending_bio_inc(sctx);
2782         raid56_parity_submit_scrub_rbio(rbio);
2783         return;
2784
2785 rbio_out:
2786         bio_put(bio);
2787 bbio_out:
2788         btrfs_bio_counter_dec(fs_info);
2789         btrfs_put_bbio(bbio);
2790         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2791                   sparity->nsectors);
2792         spin_lock(&sctx->stat_lock);
2793         sctx->stat.malloc_errors++;
2794         spin_unlock(&sctx->stat_lock);
2795 out:
2796         scrub_free_parity(sparity);
2797 }
2798
2799 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2800 {
2801         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2802 }
2803
2804 static void scrub_parity_get(struct scrub_parity *sparity)
2805 {
2806         refcount_inc(&sparity->refs);
2807 }
2808
2809 static void scrub_parity_put(struct scrub_parity *sparity)
2810 {
2811         if (!refcount_dec_and_test(&sparity->refs))
2812                 return;
2813
2814         scrub_parity_check_and_repair(sparity);
2815 }
2816
2817 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2818                                                   struct map_lookup *map,
2819                                                   struct btrfs_device *sdev,
2820                                                   struct btrfs_path *path,
2821                                                   u64 logic_start,
2822                                                   u64 logic_end)
2823 {
2824         struct btrfs_fs_info *fs_info = sctx->fs_info;
2825         struct btrfs_root *root = fs_info->extent_root;
2826         struct btrfs_root *csum_root = fs_info->csum_root;
2827         struct btrfs_extent_item *extent;
2828         struct btrfs_bio *bbio = NULL;
2829         u64 flags;
2830         int ret;
2831         int slot;
2832         struct extent_buffer *l;
2833         struct btrfs_key key;
2834         u64 generation;
2835         u64 extent_logical;
2836         u64 extent_physical;
2837         u64 extent_len;
2838         u64 mapped_length;
2839         struct btrfs_device *extent_dev;
2840         struct scrub_parity *sparity;
2841         int nsectors;
2842         int bitmap_len;
2843         int extent_mirror_num;
2844         int stop_loop = 0;
2845
2846         nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2847         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2848         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2849                           GFP_NOFS);
2850         if (!sparity) {
2851                 spin_lock(&sctx->stat_lock);
2852                 sctx->stat.malloc_errors++;
2853                 spin_unlock(&sctx->stat_lock);
2854                 return -ENOMEM;
2855         }
2856
2857         sparity->stripe_len = map->stripe_len;
2858         sparity->nsectors = nsectors;
2859         sparity->sctx = sctx;
2860         sparity->scrub_dev = sdev;
2861         sparity->logic_start = logic_start;
2862         sparity->logic_end = logic_end;
2863         refcount_set(&sparity->refs, 1);
2864         INIT_LIST_HEAD(&sparity->spages);
2865         sparity->dbitmap = sparity->bitmap;
2866         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2867
2868         ret = 0;
2869         while (logic_start < logic_end) {
2870                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2871                         key.type = BTRFS_METADATA_ITEM_KEY;
2872                 else
2873                         key.type = BTRFS_EXTENT_ITEM_KEY;
2874                 key.objectid = logic_start;
2875                 key.offset = (u64)-1;
2876
2877                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2878                 if (ret < 0)
2879                         goto out;
2880
2881                 if (ret > 0) {
2882                         ret = btrfs_previous_extent_item(root, path, 0);
2883                         if (ret < 0)
2884                                 goto out;
2885                         if (ret > 0) {
2886                                 btrfs_release_path(path);
2887                                 ret = btrfs_search_slot(NULL, root, &key,
2888                                                         path, 0, 0);
2889                                 if (ret < 0)
2890                                         goto out;
2891                         }
2892                 }
2893
2894                 stop_loop = 0;
2895                 while (1) {
2896                         u64 bytes;
2897
2898                         l = path->nodes[0];
2899                         slot = path->slots[0];
2900                         if (slot >= btrfs_header_nritems(l)) {
2901                                 ret = btrfs_next_leaf(root, path);
2902                                 if (ret == 0)
2903                                         continue;
2904                                 if (ret < 0)
2905                                         goto out;
2906
2907                                 stop_loop = 1;
2908                                 break;
2909                         }
2910                         btrfs_item_key_to_cpu(l, &key, slot);
2911
2912                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2913                             key.type != BTRFS_METADATA_ITEM_KEY)
2914                                 goto next;
2915
2916                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2917                                 bytes = fs_info->nodesize;
2918                         else
2919                                 bytes = key.offset;
2920
2921                         if (key.objectid + bytes <= logic_start)
2922                                 goto next;
2923
2924                         if (key.objectid >= logic_end) {
2925                                 stop_loop = 1;
2926                                 break;
2927                         }
2928
2929                         while (key.objectid >= logic_start + map->stripe_len)
2930                                 logic_start += map->stripe_len;
2931
2932                         extent = btrfs_item_ptr(l, slot,
2933                                                 struct btrfs_extent_item);
2934                         flags = btrfs_extent_flags(l, extent);
2935                         generation = btrfs_extent_generation(l, extent);
2936
2937                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2938                             (key.objectid < logic_start ||
2939                              key.objectid + bytes >
2940                              logic_start + map->stripe_len)) {
2941                                 btrfs_err(fs_info,
2942                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2943                                           key.objectid, logic_start);
2944                                 spin_lock(&sctx->stat_lock);
2945                                 sctx->stat.uncorrectable_errors++;
2946                                 spin_unlock(&sctx->stat_lock);
2947                                 goto next;
2948                         }
2949 again:
2950                         extent_logical = key.objectid;
2951                         extent_len = bytes;
2952
2953                         if (extent_logical < logic_start) {
2954                                 extent_len -= logic_start - extent_logical;
2955                                 extent_logical = logic_start;
2956                         }
2957
2958                         if (extent_logical + extent_len >
2959                             logic_start + map->stripe_len)
2960                                 extent_len = logic_start + map->stripe_len -
2961                                              extent_logical;
2962
2963                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2964                                                        extent_len);
2965
2966                         mapped_length = extent_len;
2967                         bbio = NULL;
2968                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2969                                         extent_logical, &mapped_length, &bbio,
2970                                         0);
2971                         if (!ret) {
2972                                 if (!bbio || mapped_length < extent_len)
2973                                         ret = -EIO;
2974                         }
2975                         if (ret) {
2976                                 btrfs_put_bbio(bbio);
2977                                 goto out;
2978                         }
2979                         extent_physical = bbio->stripes[0].physical;
2980                         extent_mirror_num = bbio->mirror_num;
2981                         extent_dev = bbio->stripes[0].dev;
2982                         btrfs_put_bbio(bbio);
2983
2984                         ret = btrfs_lookup_csums_range(csum_root,
2985                                                 extent_logical,
2986                                                 extent_logical + extent_len - 1,
2987                                                 &sctx->csum_list, 1);
2988                         if (ret)
2989                                 goto out;
2990
2991                         ret = scrub_extent_for_parity(sparity, extent_logical,
2992                                                       extent_len,
2993                                                       extent_physical,
2994                                                       extent_dev, flags,
2995                                                       generation,
2996                                                       extent_mirror_num);
2997
2998                         scrub_free_csums(sctx);
2999
3000                         if (ret)
3001                                 goto out;
3002
3003                         if (extent_logical + extent_len <
3004                             key.objectid + bytes) {
3005                                 logic_start += map->stripe_len;
3006
3007                                 if (logic_start >= logic_end) {
3008                                         stop_loop = 1;
3009                                         break;
3010                                 }
3011
3012                                 if (logic_start < key.objectid + bytes) {
3013                                         cond_resched();
3014                                         goto again;
3015                                 }
3016                         }
3017 next:
3018                         path->slots[0]++;
3019                 }
3020
3021                 btrfs_release_path(path);
3022
3023                 if (stop_loop)
3024                         break;
3025
3026                 logic_start += map->stripe_len;
3027         }
3028 out:
3029         if (ret < 0)
3030                 scrub_parity_mark_sectors_error(sparity, logic_start,
3031                                                 logic_end - logic_start);
3032         scrub_parity_put(sparity);
3033         scrub_submit(sctx);
3034         mutex_lock(&sctx->wr_lock);
3035         scrub_wr_submit(sctx);
3036         mutex_unlock(&sctx->wr_lock);
3037
3038         btrfs_release_path(path);
3039         return ret < 0 ? ret : 0;
3040 }
3041
3042 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3043                                            struct map_lookup *map,
3044                                            struct btrfs_device *scrub_dev,
3045                                            int num, u64 base, u64 length,
3046                                            struct btrfs_block_group *cache)
3047 {
3048         struct btrfs_path *path, *ppath;
3049         struct btrfs_fs_info *fs_info = sctx->fs_info;
3050         struct btrfs_root *root = fs_info->extent_root;
3051         struct btrfs_root *csum_root = fs_info->csum_root;
3052         struct btrfs_extent_item *extent;
3053         struct blk_plug plug;
3054         u64 flags;
3055         int ret;
3056         int slot;
3057         u64 nstripes;
3058         struct extent_buffer *l;
3059         u64 physical;
3060         u64 logical;
3061         u64 logic_end;
3062         u64 physical_end;
3063         u64 generation;
3064         int mirror_num;
3065         struct reada_control *reada1;
3066         struct reada_control *reada2;
3067         struct btrfs_key key;
3068         struct btrfs_key key_end;
3069         u64 increment = map->stripe_len;
3070         u64 offset;
3071         u64 extent_logical;
3072         u64 extent_physical;
3073         u64 extent_len;
3074         u64 stripe_logical;
3075         u64 stripe_end;
3076         struct btrfs_device *extent_dev;
3077         int extent_mirror_num;
3078         int stop_loop = 0;
3079
3080         physical = map->stripes[num].physical;
3081         offset = 0;
3082         nstripes = div64_u64(length, map->stripe_len);
3083         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3084                 offset = map->stripe_len * num;
3085                 increment = map->stripe_len * map->num_stripes;
3086                 mirror_num = 1;
3087         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3088                 int factor = map->num_stripes / map->sub_stripes;
3089                 offset = map->stripe_len * (num / map->sub_stripes);
3090                 increment = map->stripe_len * factor;
3091                 mirror_num = num % map->sub_stripes + 1;
3092         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1_MASK) {
3093                 increment = map->stripe_len;
3094                 mirror_num = num % map->num_stripes + 1;
3095         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3096                 increment = map->stripe_len;
3097                 mirror_num = num % map->num_stripes + 1;
3098         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3099                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3100                 increment = map->stripe_len * nr_data_stripes(map);
3101                 mirror_num = 1;
3102         } else {
3103                 increment = map->stripe_len;
3104                 mirror_num = 1;
3105         }
3106
3107         path = btrfs_alloc_path();
3108         if (!path)
3109                 return -ENOMEM;
3110
3111         ppath = btrfs_alloc_path();
3112         if (!ppath) {
3113                 btrfs_free_path(path);
3114                 return -ENOMEM;
3115         }
3116
3117         /*
3118          * work on commit root. The related disk blocks are static as
3119          * long as COW is applied. This means, it is save to rewrite
3120          * them to repair disk errors without any race conditions
3121          */
3122         path->search_commit_root = 1;
3123         path->skip_locking = 1;
3124
3125         ppath->search_commit_root = 1;
3126         ppath->skip_locking = 1;
3127         /*
3128          * trigger the readahead for extent tree csum tree and wait for
3129          * completion. During readahead, the scrub is officially paused
3130          * to not hold off transaction commits
3131          */
3132         logical = base + offset;
3133         physical_end = physical + nstripes * map->stripe_len;
3134         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3135                 get_raid56_logic_offset(physical_end, num,
3136                                         map, &logic_end, NULL);
3137                 logic_end += base;
3138         } else {
3139                 logic_end = logical + increment * nstripes;
3140         }
3141         wait_event(sctx->list_wait,
3142                    atomic_read(&sctx->bios_in_flight) == 0);
3143         scrub_blocked_if_needed(fs_info);
3144
3145         /* FIXME it might be better to start readahead at commit root */
3146         key.objectid = logical;
3147         key.type = BTRFS_EXTENT_ITEM_KEY;
3148         key.offset = (u64)0;
3149         key_end.objectid = logic_end;
3150         key_end.type = BTRFS_METADATA_ITEM_KEY;
3151         key_end.offset = (u64)-1;
3152         reada1 = btrfs_reada_add(root, &key, &key_end);
3153
3154         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3155         key.type = BTRFS_EXTENT_CSUM_KEY;
3156         key.offset = logical;
3157         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3158         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3159         key_end.offset = logic_end;
3160         reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3161
3162         if (!IS_ERR(reada1))
3163                 btrfs_reada_wait(reada1);
3164         if (!IS_ERR(reada2))
3165                 btrfs_reada_wait(reada2);
3166
3167
3168         /*
3169          * collect all data csums for the stripe to avoid seeking during
3170          * the scrub. This might currently (crc32) end up to be about 1MB
3171          */
3172         blk_start_plug(&plug);
3173
3174         /*
3175          * now find all extents for each stripe and scrub them
3176          */
3177         ret = 0;
3178         while (physical < physical_end) {
3179                 /*
3180                  * canceled?
3181                  */
3182                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3183                     atomic_read(&sctx->cancel_req)) {
3184                         ret = -ECANCELED;
3185                         goto out;
3186                 }
3187                 /*
3188                  * check to see if we have to pause
3189                  */
3190                 if (atomic_read(&fs_info->scrub_pause_req)) {
3191                         /* push queued extents */
3192                         sctx->flush_all_writes = true;
3193                         scrub_submit(sctx);
3194                         mutex_lock(&sctx->wr_lock);
3195                         scrub_wr_submit(sctx);
3196                         mutex_unlock(&sctx->wr_lock);
3197                         wait_event(sctx->list_wait,
3198                                    atomic_read(&sctx->bios_in_flight) == 0);
3199                         sctx->flush_all_writes = false;
3200                         scrub_blocked_if_needed(fs_info);
3201                 }
3202
3203                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3204                         ret = get_raid56_logic_offset(physical, num, map,
3205                                                       &logical,
3206                                                       &stripe_logical);
3207                         logical += base;
3208                         if (ret) {
3209                                 /* it is parity strip */
3210                                 stripe_logical += base;
3211                                 stripe_end = stripe_logical + increment;
3212                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3213                                                           ppath, stripe_logical,
3214                                                           stripe_end);
3215                                 if (ret)
3216                                         goto out;
3217                                 goto skip;
3218                         }
3219                 }
3220
3221                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3222                         key.type = BTRFS_METADATA_ITEM_KEY;
3223                 else
3224                         key.type = BTRFS_EXTENT_ITEM_KEY;
3225                 key.objectid = logical;
3226                 key.offset = (u64)-1;
3227
3228                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3229                 if (ret < 0)
3230                         goto out;
3231
3232                 if (ret > 0) {
3233                         ret = btrfs_previous_extent_item(root, path, 0);
3234                         if (ret < 0)
3235                                 goto out;
3236                         if (ret > 0) {
3237                                 /* there's no smaller item, so stick with the
3238                                  * larger one */
3239                                 btrfs_release_path(path);
3240                                 ret = btrfs_search_slot(NULL, root, &key,
3241                                                         path, 0, 0);
3242                                 if (ret < 0)
3243                                         goto out;
3244                         }
3245                 }
3246
3247                 stop_loop = 0;
3248                 while (1) {
3249                         u64 bytes;
3250
3251                         l = path->nodes[0];
3252                         slot = path->slots[0];
3253                         if (slot >= btrfs_header_nritems(l)) {
3254                                 ret = btrfs_next_leaf(root, path);
3255                                 if (ret == 0)
3256                                         continue;
3257                                 if (ret < 0)
3258                                         goto out;
3259
3260                                 stop_loop = 1;
3261                                 break;
3262                         }
3263                         btrfs_item_key_to_cpu(l, &key, slot);
3264
3265                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3266                             key.type != BTRFS_METADATA_ITEM_KEY)
3267                                 goto next;
3268
3269                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3270                                 bytes = fs_info->nodesize;
3271                         else
3272                                 bytes = key.offset;
3273
3274                         if (key.objectid + bytes <= logical)
3275                                 goto next;
3276
3277                         if (key.objectid >= logical + map->stripe_len) {
3278                                 /* out of this device extent */
3279                                 if (key.objectid >= logic_end)
3280                                         stop_loop = 1;
3281                                 break;
3282                         }
3283
3284                         /*
3285                          * If our block group was removed in the meanwhile, just
3286                          * stop scrubbing since there is no point in continuing.
3287                          * Continuing would prevent reusing its device extents
3288                          * for new block groups for a long time.
3289                          */
3290                         spin_lock(&cache->lock);
3291                         if (cache->removed) {
3292                                 spin_unlock(&cache->lock);
3293                                 ret = 0;
3294                                 goto out;
3295                         }
3296                         spin_unlock(&cache->lock);
3297
3298                         extent = btrfs_item_ptr(l, slot,
3299                                                 struct btrfs_extent_item);
3300                         flags = btrfs_extent_flags(l, extent);
3301                         generation = btrfs_extent_generation(l, extent);
3302
3303                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3304                             (key.objectid < logical ||
3305                              key.objectid + bytes >
3306                              logical + map->stripe_len)) {
3307                                 btrfs_err(fs_info,
3308                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3309                                        key.objectid, logical);
3310                                 spin_lock(&sctx->stat_lock);
3311                                 sctx->stat.uncorrectable_errors++;
3312                                 spin_unlock(&sctx->stat_lock);
3313                                 goto next;
3314                         }
3315
3316 again:
3317                         extent_logical = key.objectid;
3318                         extent_len = bytes;
3319
3320                         /*
3321                          * trim extent to this stripe
3322                          */
3323                         if (extent_logical < logical) {
3324                                 extent_len -= logical - extent_logical;
3325                                 extent_logical = logical;
3326                         }
3327                         if (extent_logical + extent_len >
3328                             logical + map->stripe_len) {
3329                                 extent_len = logical + map->stripe_len -
3330                                              extent_logical;
3331                         }
3332
3333                         extent_physical = extent_logical - logical + physical;
3334                         extent_dev = scrub_dev;
3335                         extent_mirror_num = mirror_num;
3336                         if (sctx->is_dev_replace)
3337                                 scrub_remap_extent(fs_info, extent_logical,
3338                                                    extent_len, &extent_physical,
3339                                                    &extent_dev,
3340                                                    &extent_mirror_num);
3341
3342                         if (flags & BTRFS_EXTENT_FLAG_DATA) {
3343                                 ret = btrfs_lookup_csums_range(csum_root,
3344                                                 extent_logical,
3345                                                 extent_logical + extent_len - 1,
3346                                                 &sctx->csum_list, 1);
3347                                 if (ret)
3348                                         goto out;
3349                         }
3350
3351                         ret = scrub_extent(sctx, map, extent_logical, extent_len,
3352                                            extent_physical, extent_dev, flags,
3353                                            generation, extent_mirror_num,
3354                                            extent_logical - logical + physical);
3355
3356                         scrub_free_csums(sctx);
3357
3358                         if (ret)
3359                                 goto out;
3360
3361                         if (extent_logical + extent_len <
3362                             key.objectid + bytes) {
3363                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3364                                         /*
3365                                          * loop until we find next data stripe
3366                                          * or we have finished all stripes.
3367                                          */
3368 loop:
3369                                         physical += map->stripe_len;
3370                                         ret = get_raid56_logic_offset(physical,
3371                                                         num, map, &logical,
3372                                                         &stripe_logical);
3373                                         logical += base;
3374
3375                                         if (ret && physical < physical_end) {
3376                                                 stripe_logical += base;
3377                                                 stripe_end = stripe_logical +
3378                                                                 increment;
3379                                                 ret = scrub_raid56_parity(sctx,
3380                                                         map, scrub_dev, ppath,
3381                                                         stripe_logical,
3382                                                         stripe_end);
3383                                                 if (ret)
3384                                                         goto out;
3385                                                 goto loop;
3386                                         }
3387                                 } else {
3388                                         physical += map->stripe_len;
3389                                         logical += increment;
3390                                 }
3391                                 if (logical < key.objectid + bytes) {
3392                                         cond_resched();
3393                                         goto again;
3394                                 }
3395
3396                                 if (physical >= physical_end) {
3397                                         stop_loop = 1;
3398                                         break;
3399                                 }
3400                         }
3401 next:
3402                         path->slots[0]++;
3403                 }
3404                 btrfs_release_path(path);
3405 skip:
3406                 logical += increment;
3407                 physical += map->stripe_len;
3408                 spin_lock(&sctx->stat_lock);
3409                 if (stop_loop)
3410                         sctx->stat.last_physical = map->stripes[num].physical +
3411                                                    length;
3412                 else
3413                         sctx->stat.last_physical = physical;
3414                 spin_unlock(&sctx->stat_lock);
3415                 if (stop_loop)
3416                         break;
3417         }
3418 out:
3419         /* push queued extents */
3420         scrub_submit(sctx);
3421         mutex_lock(&sctx->wr_lock);
3422         scrub_wr_submit(sctx);
3423         mutex_unlock(&sctx->wr_lock);
3424
3425         blk_finish_plug(&plug);
3426         btrfs_free_path(path);
3427         btrfs_free_path(ppath);
3428         return ret < 0 ? ret : 0;
3429 }
3430
3431 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3432                                           struct btrfs_device *scrub_dev,
3433                                           u64 chunk_offset, u64 length,
3434                                           u64 dev_offset,
3435                                           struct btrfs_block_group *cache)
3436 {
3437         struct btrfs_fs_info *fs_info = sctx->fs_info;
3438         struct extent_map_tree *map_tree = &fs_info->mapping_tree;
3439         struct map_lookup *map;
3440         struct extent_map *em;
3441         int i;
3442         int ret = 0;
3443
3444         read_lock(&map_tree->lock);
3445         em = lookup_extent_mapping(map_tree, chunk_offset, 1);
3446         read_unlock(&map_tree->lock);
3447
3448         if (!em) {
3449                 /*
3450                  * Might have been an unused block group deleted by the cleaner
3451                  * kthread or relocation.
3452                  */
3453                 spin_lock(&cache->lock);
3454                 if (!cache->removed)
3455                         ret = -EINVAL;
3456                 spin_unlock(&cache->lock);
3457
3458                 return ret;
3459         }
3460
3461         map = em->map_lookup;
3462         if (em->start != chunk_offset)
3463                 goto out;
3464
3465         if (em->len < length)
3466                 goto out;
3467
3468         for (i = 0; i < map->num_stripes; ++i) {
3469                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3470                     map->stripes[i].physical == dev_offset) {
3471                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3472                                            chunk_offset, length, cache);
3473                         if (ret)
3474                                 goto out;
3475                 }
3476         }
3477 out:
3478         free_extent_map(em);
3479
3480         return ret;
3481 }
3482
3483 static noinline_for_stack
3484 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3485                            struct btrfs_device *scrub_dev, u64 start, u64 end)
3486 {
3487         struct btrfs_dev_extent *dev_extent = NULL;
3488         struct btrfs_path *path;
3489         struct btrfs_fs_info *fs_info = sctx->fs_info;
3490         struct btrfs_root *root = fs_info->dev_root;
3491         u64 length;
3492         u64 chunk_offset;
3493         int ret = 0;
3494         int ro_set;
3495         int slot;
3496         struct extent_buffer *l;
3497         struct btrfs_key key;
3498         struct btrfs_key found_key;
3499         struct btrfs_block_group *cache;
3500         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3501
3502         path = btrfs_alloc_path();
3503         if (!path)
3504                 return -ENOMEM;
3505
3506         path->reada = READA_FORWARD;
3507         path->search_commit_root = 1;
3508         path->skip_locking = 1;
3509
3510         key.objectid = scrub_dev->devid;
3511         key.offset = 0ull;
3512         key.type = BTRFS_DEV_EXTENT_KEY;
3513
3514         while (1) {
3515                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3516                 if (ret < 0)
3517                         break;
3518                 if (ret > 0) {
3519                         if (path->slots[0] >=
3520                             btrfs_header_nritems(path->nodes[0])) {
3521                                 ret = btrfs_next_leaf(root, path);
3522                                 if (ret < 0)
3523                                         break;
3524                                 if (ret > 0) {
3525                                         ret = 0;
3526                                         break;
3527                                 }
3528                         } else {
3529                                 ret = 0;
3530                         }
3531                 }
3532
3533                 l = path->nodes[0];
3534                 slot = path->slots[0];
3535
3536                 btrfs_item_key_to_cpu(l, &found_key, slot);
3537
3538                 if (found_key.objectid != scrub_dev->devid)
3539                         break;
3540
3541                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3542                         break;
3543
3544                 if (found_key.offset >= end)
3545                         break;
3546
3547                 if (found_key.offset < key.offset)
3548                         break;
3549
3550                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3551                 length = btrfs_dev_extent_length(l, dev_extent);
3552
3553                 if (found_key.offset + length <= start)
3554                         goto skip;
3555
3556                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3557
3558                 /*
3559                  * get a reference on the corresponding block group to prevent
3560                  * the chunk from going away while we scrub it
3561                  */
3562                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3563
3564                 /* some chunks are removed but not committed to disk yet,
3565                  * continue scrubbing */
3566                 if (!cache)
3567                         goto skip;
3568
3569                 /*
3570                  * Make sure that while we are scrubbing the corresponding block
3571                  * group doesn't get its logical address and its device extents
3572                  * reused for another block group, which can possibly be of a
3573                  * different type and different profile. We do this to prevent
3574                  * false error detections and crashes due to bogus attempts to
3575                  * repair extents.
3576                  */
3577                 spin_lock(&cache->lock);
3578                 if (cache->removed) {
3579                         spin_unlock(&cache->lock);
3580                         btrfs_put_block_group(cache);
3581                         goto skip;
3582                 }
3583                 btrfs_freeze_block_group(cache);
3584                 spin_unlock(&cache->lock);
3585
3586                 /*
3587                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3588                  * to avoid deadlock caused by:
3589                  * btrfs_inc_block_group_ro()
3590                  * -> btrfs_wait_for_commit()
3591                  * -> btrfs_commit_transaction()
3592                  * -> btrfs_scrub_pause()
3593                  */
3594                 scrub_pause_on(fs_info);
3595
3596                 /*
3597                  * Don't do chunk preallocation for scrub.
3598                  *
3599                  * This is especially important for SYSTEM bgs, or we can hit
3600                  * -EFBIG from btrfs_finish_chunk_alloc() like:
3601                  * 1. The only SYSTEM bg is marked RO.
3602                  *    Since SYSTEM bg is small, that's pretty common.
3603                  * 2. New SYSTEM bg will be allocated
3604                  *    Due to regular version will allocate new chunk.
3605                  * 3. New SYSTEM bg is empty and will get cleaned up
3606                  *    Before cleanup really happens, it's marked RO again.
3607                  * 4. Empty SYSTEM bg get scrubbed
3608                  *    We go back to 2.
3609                  *
3610                  * This can easily boost the amount of SYSTEM chunks if cleaner
3611                  * thread can't be triggered fast enough, and use up all space
3612                  * of btrfs_super_block::sys_chunk_array
3613                  *
3614                  * While for dev replace, we need to try our best to mark block
3615                  * group RO, to prevent race between:
3616                  * - Write duplication
3617                  *   Contains latest data
3618                  * - Scrub copy
3619                  *   Contains data from commit tree
3620                  *
3621                  * If target block group is not marked RO, nocow writes can
3622                  * be overwritten by scrub copy, causing data corruption.
3623                  * So for dev-replace, it's not allowed to continue if a block
3624                  * group is not RO.
3625                  */
3626                 ret = btrfs_inc_block_group_ro(cache, sctx->is_dev_replace);
3627                 if (ret == 0) {
3628                         ro_set = 1;
3629                 } else if (ret == -ENOSPC && !sctx->is_dev_replace) {
3630                         /*
3631                          * btrfs_inc_block_group_ro return -ENOSPC when it
3632                          * failed in creating new chunk for metadata.
3633                          * It is not a problem for scrub, because
3634                          * metadata are always cowed, and our scrub paused
3635                          * commit_transactions.
3636                          */
3637                         ro_set = 0;
3638                 } else {
3639                         btrfs_warn(fs_info,
3640                                    "failed setting block group ro: %d", ret);
3641                         btrfs_unfreeze_block_group(cache);
3642                         btrfs_put_block_group(cache);
3643                         scrub_pause_off(fs_info);
3644                         break;
3645                 }
3646
3647                 /*
3648                  * Now the target block is marked RO, wait for nocow writes to
3649                  * finish before dev-replace.
3650                  * COW is fine, as COW never overwrites extents in commit tree.
3651                  */
3652                 if (sctx->is_dev_replace) {
3653                         btrfs_wait_nocow_writers(cache);
3654                         btrfs_wait_ordered_roots(fs_info, U64_MAX, cache->start,
3655                                         cache->length);
3656                 }
3657
3658                 scrub_pause_off(fs_info);
3659                 down_write(&dev_replace->rwsem);
3660                 dev_replace->cursor_right = found_key.offset + length;
3661                 dev_replace->cursor_left = found_key.offset;
3662                 dev_replace->item_needs_writeback = 1;
3663                 up_write(&dev_replace->rwsem);
3664
3665                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3666                                   found_key.offset, cache);
3667
3668                 /*
3669                  * flush, submit all pending read and write bios, afterwards
3670                  * wait for them.
3671                  * Note that in the dev replace case, a read request causes
3672                  * write requests that are submitted in the read completion
3673                  * worker. Therefore in the current situation, it is required
3674                  * that all write requests are flushed, so that all read and
3675                  * write requests are really completed when bios_in_flight
3676                  * changes to 0.
3677                  */
3678                 sctx->flush_all_writes = true;
3679                 scrub_submit(sctx);
3680                 mutex_lock(&sctx->wr_lock);
3681                 scrub_wr_submit(sctx);
3682                 mutex_unlock(&sctx->wr_lock);
3683
3684                 wait_event(sctx->list_wait,
3685                            atomic_read(&sctx->bios_in_flight) == 0);
3686
3687                 scrub_pause_on(fs_info);
3688
3689                 /*
3690                  * must be called before we decrease @scrub_paused.
3691                  * make sure we don't block transaction commit while
3692                  * we are waiting pending workers finished.
3693                  */
3694                 wait_event(sctx->list_wait,
3695                            atomic_read(&sctx->workers_pending) == 0);
3696                 sctx->flush_all_writes = false;
3697
3698                 scrub_pause_off(fs_info);
3699
3700                 down_write(&dev_replace->rwsem);
3701                 dev_replace->cursor_left = dev_replace->cursor_right;
3702                 dev_replace->item_needs_writeback = 1;
3703                 up_write(&dev_replace->rwsem);
3704
3705                 if (ro_set)
3706                         btrfs_dec_block_group_ro(cache);
3707
3708                 /*
3709                  * We might have prevented the cleaner kthread from deleting
3710                  * this block group if it was already unused because we raced
3711                  * and set it to RO mode first. So add it back to the unused
3712                  * list, otherwise it might not ever be deleted unless a manual
3713                  * balance is triggered or it becomes used and unused again.
3714                  */
3715                 spin_lock(&cache->lock);
3716                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3717                     cache->used == 0) {
3718                         spin_unlock(&cache->lock);
3719                         if (btrfs_test_opt(fs_info, DISCARD_ASYNC))
3720                                 btrfs_discard_queue_work(&fs_info->discard_ctl,
3721                                                          cache);
3722                         else
3723                                 btrfs_mark_bg_unused(cache);
3724                 } else {
3725                         spin_unlock(&cache->lock);
3726                 }
3727
3728                 btrfs_unfreeze_block_group(cache);
3729                 btrfs_put_block_group(cache);
3730                 if (ret)
3731                         break;
3732                 if (sctx->is_dev_replace &&
3733                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3734                         ret = -EIO;
3735                         break;
3736                 }
3737                 if (sctx->stat.malloc_errors > 0) {
3738                         ret = -ENOMEM;
3739                         break;
3740                 }
3741 skip:
3742                 key.offset = found_key.offset + length;
3743                 btrfs_release_path(path);
3744         }
3745
3746         btrfs_free_path(path);
3747
3748         return ret;
3749 }
3750
3751 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3752                                            struct btrfs_device *scrub_dev)
3753 {
3754         int     i;
3755         u64     bytenr;
3756         u64     gen;
3757         int     ret;
3758         struct btrfs_fs_info *fs_info = sctx->fs_info;
3759
3760         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3761                 return -EIO;
3762
3763         /* Seed devices of a new filesystem has their own generation. */
3764         if (scrub_dev->fs_devices != fs_info->fs_devices)
3765                 gen = scrub_dev->generation;
3766         else
3767                 gen = fs_info->last_trans_committed;
3768
3769         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3770                 bytenr = btrfs_sb_offset(i);
3771                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3772                     scrub_dev->commit_total_bytes)
3773                         break;
3774
3775                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3776                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3777                                   NULL, 1, bytenr);
3778                 if (ret)
3779                         return ret;
3780         }
3781         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3782
3783         return 0;
3784 }
3785
3786 /*
3787  * get a reference count on fs_info->scrub_workers. start worker if necessary
3788  */
3789 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3790                                                 int is_dev_replace)
3791 {
3792         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3793         int max_active = fs_info->thread_pool_size;
3794
3795         lockdep_assert_held(&fs_info->scrub_lock);
3796
3797         if (refcount_read(&fs_info->scrub_workers_refcnt) == 0) {
3798                 ASSERT(fs_info->scrub_workers == NULL);
3799                 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3800                                 flags, is_dev_replace ? 1 : max_active, 4);
3801                 if (!fs_info->scrub_workers)
3802                         goto fail_scrub_workers;
3803
3804                 ASSERT(fs_info->scrub_wr_completion_workers == NULL);
3805                 fs_info->scrub_wr_completion_workers =
3806                         btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3807                                               max_active, 2);
3808                 if (!fs_info->scrub_wr_completion_workers)
3809                         goto fail_scrub_wr_completion_workers;
3810
3811                 ASSERT(fs_info->scrub_parity_workers == NULL);
3812                 fs_info->scrub_parity_workers =
3813                         btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3814                                               max_active, 2);
3815                 if (!fs_info->scrub_parity_workers)
3816                         goto fail_scrub_parity_workers;
3817
3818                 refcount_set(&fs_info->scrub_workers_refcnt, 1);
3819         } else {
3820                 refcount_inc(&fs_info->scrub_workers_refcnt);
3821         }
3822         return 0;
3823
3824 fail_scrub_parity_workers:
3825         btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3826 fail_scrub_wr_completion_workers:
3827         btrfs_destroy_workqueue(fs_info->scrub_workers);
3828 fail_scrub_workers:
3829         return -ENOMEM;
3830 }
3831
3832 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3833                     u64 end, struct btrfs_scrub_progress *progress,
3834                     int readonly, int is_dev_replace)
3835 {
3836         struct scrub_ctx *sctx;
3837         int ret;
3838         struct btrfs_device *dev;
3839         unsigned int nofs_flag;
3840         struct btrfs_workqueue *scrub_workers = NULL;
3841         struct btrfs_workqueue *scrub_wr_comp = NULL;
3842         struct btrfs_workqueue *scrub_parity = NULL;
3843
3844         if (btrfs_fs_closing(fs_info))
3845                 return -EAGAIN;
3846
3847         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3848                 /*
3849                  * in this case scrub is unable to calculate the checksum
3850                  * the way scrub is implemented. Do not handle this
3851                  * situation at all because it won't ever happen.
3852                  */
3853                 btrfs_err(fs_info,
3854                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3855                        fs_info->nodesize,
3856                        BTRFS_STRIPE_LEN);
3857                 return -EINVAL;
3858         }
3859
3860         if (fs_info->sectorsize != PAGE_SIZE) {
3861                 /* not supported for data w/o checksums */
3862                 btrfs_err_rl(fs_info,
3863                            "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3864                        fs_info->sectorsize, PAGE_SIZE);
3865                 return -EINVAL;
3866         }
3867
3868         if (fs_info->nodesize >
3869             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3870             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3871                 /*
3872                  * would exhaust the array bounds of pagev member in
3873                  * struct scrub_block
3874                  */
3875                 btrfs_err(fs_info,
3876                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3877                        fs_info->nodesize,
3878                        SCRUB_MAX_PAGES_PER_BLOCK,
3879                        fs_info->sectorsize,
3880                        SCRUB_MAX_PAGES_PER_BLOCK);
3881                 return -EINVAL;
3882         }
3883
3884         /* Allocate outside of device_list_mutex */
3885         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3886         if (IS_ERR(sctx))
3887                 return PTR_ERR(sctx);
3888
3889         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3890         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
3891         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3892                      !is_dev_replace)) {
3893                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3894                 ret = -ENODEV;
3895                 goto out_free_ctx;
3896         }
3897
3898         if (!is_dev_replace && !readonly &&
3899             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3900                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3901                 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3902                                 rcu_str_deref(dev->name));
3903                 ret = -EROFS;
3904                 goto out_free_ctx;
3905         }
3906
3907         mutex_lock(&fs_info->scrub_lock);
3908         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3909             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3910                 mutex_unlock(&fs_info->scrub_lock);
3911                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3912                 ret = -EIO;
3913                 goto out_free_ctx;
3914         }
3915
3916         down_read(&fs_info->dev_replace.rwsem);
3917         if (dev->scrub_ctx ||
3918             (!is_dev_replace &&
3919              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3920                 up_read(&fs_info->dev_replace.rwsem);
3921                 mutex_unlock(&fs_info->scrub_lock);
3922                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3923                 ret = -EINPROGRESS;
3924                 goto out_free_ctx;
3925         }
3926         up_read(&fs_info->dev_replace.rwsem);
3927
3928         ret = scrub_workers_get(fs_info, is_dev_replace);
3929         if (ret) {
3930                 mutex_unlock(&fs_info->scrub_lock);
3931                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3932                 goto out_free_ctx;
3933         }
3934
3935         sctx->readonly = readonly;
3936         dev->scrub_ctx = sctx;
3937         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3938
3939         /*
3940          * checking @scrub_pause_req here, we can avoid
3941          * race between committing transaction and scrubbing.
3942          */
3943         __scrub_blocked_if_needed(fs_info);
3944         atomic_inc(&fs_info->scrubs_running);
3945         mutex_unlock(&fs_info->scrub_lock);
3946
3947         /*
3948          * In order to avoid deadlock with reclaim when there is a transaction
3949          * trying to pause scrub, make sure we use GFP_NOFS for all the
3950          * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3951          * invoked by our callees. The pausing request is done when the
3952          * transaction commit starts, and it blocks the transaction until scrub
3953          * is paused (done at specific points at scrub_stripe() or right above
3954          * before incrementing fs_info->scrubs_running).
3955          */
3956         nofs_flag = memalloc_nofs_save();
3957         if (!is_dev_replace) {
3958                 btrfs_info(fs_info, "scrub: started on devid %llu", devid);
3959                 /*
3960                  * by holding device list mutex, we can
3961                  * kick off writing super in log tree sync.
3962                  */
3963                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3964                 ret = scrub_supers(sctx, dev);
3965                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3966         }
3967
3968         if (!ret)
3969                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3970         memalloc_nofs_restore(nofs_flag);
3971
3972         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3973         atomic_dec(&fs_info->scrubs_running);
3974         wake_up(&fs_info->scrub_pause_wait);
3975
3976         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3977
3978         if (progress)
3979                 memcpy(progress, &sctx->stat, sizeof(*progress));
3980
3981         if (!is_dev_replace)
3982                 btrfs_info(fs_info, "scrub: %s on devid %llu with status: %d",
3983                         ret ? "not finished" : "finished", devid, ret);
3984
3985         mutex_lock(&fs_info->scrub_lock);
3986         dev->scrub_ctx = NULL;
3987         if (refcount_dec_and_test(&fs_info->scrub_workers_refcnt)) {
3988                 scrub_workers = fs_info->scrub_workers;
3989                 scrub_wr_comp = fs_info->scrub_wr_completion_workers;
3990                 scrub_parity = fs_info->scrub_parity_workers;
3991
3992                 fs_info->scrub_workers = NULL;
3993                 fs_info->scrub_wr_completion_workers = NULL;
3994                 fs_info->scrub_parity_workers = NULL;
3995         }
3996         mutex_unlock(&fs_info->scrub_lock);
3997
3998         btrfs_destroy_workqueue(scrub_workers);
3999         btrfs_destroy_workqueue(scrub_wr_comp);
4000         btrfs_destroy_workqueue(scrub_parity);
4001         scrub_put_ctx(sctx);
4002
4003         return ret;
4004
4005 out_free_ctx:
4006         scrub_free_ctx(sctx);
4007
4008         return ret;
4009 }
4010
4011 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4012 {
4013         mutex_lock(&fs_info->scrub_lock);
4014         atomic_inc(&fs_info->scrub_pause_req);
4015         while (atomic_read(&fs_info->scrubs_paused) !=
4016                atomic_read(&fs_info->scrubs_running)) {
4017                 mutex_unlock(&fs_info->scrub_lock);
4018                 wait_event(fs_info->scrub_pause_wait,
4019                            atomic_read(&fs_info->scrubs_paused) ==
4020                            atomic_read(&fs_info->scrubs_running));
4021                 mutex_lock(&fs_info->scrub_lock);
4022         }
4023         mutex_unlock(&fs_info->scrub_lock);
4024 }
4025
4026 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4027 {
4028         atomic_dec(&fs_info->scrub_pause_req);
4029         wake_up(&fs_info->scrub_pause_wait);
4030 }
4031
4032 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4033 {
4034         mutex_lock(&fs_info->scrub_lock);
4035         if (!atomic_read(&fs_info->scrubs_running)) {
4036                 mutex_unlock(&fs_info->scrub_lock);
4037                 return -ENOTCONN;
4038         }
4039
4040         atomic_inc(&fs_info->scrub_cancel_req);
4041         while (atomic_read(&fs_info->scrubs_running)) {
4042                 mutex_unlock(&fs_info->scrub_lock);
4043                 wait_event(fs_info->scrub_pause_wait,
4044                            atomic_read(&fs_info->scrubs_running) == 0);
4045                 mutex_lock(&fs_info->scrub_lock);
4046         }
4047         atomic_dec(&fs_info->scrub_cancel_req);
4048         mutex_unlock(&fs_info->scrub_lock);
4049
4050         return 0;
4051 }
4052
4053 int btrfs_scrub_cancel_dev(struct btrfs_device *dev)
4054 {
4055         struct btrfs_fs_info *fs_info = dev->fs_info;
4056         struct scrub_ctx *sctx;
4057
4058         mutex_lock(&fs_info->scrub_lock);
4059         sctx = dev->scrub_ctx;
4060         if (!sctx) {
4061                 mutex_unlock(&fs_info->scrub_lock);
4062                 return -ENOTCONN;
4063         }
4064         atomic_inc(&sctx->cancel_req);
4065         while (dev->scrub_ctx) {
4066                 mutex_unlock(&fs_info->scrub_lock);
4067                 wait_event(fs_info->scrub_pause_wait,
4068                            dev->scrub_ctx == NULL);
4069                 mutex_lock(&fs_info->scrub_lock);
4070         }
4071         mutex_unlock(&fs_info->scrub_lock);
4072
4073         return 0;
4074 }
4075
4076 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4077                          struct btrfs_scrub_progress *progress)
4078 {
4079         struct btrfs_device *dev;
4080         struct scrub_ctx *sctx = NULL;
4081
4082         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4083         dev = btrfs_find_device(fs_info->fs_devices, devid, NULL, NULL, true);
4084         if (dev)
4085                 sctx = dev->scrub_ctx;
4086         if (sctx)
4087                 memcpy(progress, &sctx->stat, sizeof(*progress));
4088         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4089
4090         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4091 }
4092
4093 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4094                                u64 extent_logical, u64 extent_len,
4095                                u64 *extent_physical,
4096                                struct btrfs_device **extent_dev,
4097                                int *extent_mirror_num)
4098 {
4099         u64 mapped_length;
4100         struct btrfs_bio *bbio = NULL;
4101         int ret;
4102
4103         mapped_length = extent_len;
4104         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4105                               &mapped_length, &bbio, 0);
4106         if (ret || !bbio || mapped_length < extent_len ||
4107             !bbio->stripes[0].dev->bdev) {
4108                 btrfs_put_bbio(bbio);
4109                 return;
4110         }
4111
4112         *extent_physical = bbio->stripes[0].physical;
4113         *extent_mirror_num = bbio->mirror_num;
4114         *extent_dev = bbio->stripes[0].dev;
4115         btrfs_put_bbio(bbio);
4116 }