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 "ctree.h"
  10 #include "volumes.h"
  11 #include "disk-io.h"
  12 #include "ordered-data.h"
  13 #include "transaction.h"
  14 #include "backref.h"
  15 #include "extent_io.h"
  16 #include "dev-replace.h"
  17 #include "check-integrity.h"
  18 #include "rcu-string.h"
  19 #include "raid56.h"
  20
  21 /*
  22  * This is only the first step towards a full-features scrub. It reads all
  23  * extent and super block and verifies the checksums. In case a bad checksum
  24  * is found or the extent cannot be read, good data will be written back if
  25  * any can be found.
  26  *
  27  * Future enhancements:
  28  *  - In case an unrepairable extent is encountered, track which files are
  29  *    affected and report them
  30  *  - track and record media errors, throw out bad devices
  31  *  - add a mode to also read unallocated space
  32  */
  33
  34 struct scrub_block;
  35 struct scrub_ctx;
  36
  37 /*
  38  * the following three values only influence the performance.
  39  * The last one configures the number of parallel and outstanding I/O
  40  * operations. The first two values configure an upper limit for the number
  41  * of (dynamically allocated) pages that are added to a bio.
  42  */
  43 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  44 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  45 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  46
  47 /*
  48  * the following value times PAGE_SIZE needs to be large enough to match the
  49  * largest node/leaf/sector size that shall be supported.
  50  * Values larger than BTRFS_STRIPE_LEN are not supported.
  51  */
  52 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  53
  54 struct scrub_recover {
  55         refcount_t              refs;
  56         struct btrfs_bio        *bbio;
  57         u64                     map_length;
  58 };
  59
  60 struct scrub_page {
  61         struct scrub_block      *sblock;
  62         struct page             *page;
  63         struct btrfs_device     *dev;
  64         struct list_head        list;
  65         u64                     flags;  /* extent flags */
  66         u64                     generation;
  67         u64                     logical;
  68         u64                     physical;
  69         u64                     physical_for_dev_replace;
  70         atomic_t                refs;
  71         struct {
  72                 unsigned int    mirror_num:8;
  73                 unsigned int    have_csum:1;
  74                 unsigned int    io_error:1;
  75         };
  76         u8                      csum[BTRFS_CSUM_SIZE];
  77
  78         struct scrub_recover    *recover;
  79 };
  80
  81 struct scrub_bio {
  82         int                     index;
  83         struct scrub_ctx        *sctx;
  84         struct btrfs_device     *dev;
  85         struct bio              *bio;
  86         blk_status_t            status;
  87         u64                     logical;
  88         u64                     physical;
  89 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  90         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  91 #else
  92         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  93 #endif
  94         int                     page_count;
  95         int                     next_free;
  96         struct btrfs_work       work;
  97 };
  98
  99 struct scrub_block {
 100         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 101         int                     page_count;
 102         atomic_t                outstanding_pages;
 103         refcount_t              refs; /* free mem on transition to zero */
 104         struct scrub_ctx        *sctx;
 105         struct scrub_parity     *sparity;
 106         struct {
 107                 unsigned int    header_error:1;
 108                 unsigned int    checksum_error:1;
 109                 unsigned int    no_io_error_seen:1;
 110                 unsigned int    generation_error:1; /* also sets header_error */
 111
 112                 /* The following is for the data used to check parity */
 113                 /* It is for the data with checksum */
 114                 unsigned int    data_corrected:1;
 115         };
 116         struct btrfs_work       work;
 117 };
 118
 119 /* Used for the chunks with parity stripe such RAID5/6 */
 120 struct scrub_parity {
 121         struct scrub_ctx        *sctx;
 122
 123         struct btrfs_device     *scrub_dev;
 124
 125         u64                     logic_start;
 126
 127         u64                     logic_end;
 128
 129         int                     nsectors;
 130
 131         u64                     stripe_len;
 132
 133         refcount_t              refs;
 134
 135         struct list_head        spages;
 136
 137         /* Work of parity check and repair */
 138         struct btrfs_work       work;
 139
 140         /* Mark the parity blocks which have data */
 141         unsigned long           *dbitmap;
 142
 143         /*
 144          * Mark the parity blocks which have data, but errors happen when
 145          * read data or check data
 146          */
 147         unsigned long           *ebitmap;
 148
 149         unsigned long           bitmap[0];
 150 };
 151
 152 struct scrub_ctx {
 153         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 154         struct btrfs_fs_info    *fs_info;
 155         int                     first_free;
 156         int                     curr;
 157         atomic_t                bios_in_flight;
 158         atomic_t                workers_pending;
 159         spinlock_t              list_lock;
 160         wait_queue_head_t       list_wait;
 161         u16                     csum_size;
 162         struct list_head        csum_list;
 163         atomic_t                cancel_req;
 164         int                     readonly;
 165         int                     pages_per_rd_bio;
 166
 167         int                     is_dev_replace;
 168
 169         struct scrub_bio        *wr_curr_bio;
 170         struct mutex            wr_lock;
 171         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 172         struct btrfs_device     *wr_tgtdev;
 173         bool                    flush_all_writes;
 174
 175         /*
 176          * statistics
 177          */
 178         struct btrfs_scrub_progress stat;
 179         spinlock_t              stat_lock;
 180
 181         /*
 182          * Use a ref counter to avoid use-after-free issues. Scrub workers
 183          * decrement bios_in_flight and workers_pending and then do a wakeup
 184          * on the list_wait wait queue. We must ensure the main scrub task
 185          * doesn't free the scrub context before or while the workers are
 186          * doing the wakeup() call.
 187          */
 188         refcount_t              refs;
 189 };
 190
 191 struct scrub_warning {
 192         struct btrfs_path       *path;
 193         u64                     extent_item_size;
 194         const char              *errstr;
 195         u64                     physical;
 196         u64                     logical;
 197         struct btrfs_device     *dev;
 198 };
 199
 200 struct full_stripe_lock {
 201         struct rb_node node;
 202         u64 logical;
 203         u64 refs;
 204         struct mutex mutex;
 205 };
 206
 207 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 208 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 209 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 210 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 211                                      struct scrub_block *sblocks_for_recheck);
 212 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 213                                 struct scrub_block *sblock,
 214                                 int retry_failed_mirror);
 215 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 216 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 217                                              struct scrub_block *sblock_good);
 218 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 219                                             struct scrub_block *sblock_good,
 220                                             int page_num, int force_write);
 221 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 222 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 223                                            int page_num);
 224 static int scrub_checksum_data(struct scrub_block *sblock);
 225 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 226 static int scrub_checksum_super(struct scrub_block *sblock);
 227 static void scrub_block_get(struct scrub_block *sblock);
 228 static void scrub_block_put(struct scrub_block *sblock);
 229 static void scrub_page_get(struct scrub_page *spage);
 230 static void scrub_page_put(struct scrub_page *spage);
 231 static void scrub_parity_get(struct scrub_parity *sparity);
 232 static void scrub_parity_put(struct scrub_parity *sparity);
 233 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 234                                     struct scrub_page *spage);
 235 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 236                        u64 physical, struct btrfs_device *dev, u64 flags,
 237                        u64 gen, int mirror_num, u8 *csum, int force,
 238                        u64 physical_for_dev_replace);
 239 static void scrub_bio_end_io(struct bio *bio);
 240 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 241 static void scrub_block_complete(struct scrub_block *sblock);
 242 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 243                                u64 extent_logical, u64 extent_len,
 244                                u64 *extent_physical,
 245                                struct btrfs_device **extent_dev,
 246                                int *extent_mirror_num);
 247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 248                                     struct scrub_page *spage);
 249 static void scrub_wr_submit(struct scrub_ctx *sctx);
 250 static void scrub_wr_bio_end_io(struct bio *bio);
 251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 252 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 253 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 254 static void scrub_put_ctx(struct scrub_ctx *sctx);
 255
 256 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
 257 {
 258         return page->recover &&
 259                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 260 }
 261
 262 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 263 {
 264         refcount_inc(&sctx->refs);
 265         atomic_inc(&sctx->bios_in_flight);
 266 }
 267
 268 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 269 {
 270         atomic_dec(&sctx->bios_in_flight);
 271         wake_up(&sctx->list_wait);
 272         scrub_put_ctx(sctx);
 273 }
 274
 275 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 276 {
 277         while (atomic_read(&fs_info->scrub_pause_req)) {
 278                 mutex_unlock(&fs_info->scrub_lock);
 279                 wait_event(fs_info->scrub_pause_wait,
 280                    atomic_read(&fs_info->scrub_pause_req) == 0);
 281                 mutex_lock(&fs_info->scrub_lock);
 282         }
 283 }
 284
 285 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 286 {
 287         atomic_inc(&fs_info->scrubs_paused);
 288         wake_up(&fs_info->scrub_pause_wait);
 289 }
 290
 291 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 292 {
 293         mutex_lock(&fs_info->scrub_lock);
 294         __scrub_blocked_if_needed(fs_info);
 295         atomic_dec(&fs_info->scrubs_paused);
 296         mutex_unlock(&fs_info->scrub_lock);
 297
 298         wake_up(&fs_info->scrub_pause_wait);
 299 }
 300
 301 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 302 {
 303         scrub_pause_on(fs_info);
 304         scrub_pause_off(fs_info);
 305 }
 306
 307 /*
 308  * Insert new full stripe lock into full stripe locks tree
 309  *
 310  * Return pointer to existing or newly inserted full_stripe_lock structure if
 311  * everything works well.
 312  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 313  *
 314  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 315  * function
 316  */
 317 static struct full_stripe_lock *insert_full_stripe_lock(
 318                 struct btrfs_full_stripe_locks_tree *locks_root,
 319                 u64 fstripe_logical)
 320 {
 321         struct rb_node **p;
 322         struct rb_node *parent = NULL;
 323         struct full_stripe_lock *entry;
 324         struct full_stripe_lock *ret;
 325
 326         lockdep_assert_held(&locks_root->lock);
 327
 328         p = &locks_root->root.rb_node;
 329         while (*p) {
 330                 parent = *p;
 331                 entry = rb_entry(parent, struct full_stripe_lock, node);
 332                 if (fstripe_logical < entry->logical) {
 333                         p = &(*p)->rb_left;
 334                 } else if (fstripe_logical > entry->logical) {
 335                         p = &(*p)->rb_right;
 336                 } else {
 337                         entry->refs++;
 338                         return entry;
 339                 }
 340         }
 341
 342         /*
 343          * Insert new lock.
 344          */
 345         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 346         if (!ret)
 347                 return ERR_PTR(-ENOMEM);
 348         ret->logical = fstripe_logical;
 349         ret->refs = 1;
 350         mutex_init(&ret->mutex);
 351
 352         rb_link_node(&ret->node, parent, p);
 353         rb_insert_color(&ret->node, &locks_root->root);
 354         return ret;
 355 }
 356
 357 /*
 358  * Search for a full stripe lock of a block group
 359  *
 360  * Return pointer to existing full stripe lock if found
 361  * Return NULL if not found
 362  */
 363 static struct full_stripe_lock *search_full_stripe_lock(
 364                 struct btrfs_full_stripe_locks_tree *locks_root,
 365                 u64 fstripe_logical)
 366 {
 367         struct rb_node *node;
 368         struct full_stripe_lock *entry;
 369
 370         lockdep_assert_held(&locks_root->lock);
 371
 372         node = locks_root->root.rb_node;
 373         while (node) {
 374                 entry = rb_entry(node, struct full_stripe_lock, node);
 375                 if (fstripe_logical < entry->logical)
 376                         node = node->rb_left;
 377                 else if (fstripe_logical > entry->logical)
 378                         node = node->rb_right;
 379                 else
 380                         return entry;
 381         }
 382         return NULL;
 383 }
 384
 385 /*
 386  * Helper to get full stripe logical from a normal bytenr.
 387  *
 388  * Caller must ensure @cache is a RAID56 block group.
 389  */
 390 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
 391                                    u64 bytenr)
 392 {
 393         u64 ret;
 394
 395         /*
 396          * Due to chunk item size limit, full stripe length should not be
 397          * larger than U32_MAX. Just a sanity check here.
 398          */
 399         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 400
 401         /*
 402          * round_down() can only handle power of 2, while RAID56 full
 403          * stripe length can be 64KiB * n, so we need to manually round down.
 404          */
 405         ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
 406                 cache->full_stripe_len + cache->key.objectid;
 407         return ret;
 408 }
 409
 410 /*
 411  * Lock a full stripe to avoid concurrency of recovery and read
 412  *
 413  * It's only used for profiles with parities (RAID5/6), for other profiles it
 414  * does nothing.
 415  *
 416  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 417  * So caller must call unlock_full_stripe() at the same context.
 418  *
 419  * Return <0 if encounters error.
 420  */
 421 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 422                             bool *locked_ret)
 423 {
 424         struct btrfs_block_group_cache *bg_cache;
 425         struct btrfs_full_stripe_locks_tree *locks_root;
 426         struct full_stripe_lock *existing;
 427         u64 fstripe_start;
 428         int ret = 0;
 429
 430         *locked_ret = false;
 431         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 432         if (!bg_cache) {
 433                 ASSERT(0);
 434                 return -ENOENT;
 435         }
 436
 437         /* Profiles not based on parity don't need full stripe lock */
 438         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 439                 goto out;
 440         locks_root = &bg_cache->full_stripe_locks_root;
 441
 442         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 443
 444         /* Now insert the full stripe lock */
 445         mutex_lock(&locks_root->lock);
 446         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 447         mutex_unlock(&locks_root->lock);
 448         if (IS_ERR(existing)) {
 449                 ret = PTR_ERR(existing);
 450                 goto out;
 451         }
 452         mutex_lock(&existing->mutex);
 453         *locked_ret = true;
 454 out:
 455         btrfs_put_block_group(bg_cache);
 456         return ret;
 457 }
 458
 459 /*
 460  * Unlock a full stripe.
 461  *
 462  * NOTE: Caller must ensure it's the same context calling corresponding
 463  * lock_full_stripe().
 464  *
 465  * Return 0 if we unlock full stripe without problem.
 466  * Return <0 for error
 467  */
 468 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 469                               bool locked)
 470 {
 471         struct btrfs_block_group_cache *bg_cache;
 472         struct btrfs_full_stripe_locks_tree *locks_root;
 473         struct full_stripe_lock *fstripe_lock;
 474         u64 fstripe_start;
 475         bool freeit = false;
 476         int ret = 0;
 477
 478         /* If we didn't acquire full stripe lock, no need to continue */
 479         if (!locked)
 480                 return 0;
 481
 482         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 483         if (!bg_cache) {
 484                 ASSERT(0);
 485                 return -ENOENT;
 486         }
 487         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 488                 goto out;
 489
 490         locks_root = &bg_cache->full_stripe_locks_root;
 491         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 492
 493         mutex_lock(&locks_root->lock);
 494         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 495         /* Unpaired unlock_full_stripe() detected */
 496         if (!fstripe_lock) {
 497                 WARN_ON(1);
 498                 ret = -ENOENT;
 499                 mutex_unlock(&locks_root->lock);
 500                 goto out;
 501         }
 502
 503         if (fstripe_lock->refs == 0) {
 504                 WARN_ON(1);
 505                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 506                         fstripe_lock->logical);
 507         } else {
 508                 fstripe_lock->refs--;
 509         }
 510
 511         if (fstripe_lock->refs == 0) {
 512                 rb_erase(&fstripe_lock->node, &locks_root->root);
 513                 freeit = true;
 514         }
 515         mutex_unlock(&locks_root->lock);
 516
 517         mutex_unlock(&fstripe_lock->mutex);
 518         if (freeit)
 519                 kfree(fstripe_lock);
 520 out:
 521         btrfs_put_block_group(bg_cache);
 522         return ret;
 523 }
 524
 525 static void scrub_free_csums(struct scrub_ctx *sctx)
 526 {
 527         while (!list_empty(&sctx->csum_list)) {
 528                 struct btrfs_ordered_sum *sum;
 529                 sum = list_first_entry(&sctx->csum_list,
 530                                        struct btrfs_ordered_sum, list);
 531                 list_del(&sum->list);
 532                 kfree(sum);
 533         }
 534 }
 535
 536 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 537 {
 538         int i;
 539
 540         if (!sctx)
 541                 return;
 542
 543         /* this can happen when scrub is cancelled */
 544         if (sctx->curr != -1) {
 545                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 546
 547                 for (i = 0; i < sbio->page_count; i++) {
 548                         WARN_ON(!sbio->pagev[i]->page);
 549                         scrub_block_put(sbio->pagev[i]->sblock);
 550                 }
 551                 bio_put(sbio->bio);
 552         }
 553
 554         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 555                 struct scrub_bio *sbio = sctx->bios[i];
 556
 557                 if (!sbio)
 558                         break;
 559                 kfree(sbio);
 560         }
 561
 562         kfree(sctx->wr_curr_bio);
 563         scrub_free_csums(sctx);
 564         kfree(sctx);
 565 }
 566
 567 static void scrub_put_ctx(struct scrub_ctx *sctx)
 568 {
 569         if (refcount_dec_and_test(&sctx->refs))
 570                 scrub_free_ctx(sctx);
 571 }
 572
 573 static noinline_for_stack struct scrub_ctx *scrub_setup_ctx(
 574                 struct btrfs_fs_info *fs_info, int is_dev_replace)
 575 {
 576         struct scrub_ctx *sctx;
 577         int             i;
 578
 579         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 580         if (!sctx)
 581                 goto nomem;
 582         refcount_set(&sctx->refs, 1);
 583         sctx->is_dev_replace = is_dev_replace;
 584         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 585         sctx->curr = -1;
 586         sctx->fs_info = fs_info;
 587         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 588                 struct scrub_bio *sbio;
 589
 590                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 591                 if (!sbio)
 592                         goto nomem;
 593                 sctx->bios[i] = sbio;
 594
 595                 sbio->index = i;
 596                 sbio->sctx = sctx;
 597                 sbio->page_count = 0;
 598                 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
 599                                 scrub_bio_end_io_worker, NULL, NULL);
 600
 601                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 602                         sctx->bios[i]->next_free = i + 1;
 603                 else
 604                         sctx->bios[i]->next_free = -1;
 605         }
 606         sctx->first_free = 0;
 607         atomic_set(&sctx->bios_in_flight, 0);
 608         atomic_set(&sctx->workers_pending, 0);
 609         atomic_set(&sctx->cancel_req, 0);
 610         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 611         INIT_LIST_HEAD(&sctx->csum_list);
 612
 613         spin_lock_init(&sctx->list_lock);
 614         spin_lock_init(&sctx->stat_lock);
 615         init_waitqueue_head(&sctx->list_wait);
 616
 617         WARN_ON(sctx->wr_curr_bio != NULL);
 618         mutex_init(&sctx->wr_lock);
 619         sctx->wr_curr_bio = NULL;
 620         if (is_dev_replace) {
 621                 WARN_ON(!fs_info->dev_replace.tgtdev);
 622                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 623                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 624                 sctx->flush_all_writes = false;
 625         }
 626
 627         return sctx;
 628
 629 nomem:
 630         scrub_free_ctx(sctx);
 631         return ERR_PTR(-ENOMEM);
 632 }
 633
 634 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 635                                      void *warn_ctx)
 636 {
 637         u64 isize;
 638         u32 nlink;
 639         int ret;
 640         int i;
 641         unsigned nofs_flag;
 642         struct extent_buffer *eb;
 643         struct btrfs_inode_item *inode_item;
 644         struct scrub_warning *swarn = warn_ctx;
 645         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 646         struct inode_fs_paths *ipath = NULL;
 647         struct btrfs_root *local_root;
 648         struct btrfs_key root_key;
 649         struct btrfs_key key;
 650
 651         root_key.objectid = root;
 652         root_key.type = BTRFS_ROOT_ITEM_KEY;
 653         root_key.offset = (u64)-1;
 654         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 655         if (IS_ERR(local_root)) {
 656                 ret = PTR_ERR(local_root);
 657                 goto err;
 658         }
 659
 660         /*
 661          * this makes the path point to (inum INODE_ITEM ioff)
 662          */
 663         key.objectid = inum;
 664         key.type = BTRFS_INODE_ITEM_KEY;
 665         key.offset = 0;
 666
 667         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 668         if (ret) {
 669                 btrfs_release_path(swarn->path);
 670                 goto err;
 671         }
 672
 673         eb = swarn->path->nodes[0];
 674         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 675                                         struct btrfs_inode_item);
 676         isize = btrfs_inode_size(eb, inode_item);
 677         nlink = btrfs_inode_nlink(eb, inode_item);
 678         btrfs_release_path(swarn->path);
 679
 680         /*
 681          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 682          * uses GFP_NOFS in this context, so we keep it consistent but it does
 683          * not seem to be strictly necessary.
 684          */
 685         nofs_flag = memalloc_nofs_save();
 686         ipath = init_ipath(4096, local_root, swarn->path);
 687         memalloc_nofs_restore(nofs_flag);
 688         if (IS_ERR(ipath)) {
 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         free_ipath(ipath);
 713         return 0;
 714
 715 err:
 716         btrfs_warn_in_rcu(fs_info,
 717                           "%s at logical %llu on dev %s, physical %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 718                           swarn->errstr, swarn->logical,
 719                           rcu_str_deref(swarn->dev->name),
 720                           swarn->physical,
 721                           root, inum, offset, ret);
 722
 723         free_ipath(ipath);
 724         return 0;
 725 }
 726
 727 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 728 {
 729         struct btrfs_device *dev;
 730         struct btrfs_fs_info *fs_info;
 731         struct btrfs_path *path;
 732         struct btrfs_key found_key;
 733         struct extent_buffer *eb;
 734         struct btrfs_extent_item *ei;
 735         struct scrub_warning swarn;
 736         unsigned long ptr = 0;
 737         u64 extent_item_pos;
 738         u64 flags = 0;
 739         u64 ref_root;
 740         u32 item_size;
 741         u8 ref_level = 0;
 742         int ret;
 743
 744         WARN_ON(sblock->page_count < 1);
 745         dev = sblock->pagev[0]->dev;
 746         fs_info = sblock->sctx->fs_info;
 747
 748         path = btrfs_alloc_path();
 749         if (!path)
 750                 return;
 751
 752         swarn.physical = sblock->pagev[0]->physical;
 753         swarn.logical = sblock->pagev[0]->logical;
 754         swarn.errstr = errstr;
 755         swarn.dev = NULL;
 756
 757         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 758                                   &flags);
 759         if (ret < 0)
 760                 goto out;
 761
 762         extent_item_pos = swarn.logical - found_key.objectid;
 763         swarn.extent_item_size = found_key.offset;
 764
 765         eb = path->nodes[0];
 766         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 767         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 768
 769         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 770                 do {
 771                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 772                                                       item_size, &ref_root,
 773                                                       &ref_level);
 774                         btrfs_warn_in_rcu(fs_info,
 775 "%s at logical %llu on dev %s, physical %llu: metadata %s (level %d) in tree %llu",
 776                                 errstr, swarn.logical,
 777                                 rcu_str_deref(dev->name),
 778                                 swarn.physical,
 779                                 ref_level ? "node" : "leaf",
 780                                 ret < 0 ? -1 : ref_level,
 781                                 ret < 0 ? -1 : ref_root);
 782                 } while (ret != 1);
 783                 btrfs_release_path(path);
 784         } else {
 785                 btrfs_release_path(path);
 786                 swarn.path = path;
 787                 swarn.dev = dev;
 788                 iterate_extent_inodes(fs_info, found_key.objectid,
 789                                         extent_item_pos, 1,
 790                                         scrub_print_warning_inode, &swarn, false);
 791         }
 792
 793 out:
 794         btrfs_free_path(path);
 795 }
 796
 797 static inline void scrub_get_recover(struct scrub_recover *recover)
 798 {
 799         refcount_inc(&recover->refs);
 800 }
 801
 802 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
 803                                      struct scrub_recover *recover)
 804 {
 805         if (refcount_dec_and_test(&recover->refs)) {
 806                 btrfs_bio_counter_dec(fs_info);
 807                 btrfs_put_bbio(recover->bbio);
 808                 kfree(recover);
 809         }
 810 }
 811
 812 /*
 813  * scrub_handle_errored_block gets called when either verification of the
 814  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 815  * case, this function handles all pages in the bio, even though only one
 816  * may be bad.
 817  * The goal of this function is to repair the errored block by using the
 818  * contents of one of the mirrors.
 819  */
 820 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 821 {
 822         struct scrub_ctx *sctx = sblock_to_check->sctx;
 823         struct btrfs_device *dev;
 824         struct btrfs_fs_info *fs_info;
 825         u64 logical;
 826         unsigned int failed_mirror_index;
 827         unsigned int is_metadata;
 828         unsigned int have_csum;
 829         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 830         struct scrub_block *sblock_bad;
 831         int ret;
 832         int mirror_index;
 833         int page_num;
 834         int success;
 835         bool full_stripe_locked;
 836         unsigned int nofs_flag;
 837         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 838                                       DEFAULT_RATELIMIT_BURST);
 839
 840         BUG_ON(sblock_to_check->page_count < 1);
 841         fs_info = sctx->fs_info;
 842         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 843                 /*
 844                  * if we find an error in a super block, we just report it.
 845                  * They will get written with the next transaction commit
 846                  * anyway
 847                  */
 848                 spin_lock(&sctx->stat_lock);
 849                 ++sctx->stat.super_errors;
 850                 spin_unlock(&sctx->stat_lock);
 851                 return 0;
 852         }
 853         logical = sblock_to_check->pagev[0]->logical;
 854         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 855         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 856         is_metadata = !(sblock_to_check->pagev[0]->flags &
 857                         BTRFS_EXTENT_FLAG_DATA);
 858         have_csum = sblock_to_check->pagev[0]->have_csum;
 859         dev = sblock_to_check->pagev[0]->dev;
 860
 861         /*
 862          * We must use GFP_NOFS because the scrub task might be waiting for a
 863          * worker task executing this function and in turn a transaction commit
 864          * might be waiting the scrub task to pause (which needs to wait for all
 865          * the worker tasks to complete before pausing).
 866          * We do allocations in the workers through insert_full_stripe_lock()
 867          * and scrub_add_page_to_wr_bio(), which happens down the call chain of
 868          * this function.
 869          */
 870         nofs_flag = memalloc_nofs_save();
 871         /*
 872          * For RAID5/6, race can happen for a different device scrub thread.
 873          * For data corruption, Parity and Data threads will both try
 874          * to recovery the data.
 875          * Race can lead to doubly added csum error, or even unrecoverable
 876          * error.
 877          */
 878         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
 879         if (ret < 0) {
 880                 memalloc_nofs_restore(nofs_flag);
 881                 spin_lock(&sctx->stat_lock);
 882                 if (ret == -ENOMEM)
 883                         sctx->stat.malloc_errors++;
 884                 sctx->stat.read_errors++;
 885                 sctx->stat.uncorrectable_errors++;
 886                 spin_unlock(&sctx->stat_lock);
 887                 return ret;
 888         }
 889
 890         /*
 891          * read all mirrors one after the other. This includes to
 892          * re-read the extent or metadata block that failed (that was
 893          * the cause that this fixup code is called) another time,
 894          * page by page this time in order to know which pages
 895          * caused I/O errors and which ones are good (for all mirrors).
 896          * It is the goal to handle the situation when more than one
 897          * mirror contains I/O errors, but the errors do not
 898          * overlap, i.e. the data can be repaired by selecting the
 899          * pages from those mirrors without I/O error on the
 900          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 901          * would be that mirror #1 has an I/O error on the first page,
 902          * the second page is good, and mirror #2 has an I/O error on
 903          * the second page, but the first page is good.
 904          * Then the first page of the first mirror can be repaired by
 905          * taking the first page of the second mirror, and the
 906          * second page of the second mirror can be repaired by
 907          * copying the contents of the 2nd page of the 1st mirror.
 908          * One more note: if the pages of one mirror contain I/O
 909          * errors, the checksum cannot be verified. In order to get
 910          * the best data for repairing, the first attempt is to find
 911          * a mirror without I/O errors and with a validated checksum.
 912          * Only if this is not possible, the pages are picked from
 913          * mirrors with I/O errors without considering the checksum.
 914          * If the latter is the case, at the end, the checksum of the
 915          * repaired area is verified in order to correctly maintain
 916          * the statistics.
 917          */
 918
 919         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 920                                       sizeof(*sblocks_for_recheck), GFP_KERNEL);
 921         if (!sblocks_for_recheck) {
 922                 spin_lock(&sctx->stat_lock);
 923                 sctx->stat.malloc_errors++;
 924                 sctx->stat.read_errors++;
 925                 sctx->stat.uncorrectable_errors++;
 926                 spin_unlock(&sctx->stat_lock);
 927                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 928                 goto out;
 929         }
 930
 931         /* setup the context, map the logical blocks and alloc the pages */
 932         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 933         if (ret) {
 934                 spin_lock(&sctx->stat_lock);
 935                 sctx->stat.read_errors++;
 936                 sctx->stat.uncorrectable_errors++;
 937                 spin_unlock(&sctx->stat_lock);
 938                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 939                 goto out;
 940         }
 941         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 942         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 943
 944         /* build and submit the bios for the failed mirror, check checksums */
 945         scrub_recheck_block(fs_info, sblock_bad, 1);
 946
 947         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 948             sblock_bad->no_io_error_seen) {
 949                 /*
 950                  * the error disappeared after reading page by page, or
 951                  * the area was part of a huge bio and other parts of the
 952                  * bio caused I/O errors, or the block layer merged several
 953                  * read requests into one and the error is caused by a
 954                  * different bio (usually one of the two latter cases is
 955                  * the cause)
 956                  */
 957                 spin_lock(&sctx->stat_lock);
 958                 sctx->stat.unverified_errors++;
 959                 sblock_to_check->data_corrected = 1;
 960                 spin_unlock(&sctx->stat_lock);
 961
 962                 if (sctx->is_dev_replace)
 963                         scrub_write_block_to_dev_replace(sblock_bad);
 964                 goto out;
 965         }
 966
 967         if (!sblock_bad->no_io_error_seen) {
 968                 spin_lock(&sctx->stat_lock);
 969                 sctx->stat.read_errors++;
 970                 spin_unlock(&sctx->stat_lock);
 971                 if (__ratelimit(&_rs))
 972                         scrub_print_warning("i/o error", sblock_to_check);
 973                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 974         } else if (sblock_bad->checksum_error) {
 975                 spin_lock(&sctx->stat_lock);
 976                 sctx->stat.csum_errors++;
 977                 spin_unlock(&sctx->stat_lock);
 978                 if (__ratelimit(&_rs))
 979                         scrub_print_warning("checksum error", sblock_to_check);
 980                 btrfs_dev_stat_inc_and_print(dev,
 981                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 982         } else if (sblock_bad->header_error) {
 983                 spin_lock(&sctx->stat_lock);
 984                 sctx->stat.verify_errors++;
 985                 spin_unlock(&sctx->stat_lock);
 986                 if (__ratelimit(&_rs))
 987                         scrub_print_warning("checksum/header error",
 988                                             sblock_to_check);
 989                 if (sblock_bad->generation_error)
 990                         btrfs_dev_stat_inc_and_print(dev,
 991                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 992                 else
 993                         btrfs_dev_stat_inc_and_print(dev,
 994                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 995         }
 996
 997         if (sctx->readonly) {
 998                 ASSERT(!sctx->is_dev_replace);
 999                 goto out;
1000         }
1001
1002         /*
1003          * now build and submit the bios for the other mirrors, check
1004          * checksums.
1005          * First try to pick the mirror which is completely without I/O
1006          * errors and also does not have a checksum error.
1007          * If one is found, and if a checksum is present, the full block
1008          * that is known to contain an error is rewritten. Afterwards
1009          * the block is known to be corrected.
1010          * If a mirror is found which is completely correct, and no
1011          * checksum is present, only those pages are rewritten that had
1012          * an I/O error in the block to be repaired, since it cannot be
1013          * determined, which copy of the other pages is better (and it
1014          * could happen otherwise that a correct page would be
1015          * overwritten by a bad one).
1016          */
1017         for (mirror_index = 0; ;mirror_index++) {
1018                 struct scrub_block *sblock_other;
1019
1020                 if (mirror_index == failed_mirror_index)
1021                         continue;
1022
1023                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1024                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1025                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1026                                 break;
1027                         if (!sblocks_for_recheck[mirror_index].page_count)
1028                                 break;
1029
1030                         sblock_other = sblocks_for_recheck + mirror_index;
1031                 } else {
1032                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1033                         int max_allowed = r->bbio->num_stripes -
1034                                                 r->bbio->num_tgtdevs;
1035
1036                         if (mirror_index >= max_allowed)
1037                                 break;
1038                         if (!sblocks_for_recheck[1].page_count)
1039                                 break;
1040
1041                         ASSERT(failed_mirror_index == 0);
1042                         sblock_other = sblocks_for_recheck + 1;
1043                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1044                 }
1045
1046                 /* build and submit the bios, check checksums */
1047                 scrub_recheck_block(fs_info, sblock_other, 0);
1048
1049                 if (!sblock_other->header_error &&
1050                     !sblock_other->checksum_error &&
1051                     sblock_other->no_io_error_seen) {
1052                         if (sctx->is_dev_replace) {
1053                                 scrub_write_block_to_dev_replace(sblock_other);
1054                                 goto corrected_error;
1055                         } else {
1056                                 ret = scrub_repair_block_from_good_copy(
1057                                                 sblock_bad, sblock_other);
1058                                 if (!ret)
1059                                         goto corrected_error;
1060                         }
1061                 }
1062         }
1063
1064         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1065                 goto did_not_correct_error;
1066
1067         /*
1068          * In case of I/O errors in the area that is supposed to be
1069          * repaired, continue by picking good copies of those pages.
1070          * Select the good pages from mirrors to rewrite bad pages from
1071          * the area to fix. Afterwards verify the checksum of the block
1072          * that is supposed to be repaired. This verification step is
1073          * only done for the purpose of statistic counting and for the
1074          * final scrub report, whether errors remain.
1075          * A perfect algorithm could make use of the checksum and try
1076          * all possible combinations of pages from the different mirrors
1077          * until the checksum verification succeeds. For example, when
1078          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1079          * of mirror #2 is readable but the final checksum test fails,
1080          * then the 2nd page of mirror #3 could be tried, whether now
1081          * the final checksum succeeds. But this would be a rare
1082          * exception and is therefore not implemented. At least it is
1083          * avoided that the good copy is overwritten.
1084          * A more useful improvement would be to pick the sectors
1085          * without I/O error based on sector sizes (512 bytes on legacy
1086          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1087          * mirror could be repaired by taking 512 byte of a different
1088          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1089          * area are unreadable.
1090          */
1091         success = 1;
1092         for (page_num = 0; page_num < sblock_bad->page_count;
1093              page_num++) {
1094                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1095                 struct scrub_block *sblock_other = NULL;
1096
1097                 /* skip no-io-error page in scrub */
1098                 if (!page_bad->io_error && !sctx->is_dev_replace)
1099                         continue;
1100
1101                 if (scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1102                         /*
1103                          * In case of dev replace, if raid56 rebuild process
1104                          * didn't work out correct data, then copy the content
1105                          * in sblock_bad to make sure target device is identical
1106                          * to source device, instead of writing garbage data in
1107                          * sblock_for_recheck array to target device.
1108                          */
1109                         sblock_other = NULL;
1110                 } else if (page_bad->io_error) {
1111                         /* try to find no-io-error page in mirrors */
1112                         for (mirror_index = 0;
1113                              mirror_index < BTRFS_MAX_MIRRORS &&
1114                              sblocks_for_recheck[mirror_index].page_count > 0;
1115                              mirror_index++) {
1116                                 if (!sblocks_for_recheck[mirror_index].
1117                                     pagev[page_num]->io_error) {
1118                                         sblock_other = sblocks_for_recheck +
1119                                                        mirror_index;
1120                                         break;
1121                                 }
1122                         }
1123                         if (!sblock_other)
1124                                 success = 0;
1125                 }
1126
1127                 if (sctx->is_dev_replace) {
1128                         /*
1129                          * did not find a mirror to fetch the page
1130                          * from. scrub_write_page_to_dev_replace()
1131                          * handles this case (page->io_error), by
1132                          * filling the block with zeros before
1133                          * submitting the write request
1134                          */
1135                         if (!sblock_other)
1136                                 sblock_other = sblock_bad;
1137
1138                         if (scrub_write_page_to_dev_replace(sblock_other,
1139                                                             page_num) != 0) {
1140                                 atomic64_inc(
1141                                         &fs_info->dev_replace.num_write_errors);
1142                                 success = 0;
1143                         }
1144                 } else if (sblock_other) {
1145                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1146                                                                sblock_other,
1147                                                                page_num, 0);
1148                         if (0 == ret)
1149                                 page_bad->io_error = 0;
1150                         else
1151                                 success = 0;
1152                 }
1153         }
1154
1155         if (success && !sctx->is_dev_replace) {
1156                 if (is_metadata || have_csum) {
1157                         /*
1158                          * need to verify the checksum now that all
1159                          * sectors on disk are repaired (the write
1160                          * request for data to be repaired is on its way).
1161                          * Just be lazy and use scrub_recheck_block()
1162                          * which re-reads the data before the checksum
1163                          * is verified, but most likely the data comes out
1164                          * of the page cache.
1165                          */
1166                         scrub_recheck_block(fs_info, sblock_bad, 1);
1167                         if (!sblock_bad->header_error &&
1168                             !sblock_bad->checksum_error &&
1169                             sblock_bad->no_io_error_seen)
1170                                 goto corrected_error;
1171                         else
1172                                 goto did_not_correct_error;
1173                 } else {
1174 corrected_error:
1175                         spin_lock(&sctx->stat_lock);
1176                         sctx->stat.corrected_errors++;
1177                         sblock_to_check->data_corrected = 1;
1178                         spin_unlock(&sctx->stat_lock);
1179                         btrfs_err_rl_in_rcu(fs_info,
1180                                 "fixed up error at logical %llu on dev %s",
1181                                 logical, rcu_str_deref(dev->name));
1182                 }
1183         } else {
1184 did_not_correct_error:
1185                 spin_lock(&sctx->stat_lock);
1186                 sctx->stat.uncorrectable_errors++;
1187                 spin_unlock(&sctx->stat_lock);
1188                 btrfs_err_rl_in_rcu(fs_info,
1189                         "unable to fixup (regular) error at logical %llu on dev %s",
1190                         logical, rcu_str_deref(dev->name));
1191         }
1192
1193 out:
1194         if (sblocks_for_recheck) {
1195                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1196                      mirror_index++) {
1197                         struct scrub_block *sblock = sblocks_for_recheck +
1198                                                      mirror_index;
1199                         struct scrub_recover *recover;
1200                         int page_index;
1201
1202                         for (page_index = 0; page_index < sblock->page_count;
1203                              page_index++) {
1204                                 sblock->pagev[page_index]->sblock = NULL;
1205                                 recover = sblock->pagev[page_index]->recover;
1206                                 if (recover) {
1207                                         scrub_put_recover(fs_info, recover);
1208                                         sblock->pagev[page_index]->recover =
1209                                                                         NULL;
1210                                 }
1211                                 scrub_page_put(sblock->pagev[page_index]);
1212                         }
1213                 }
1214                 kfree(sblocks_for_recheck);
1215         }
1216
1217         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1218         memalloc_nofs_restore(nofs_flag);
1219         if (ret < 0)
1220                 return ret;
1221         return 0;
1222 }
1223
1224 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1225 {
1226         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1227                 return 2;
1228         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1229                 return 3;
1230         else
1231                 return (int)bbio->num_stripes;
1232 }
1233
1234 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1235                                                  u64 *raid_map,
1236                                                  u64 mapped_length,
1237                                                  int nstripes, int mirror,
1238                                                  int *stripe_index,
1239                                                  u64 *stripe_offset)
1240 {
1241         int i;
1242
1243         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1244                 /* RAID5/6 */
1245                 for (i = 0; i < nstripes; i++) {
1246                         if (raid_map[i] == RAID6_Q_STRIPE ||
1247                             raid_map[i] == RAID5_P_STRIPE)
1248                                 continue;
1249
1250                         if (logical >= raid_map[i] &&
1251                             logical < raid_map[i] + mapped_length)
1252                                 break;
1253                 }
1254
1255                 *stripe_index = i;
1256                 *stripe_offset = logical - raid_map[i];
1257         } else {
1258                 /* The other RAID type */
1259                 *stripe_index = mirror;
1260                 *stripe_offset = 0;
1261         }
1262 }
1263
1264 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1265                                      struct scrub_block *sblocks_for_recheck)
1266 {
1267         struct scrub_ctx *sctx = original_sblock->sctx;
1268         struct btrfs_fs_info *fs_info = sctx->fs_info;
1269         u64 length = original_sblock->page_count * PAGE_SIZE;
1270         u64 logical = original_sblock->pagev[0]->logical;
1271         u64 generation = original_sblock->pagev[0]->generation;
1272         u64 flags = original_sblock->pagev[0]->flags;
1273         u64 have_csum = original_sblock->pagev[0]->have_csum;
1274         struct scrub_recover *recover;
1275         struct btrfs_bio *bbio;
1276         u64 sublen;
1277         u64 mapped_length;
1278         u64 stripe_offset;
1279         int stripe_index;
1280         int page_index = 0;
1281         int mirror_index;
1282         int nmirrors;
1283         int ret;
1284
1285         /*
1286          * note: the two members refs and outstanding_pages
1287          * are not used (and not set) in the blocks that are used for
1288          * the recheck procedure
1289          */
1290
1291         while (length > 0) {
1292                 sublen = min_t(u64, length, PAGE_SIZE);
1293                 mapped_length = sublen;
1294                 bbio = NULL;
1295
1296                 /*
1297                  * with a length of PAGE_SIZE, each returned stripe
1298                  * represents one mirror
1299                  */
1300                 btrfs_bio_counter_inc_blocked(fs_info);
1301                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1302                                 logical, &mapped_length, &bbio);
1303                 if (ret || !bbio || mapped_length < sublen) {
1304                         btrfs_put_bbio(bbio);
1305                         btrfs_bio_counter_dec(fs_info);
1306                         return -EIO;
1307                 }
1308
1309                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1310                 if (!recover) {
1311                         btrfs_put_bbio(bbio);
1312                         btrfs_bio_counter_dec(fs_info);
1313                         return -ENOMEM;
1314                 }
1315
1316                 refcount_set(&recover->refs, 1);
1317                 recover->bbio = bbio;
1318                 recover->map_length = mapped_length;
1319
1320                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1321
1322                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1323
1324                 for (mirror_index = 0; mirror_index < nmirrors;
1325                      mirror_index++) {
1326                         struct scrub_block *sblock;
1327                         struct scrub_page *page;
1328
1329                         sblock = sblocks_for_recheck + mirror_index;
1330                         sblock->sctx = sctx;
1331
1332                         page = kzalloc(sizeof(*page), GFP_NOFS);
1333                         if (!page) {
1334 leave_nomem:
1335                                 spin_lock(&sctx->stat_lock);
1336                                 sctx->stat.malloc_errors++;
1337                                 spin_unlock(&sctx->stat_lock);
1338                                 scrub_put_recover(fs_info, recover);
1339                                 return -ENOMEM;
1340                         }
1341                         scrub_page_get(page);
1342                         sblock->pagev[page_index] = page;
1343                         page->sblock = sblock;
1344                         page->flags = flags;
1345                         page->generation = generation;
1346                         page->logical = logical;
1347                         page->have_csum = have_csum;
1348                         if (have_csum)
1349                                 memcpy(page->csum,
1350                                        original_sblock->pagev[0]->csum,
1351                                        sctx->csum_size);
1352
1353                         scrub_stripe_index_and_offset(logical,
1354                                                       bbio->map_type,
1355                                                       bbio->raid_map,
1356                                                       mapped_length,
1357                                                       bbio->num_stripes -
1358                                                       bbio->num_tgtdevs,
1359                                                       mirror_index,
1360                                                       &stripe_index,
1361                                                       &stripe_offset);
1362                         page->physical = bbio->stripes[stripe_index].physical +
1363                                          stripe_offset;
1364                         page->dev = bbio->stripes[stripe_index].dev;
1365
1366                         BUG_ON(page_index >= original_sblock->page_count);
1367                         page->physical_for_dev_replace =
1368                                 original_sblock->pagev[page_index]->
1369                                 physical_for_dev_replace;
1370                         /* for missing devices, dev->bdev is NULL */
1371                         page->mirror_num = mirror_index + 1;
1372                         sblock->page_count++;
1373                         page->page = alloc_page(GFP_NOFS);
1374                         if (!page->page)
1375                                 goto leave_nomem;
1376
1377                         scrub_get_recover(recover);
1378                         page->recover = recover;
1379                 }
1380                 scrub_put_recover(fs_info, recover);
1381                 length -= sublen;
1382                 logical += sublen;
1383                 page_index++;
1384         }
1385
1386         return 0;
1387 }
1388
1389 static void scrub_bio_wait_endio(struct bio *bio)
1390 {
1391         complete(bio->bi_private);
1392 }
1393
1394 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1395                                         struct bio *bio,
1396                                         struct scrub_page *page)
1397 {
1398         DECLARE_COMPLETION_ONSTACK(done);
1399         int ret;
1400         int mirror_num;
1401
1402         bio->bi_iter.bi_sector = page->logical >> 9;
1403         bio->bi_private = &done;
1404         bio->bi_end_io = scrub_bio_wait_endio;
1405
1406         mirror_num = page->sblock->pagev[0]->mirror_num;
1407         ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1408                                     page->recover->map_length,
1409                                     mirror_num, 0);
1410         if (ret)
1411                 return ret;
1412
1413         wait_for_completion_io(&done);
1414         return blk_status_to_errno(bio->bi_status);
1415 }
1416
1417 static void scrub_recheck_block_on_raid56(struct btrfs_fs_info *fs_info,
1418                                           struct scrub_block *sblock)
1419 {
1420         struct scrub_page *first_page = sblock->pagev[0];
1421         struct bio *bio;
1422         int page_num;
1423
1424         /* All pages in sblock belong to the same stripe on the same device. */
1425         ASSERT(first_page->dev);
1426         if (!first_page->dev->bdev)
1427                 goto out;
1428
1429         bio = btrfs_io_bio_alloc(BIO_MAX_PAGES);
1430         bio_set_dev(bio, first_page->dev->bdev);
1431
1432         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1433                 struct scrub_page *page = sblock->pagev[page_num];
1434
1435                 WARN_ON(!page->page);
1436                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1437         }
1438
1439         if (scrub_submit_raid56_bio_wait(fs_info, bio, first_page)) {
1440                 bio_put(bio);
1441                 goto out;
1442         }
1443
1444         bio_put(bio);
1445
1446         scrub_recheck_block_checksum(sblock);
1447
1448         return;
1449 out:
1450         for (page_num = 0; page_num < sblock->page_count; page_num++)
1451                 sblock->pagev[page_num]->io_error = 1;
1452
1453         sblock->no_io_error_seen = 0;
1454 }
1455
1456 /*
1457  * this function will check the on disk data for checksum errors, header
1458  * errors and read I/O errors. If any I/O errors happen, the exact pages
1459  * which are errored are marked as being bad. The goal is to enable scrub
1460  * to take those pages that are not errored from all the mirrors so that
1461  * the pages that are errored in the just handled mirror can be repaired.
1462  */
1463 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1464                                 struct scrub_block *sblock,
1465                                 int retry_failed_mirror)
1466 {
1467         int page_num;
1468
1469         sblock->no_io_error_seen = 1;
1470
1471         /* short cut for raid56 */
1472         if (!retry_failed_mirror && scrub_is_page_on_raid56(sblock->pagev[0]))
1473                 return scrub_recheck_block_on_raid56(fs_info, sblock);
1474
1475         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1476                 struct bio *bio;
1477                 struct scrub_page *page = sblock->pagev[page_num];
1478
1479                 if (page->dev->bdev == NULL) {
1480                         page->io_error = 1;
1481                         sblock->no_io_error_seen = 0;
1482                         continue;
1483                 }
1484
1485                 WARN_ON(!page->page);
1486                 bio = btrfs_io_bio_alloc(1);
1487                 bio_set_dev(bio, page->dev->bdev);
1488
1489                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1490                 bio->bi_iter.bi_sector = page->physical >> 9;
1491                 bio->bi_opf = REQ_OP_READ;
1492
1493                 if (btrfsic_submit_bio_wait(bio)) {
1494                         page->io_error = 1;
1495                         sblock->no_io_error_seen = 0;
1496                 }
1497
1498                 bio_put(bio);
1499         }
1500
1501         if (sblock->no_io_error_seen)
1502                 scrub_recheck_block_checksum(sblock);
1503 }
1504
1505 static inline int scrub_check_fsid(u8 fsid[],
1506                                    struct scrub_page *spage)
1507 {
1508         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1509         int ret;
1510
1511         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1512         return !ret;
1513 }
1514
1515 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1516 {
1517         sblock->header_error = 0;
1518         sblock->checksum_error = 0;
1519         sblock->generation_error = 0;
1520
1521         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1522                 scrub_checksum_data(sblock);
1523         else
1524                 scrub_checksum_tree_block(sblock);
1525 }
1526
1527 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1528                                              struct scrub_block *sblock_good)
1529 {
1530         int page_num;
1531         int ret = 0;
1532
1533         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1534                 int ret_sub;
1535
1536                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1537                                                            sblock_good,
1538                                                            page_num, 1);
1539                 if (ret_sub)
1540                         ret = ret_sub;
1541         }
1542
1543         return ret;
1544 }
1545
1546 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1547                                             struct scrub_block *sblock_good,
1548                                             int page_num, int force_write)
1549 {
1550         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1551         struct scrub_page *page_good = sblock_good->pagev[page_num];
1552         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1553
1554         BUG_ON(page_bad->page == NULL);
1555         BUG_ON(page_good->page == NULL);
1556         if (force_write || sblock_bad->header_error ||
1557             sblock_bad->checksum_error || page_bad->io_error) {
1558                 struct bio *bio;
1559                 int ret;
1560
1561                 if (!page_bad->dev->bdev) {
1562                         btrfs_warn_rl(fs_info,
1563                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1564                         return -EIO;
1565                 }
1566
1567                 bio = btrfs_io_bio_alloc(1);
1568                 bio_set_dev(bio, page_bad->dev->bdev);
1569                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1570                 bio->bi_opf = REQ_OP_WRITE;
1571
1572                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1573                 if (PAGE_SIZE != ret) {
1574                         bio_put(bio);
1575                         return -EIO;
1576                 }
1577
1578                 if (btrfsic_submit_bio_wait(bio)) {
1579                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1580                                 BTRFS_DEV_STAT_WRITE_ERRS);
1581                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1582                         bio_put(bio);
1583                         return -EIO;
1584                 }
1585                 bio_put(bio);
1586         }
1587
1588         return 0;
1589 }
1590
1591 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1592 {
1593         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1594         int page_num;
1595
1596         /*
1597          * This block is used for the check of the parity on the source device,
1598          * so the data needn't be written into the destination device.
1599          */
1600         if (sblock->sparity)
1601                 return;
1602
1603         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1604                 int ret;
1605
1606                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1607                 if (ret)
1608                         atomic64_inc(&fs_info->dev_replace.num_write_errors);
1609         }
1610 }
1611
1612 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1613                                            int page_num)
1614 {
1615         struct scrub_page *spage = sblock->pagev[page_num];
1616
1617         BUG_ON(spage->page == NULL);
1618         if (spage->io_error) {
1619                 void *mapped_buffer = kmap_atomic(spage->page);
1620
1621                 clear_page(mapped_buffer);
1622                 flush_dcache_page(spage->page);
1623                 kunmap_atomic(mapped_buffer);
1624         }
1625         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1626 }
1627
1628 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1629                                     struct scrub_page *spage)
1630 {
1631         struct scrub_bio *sbio;
1632         int ret;
1633
1634         mutex_lock(&sctx->wr_lock);
1635 again:
1636         if (!sctx->wr_curr_bio) {
1637                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1638                                               GFP_KERNEL);
1639                 if (!sctx->wr_curr_bio) {
1640                         mutex_unlock(&sctx->wr_lock);
1641                         return -ENOMEM;
1642                 }
1643                 sctx->wr_curr_bio->sctx = sctx;
1644                 sctx->wr_curr_bio->page_count = 0;
1645         }
1646         sbio = sctx->wr_curr_bio;
1647         if (sbio->page_count == 0) {
1648                 struct bio *bio;
1649
1650                 sbio->physical = spage->physical_for_dev_replace;
1651                 sbio->logical = spage->logical;
1652                 sbio->dev = sctx->wr_tgtdev;
1653                 bio = sbio->bio;
1654                 if (!bio) {
1655                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1656                         sbio->bio = bio;
1657                 }
1658
1659                 bio->bi_private = sbio;
1660                 bio->bi_end_io = scrub_wr_bio_end_io;
1661                 bio_set_dev(bio, sbio->dev->bdev);
1662                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1663                 bio->bi_opf = REQ_OP_WRITE;
1664                 sbio->status = 0;
1665         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1666                    spage->physical_for_dev_replace ||
1667                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1668                    spage->logical) {
1669                 scrub_wr_submit(sctx);
1670                 goto again;
1671         }
1672
1673         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1674         if (ret != PAGE_SIZE) {
1675                 if (sbio->page_count < 1) {
1676                         bio_put(sbio->bio);
1677                         sbio->bio = NULL;
1678                         mutex_unlock(&sctx->wr_lock);
1679                         return -EIO;
1680                 }
1681                 scrub_wr_submit(sctx);
1682                 goto again;
1683         }
1684
1685         sbio->pagev[sbio->page_count] = spage;
1686         scrub_page_get(spage);
1687         sbio->page_count++;
1688         if (sbio->page_count == sctx->pages_per_wr_bio)
1689                 scrub_wr_submit(sctx);
1690         mutex_unlock(&sctx->wr_lock);
1691
1692         return 0;
1693 }
1694
1695 static void scrub_wr_submit(struct scrub_ctx *sctx)
1696 {
1697         struct scrub_bio *sbio;
1698
1699         if (!sctx->wr_curr_bio)
1700                 return;
1701
1702         sbio = sctx->wr_curr_bio;
1703         sctx->wr_curr_bio = NULL;
1704         WARN_ON(!sbio->bio->bi_disk);
1705         scrub_pending_bio_inc(sctx);
1706         /* process all writes in a single worker thread. Then the block layer
1707          * orders the requests before sending them to the driver which
1708          * doubled the write performance on spinning disks when measured
1709          * with Linux 3.5 */
1710         btrfsic_submit_bio(sbio->bio);
1711 }
1712
1713 static void scrub_wr_bio_end_io(struct bio *bio)
1714 {
1715         struct scrub_bio *sbio = bio->bi_private;
1716         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
1717
1718         sbio->status = bio->bi_status;
1719         sbio->bio = bio;
1720
1721         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1722                          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         u8 csum[BTRFS_CSUM_SIZE];
1791         u8 *on_disk_csum;
1792         struct page *page;
1793         void *buffer;
1794         u32 crc = ~(u32)0;
1795         u64 len;
1796         int index;
1797
1798         BUG_ON(sblock->page_count < 1);
1799         if (!sblock->pagev[0]->have_csum)
1800                 return 0;
1801
1802         on_disk_csum = sblock->pagev[0]->csum;
1803         page = sblock->pagev[0]->page;
1804         buffer = kmap_atomic(page);
1805
1806         len = sctx->fs_info->sectorsize;
1807         index = 0;
1808         for (;;) {
1809                 u64 l = min_t(u64, len, PAGE_SIZE);
1810
1811                 crc = btrfs_csum_data(buffer, crc, l);
1812                 kunmap_atomic(buffer);
1813                 len -= l;
1814                 if (len == 0)
1815                         break;
1816                 index++;
1817                 BUG_ON(index >= sblock->page_count);
1818                 BUG_ON(!sblock->pagev[index]->page);
1819                 page = sblock->pagev[index]->page;
1820                 buffer = kmap_atomic(page);
1821         }
1822
1823         btrfs_csum_final(crc, csum);
1824         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1825                 sblock->checksum_error = 1;
1826
1827         return sblock->checksum_error;
1828 }
1829
1830 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1831 {
1832         struct scrub_ctx *sctx = sblock->sctx;
1833         struct btrfs_header *h;
1834         struct btrfs_fs_info *fs_info = sctx->fs_info;
1835         u8 calculated_csum[BTRFS_CSUM_SIZE];
1836         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1837         struct page *page;
1838         void *mapped_buffer;
1839         u64 mapped_size;
1840         void *p;
1841         u32 crc = ~(u32)0;
1842         u64 len;
1843         int index;
1844
1845         BUG_ON(sblock->page_count < 1);
1846         page = sblock->pagev[0]->page;
1847         mapped_buffer = kmap_atomic(page);
1848         h = (struct btrfs_header *)mapped_buffer;
1849         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1850
1851         /*
1852          * we don't use the getter functions here, as we
1853          * a) don't have an extent buffer and
1854          * b) the page is already kmapped
1855          */
1856         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1857                 sblock->header_error = 1;
1858
1859         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1860                 sblock->header_error = 1;
1861                 sblock->generation_error = 1;
1862         }
1863
1864         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1865                 sblock->header_error = 1;
1866
1867         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1868                    BTRFS_UUID_SIZE))
1869                 sblock->header_error = 1;
1870
1871         len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
1872         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1873         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1874         index = 0;
1875         for (;;) {
1876                 u64 l = min_t(u64, len, mapped_size);
1877
1878                 crc = btrfs_csum_data(p, crc, l);
1879                 kunmap_atomic(mapped_buffer);
1880                 len -= l;
1881                 if (len == 0)
1882                         break;
1883                 index++;
1884                 BUG_ON(index >= sblock->page_count);
1885                 BUG_ON(!sblock->pagev[index]->page);
1886                 page = sblock->pagev[index]->page;
1887                 mapped_buffer = kmap_atomic(page);
1888                 mapped_size = PAGE_SIZE;
1889                 p = mapped_buffer;
1890         }
1891
1892         btrfs_csum_final(crc, calculated_csum);
1893         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1894                 sblock->checksum_error = 1;
1895
1896         return sblock->header_error || sblock->checksum_error;
1897 }
1898
1899 static int scrub_checksum_super(struct scrub_block *sblock)
1900 {
1901         struct btrfs_super_block *s;
1902         struct scrub_ctx *sctx = sblock->sctx;
1903         u8 calculated_csum[BTRFS_CSUM_SIZE];
1904         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1905         struct page *page;
1906         void *mapped_buffer;
1907         u64 mapped_size;
1908         void *p;
1909         u32 crc = ~(u32)0;
1910         int fail_gen = 0;
1911         int fail_cor = 0;
1912         u64 len;
1913         int index;
1914
1915         BUG_ON(sblock->page_count < 1);
1916         page = sblock->pagev[0]->page;
1917         mapped_buffer = kmap_atomic(page);
1918         s = (struct btrfs_super_block *)mapped_buffer;
1919         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1920
1921         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1922                 ++fail_cor;
1923
1924         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1925                 ++fail_gen;
1926
1927         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1928                 ++fail_cor;
1929
1930         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1931         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1932         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1933         index = 0;
1934         for (;;) {
1935                 u64 l = min_t(u64, len, mapped_size);
1936
1937                 crc = btrfs_csum_data(p, crc, l);
1938                 kunmap_atomic(mapped_buffer);
1939                 len -= l;
1940                 if (len == 0)
1941                         break;
1942                 index++;
1943                 BUG_ON(index >= sblock->page_count);
1944                 BUG_ON(!sblock->pagev[index]->page);
1945                 page = sblock->pagev[index]->page;
1946                 mapped_buffer = kmap_atomic(page);
1947                 mapped_size = PAGE_SIZE;
1948                 p = mapped_buffer;
1949         }
1950
1951         btrfs_csum_final(crc, calculated_csum);
1952         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1953                 ++fail_cor;
1954
1955         if (fail_cor + fail_gen) {
1956                 /*
1957                  * if we find an error in a super block, we just report it.
1958                  * They will get written with the next transaction commit
1959                  * anyway
1960                  */
1961                 spin_lock(&sctx->stat_lock);
1962                 ++sctx->stat.super_errors;
1963                 spin_unlock(&sctx->stat_lock);
1964                 if (fail_cor)
1965                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1966                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1967                 else
1968                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1969                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1970         }
1971
1972         return fail_cor + fail_gen;
1973 }
1974
1975 static void scrub_block_get(struct scrub_block *sblock)
1976 {
1977         refcount_inc(&sblock->refs);
1978 }
1979
1980 static void scrub_block_put(struct scrub_block *sblock)
1981 {
1982         if (refcount_dec_and_test(&sblock->refs)) {
1983                 int i;
1984
1985                 if (sblock->sparity)
1986                         scrub_parity_put(sblock->sparity);
1987
1988                 for (i = 0; i < sblock->page_count; i++)
1989                         scrub_page_put(sblock->pagev[i]);
1990                 kfree(sblock);
1991         }
1992 }
1993
1994 static void scrub_page_get(struct scrub_page *spage)
1995 {
1996         atomic_inc(&spage->refs);
1997 }
1998
1999 static void scrub_page_put(struct scrub_page *spage)
2000 {
2001         if (atomic_dec_and_test(&spage->refs)) {
2002                 if (spage->page)
2003                         __free_page(spage->page);
2004                 kfree(spage);
2005         }
2006 }
2007
2008 static void scrub_submit(struct scrub_ctx *sctx)
2009 {
2010         struct scrub_bio *sbio;
2011
2012         if (sctx->curr == -1)
2013                 return;
2014
2015         sbio = sctx->bios[sctx->curr];
2016         sctx->curr = -1;
2017         scrub_pending_bio_inc(sctx);
2018         btrfsic_submit_bio(sbio->bio);
2019 }
2020
2021 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2022                                     struct scrub_page *spage)
2023 {
2024         struct scrub_block *sblock = spage->sblock;
2025         struct scrub_bio *sbio;
2026         int ret;
2027
2028 again:
2029         /*
2030          * grab a fresh bio or wait for one to become available
2031          */
2032         while (sctx->curr == -1) {
2033                 spin_lock(&sctx->list_lock);
2034                 sctx->curr = sctx->first_free;
2035                 if (sctx->curr != -1) {
2036                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2037                         sctx->bios[sctx->curr]->next_free = -1;
2038                         sctx->bios[sctx->curr]->page_count = 0;
2039                         spin_unlock(&sctx->list_lock);
2040                 } else {
2041                         spin_unlock(&sctx->list_lock);
2042                         wait_event(sctx->list_wait, sctx->first_free != -1);
2043                 }
2044         }
2045         sbio = sctx->bios[sctx->curr];
2046         if (sbio->page_count == 0) {
2047                 struct bio *bio;
2048
2049                 sbio->physical = spage->physical;
2050                 sbio->logical = spage->logical;
2051                 sbio->dev = spage->dev;
2052                 bio = sbio->bio;
2053                 if (!bio) {
2054                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2055                         sbio->bio = bio;
2056                 }
2057
2058                 bio->bi_private = sbio;
2059                 bio->bi_end_io = scrub_bio_end_io;
2060                 bio_set_dev(bio, sbio->dev->bdev);
2061                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2062                 bio->bi_opf = REQ_OP_READ;
2063                 sbio->status = 0;
2064         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2065                    spage->physical ||
2066                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2067                    spage->logical ||
2068                    sbio->dev != spage->dev) {
2069                 scrub_submit(sctx);
2070                 goto again;
2071         }
2072
2073         sbio->pagev[sbio->page_count] = spage;
2074         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2075         if (ret != PAGE_SIZE) {
2076                 if (sbio->page_count < 1) {
2077                         bio_put(sbio->bio);
2078                         sbio->bio = NULL;
2079                         return -EIO;
2080                 }
2081                 scrub_submit(sctx);
2082                 goto again;
2083         }
2084
2085         scrub_block_get(sblock); /* one for the page added to the bio */
2086         atomic_inc(&sblock->outstanding_pages);
2087         sbio->page_count++;
2088         if (sbio->page_count == sctx->pages_per_rd_bio)
2089                 scrub_submit(sctx);
2090
2091         return 0;
2092 }
2093
2094 static void scrub_missing_raid56_end_io(struct bio *bio)
2095 {
2096         struct scrub_block *sblock = bio->bi_private;
2097         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2098
2099         if (bio->bi_status)
2100                 sblock->no_io_error_seen = 0;
2101
2102         bio_put(bio);
2103
2104         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2105 }
2106
2107 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2108 {
2109         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2110         struct scrub_ctx *sctx = sblock->sctx;
2111         struct btrfs_fs_info *fs_info = sctx->fs_info;
2112         u64 logical;
2113         struct btrfs_device *dev;
2114
2115         logical = sblock->pagev[0]->logical;
2116         dev = sblock->pagev[0]->dev;
2117
2118         if (sblock->no_io_error_seen)
2119                 scrub_recheck_block_checksum(sblock);
2120
2121         if (!sblock->no_io_error_seen) {
2122                 spin_lock(&sctx->stat_lock);
2123                 sctx->stat.read_errors++;
2124                 spin_unlock(&sctx->stat_lock);
2125                 btrfs_err_rl_in_rcu(fs_info,
2126                         "IO error rebuilding logical %llu for dev %s",
2127                         logical, rcu_str_deref(dev->name));
2128         } else if (sblock->header_error || sblock->checksum_error) {
2129                 spin_lock(&sctx->stat_lock);
2130                 sctx->stat.uncorrectable_errors++;
2131                 spin_unlock(&sctx->stat_lock);
2132                 btrfs_err_rl_in_rcu(fs_info,
2133                         "failed to rebuild valid logical %llu for dev %s",
2134                         logical, rcu_str_deref(dev->name));
2135         } else {
2136                 scrub_write_block_to_dev_replace(sblock);
2137         }
2138
2139         scrub_block_put(sblock);
2140
2141         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2142                 mutex_lock(&sctx->wr_lock);
2143                 scrub_wr_submit(sctx);
2144                 mutex_unlock(&sctx->wr_lock);
2145         }
2146
2147         scrub_pending_bio_dec(sctx);
2148 }
2149
2150 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2151 {
2152         struct scrub_ctx *sctx = sblock->sctx;
2153         struct btrfs_fs_info *fs_info = sctx->fs_info;
2154         u64 length = sblock->page_count * PAGE_SIZE;
2155         u64 logical = sblock->pagev[0]->logical;
2156         struct btrfs_bio *bbio = NULL;
2157         struct bio *bio;
2158         struct btrfs_raid_bio *rbio;
2159         int ret;
2160         int i;
2161
2162         btrfs_bio_counter_inc_blocked(fs_info);
2163         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2164                         &length, &bbio);
2165         if (ret || !bbio || !bbio->raid_map)
2166                 goto bbio_out;
2167
2168         if (WARN_ON(!sctx->is_dev_replace ||
2169                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2170                 /*
2171                  * We shouldn't be scrubbing a missing device. Even for dev
2172                  * replace, we should only get here for RAID 5/6. We either
2173                  * managed to mount something with no mirrors remaining or
2174                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2175                  */
2176                 goto bbio_out;
2177         }
2178
2179         bio = btrfs_io_bio_alloc(0);
2180         bio->bi_iter.bi_sector = logical >> 9;
2181         bio->bi_private = sblock;
2182         bio->bi_end_io = scrub_missing_raid56_end_io;
2183
2184         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2185         if (!rbio)
2186                 goto rbio_out;
2187
2188         for (i = 0; i < sblock->page_count; i++) {
2189                 struct scrub_page *spage = sblock->pagev[i];
2190
2191                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2192         }
2193
2194         btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2195                         scrub_missing_raid56_worker, NULL, NULL);
2196         scrub_block_get(sblock);
2197         scrub_pending_bio_inc(sctx);
2198         raid56_submit_missing_rbio(rbio);
2199         return;
2200
2201 rbio_out:
2202         bio_put(bio);
2203 bbio_out:
2204         btrfs_bio_counter_dec(fs_info);
2205         btrfs_put_bbio(bbio);
2206         spin_lock(&sctx->stat_lock);
2207         sctx->stat.malloc_errors++;
2208         spin_unlock(&sctx->stat_lock);
2209 }
2210
2211 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2212                        u64 physical, struct btrfs_device *dev, u64 flags,
2213                        u64 gen, int mirror_num, u8 *csum, int force,
2214                        u64 physical_for_dev_replace)
2215 {
2216         struct scrub_block *sblock;
2217         int index;
2218
2219         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2220         if (!sblock) {
2221                 spin_lock(&sctx->stat_lock);
2222                 sctx->stat.malloc_errors++;
2223                 spin_unlock(&sctx->stat_lock);
2224                 return -ENOMEM;
2225         }
2226
2227         /* one ref inside this function, plus one for each page added to
2228          * a bio later on */
2229         refcount_set(&sblock->refs, 1);
2230         sblock->sctx = sctx;
2231         sblock->no_io_error_seen = 1;
2232
2233         for (index = 0; len > 0; index++) {
2234                 struct scrub_page *spage;
2235                 u64 l = min_t(u64, len, PAGE_SIZE);
2236
2237                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2238                 if (!spage) {
2239 leave_nomem:
2240                         spin_lock(&sctx->stat_lock);
2241                         sctx->stat.malloc_errors++;
2242                         spin_unlock(&sctx->stat_lock);
2243                         scrub_block_put(sblock);
2244                         return -ENOMEM;
2245                 }
2246                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2247                 scrub_page_get(spage);
2248                 sblock->pagev[index] = spage;
2249                 spage->sblock = sblock;
2250                 spage->dev = dev;
2251                 spage->flags = flags;
2252                 spage->generation = gen;
2253                 spage->logical = logical;
2254                 spage->physical = physical;
2255                 spage->physical_for_dev_replace = physical_for_dev_replace;
2256                 spage->mirror_num = mirror_num;
2257                 if (csum) {
2258                         spage->have_csum = 1;
2259                         memcpy(spage->csum, csum, sctx->csum_size);
2260                 } else {
2261                         spage->have_csum = 0;
2262                 }
2263                 sblock->page_count++;
2264                 spage->page = alloc_page(GFP_KERNEL);
2265                 if (!spage->page)
2266                         goto leave_nomem;
2267                 len -= l;
2268                 logical += l;
2269                 physical += l;
2270                 physical_for_dev_replace += l;
2271         }
2272
2273         WARN_ON(sblock->page_count == 0);
2274         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2275                 /*
2276                  * This case should only be hit for RAID 5/6 device replace. See
2277                  * the comment in scrub_missing_raid56_pages() for details.
2278                  */
2279                 scrub_missing_raid56_pages(sblock);
2280         } else {
2281                 for (index = 0; index < sblock->page_count; index++) {
2282                         struct scrub_page *spage = sblock->pagev[index];
2283                         int ret;
2284
2285                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2286                         if (ret) {
2287                                 scrub_block_put(sblock);
2288                                 return ret;
2289                         }
2290                 }
2291
2292                 if (force)
2293                         scrub_submit(sctx);
2294         }
2295
2296         /* last one frees, either here or in bio completion for last page */
2297         scrub_block_put(sblock);
2298         return 0;
2299 }
2300
2301 static void scrub_bio_end_io(struct bio *bio)
2302 {
2303         struct scrub_bio *sbio = bio->bi_private;
2304         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2305
2306         sbio->status = bio->bi_status;
2307         sbio->bio = bio;
2308
2309         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2310 }
2311
2312 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2313 {
2314         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2315         struct scrub_ctx *sctx = sbio->sctx;
2316         int i;
2317
2318         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2319         if (sbio->status) {
2320                 for (i = 0; i < sbio->page_count; i++) {
2321                         struct scrub_page *spage = sbio->pagev[i];
2322
2323                         spage->io_error = 1;
2324                         spage->sblock->no_io_error_seen = 0;
2325                 }
2326         }
2327
2328         /* now complete the scrub_block items that have all pages completed */
2329         for (i = 0; i < sbio->page_count; i++) {
2330                 struct scrub_page *spage = sbio->pagev[i];
2331                 struct scrub_block *sblock = spage->sblock;
2332
2333                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2334                         scrub_block_complete(sblock);
2335                 scrub_block_put(sblock);
2336         }
2337
2338         bio_put(sbio->bio);
2339         sbio->bio = NULL;
2340         spin_lock(&sctx->list_lock);
2341         sbio->next_free = sctx->first_free;
2342         sctx->first_free = sbio->index;
2343         spin_unlock(&sctx->list_lock);
2344
2345         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2346                 mutex_lock(&sctx->wr_lock);
2347                 scrub_wr_submit(sctx);
2348                 mutex_unlock(&sctx->wr_lock);
2349         }
2350
2351         scrub_pending_bio_dec(sctx);
2352 }
2353
2354 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2355                                        unsigned long *bitmap,
2356                                        u64 start, u64 len)
2357 {
2358         u64 offset;
2359         u64 nsectors64;
2360         u32 nsectors;
2361         int sectorsize = sparity->sctx->fs_info->sectorsize;
2362
2363         if (len >= sparity->stripe_len) {
2364                 bitmap_set(bitmap, 0, sparity->nsectors);
2365                 return;
2366         }
2367
2368         start -= sparity->logic_start;
2369         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2370         offset = div_u64(offset, sectorsize);
2371         nsectors64 = div_u64(len, sectorsize);
2372
2373         ASSERT(nsectors64 < UINT_MAX);
2374         nsectors = (u32)nsectors64;
2375
2376         if (offset + nsectors <= sparity->nsectors) {
2377                 bitmap_set(bitmap, offset, nsectors);
2378                 return;
2379         }
2380
2381         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2382         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2383 }
2384
2385 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2386                                                    u64 start, u64 len)
2387 {
2388         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2389 }
2390
2391 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2392                                                   u64 start, u64 len)
2393 {
2394         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2395 }
2396
2397 static void scrub_block_complete(struct scrub_block *sblock)
2398 {
2399         int corrupted = 0;
2400
2401         if (!sblock->no_io_error_seen) {
2402                 corrupted = 1;
2403                 scrub_handle_errored_block(sblock);
2404         } else {
2405                 /*
2406                  * if has checksum error, write via repair mechanism in
2407                  * dev replace case, otherwise write here in dev replace
2408                  * case.
2409                  */
2410                 corrupted = scrub_checksum(sblock);
2411                 if (!corrupted && sblock->sctx->is_dev_replace)
2412                         scrub_write_block_to_dev_replace(sblock);
2413         }
2414
2415         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2416                 u64 start = sblock->pagev[0]->logical;
2417                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2418                           PAGE_SIZE;
2419
2420                 scrub_parity_mark_sectors_error(sblock->sparity,
2421                                                 start, end - start);
2422         }
2423 }
2424
2425 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2426 {
2427         struct btrfs_ordered_sum *sum = NULL;
2428         unsigned long index;
2429         unsigned long num_sectors;
2430
2431         while (!list_empty(&sctx->csum_list)) {
2432                 sum = list_first_entry(&sctx->csum_list,
2433                                        struct btrfs_ordered_sum, list);
2434                 if (sum->bytenr > logical)
2435                         return 0;
2436                 if (sum->bytenr + sum->len > logical)
2437                         break;
2438
2439                 ++sctx->stat.csum_discards;
2440                 list_del(&sum->list);
2441                 kfree(sum);
2442                 sum = NULL;
2443         }
2444         if (!sum)
2445                 return 0;
2446
2447         index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2448         ASSERT(index < UINT_MAX);
2449
2450         num_sectors = sum->len / sctx->fs_info->sectorsize;
2451         memcpy(csum, sum->sums + index, sctx->csum_size);
2452         if (index == num_sectors - 1) {
2453                 list_del(&sum->list);
2454                 kfree(sum);
2455         }
2456         return 1;
2457 }
2458
2459 /* scrub extent tries to collect up to 64 kB for each bio */
2460 static int scrub_extent(struct scrub_ctx *sctx, struct map_lookup *map,
2461                         u64 logical, u64 len,
2462                         u64 physical, struct btrfs_device *dev, u64 flags,
2463                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2464 {
2465         int ret;
2466         u8 csum[BTRFS_CSUM_SIZE];
2467         u32 blocksize;
2468
2469         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2470                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2471                         blocksize = map->stripe_len;
2472                 else
2473                         blocksize = sctx->fs_info->sectorsize;
2474                 spin_lock(&sctx->stat_lock);
2475                 sctx->stat.data_extents_scrubbed++;
2476                 sctx->stat.data_bytes_scrubbed += len;
2477                 spin_unlock(&sctx->stat_lock);
2478         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2479                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2480                         blocksize = map->stripe_len;
2481                 else
2482                         blocksize = sctx->fs_info->nodesize;
2483                 spin_lock(&sctx->stat_lock);
2484                 sctx->stat.tree_extents_scrubbed++;
2485                 sctx->stat.tree_bytes_scrubbed += len;
2486                 spin_unlock(&sctx->stat_lock);
2487         } else {
2488                 blocksize = sctx->fs_info->sectorsize;
2489                 WARN_ON(1);
2490         }
2491
2492         while (len) {
2493                 u64 l = min_t(u64, len, blocksize);
2494                 int have_csum = 0;
2495
2496                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2497                         /* push csums to sbio */
2498                         have_csum = scrub_find_csum(sctx, logical, csum);
2499                         if (have_csum == 0)
2500                                 ++sctx->stat.no_csum;
2501                 }
2502                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2503                                   mirror_num, have_csum ? csum : NULL, 0,
2504                                   physical_for_dev_replace);
2505                 if (ret)
2506                         return ret;
2507                 len -= l;
2508                 logical += l;
2509                 physical += l;
2510                 physical_for_dev_replace += l;
2511         }
2512         return 0;
2513 }
2514
2515 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2516                                   u64 logical, u64 len,
2517                                   u64 physical, struct btrfs_device *dev,
2518                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2519 {
2520         struct scrub_ctx *sctx = sparity->sctx;
2521         struct scrub_block *sblock;
2522         int index;
2523
2524         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2525         if (!sblock) {
2526                 spin_lock(&sctx->stat_lock);
2527                 sctx->stat.malloc_errors++;
2528                 spin_unlock(&sctx->stat_lock);
2529                 return -ENOMEM;
2530         }
2531
2532         /* one ref inside this function, plus one for each page added to
2533          * a bio later on */
2534         refcount_set(&sblock->refs, 1);
2535         sblock->sctx = sctx;
2536         sblock->no_io_error_seen = 1;
2537         sblock->sparity = sparity;
2538         scrub_parity_get(sparity);
2539
2540         for (index = 0; len > 0; index++) {
2541                 struct scrub_page *spage;
2542                 u64 l = min_t(u64, len, PAGE_SIZE);
2543
2544                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2545                 if (!spage) {
2546 leave_nomem:
2547                         spin_lock(&sctx->stat_lock);
2548                         sctx->stat.malloc_errors++;
2549                         spin_unlock(&sctx->stat_lock);
2550                         scrub_block_put(sblock);
2551                         return -ENOMEM;
2552                 }
2553                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2554                 /* For scrub block */
2555                 scrub_page_get(spage);
2556                 sblock->pagev[index] = spage;
2557                 /* For scrub parity */
2558                 scrub_page_get(spage);
2559                 list_add_tail(&spage->list, &sparity->spages);
2560                 spage->sblock = sblock;
2561                 spage->dev = dev;
2562                 spage->flags = flags;
2563                 spage->generation = gen;
2564                 spage->logical = logical;
2565                 spage->physical = physical;
2566                 spage->mirror_num = mirror_num;
2567                 if (csum) {
2568                         spage->have_csum = 1;
2569                         memcpy(spage->csum, csum, sctx->csum_size);
2570                 } else {
2571                         spage->have_csum = 0;
2572                 }
2573                 sblock->page_count++;
2574                 spage->page = alloc_page(GFP_KERNEL);
2575                 if (!spage->page)
2576                         goto leave_nomem;
2577                 len -= l;
2578                 logical += l;
2579                 physical += l;
2580         }
2581
2582         WARN_ON(sblock->page_count == 0);
2583         for (index = 0; index < sblock->page_count; index++) {
2584                 struct scrub_page *spage = sblock->pagev[index];
2585                 int ret;
2586
2587                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2588                 if (ret) {
2589                         scrub_block_put(sblock);
2590                         return ret;
2591                 }
2592         }
2593
2594         /* last one frees, either here or in bio completion for last page */
2595         scrub_block_put(sblock);
2596         return 0;
2597 }
2598
2599 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2600                                    u64 logical, u64 len,
2601                                    u64 physical, struct btrfs_device *dev,
2602                                    u64 flags, u64 gen, int mirror_num)
2603 {
2604         struct scrub_ctx *sctx = sparity->sctx;
2605         int ret;
2606         u8 csum[BTRFS_CSUM_SIZE];
2607         u32 blocksize;
2608
2609         if (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state)) {
2610                 scrub_parity_mark_sectors_error(sparity, logical, len);
2611                 return 0;
2612         }
2613
2614         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2615                 blocksize = sparity->stripe_len;
2616         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2617                 blocksize = sparity->stripe_len;
2618         } else {
2619                 blocksize = sctx->fs_info->sectorsize;
2620                 WARN_ON(1);
2621         }
2622
2623         while (len) {
2624                 u64 l = min_t(u64, len, blocksize);
2625                 int have_csum = 0;
2626
2627                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2628                         /* push csums to sbio */
2629                         have_csum = scrub_find_csum(sctx, logical, csum);
2630                         if (have_csum == 0)
2631                                 goto skip;
2632                 }
2633                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2634                                              flags, gen, mirror_num,
2635                                              have_csum ? csum : NULL);
2636                 if (ret)
2637                         return ret;
2638 skip:
2639                 len -= l;
2640                 logical += l;
2641                 physical += l;
2642         }
2643         return 0;
2644 }
2645
2646 /*
2647  * Given a physical address, this will calculate it's
2648  * logical offset. if this is a parity stripe, it will return
2649  * the most left data stripe's logical offset.
2650  *
2651  * return 0 if it is a data stripe, 1 means parity stripe.
2652  */
2653 static int get_raid56_logic_offset(u64 physical, int num,
2654                                    struct map_lookup *map, u64 *offset,
2655                                    u64 *stripe_start)
2656 {
2657         int i;
2658         int j = 0;
2659         u64 stripe_nr;
2660         u64 last_offset;
2661         u32 stripe_index;
2662         u32 rot;
2663
2664         last_offset = (physical - map->stripes[num].physical) *
2665                       nr_data_stripes(map);
2666         if (stripe_start)
2667                 *stripe_start = last_offset;
2668
2669         *offset = last_offset;
2670         for (i = 0; i < nr_data_stripes(map); i++) {
2671                 *offset = last_offset + i * map->stripe_len;
2672
2673                 stripe_nr = div64_u64(*offset, map->stripe_len);
2674                 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2675
2676                 /* Work out the disk rotation on this stripe-set */
2677                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2678                 /* calculate which stripe this data locates */
2679                 rot += i;
2680                 stripe_index = rot % map->num_stripes;
2681                 if (stripe_index == num)
2682                         return 0;
2683                 if (stripe_index < num)
2684                         j++;
2685         }
2686         *offset = last_offset + j * map->stripe_len;
2687         return 1;
2688 }
2689
2690 static void scrub_free_parity(struct scrub_parity *sparity)
2691 {
2692         struct scrub_ctx *sctx = sparity->sctx;
2693         struct scrub_page *curr, *next;
2694         int nbits;
2695
2696         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2697         if (nbits) {
2698                 spin_lock(&sctx->stat_lock);
2699                 sctx->stat.read_errors += nbits;
2700                 sctx->stat.uncorrectable_errors += nbits;
2701                 spin_unlock(&sctx->stat_lock);
2702         }
2703
2704         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2705                 list_del_init(&curr->list);
2706                 scrub_page_put(curr);
2707         }
2708
2709         kfree(sparity);
2710 }
2711
2712 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2713 {
2714         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2715                                                     work);
2716         struct scrub_ctx *sctx = sparity->sctx;
2717
2718         scrub_free_parity(sparity);
2719         scrub_pending_bio_dec(sctx);
2720 }
2721
2722 static void scrub_parity_bio_endio(struct bio *bio)
2723 {
2724         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2725         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
2726
2727         if (bio->bi_status)
2728                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2729                           sparity->nsectors);
2730
2731         bio_put(bio);
2732
2733         btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2734                         scrub_parity_bio_endio_worker, NULL, NULL);
2735         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
2736 }
2737
2738 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2739 {
2740         struct scrub_ctx *sctx = sparity->sctx;
2741         struct btrfs_fs_info *fs_info = sctx->fs_info;
2742         struct bio *bio;
2743         struct btrfs_raid_bio *rbio;
2744         struct btrfs_bio *bbio = NULL;
2745         u64 length;
2746         int ret;
2747
2748         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2749                            sparity->nsectors))
2750                 goto out;
2751
2752         length = sparity->logic_end - sparity->logic_start;
2753
2754         btrfs_bio_counter_inc_blocked(fs_info);
2755         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
2756                                &length, &bbio);
2757         if (ret || !bbio || !bbio->raid_map)
2758                 goto bbio_out;
2759
2760         bio = btrfs_io_bio_alloc(0);
2761         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2762         bio->bi_private = sparity;
2763         bio->bi_end_io = scrub_parity_bio_endio;
2764
2765         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
2766                                               length, sparity->scrub_dev,
2767                                               sparity->dbitmap,
2768                                               sparity->nsectors);
2769         if (!rbio)
2770                 goto rbio_out;
2771
2772         scrub_pending_bio_inc(sctx);
2773         raid56_parity_submit_scrub_rbio(rbio);
2774         return;
2775
2776 rbio_out:
2777         bio_put(bio);
2778 bbio_out:
2779         btrfs_bio_counter_dec(fs_info);
2780         btrfs_put_bbio(bbio);
2781         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2782                   sparity->nsectors);
2783         spin_lock(&sctx->stat_lock);
2784         sctx->stat.malloc_errors++;
2785         spin_unlock(&sctx->stat_lock);
2786 out:
2787         scrub_free_parity(sparity);
2788 }
2789
2790 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2791 {
2792         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
2793 }
2794
2795 static void scrub_parity_get(struct scrub_parity *sparity)
2796 {
2797         refcount_inc(&sparity->refs);
2798 }
2799
2800 static void scrub_parity_put(struct scrub_parity *sparity)
2801 {
2802         if (!refcount_dec_and_test(&sparity->refs))
2803                 return;
2804
2805         scrub_parity_check_and_repair(sparity);
2806 }
2807
2808 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2809                                                   struct map_lookup *map,
2810                                                   struct btrfs_device *sdev,
2811                                                   struct btrfs_path *path,
2812                                                   u64 logic_start,
2813                                                   u64 logic_end)
2814 {
2815         struct btrfs_fs_info *fs_info = sctx->fs_info;
2816         struct btrfs_root *root = fs_info->extent_root;
2817         struct btrfs_root *csum_root = fs_info->csum_root;
2818         struct btrfs_extent_item *extent;
2819         struct btrfs_bio *bbio = NULL;
2820         u64 flags;
2821         int ret;
2822         int slot;
2823         struct extent_buffer *l;
2824         struct btrfs_key key;
2825         u64 generation;
2826         u64 extent_logical;
2827         u64 extent_physical;
2828         u64 extent_len;
2829         u64 mapped_length;
2830         struct btrfs_device *extent_dev;
2831         struct scrub_parity *sparity;
2832         int nsectors;
2833         int bitmap_len;
2834         int extent_mirror_num;
2835         int stop_loop = 0;
2836
2837         nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
2838         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2839         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2840                           GFP_NOFS);
2841         if (!sparity) {
2842                 spin_lock(&sctx->stat_lock);
2843                 sctx->stat.malloc_errors++;
2844                 spin_unlock(&sctx->stat_lock);
2845                 return -ENOMEM;
2846         }
2847
2848         sparity->stripe_len = map->stripe_len;
2849         sparity->nsectors = nsectors;
2850         sparity->sctx = sctx;
2851         sparity->scrub_dev = sdev;
2852         sparity->logic_start = logic_start;
2853         sparity->logic_end = logic_end;
2854         refcount_set(&sparity->refs, 1);
2855         INIT_LIST_HEAD(&sparity->spages);
2856         sparity->dbitmap = sparity->bitmap;
2857         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2858
2859         ret = 0;
2860         while (logic_start < logic_end) {
2861                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2862                         key.type = BTRFS_METADATA_ITEM_KEY;
2863                 else
2864                         key.type = BTRFS_EXTENT_ITEM_KEY;
2865                 key.objectid = logic_start;
2866                 key.offset = (u64)-1;
2867
2868                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2869                 if (ret < 0)
2870                         goto out;
2871
2872                 if (ret > 0) {
2873                         ret = btrfs_previous_extent_item(root, path, 0);
2874                         if (ret < 0)
2875                                 goto out;
2876                         if (ret > 0) {
2877                                 btrfs_release_path(path);
2878                                 ret = btrfs_search_slot(NULL, root, &key,
2879                                                         path, 0, 0);
2880                                 if (ret < 0)
2881                                         goto out;
2882                         }
2883                 }
2884
2885                 stop_loop = 0;
2886                 while (1) {
2887                         u64 bytes;
2888
2889                         l = path->nodes[0];
2890                         slot = path->slots[0];
2891                         if (slot >= btrfs_header_nritems(l)) {
2892                                 ret = btrfs_next_leaf(root, path);
2893                                 if (ret == 0)
2894                                         continue;
2895                                 if (ret < 0)
2896                                         goto out;
2897
2898                                 stop_loop = 1;
2899                                 break;
2900                         }
2901                         btrfs_item_key_to_cpu(l, &key, slot);
2902
2903                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2904                             key.type != BTRFS_METADATA_ITEM_KEY)
2905                                 goto next;
2906
2907                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2908                                 bytes = fs_info->nodesize;
2909                         else
2910                                 bytes = key.offset;
2911
2912                         if (key.objectid + bytes <= logic_start)
2913                                 goto next;
2914
2915                         if (key.objectid >= logic_end) {
2916                                 stop_loop = 1;
2917                                 break;
2918                         }
2919
2920                         while (key.objectid >= logic_start + map->stripe_len)
2921                                 logic_start += map->stripe_len;
2922
2923                         extent = btrfs_item_ptr(l, slot,
2924                                                 struct btrfs_extent_item);
2925                         flags = btrfs_extent_flags(l, extent);
2926                         generation = btrfs_extent_generation(l, extent);
2927
2928                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2929                             (key.objectid < logic_start ||
2930                              key.objectid + bytes >
2931                              logic_start + map->stripe_len)) {
2932                                 btrfs_err(fs_info,
2933                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2934                                           key.objectid, logic_start);
2935                                 spin_lock(&sctx->stat_lock);
2936                                 sctx->stat.uncorrectable_errors++;
2937                                 spin_unlock(&sctx->stat_lock);
2938                                 goto next;
2939                         }
2940 again:
2941                         extent_logical = key.objectid;
2942                         extent_len = bytes;
2943
2944                         if (extent_logical < logic_start) {
2945                                 extent_len -= logic_start - extent_logical;
2946                                 extent_logical = logic_start;
2947                         }
2948
2949                         if (extent_logical + extent_len >
2950                             logic_start + map->stripe_len)
2951                                 extent_len = logic_start + map->stripe_len -
2952                                              extent_logical;
2953
2954                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2955                                                        extent_len);
2956
2957                         mapped_length = extent_len;
2958                         bbio = NULL;
2959                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
2960                                         extent_logical, &mapped_length, &bbio,
2961                                         0);
2962                         if (!ret) {
2963                                 if (!bbio || mapped_length < extent_len)
2964                                         ret = -EIO;
2965                         }
2966                         if (ret) {
2967                                 btrfs_put_bbio(bbio);
2968                                 goto out;
2969                         }
2970                         extent_physical = bbio->stripes[0].physical;
2971                         extent_mirror_num = bbio->mirror_num;
2972                         extent_dev = bbio->stripes[0].dev;
2973                         btrfs_put_bbio(bbio);
2974
2975                         ret = btrfs_lookup_csums_range(csum_root,
2976                                                 extent_logical,
2977                                                 extent_logical + extent_len - 1,
2978                                                 &sctx->csum_list, 1);
2979                         if (ret)
2980                                 goto out;
2981
2982                         ret = scrub_extent_for_parity(sparity, extent_logical,
2983                                                       extent_len,
2984                                                       extent_physical,
2985                                                       extent_dev, flags,
2986                                                       generation,
2987                                                       extent_mirror_num);
2988
2989                         scrub_free_csums(sctx);
2990
2991                         if (ret)
2992                                 goto out;
2993
2994                         if (extent_logical + extent_len <
2995                             key.objectid + bytes) {
2996                                 logic_start += map->stripe_len;
2997
2998                                 if (logic_start >= logic_end) {
2999                                         stop_loop = 1;
3000                                         break;
3001                                 }
3002
3003                                 if (logic_start < key.objectid + bytes) {
3004                                         cond_resched();
3005                                         goto again;
3006                                 }
3007                         }
3008 next:
3009                         path->slots[0]++;
3010                 }
3011
3012                 btrfs_release_path(path);
3013
3014                 if (stop_loop)
3015                         break;
3016
3017                 logic_start += map->stripe_len;
3018         }
3019 out:
3020         if (ret < 0)
3021                 scrub_parity_mark_sectors_error(sparity, logic_start,
3022                                                 logic_end - logic_start);
3023         scrub_parity_put(sparity);
3024         scrub_submit(sctx);
3025         mutex_lock(&sctx->wr_lock);
3026         scrub_wr_submit(sctx);
3027         mutex_unlock(&sctx->wr_lock);
3028
3029         btrfs_release_path(path);
3030         return ret < 0 ? ret : 0;
3031 }
3032
3033 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3034                                            struct map_lookup *map,
3035                                            struct btrfs_device *scrub_dev,
3036                                            int num, u64 base, u64 length)
3037 {
3038         struct btrfs_path *path, *ppath;
3039         struct btrfs_fs_info *fs_info = sctx->fs_info;
3040         struct btrfs_root *root = fs_info->extent_root;
3041         struct btrfs_root *csum_root = fs_info->csum_root;
3042         struct btrfs_extent_item *extent;
3043         struct blk_plug plug;
3044         u64 flags;
3045         int ret;
3046         int slot;
3047         u64 nstripes;
3048         struct extent_buffer *l;
3049         u64 physical;
3050         u64 logical;
3051         u64 logic_end;
3052         u64 physical_end;
3053         u64 generation;
3054         int mirror_num;
3055         struct reada_control *reada1;
3056         struct reada_control *reada2;
3057         struct btrfs_key key;
3058         struct btrfs_key key_end;
3059         u64 increment = map->stripe_len;
3060         u64 offset;
3061         u64 extent_logical;
3062         u64 extent_physical;
3063         u64 extent_len;
3064         u64 stripe_logical;
3065         u64 stripe_end;
3066         struct btrfs_device *extent_dev;
3067         int extent_mirror_num;
3068         int stop_loop = 0;
3069
3070         physical = map->stripes[num].physical;
3071         offset = 0;
3072         nstripes = div64_u64(length, map->stripe_len);
3073         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3074                 offset = map->stripe_len * num;
3075                 increment = map->stripe_len * map->num_stripes;
3076                 mirror_num = 1;
3077         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3078                 int factor = map->num_stripes / map->sub_stripes;
3079                 offset = map->stripe_len * (num / map->sub_stripes);
3080                 increment = map->stripe_len * factor;
3081                 mirror_num = num % map->sub_stripes + 1;
3082         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3083                 increment = map->stripe_len;
3084                 mirror_num = num % map->num_stripes + 1;
3085         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3086                 increment = map->stripe_len;
3087                 mirror_num = num % map->num_stripes + 1;
3088         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3089                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3090                 increment = map->stripe_len * nr_data_stripes(map);
3091                 mirror_num = 1;
3092         } else {
3093                 increment = map->stripe_len;
3094                 mirror_num = 1;
3095         }
3096
3097         path = btrfs_alloc_path();
3098         if (!path)
3099                 return -ENOMEM;
3100
3101         ppath = btrfs_alloc_path();
3102         if (!ppath) {
3103                 btrfs_free_path(path);
3104                 return -ENOMEM;
3105         }
3106
3107         /*
3108          * work on commit root. The related disk blocks are static as
3109          * long as COW is applied. This means, it is save to rewrite
3110          * them to repair disk errors without any race conditions
3111          */
3112         path->search_commit_root = 1;
3113         path->skip_locking = 1;
3114
3115         ppath->search_commit_root = 1;
3116         ppath->skip_locking = 1;
3117         /*
3118          * trigger the readahead for extent tree csum tree and wait for
3119          * completion. During readahead, the scrub is officially paused
3120          * to not hold off transaction commits
3121          */
3122         logical = base + offset;
3123         physical_end = physical + nstripes * map->stripe_len;
3124         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3125                 get_raid56_logic_offset(physical_end, num,
3126                                         map, &logic_end, NULL);
3127                 logic_end += base;
3128         } else {
3129                 logic_end = logical + increment * nstripes;
3130         }
3131         wait_event(sctx->list_wait,
3132                    atomic_read(&sctx->bios_in_flight) == 0);
3133         scrub_blocked_if_needed(fs_info);
3134
3135         /* FIXME it might be better to start readahead at commit root */
3136         key.objectid = logical;
3137         key.type = BTRFS_EXTENT_ITEM_KEY;
3138         key.offset = (u64)0;
3139         key_end.objectid = logic_end;
3140         key_end.type = BTRFS_METADATA_ITEM_KEY;
3141         key_end.offset = (u64)-1;
3142         reada1 = btrfs_reada_add(root, &key, &key_end);
3143
3144         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3145         key.type = BTRFS_EXTENT_CSUM_KEY;
3146         key.offset = logical;
3147         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3148         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3149         key_end.offset = logic_end;
3150         reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3151
3152         if (!IS_ERR(reada1))
3153                 btrfs_reada_wait(reada1);
3154         if (!IS_ERR(reada2))
3155                 btrfs_reada_wait(reada2);
3156
3157
3158         /*
3159          * collect all data csums for the stripe to avoid seeking during
3160          * the scrub. This might currently (crc32) end up to be about 1MB
3161          */
3162         blk_start_plug(&plug);
3163
3164         /*
3165          * now find all extents for each stripe and scrub them
3166          */
3167         ret = 0;
3168         while (physical < physical_end) {
3169                 /*
3170                  * canceled?
3171                  */
3172                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3173                     atomic_read(&sctx->cancel_req)) {
3174                         ret = -ECANCELED;
3175                         goto out;
3176                 }
3177                 /*
3178                  * check to see if we have to pause
3179                  */
3180                 if (atomic_read(&fs_info->scrub_pause_req)) {
3181                         /* push queued extents */
3182                         sctx->flush_all_writes = true;
3183                         scrub_submit(sctx);
3184                         mutex_lock(&sctx->wr_lock);
3185                         scrub_wr_submit(sctx);
3186                         mutex_unlock(&sctx->wr_lock);
3187                         wait_event(sctx->list_wait,
3188                                    atomic_read(&sctx->bios_in_flight) == 0);
3189                         sctx->flush_all_writes = false;
3190                         scrub_blocked_if_needed(fs_info);
3191                 }
3192
3193                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3194                         ret = get_raid56_logic_offset(physical, num, map,
3195                                                       &logical,
3196                                                       &stripe_logical);
3197                         logical += base;
3198                         if (ret) {
3199                                 /* it is parity strip */
3200                                 stripe_logical += base;
3201                                 stripe_end = stripe_logical + increment;
3202                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3203                                                           ppath, stripe_logical,
3204                                                           stripe_end);
3205                                 if (ret)
3206                                         goto out;
3207                                 goto skip;
3208                         }
3209                 }
3210
3211                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3212                         key.type = BTRFS_METADATA_ITEM_KEY;
3213                 else
3214                         key.type = BTRFS_EXTENT_ITEM_KEY;
3215                 key.objectid = logical;
3216                 key.offset = (u64)-1;
3217
3218                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3219                 if (ret < 0)
3220                         goto out;
3221
3222                 if (ret > 0) {
3223                         ret = btrfs_previous_extent_item(root, path, 0);
3224                         if (ret < 0)
3225                                 goto out;
3226                         if (ret > 0) {
3227                                 /* there's no smaller item, so stick with the
3228                                  * larger one */
3229                                 btrfs_release_path(path);
3230                                 ret = btrfs_search_slot(NULL, root, &key,
3231                                                         path, 0, 0);
3232                                 if (ret < 0)
3233                                         goto out;
3234                         }
3235                 }
3236
3237                 stop_loop = 0;
3238                 while (1) {
3239                         u64 bytes;
3240
3241                         l = path->nodes[0];
3242                         slot = path->slots[0];
3243                         if (slot >= btrfs_header_nritems(l)) {
3244                                 ret = btrfs_next_leaf(root, path);
3245                                 if (ret == 0)
3246                                         continue;
3247                                 if (ret < 0)
3248                                         goto out;
3249
3250                                 stop_loop = 1;
3251                                 break;
3252                         }
3253                         btrfs_item_key_to_cpu(l, &key, slot);
3254
3255                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3256                             key.type != BTRFS_METADATA_ITEM_KEY)
3257                                 goto next;
3258
3259                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3260                                 bytes = fs_info->nodesize;
3261                         else
3262                                 bytes = key.offset;
3263
3264                         if (key.objectid + bytes <= logical)
3265                                 goto next;
3266
3267                         if (key.objectid >= logical + map->stripe_len) {
3268                                 /* out of this device extent */
3269                                 if (key.objectid >= logic_end)
3270                                         stop_loop = 1;
3271                                 break;
3272                         }
3273
3274                         extent = btrfs_item_ptr(l, slot,
3275                                                 struct btrfs_extent_item);
3276                         flags = btrfs_extent_flags(l, extent);
3277                         generation = btrfs_extent_generation(l, extent);
3278
3279                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3280                             (key.objectid < logical ||
3281                              key.objectid + bytes >
3282                              logical + map->stripe_len)) {
3283                                 btrfs_err(fs_info,
3284                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3285                                        key.objectid, logical);
3286                                 spin_lock(&sctx->stat_lock);
3287                                 sctx->stat.uncorrectable_errors++;
3288                                 spin_unlock(&sctx->stat_lock);
3289                                 goto next;
3290                         }
3291
3292 again:
3293                         extent_logical = key.objectid;
3294                         extent_len = bytes;
3295
3296                         /*
3297                          * trim extent to this stripe
3298                          */
3299                         if (extent_logical < logical) {
3300                                 extent_len -= logical - extent_logical;
3301                                 extent_logical = logical;
3302                         }
3303                         if (extent_logical + extent_len >
3304                             logical + map->stripe_len) {
3305                                 extent_len = logical + map->stripe_len -
3306                                              extent_logical;
3307                         }
3308
3309                         extent_physical = extent_logical - logical + physical;
3310                         extent_dev = scrub_dev;
3311                         extent_mirror_num = mirror_num;
3312                         if (sctx->is_dev_replace)
3313                                 scrub_remap_extent(fs_info, extent_logical,
3314                                                    extent_len, &extent_physical,
3315                                                    &extent_dev,
3316                                                    &extent_mirror_num);
3317
3318                         ret = btrfs_lookup_csums_range(csum_root,
3319                                                        extent_logical,
3320                                                        extent_logical +
3321                                                        extent_len - 1,
3322                                                        &sctx->csum_list, 1);
3323                         if (ret)
3324                                 goto out;
3325
3326                         ret = scrub_extent(sctx, map, extent_logical, extent_len,
3327                                            extent_physical, extent_dev, flags,
3328                                            generation, extent_mirror_num,
3329                                            extent_logical - logical + physical);
3330
3331                         scrub_free_csums(sctx);
3332
3333                         if (ret)
3334                                 goto out;
3335
3336                         if (extent_logical + extent_len <
3337                             key.objectid + bytes) {
3338                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3339                                         /*
3340                                          * loop until we find next data stripe
3341                                          * or we have finished all stripes.
3342                                          */
3343 loop:
3344                                         physical += map->stripe_len;
3345                                         ret = get_raid56_logic_offset(physical,
3346                                                         num, map, &logical,
3347                                                         &stripe_logical);
3348                                         logical += base;
3349
3350                                         if (ret && physical < physical_end) {
3351                                                 stripe_logical += base;
3352                                                 stripe_end = stripe_logical +
3353                                                                 increment;
3354                                                 ret = scrub_raid56_parity(sctx,
3355                                                         map, scrub_dev, ppath,
3356                                                         stripe_logical,
3357                                                         stripe_end);
3358                                                 if (ret)
3359                                                         goto out;
3360                                                 goto loop;
3361                                         }
3362                                 } else {
3363                                         physical += map->stripe_len;
3364                                         logical += increment;
3365                                 }
3366                                 if (logical < key.objectid + bytes) {
3367                                         cond_resched();
3368                                         goto again;
3369                                 }
3370
3371                                 if (physical >= physical_end) {
3372                                         stop_loop = 1;
3373                                         break;
3374                                 }
3375                         }
3376 next:
3377                         path->slots[0]++;
3378                 }
3379                 btrfs_release_path(path);
3380 skip:
3381                 logical += increment;
3382                 physical += map->stripe_len;
3383                 spin_lock(&sctx->stat_lock);
3384                 if (stop_loop)
3385                         sctx->stat.last_physical = map->stripes[num].physical +
3386                                                    length;
3387                 else
3388                         sctx->stat.last_physical = physical;
3389                 spin_unlock(&sctx->stat_lock);
3390                 if (stop_loop)
3391                         break;
3392         }
3393 out:
3394         /* push queued extents */
3395         scrub_submit(sctx);
3396         mutex_lock(&sctx->wr_lock);
3397         scrub_wr_submit(sctx);
3398         mutex_unlock(&sctx->wr_lock);
3399
3400         blk_finish_plug(&plug);
3401         btrfs_free_path(path);
3402         btrfs_free_path(ppath);
3403         return ret < 0 ? ret : 0;
3404 }
3405
3406 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3407                                           struct btrfs_device *scrub_dev,
3408                                           u64 chunk_offset, u64 length,
3409                                           u64 dev_offset,
3410                                           struct btrfs_block_group_cache *cache)
3411 {
3412         struct btrfs_fs_info *fs_info = sctx->fs_info;
3413         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3414         struct map_lookup *map;
3415         struct extent_map *em;
3416         int i;
3417         int ret = 0;
3418
3419         read_lock(&map_tree->map_tree.lock);
3420         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3421         read_unlock(&map_tree->map_tree.lock);
3422
3423         if (!em) {
3424                 /*
3425                  * Might have been an unused block group deleted by the cleaner
3426                  * kthread or relocation.
3427                  */
3428                 spin_lock(&cache->lock);
3429                 if (!cache->removed)
3430                         ret = -EINVAL;
3431                 spin_unlock(&cache->lock);
3432
3433                 return ret;
3434         }
3435
3436         map = em->map_lookup;
3437         if (em->start != chunk_offset)
3438                 goto out;
3439
3440         if (em->len < length)
3441                 goto out;
3442
3443         for (i = 0; i < map->num_stripes; ++i) {
3444                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3445                     map->stripes[i].physical == dev_offset) {
3446                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3447                                            chunk_offset, length);
3448                         if (ret)
3449                                 goto out;
3450                 }
3451         }
3452 out:
3453         free_extent_map(em);
3454
3455         return ret;
3456 }
3457
3458 static noinline_for_stack
3459 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3460                            struct btrfs_device *scrub_dev, u64 start, u64 end)
3461 {
3462         struct btrfs_dev_extent *dev_extent = NULL;
3463         struct btrfs_path *path;
3464         struct btrfs_fs_info *fs_info = sctx->fs_info;
3465         struct btrfs_root *root = fs_info->dev_root;
3466         u64 length;
3467         u64 chunk_offset;
3468         int ret = 0;
3469         int ro_set;
3470         int slot;
3471         struct extent_buffer *l;
3472         struct btrfs_key key;
3473         struct btrfs_key found_key;
3474         struct btrfs_block_group_cache *cache;
3475         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3476
3477         path = btrfs_alloc_path();
3478         if (!path)
3479                 return -ENOMEM;
3480
3481         path->reada = READA_FORWARD;
3482         path->search_commit_root = 1;
3483         path->skip_locking = 1;
3484
3485         key.objectid = scrub_dev->devid;
3486         key.offset = 0ull;
3487         key.type = BTRFS_DEV_EXTENT_KEY;
3488
3489         while (1) {
3490                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3491                 if (ret < 0)
3492                         break;
3493                 if (ret > 0) {
3494                         if (path->slots[0] >=
3495                             btrfs_header_nritems(path->nodes[0])) {
3496                                 ret = btrfs_next_leaf(root, path);
3497                                 if (ret < 0)
3498                                         break;
3499                                 if (ret > 0) {
3500                                         ret = 0;
3501                                         break;
3502                                 }
3503                         } else {
3504                                 ret = 0;
3505                         }
3506                 }
3507
3508                 l = path->nodes[0];
3509                 slot = path->slots[0];
3510
3511                 btrfs_item_key_to_cpu(l, &found_key, slot);
3512
3513                 if (found_key.objectid != scrub_dev->devid)
3514                         break;
3515
3516                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3517                         break;
3518
3519                 if (found_key.offset >= end)
3520                         break;
3521
3522                 if (found_key.offset < key.offset)
3523                         break;
3524
3525                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3526                 length = btrfs_dev_extent_length(l, dev_extent);
3527
3528                 if (found_key.offset + length <= start)
3529                         goto skip;
3530
3531                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3532
3533                 /*
3534                  * get a reference on the corresponding block group to prevent
3535                  * the chunk from going away while we scrub it
3536                  */
3537                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3538
3539                 /* some chunks are removed but not committed to disk yet,
3540                  * continue scrubbing */
3541                 if (!cache)
3542                         goto skip;
3543
3544                 /*
3545                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3546                  * to avoid deadlock caused by:
3547                  * btrfs_inc_block_group_ro()
3548                  * -> btrfs_wait_for_commit()
3549                  * -> btrfs_commit_transaction()
3550                  * -> btrfs_scrub_pause()
3551                  */
3552                 scrub_pause_on(fs_info);
3553                 ret = btrfs_inc_block_group_ro(cache);
3554                 if (!ret && sctx->is_dev_replace) {
3555                         /*
3556                          * If we are doing a device replace wait for any tasks
3557                          * that started delalloc right before we set the block
3558                          * group to RO mode, as they might have just allocated
3559                          * an extent from it or decided they could do a nocow
3560                          * write. And if any such tasks did that, wait for their
3561                          * ordered extents to complete and then commit the
3562                          * current transaction, so that we can later see the new
3563                          * extent items in the extent tree - the ordered extents
3564                          * create delayed data references (for cow writes) when
3565                          * they complete, which will be run and insert the
3566                          * corresponding extent items into the extent tree when
3567                          * we commit the transaction they used when running
3568                          * inode.c:btrfs_finish_ordered_io(). We later use
3569                          * the commit root of the extent tree to find extents
3570                          * to copy from the srcdev into the tgtdev, and we don't
3571                          * want to miss any new extents.
3572                          */
3573                         btrfs_wait_block_group_reservations(cache);
3574                         btrfs_wait_nocow_writers(cache);
3575                         ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3576                                                        cache->key.objectid,
3577                                                        cache->key.offset);
3578                         if (ret > 0) {
3579                                 struct btrfs_trans_handle *trans;
3580
3581                                 trans = btrfs_join_transaction(root);
3582                                 if (IS_ERR(trans))
3583                                         ret = PTR_ERR(trans);
3584                                 else
3585                                         ret = btrfs_commit_transaction(trans);
3586                                 if (ret) {
3587                                         scrub_pause_off(fs_info);
3588                                         btrfs_put_block_group(cache);
3589                                         break;
3590                                 }
3591                         }
3592                 }
3593                 scrub_pause_off(fs_info);
3594
3595                 if (ret == 0) {
3596                         ro_set = 1;
3597                 } else if (ret == -ENOSPC) {
3598                         /*
3599                          * btrfs_inc_block_group_ro return -ENOSPC when it
3600                          * failed in creating new chunk for metadata.
3601                          * It is not a problem for scrub/replace, because
3602                          * metadata are always cowed, and our scrub paused
3603                          * commit_transactions.
3604                          */
3605                         ro_set = 0;
3606                 } else {
3607                         btrfs_warn(fs_info,
3608                                    "failed setting block group ro: %d", ret);
3609                         btrfs_put_block_group(cache);
3610                         break;
3611                 }
3612
3613                 down_write(&fs_info->dev_replace.rwsem);
3614                 dev_replace->cursor_right = found_key.offset + length;
3615                 dev_replace->cursor_left = found_key.offset;
3616                 dev_replace->item_needs_writeback = 1;
3617                 up_write(&dev_replace->rwsem);
3618
3619                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3620                                   found_key.offset, cache);
3621
3622                 /*
3623                  * flush, submit all pending read and write bios, afterwards
3624                  * wait for them.
3625                  * Note that in the dev replace case, a read request causes
3626                  * write requests that are submitted in the read completion
3627                  * worker. Therefore in the current situation, it is required
3628                  * that all write requests are flushed, so that all read and
3629                  * write requests are really completed when bios_in_flight
3630                  * changes to 0.
3631                  */
3632                 sctx->flush_all_writes = true;
3633                 scrub_submit(sctx);
3634                 mutex_lock(&sctx->wr_lock);
3635                 scrub_wr_submit(sctx);
3636                 mutex_unlock(&sctx->wr_lock);
3637
3638                 wait_event(sctx->list_wait,
3639                            atomic_read(&sctx->bios_in_flight) == 0);
3640
3641                 scrub_pause_on(fs_info);
3642
3643                 /*
3644                  * must be called before we decrease @scrub_paused.
3645                  * make sure we don't block transaction commit while
3646                  * we are waiting pending workers finished.
3647                  */
3648                 wait_event(sctx->list_wait,
3649                            atomic_read(&sctx->workers_pending) == 0);
3650                 sctx->flush_all_writes = false;
3651
3652                 scrub_pause_off(fs_info);
3653
3654                 down_write(&fs_info->dev_replace.rwsem);
3655                 dev_replace->cursor_left = dev_replace->cursor_right;
3656                 dev_replace->item_needs_writeback = 1;
3657                 up_write(&fs_info->dev_replace.rwsem);
3658
3659                 if (ro_set)
3660                         btrfs_dec_block_group_ro(cache);
3661
3662                 /*
3663                  * We might have prevented the cleaner kthread from deleting
3664                  * this block group if it was already unused because we raced
3665                  * and set it to RO mode first. So add it back to the unused
3666                  * list, otherwise it might not ever be deleted unless a manual
3667                  * balance is triggered or it becomes used and unused again.
3668                  */
3669                 spin_lock(&cache->lock);
3670                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3671                     btrfs_block_group_used(&cache->item) == 0) {
3672                         spin_unlock(&cache->lock);
3673                         btrfs_mark_bg_unused(cache);
3674                 } else {
3675                         spin_unlock(&cache->lock);
3676                 }
3677
3678                 btrfs_put_block_group(cache);
3679                 if (ret)
3680                         break;
3681                 if (sctx->is_dev_replace &&
3682                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3683                         ret = -EIO;
3684                         break;
3685                 }
3686                 if (sctx->stat.malloc_errors > 0) {
3687                         ret = -ENOMEM;
3688                         break;
3689                 }
3690 skip:
3691                 key.offset = found_key.offset + length;
3692                 btrfs_release_path(path);
3693         }
3694
3695         btrfs_free_path(path);
3696
3697         return ret;
3698 }
3699
3700 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3701                                            struct btrfs_device *scrub_dev)
3702 {
3703         int     i;
3704         u64     bytenr;
3705         u64     gen;
3706         int     ret;
3707         struct btrfs_fs_info *fs_info = sctx->fs_info;
3708
3709         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
3710                 return -EIO;
3711
3712         /* Seed devices of a new filesystem has their own generation. */
3713         if (scrub_dev->fs_devices != fs_info->fs_devices)
3714                 gen = scrub_dev->generation;
3715         else
3716                 gen = fs_info->last_trans_committed;
3717
3718         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3719                 bytenr = btrfs_sb_offset(i);
3720                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3721                     scrub_dev->commit_total_bytes)
3722                         break;
3723
3724                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3725                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3726                                   NULL, 1, bytenr);
3727                 if (ret)
3728                         return ret;
3729         }
3730         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3731
3732         return 0;
3733 }
3734
3735 /*
3736  * get a reference count on fs_info->scrub_workers. start worker if necessary
3737  */
3738 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3739                                                 int is_dev_replace)
3740 {
3741         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3742         int max_active = fs_info->thread_pool_size;
3743
3744         if (fs_info->scrub_workers_refcnt == 0) {
3745                 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
3746                                 flags, is_dev_replace ? 1 : max_active, 4);
3747                 if (!fs_info->scrub_workers)
3748                         goto fail_scrub_workers;
3749
3750                 fs_info->scrub_wr_completion_workers =
3751                         btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
3752                                               max_active, 2);
3753                 if (!fs_info->scrub_wr_completion_workers)
3754                         goto fail_scrub_wr_completion_workers;
3755
3756                 fs_info->scrub_parity_workers =
3757                         btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
3758                                               max_active, 2);
3759                 if (!fs_info->scrub_parity_workers)
3760                         goto fail_scrub_parity_workers;
3761         }
3762         ++fs_info->scrub_workers_refcnt;
3763         return 0;
3764
3765 fail_scrub_parity_workers:
3766         btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3767 fail_scrub_wr_completion_workers:
3768         btrfs_destroy_workqueue(fs_info->scrub_workers);
3769 fail_scrub_workers:
3770         return -ENOMEM;
3771 }
3772
3773 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3774 {
3775         if (--fs_info->scrub_workers_refcnt == 0) {
3776                 btrfs_destroy_workqueue(fs_info->scrub_workers);
3777                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3778                 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3779         }
3780         WARN_ON(fs_info->scrub_workers_refcnt < 0);
3781 }
3782
3783 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3784                     u64 end, struct btrfs_scrub_progress *progress,
3785                     int readonly, int is_dev_replace)
3786 {
3787         struct scrub_ctx *sctx;
3788         int ret;
3789         struct btrfs_device *dev;
3790         unsigned int nofs_flag;
3791
3792         if (btrfs_fs_closing(fs_info))
3793                 return -EINVAL;
3794
3795         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
3796                 /*
3797                  * in this case scrub is unable to calculate the checksum
3798                  * the way scrub is implemented. Do not handle this
3799                  * situation at all because it won't ever happen.
3800                  */
3801                 btrfs_err(fs_info,
3802                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3803                        fs_info->nodesize,
3804                        BTRFS_STRIPE_LEN);
3805                 return -EINVAL;
3806         }
3807
3808         if (fs_info->sectorsize != PAGE_SIZE) {
3809                 /* not supported for data w/o checksums */
3810                 btrfs_err_rl(fs_info,
3811                            "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
3812                        fs_info->sectorsize, PAGE_SIZE);
3813                 return -EINVAL;
3814         }
3815
3816         if (fs_info->nodesize >
3817             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3818             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3819                 /*
3820                  * would exhaust the array bounds of pagev member in
3821                  * struct scrub_block
3822                  */
3823                 btrfs_err(fs_info,
3824                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3825                        fs_info->nodesize,
3826                        SCRUB_MAX_PAGES_PER_BLOCK,
3827                        fs_info->sectorsize,
3828                        SCRUB_MAX_PAGES_PER_BLOCK);
3829                 return -EINVAL;
3830         }
3831
3832         /* Allocate outside of device_list_mutex */
3833         sctx = scrub_setup_ctx(fs_info, is_dev_replace);
3834         if (IS_ERR(sctx))
3835                 return PTR_ERR(sctx);
3836
3837         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3838         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3839         if (!dev || (test_bit(BTRFS_DEV_STATE_MISSING, &dev->dev_state) &&
3840                      !is_dev_replace)) {
3841                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3842                 ret = -ENODEV;
3843                 goto out_free_ctx;
3844         }
3845
3846         if (!is_dev_replace && !readonly &&
3847             !test_bit(BTRFS_DEV_STATE_WRITEABLE, &dev->dev_state)) {
3848                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3849                 btrfs_err_in_rcu(fs_info, "scrub: device %s is not writable",
3850                                 rcu_str_deref(dev->name));
3851                 ret = -EROFS;
3852                 goto out_free_ctx;
3853         }
3854
3855         mutex_lock(&fs_info->scrub_lock);
3856         if (!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &dev->dev_state) ||
3857             test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &dev->dev_state)) {
3858                 mutex_unlock(&fs_info->scrub_lock);
3859                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3860                 ret = -EIO;
3861                 goto out_free_ctx;
3862         }
3863
3864         down_read(&fs_info->dev_replace.rwsem);
3865         if (dev->scrub_ctx ||
3866             (!is_dev_replace &&
3867              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3868                 up_read(&fs_info->dev_replace.rwsem);
3869                 mutex_unlock(&fs_info->scrub_lock);
3870                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3871                 ret = -EINPROGRESS;
3872                 goto out_free_ctx;
3873         }
3874         up_read(&fs_info->dev_replace.rwsem);
3875
3876         ret = scrub_workers_get(fs_info, is_dev_replace);
3877         if (ret) {
3878                 mutex_unlock(&fs_info->scrub_lock);
3879                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3880                 goto out_free_ctx;
3881         }
3882
3883         sctx->readonly = readonly;
3884         dev->scrub_ctx = sctx;
3885         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3886
3887         /*
3888          * checking @scrub_pause_req here, we can avoid
3889          * race between committing transaction and scrubbing.
3890          */
3891         __scrub_blocked_if_needed(fs_info);
3892         atomic_inc(&fs_info->scrubs_running);
3893         mutex_unlock(&fs_info->scrub_lock);
3894
3895         /*
3896          * In order to avoid deadlock with reclaim when there is a transaction
3897          * trying to pause scrub, make sure we use GFP_NOFS for all the
3898          * allocations done at btrfs_scrub_pages() and scrub_pages_for_parity()
3899          * invoked by our callees. The pausing request is done when the
3900          * transaction commit starts, and it blocks the transaction until scrub
3901          * is paused (done at specific points at scrub_stripe() or right above
3902          * before incrementing fs_info->scrubs_running).
3903          */
3904         nofs_flag = memalloc_nofs_save();
3905         if (!is_dev_replace) {
3906                 /*
3907                  * by holding device list mutex, we can
3908                  * kick off writing super in log tree sync.
3909                  */
3910                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3911                 ret = scrub_supers(sctx, dev);
3912                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3913         }
3914
3915         if (!ret)
3916                 ret = scrub_enumerate_chunks(sctx, dev, start, end);
3917         memalloc_nofs_restore(nofs_flag);
3918
3919         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3920         atomic_dec(&fs_info->scrubs_running);
3921         wake_up(&fs_info->scrub_pause_wait);
3922
3923         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3924
3925         if (progress)
3926                 memcpy(progress, &sctx->stat, sizeof(*progress));
3927
3928         mutex_lock(&fs_info->scrub_lock);
3929         dev->scrub_ctx = NULL;
3930         scrub_workers_put(fs_info);
3931         mutex_unlock(&fs_info->scrub_lock);
3932
3933         scrub_put_ctx(sctx);
3934
3935         return ret;
3936
3937 out_free_ctx:
3938         scrub_free_ctx(sctx);
3939
3940         return ret;
3941 }
3942
3943 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
3944 {
3945         mutex_lock(&fs_info->scrub_lock);
3946         atomic_inc(&fs_info->scrub_pause_req);
3947         while (atomic_read(&fs_info->scrubs_paused) !=
3948                atomic_read(&fs_info->scrubs_running)) {
3949                 mutex_unlock(&fs_info->scrub_lock);
3950                 wait_event(fs_info->scrub_pause_wait,
3951                            atomic_read(&fs_info->scrubs_paused) ==
3952                            atomic_read(&fs_info->scrubs_running));
3953                 mutex_lock(&fs_info->scrub_lock);
3954         }
3955         mutex_unlock(&fs_info->scrub_lock);
3956 }
3957
3958 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
3959 {
3960         atomic_dec(&fs_info->scrub_pause_req);
3961         wake_up(&fs_info->scrub_pause_wait);
3962 }
3963
3964 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3965 {
3966         mutex_lock(&fs_info->scrub_lock);
3967         if (!atomic_read(&fs_info->scrubs_running)) {
3968                 mutex_unlock(&fs_info->scrub_lock);
3969                 return -ENOTCONN;
3970         }
3971
3972         atomic_inc(&fs_info->scrub_cancel_req);
3973         while (atomic_read(&fs_info->scrubs_running)) {
3974                 mutex_unlock(&fs_info->scrub_lock);
3975                 wait_event(fs_info->scrub_pause_wait,
3976                            atomic_read(&fs_info->scrubs_running) == 0);
3977                 mutex_lock(&fs_info->scrub_lock);
3978         }
3979         atomic_dec(&fs_info->scrub_cancel_req);
3980         mutex_unlock(&fs_info->scrub_lock);
3981
3982         return 0;
3983 }
3984
3985 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3986                            struct btrfs_device *dev)
3987 {
3988         struct scrub_ctx *sctx;
3989
3990         mutex_lock(&fs_info->scrub_lock);
3991         sctx = dev->scrub_ctx;
3992         if (!sctx) {
3993                 mutex_unlock(&fs_info->scrub_lock);
3994                 return -ENOTCONN;
3995         }
3996         atomic_inc(&sctx->cancel_req);
3997         while (dev->scrub_ctx) {
3998                 mutex_unlock(&fs_info->scrub_lock);
3999                 wait_event(fs_info->scrub_pause_wait,
4000                            dev->scrub_ctx == NULL);
4001                 mutex_lock(&fs_info->scrub_lock);
4002         }
4003         mutex_unlock(&fs_info->scrub_lock);
4004
4005         return 0;
4006 }
4007
4008 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4009                          struct btrfs_scrub_progress *progress)
4010 {
4011         struct btrfs_device *dev;
4012         struct scrub_ctx *sctx = NULL;
4013
4014         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4015         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4016         if (dev)
4017                 sctx = dev->scrub_ctx;
4018         if (sctx)
4019                 memcpy(progress, &sctx->stat, sizeof(*progress));
4020         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4021
4022         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4023 }
4024
4025 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4026                                u64 extent_logical, u64 extent_len,
4027                                u64 *extent_physical,
4028                                struct btrfs_device **extent_dev,
4029                                int *extent_mirror_num)
4030 {
4031         u64 mapped_length;
4032         struct btrfs_bio *bbio = NULL;
4033         int ret;
4034
4035         mapped_length = extent_len;
4036         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4037                               &mapped_length, &bbio, 0);
4038         if (ret || !bbio || mapped_length < extent_len ||
4039             !bbio->stripes[0].dev->bdev) {
4040                 btrfs_put_bbio(bbio);
4041                 return;
4042         }
4043
4044         *extent_physical = bbio->stripes[0].physical;
4045         *extent_mirror_num = bbio->mirror_num;
4046         *extent_dev = bbio->stripes[0].dev;
4047         btrfs_put_bbio(bbio);
4048 }