fs/btrfs/scrub.c

   1 /*
   2  * Copyright (C) 2011, 2012 STRATO.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/blkdev.h>
  20 #include <linux/ratelimit.h>
  21 #include <linux/sched/mm.h>
  22 #include "ctree.h"
  23 #include "volumes.h"
  24 #include "disk-io.h"
  25 #include "ordered-data.h"
  26 #include "transaction.h"
  27 #include "backref.h"
  28 #include "extent_io.h"
  29 #include "dev-replace.h"
  30 #include "check-integrity.h"
  31 #include "rcu-string.h"
  32 #include "raid56.h"
  33
  34 /*
  35  * This is only the first step towards a full-features scrub. It reads all
  36  * extent and super block and verifies the checksums. In case a bad checksum
  37  * is found or the extent cannot be read, good data will be written back if
  38  * any can be found.
  39  *
  40  * Future enhancements:
  41  *  - In case an unrepairable extent is encountered, track which files are
  42  *    affected and report them
  43  *  - track and record media errors, throw out bad devices
  44  *  - add a mode to also read unallocated space
  45  */
  46
  47 struct scrub_block;
  48 struct scrub_ctx;
  49
  50 /*
  51  * the following three values only influence the performance.
  52  * The last one configures the number of parallel and outstanding I/O
  53  * operations. The first two values configure an upper limit for the number
  54  * of (dynamically allocated) pages that are added to a bio.
  55  */
  56 #define SCRUB_PAGES_PER_RD_BIO  32      /* 128k per bio */
  57 #define SCRUB_PAGES_PER_WR_BIO  32      /* 128k per bio */
  58 #define SCRUB_BIOS_PER_SCTX     64      /* 8MB per device in flight */
  59
  60 /*
  61  * the following value times PAGE_SIZE needs to be large enough to match the
  62  * largest node/leaf/sector size that shall be supported.
  63  * Values larger than BTRFS_STRIPE_LEN are not supported.
  64  */
  65 #define SCRUB_MAX_PAGES_PER_BLOCK       16      /* 64k per node/leaf/sector */
  66
  67 struct scrub_recover {
  68         refcount_t              refs;
  69         struct btrfs_bio        *bbio;
  70         u64                     map_length;
  71 };
  72
  73 struct scrub_page {
  74         struct scrub_block      *sblock;
  75         struct page             *page;
  76         struct btrfs_device     *dev;
  77         struct list_head        list;
  78         u64                     flags;  /* extent flags */
  79         u64                     generation;
  80         u64                     logical;
  81         u64                     physical;
  82         u64                     physical_for_dev_replace;
  83         atomic_t                refs;
  84         struct {
  85                 unsigned int    mirror_num:8;
  86                 unsigned int    have_csum:1;
  87                 unsigned int    io_error:1;
  88         };
  89         u8                      csum[BTRFS_CSUM_SIZE];
  90
  91         struct scrub_recover    *recover;
  92 };
  93
  94 struct scrub_bio {
  95         int                     index;
  96         struct scrub_ctx        *sctx;
  97         struct btrfs_device     *dev;
  98         struct bio              *bio;
  99         blk_status_t            status;
 100         u64                     logical;
 101         u64                     physical;
 102 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
 103         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
 104 #else
 105         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
 106 #endif
 107         int                     page_count;
 108         int                     next_free;
 109         struct btrfs_work       work;
 110 };
 111
 112 struct scrub_block {
 113         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 114         int                     page_count;
 115         atomic_t                outstanding_pages;
 116         refcount_t              refs; /* free mem on transition to zero */
 117         struct scrub_ctx        *sctx;
 118         struct scrub_parity     *sparity;
 119         struct {
 120                 unsigned int    header_error:1;
 121                 unsigned int    checksum_error:1;
 122                 unsigned int    no_io_error_seen:1;
 123                 unsigned int    generation_error:1; /* also sets header_error */
 124
 125                 /* The following is for the data used to check parity */
 126                 /* It is for the data with checksum */
 127                 unsigned int    data_corrected:1;
 128         };
 129         struct btrfs_work       work;
 130 };
 131
 132 /* Used for the chunks with parity stripe such RAID5/6 */
 133 struct scrub_parity {
 134         struct scrub_ctx        *sctx;
 135
 136         struct btrfs_device     *scrub_dev;
 137
 138         u64                     logic_start;
 139
 140         u64                     logic_end;
 141
 142         int                     nsectors;
 143
 144         u64                     stripe_len;
 145
 146         refcount_t              refs;
 147
 148         struct list_head        spages;
 149
 150         /* Work of parity check and repair */
 151         struct btrfs_work       work;
 152
 153         /* Mark the parity blocks which have data */
 154         unsigned long           *dbitmap;
 155
 156         /*
 157          * Mark the parity blocks which have data, but errors happen when
 158          * read data or check data
 159          */
 160         unsigned long           *ebitmap;
 161
 162         unsigned long           bitmap[0];
 163 };
 164
 165 struct scrub_ctx {
 166         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 167         struct btrfs_fs_info    *fs_info;
 168         int                     first_free;
 169         int                     curr;
 170         atomic_t                bios_in_flight;
 171         atomic_t                workers_pending;
 172         spinlock_t              list_lock;
 173         wait_queue_head_t       list_wait;
 174         u16                     csum_size;
 175         struct list_head        csum_list;
 176         atomic_t                cancel_req;
 177         int                     readonly;
 178         int                     pages_per_rd_bio;
 179
 180         int                     is_dev_replace;
 181
 182         struct scrub_bio        *wr_curr_bio;
 183         struct mutex            wr_lock;
 184         int                     pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 185         struct btrfs_device     *wr_tgtdev;
 186         bool                    flush_all_writes;
 187
 188         /*
 189          * statistics
 190          */
 191         struct btrfs_scrub_progress stat;
 192         spinlock_t              stat_lock;
 193
 194         /*
 195          * Use a ref counter to avoid use-after-free issues. Scrub workers
 196          * decrement bios_in_flight and workers_pending and then do a wakeup
 197          * on the list_wait wait queue. We must ensure the main scrub task
 198          * doesn't free the scrub context before or while the workers are
 199          * doing the wakeup() call.
 200          */
 201         refcount_t              refs;
 202 };
 203
 204 struct scrub_fixup_nodatasum {
 205         struct scrub_ctx        *sctx;
 206         struct btrfs_device     *dev;
 207         u64                     logical;
 208         struct btrfs_root       *root;
 209         struct btrfs_work       work;
 210         int                     mirror_num;
 211 };
 212
 213 struct scrub_nocow_inode {
 214         u64                     inum;
 215         u64                     offset;
 216         u64                     root;
 217         struct list_head        list;
 218 };
 219
 220 struct scrub_copy_nocow_ctx {
 221         struct scrub_ctx        *sctx;
 222         u64                     logical;
 223         u64                     len;
 224         int                     mirror_num;
 225         u64                     physical_for_dev_replace;
 226         struct list_head        inodes;
 227         struct btrfs_work       work;
 228 };
 229
 230 struct scrub_warning {
 231         struct btrfs_path       *path;
 232         u64                     extent_item_size;
 233         const char              *errstr;
 234         sector_t                sector;
 235         u64                     logical;
 236         struct btrfs_device     *dev;
 237 };
 238
 239 struct full_stripe_lock {
 240         struct rb_node node;
 241         u64 logical;
 242         u64 refs;
 243         struct mutex mutex;
 244 };
 245
 246 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 247 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 248 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 249 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 250 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 251 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
 252                                      struct scrub_block *sblocks_for_recheck);
 253 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 254                                 struct scrub_block *sblock,
 255                                 int retry_failed_mirror);
 256 static void scrub_recheck_block_checksum(struct scrub_block *sblock);
 257 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 258                                              struct scrub_block *sblock_good);
 259 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 260                                             struct scrub_block *sblock_good,
 261                                             int page_num, int force_write);
 262 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 263 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 264                                            int page_num);
 265 static int scrub_checksum_data(struct scrub_block *sblock);
 266 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 267 static int scrub_checksum_super(struct scrub_block *sblock);
 268 static void scrub_block_get(struct scrub_block *sblock);
 269 static void scrub_block_put(struct scrub_block *sblock);
 270 static void scrub_page_get(struct scrub_page *spage);
 271 static void scrub_page_put(struct scrub_page *spage);
 272 static void scrub_parity_get(struct scrub_parity *sparity);
 273 static void scrub_parity_put(struct scrub_parity *sparity);
 274 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 275                                     struct scrub_page *spage);
 276 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 277                        u64 physical, struct btrfs_device *dev, u64 flags,
 278                        u64 gen, int mirror_num, u8 *csum, int force,
 279                        u64 physical_for_dev_replace);
 280 static void scrub_bio_end_io(struct bio *bio);
 281 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 282 static void scrub_block_complete(struct scrub_block *sblock);
 283 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 284                                u64 extent_logical, u64 extent_len,
 285                                u64 *extent_physical,
 286                                struct btrfs_device **extent_dev,
 287                                int *extent_mirror_num);
 288 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 289                                     struct scrub_page *spage);
 290 static void scrub_wr_submit(struct scrub_ctx *sctx);
 291 static void scrub_wr_bio_end_io(struct bio *bio);
 292 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 293 static int write_page_nocow(struct scrub_ctx *sctx,
 294                             u64 physical_for_dev_replace, struct page *page);
 295 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 296                                       struct scrub_copy_nocow_ctx *ctx);
 297 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 298                             int mirror_num, u64 physical_for_dev_replace);
 299 static void copy_nocow_pages_worker(struct btrfs_work *work);
 300 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 301 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 302 static void scrub_put_ctx(struct scrub_ctx *sctx);
 303
 304 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
 305 {
 306         return page->recover &&
 307                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
 308 }
 309
 310 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 311 {
 312         refcount_inc(&sctx->refs);
 313         atomic_inc(&sctx->bios_in_flight);
 314 }
 315
 316 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 317 {
 318         atomic_dec(&sctx->bios_in_flight);
 319         wake_up(&sctx->list_wait);
 320         scrub_put_ctx(sctx);
 321 }
 322
 323 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 324 {
 325         while (atomic_read(&fs_info->scrub_pause_req)) {
 326                 mutex_unlock(&fs_info->scrub_lock);
 327                 wait_event(fs_info->scrub_pause_wait,
 328                    atomic_read(&fs_info->scrub_pause_req) == 0);
 329                 mutex_lock(&fs_info->scrub_lock);
 330         }
 331 }
 332
 333 static void scrub_pause_on(struct btrfs_fs_info *fs_info)
 334 {
 335         atomic_inc(&fs_info->scrubs_paused);
 336         wake_up(&fs_info->scrub_pause_wait);
 337 }
 338
 339 static void scrub_pause_off(struct btrfs_fs_info *fs_info)
 340 {
 341         mutex_lock(&fs_info->scrub_lock);
 342         __scrub_blocked_if_needed(fs_info);
 343         atomic_dec(&fs_info->scrubs_paused);
 344         mutex_unlock(&fs_info->scrub_lock);
 345
 346         wake_up(&fs_info->scrub_pause_wait);
 347 }
 348
 349 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 350 {
 351         scrub_pause_on(fs_info);
 352         scrub_pause_off(fs_info);
 353 }
 354
 355 /*
 356  * Insert new full stripe lock into full stripe locks tree
 357  *
 358  * Return pointer to existing or newly inserted full_stripe_lock structure if
 359  * everything works well.
 360  * Return ERR_PTR(-ENOMEM) if we failed to allocate memory
 361  *
 362  * NOTE: caller must hold full_stripe_locks_root->lock before calling this
 363  * function
 364  */
 365 static struct full_stripe_lock *insert_full_stripe_lock(
 366                 struct btrfs_full_stripe_locks_tree *locks_root,
 367                 u64 fstripe_logical)
 368 {
 369         struct rb_node **p;
 370         struct rb_node *parent = NULL;
 371         struct full_stripe_lock *entry;
 372         struct full_stripe_lock *ret;
 373
 374         WARN_ON(!mutex_is_locked(&locks_root->lock));
 375
 376         p = &locks_root->root.rb_node;
 377         while (*p) {
 378                 parent = *p;
 379                 entry = rb_entry(parent, struct full_stripe_lock, node);
 380                 if (fstripe_logical < entry->logical) {
 381                         p = &(*p)->rb_left;
 382                 } else if (fstripe_logical > entry->logical) {
 383                         p = &(*p)->rb_right;
 384                 } else {
 385                         entry->refs++;
 386                         return entry;
 387                 }
 388         }
 389
 390         /* Insert new lock */
 391         ret = kmalloc(sizeof(*ret), GFP_KERNEL);
 392         if (!ret)
 393                 return ERR_PTR(-ENOMEM);
 394         ret->logical = fstripe_logical;
 395         ret->refs = 1;
 396         mutex_init(&ret->mutex);
 397
 398         rb_link_node(&ret->node, parent, p);
 399         rb_insert_color(&ret->node, &locks_root->root);
 400         return ret;
 401 }
 402
 403 /*
 404  * Search for a full stripe lock of a block group
 405  *
 406  * Return pointer to existing full stripe lock if found
 407  * Return NULL if not found
 408  */
 409 static struct full_stripe_lock *search_full_stripe_lock(
 410                 struct btrfs_full_stripe_locks_tree *locks_root,
 411                 u64 fstripe_logical)
 412 {
 413         struct rb_node *node;
 414         struct full_stripe_lock *entry;
 415
 416         WARN_ON(!mutex_is_locked(&locks_root->lock));
 417
 418         node = locks_root->root.rb_node;
 419         while (node) {
 420                 entry = rb_entry(node, struct full_stripe_lock, node);
 421                 if (fstripe_logical < entry->logical)
 422                         node = node->rb_left;
 423                 else if (fstripe_logical > entry->logical)
 424                         node = node->rb_right;
 425                 else
 426                         return entry;
 427         }
 428         return NULL;
 429 }
 430
 431 /*
 432  * Helper to get full stripe logical from a normal bytenr.
 433  *
 434  * Caller must ensure @cache is a RAID56 block group.
 435  */
 436 static u64 get_full_stripe_logical(struct btrfs_block_group_cache *cache,
 437                                    u64 bytenr)
 438 {
 439         u64 ret;
 440
 441         /*
 442          * Due to chunk item size limit, full stripe length should not be
 443          * larger than U32_MAX. Just a sanity check here.
 444          */
 445         WARN_ON_ONCE(cache->full_stripe_len >= U32_MAX);
 446
 447         /*
 448          * round_down() can only handle power of 2, while RAID56 full
 449          * stripe length can be 64KiB * n, so we need to manually round down.
 450          */
 451         ret = div64_u64(bytenr - cache->key.objectid, cache->full_stripe_len) *
 452                 cache->full_stripe_len + cache->key.objectid;
 453         return ret;
 454 }
 455
 456 /*
 457  * Lock a full stripe to avoid concurrency of recovery and read
 458  *
 459  * It's only used for profiles with parities (RAID5/6), for other profiles it
 460  * does nothing.
 461  *
 462  * Return 0 if we locked full stripe covering @bytenr, with a mutex held.
 463  * So caller must call unlock_full_stripe() at the same context.
 464  *
 465  * Return <0 if encounters error.
 466  */
 467 static int lock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 468                             bool *locked_ret)
 469 {
 470         struct btrfs_block_group_cache *bg_cache;
 471         struct btrfs_full_stripe_locks_tree *locks_root;
 472         struct full_stripe_lock *existing;
 473         u64 fstripe_start;
 474         int ret = 0;
 475
 476         *locked_ret = false;
 477         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 478         if (!bg_cache) {
 479                 ASSERT(0);
 480                 return -ENOENT;
 481         }
 482
 483         /* Profiles not based on parity don't need full stripe lock */
 484         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 485                 goto out;
 486         locks_root = &bg_cache->full_stripe_locks_root;
 487
 488         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 489
 490         /* Now insert the full stripe lock */
 491         mutex_lock(&locks_root->lock);
 492         existing = insert_full_stripe_lock(locks_root, fstripe_start);
 493         mutex_unlock(&locks_root->lock);
 494         if (IS_ERR(existing)) {
 495                 ret = PTR_ERR(existing);
 496                 goto out;
 497         }
 498         mutex_lock(&existing->mutex);
 499         *locked_ret = true;
 500 out:
 501         btrfs_put_block_group(bg_cache);
 502         return ret;
 503 }
 504
 505 /*
 506  * Unlock a full stripe.
 507  *
 508  * NOTE: Caller must ensure it's the same context calling corresponding
 509  * lock_full_stripe().
 510  *
 511  * Return 0 if we unlock full stripe without problem.
 512  * Return <0 for error
 513  */
 514 static int unlock_full_stripe(struct btrfs_fs_info *fs_info, u64 bytenr,
 515                               bool locked)
 516 {
 517         struct btrfs_block_group_cache *bg_cache;
 518         struct btrfs_full_stripe_locks_tree *locks_root;
 519         struct full_stripe_lock *fstripe_lock;
 520         u64 fstripe_start;
 521         bool freeit = false;
 522         int ret = 0;
 523
 524         /* If we didn't acquire full stripe lock, no need to continue */
 525         if (!locked)
 526                 return 0;
 527
 528         bg_cache = btrfs_lookup_block_group(fs_info, bytenr);
 529         if (!bg_cache) {
 530                 ASSERT(0);
 531                 return -ENOENT;
 532         }
 533         if (!(bg_cache->flags & BTRFS_BLOCK_GROUP_RAID56_MASK))
 534                 goto out;
 535
 536         locks_root = &bg_cache->full_stripe_locks_root;
 537         fstripe_start = get_full_stripe_logical(bg_cache, bytenr);
 538
 539         mutex_lock(&locks_root->lock);
 540         fstripe_lock = search_full_stripe_lock(locks_root, fstripe_start);
 541         /* Unpaired unlock_full_stripe() detected */
 542         if (!fstripe_lock) {
 543                 WARN_ON(1);
 544                 ret = -ENOENT;
 545                 mutex_unlock(&locks_root->lock);
 546                 goto out;
 547         }
 548
 549         if (fstripe_lock->refs == 0) {
 550                 WARN_ON(1);
 551                 btrfs_warn(fs_info, "full stripe lock at %llu refcount underflow",
 552                         fstripe_lock->logical);
 553         } else {
 554                 fstripe_lock->refs--;
 555         }
 556
 557         if (fstripe_lock->refs == 0) {
 558                 rb_erase(&fstripe_lock->node, &locks_root->root);
 559                 freeit = true;
 560         }
 561         mutex_unlock(&locks_root->lock);
 562
 563         mutex_unlock(&fstripe_lock->mutex);
 564         if (freeit)
 565                 kfree(fstripe_lock);
 566 out:
 567         btrfs_put_block_group(bg_cache);
 568         return ret;
 569 }
 570
 571 /*
 572  * used for workers that require transaction commits (i.e., for the
 573  * NOCOW case)
 574  */
 575 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 576 {
 577         struct btrfs_fs_info *fs_info = sctx->fs_info;
 578
 579         refcount_inc(&sctx->refs);
 580         /*
 581          * increment scrubs_running to prevent cancel requests from
 582          * completing as long as a worker is running. we must also
 583          * increment scrubs_paused to prevent deadlocking on pause
 584          * requests used for transactions commits (as the worker uses a
 585          * transaction context). it is safe to regard the worker
 586          * as paused for all matters practical. effectively, we only
 587          * avoid cancellation requests from completing.
 588          */
 589         mutex_lock(&fs_info->scrub_lock);
 590         atomic_inc(&fs_info->scrubs_running);
 591         atomic_inc(&fs_info->scrubs_paused);
 592         mutex_unlock(&fs_info->scrub_lock);
 593
 594         /*
 595          * check if @scrubs_running=@scrubs_paused condition
 596          * inside wait_event() is not an atomic operation.
 597          * which means we may inc/dec @scrub_running/paused
 598          * at any time. Let's wake up @scrub_pause_wait as
 599          * much as we can to let commit transaction blocked less.
 600          */
 601         wake_up(&fs_info->scrub_pause_wait);
 602
 603         atomic_inc(&sctx->workers_pending);
 604 }
 605
 606 /* used for workers that require transaction commits */
 607 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 608 {
 609         struct btrfs_fs_info *fs_info = sctx->fs_info;
 610
 611         /*
 612          * see scrub_pending_trans_workers_inc() why we're pretending
 613          * to be paused in the scrub counters
 614          */
 615         mutex_lock(&fs_info->scrub_lock);
 616         atomic_dec(&fs_info->scrubs_running);
 617         atomic_dec(&fs_info->scrubs_paused);
 618         mutex_unlock(&fs_info->scrub_lock);
 619         atomic_dec(&sctx->workers_pending);
 620         wake_up(&fs_info->scrub_pause_wait);
 621         wake_up(&sctx->list_wait);
 622         scrub_put_ctx(sctx);
 623 }
 624
 625 static void scrub_free_csums(struct scrub_ctx *sctx)
 626 {
 627         while (!list_empty(&sctx->csum_list)) {
 628                 struct btrfs_ordered_sum *sum;
 629                 sum = list_first_entry(&sctx->csum_list,
 630                                        struct btrfs_ordered_sum, list);
 631                 list_del(&sum->list);
 632                 kfree(sum);
 633         }
 634 }
 635
 636 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 637 {
 638         int i;
 639
 640         if (!sctx)
 641                 return;
 642
 643         /* this can happen when scrub is cancelled */
 644         if (sctx->curr != -1) {
 645                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 646
 647                 for (i = 0; i < sbio->page_count; i++) {
 648                         WARN_ON(!sbio->pagev[i]->page);
 649                         scrub_block_put(sbio->pagev[i]->sblock);
 650                 }
 651                 bio_put(sbio->bio);
 652         }
 653
 654         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 655                 struct scrub_bio *sbio = sctx->bios[i];
 656
 657                 if (!sbio)
 658                         break;
 659                 kfree(sbio);
 660         }
 661
 662         kfree(sctx->wr_curr_bio);
 663         scrub_free_csums(sctx);
 664         kfree(sctx);
 665 }
 666
 667 static void scrub_put_ctx(struct scrub_ctx *sctx)
 668 {
 669         if (refcount_dec_and_test(&sctx->refs))
 670                 scrub_free_ctx(sctx);
 671 }
 672
 673 static noinline_for_stack
 674 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 675 {
 676         struct scrub_ctx *sctx;
 677         int             i;
 678         struct btrfs_fs_info *fs_info = dev->fs_info;
 679
 680         sctx = kzalloc(sizeof(*sctx), GFP_KERNEL);
 681         if (!sctx)
 682                 goto nomem;
 683         refcount_set(&sctx->refs, 1);
 684         sctx->is_dev_replace = is_dev_replace;
 685         sctx->pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 686         sctx->curr = -1;
 687         sctx->fs_info = dev->fs_info;
 688         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 689                 struct scrub_bio *sbio;
 690
 691                 sbio = kzalloc(sizeof(*sbio), GFP_KERNEL);
 692                 if (!sbio)
 693                         goto nomem;
 694                 sctx->bios[i] = sbio;
 695
 696                 sbio->index = i;
 697                 sbio->sctx = sctx;
 698                 sbio->page_count = 0;
 699                 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
 700                                 scrub_bio_end_io_worker, NULL, NULL);
 701
 702                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 703                         sctx->bios[i]->next_free = i + 1;
 704                 else
 705                         sctx->bios[i]->next_free = -1;
 706         }
 707         sctx->first_free = 0;
 708         atomic_set(&sctx->bios_in_flight, 0);
 709         atomic_set(&sctx->workers_pending, 0);
 710         atomic_set(&sctx->cancel_req, 0);
 711         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 712         INIT_LIST_HEAD(&sctx->csum_list);
 713
 714         spin_lock_init(&sctx->list_lock);
 715         spin_lock_init(&sctx->stat_lock);
 716         init_waitqueue_head(&sctx->list_wait);
 717
 718         WARN_ON(sctx->wr_curr_bio != NULL);
 719         mutex_init(&sctx->wr_lock);
 720         sctx->wr_curr_bio = NULL;
 721         if (is_dev_replace) {
 722                 WARN_ON(!fs_info->dev_replace.tgtdev);
 723                 sctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
 724                 sctx->wr_tgtdev = fs_info->dev_replace.tgtdev;
 725                 sctx->flush_all_writes = false;
 726         }
 727
 728         return sctx;
 729
 730 nomem:
 731         scrub_free_ctx(sctx);
 732         return ERR_PTR(-ENOMEM);
 733 }
 734
 735 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 736                                      void *warn_ctx)
 737 {
 738         u64 isize;
 739         u32 nlink;
 740         int ret;
 741         int i;
 742         unsigned nofs_flag;
 743         struct extent_buffer *eb;
 744         struct btrfs_inode_item *inode_item;
 745         struct scrub_warning *swarn = warn_ctx;
 746         struct btrfs_fs_info *fs_info = swarn->dev->fs_info;
 747         struct inode_fs_paths *ipath = NULL;
 748         struct btrfs_root *local_root;
 749         struct btrfs_key root_key;
 750         struct btrfs_key key;
 751
 752         root_key.objectid = root;
 753         root_key.type = BTRFS_ROOT_ITEM_KEY;
 754         root_key.offset = (u64)-1;
 755         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 756         if (IS_ERR(local_root)) {
 757                 ret = PTR_ERR(local_root);
 758                 goto err;
 759         }
 760
 761         /*
 762          * this makes the path point to (inum INODE_ITEM ioff)
 763          */
 764         key.objectid = inum;
 765         key.type = BTRFS_INODE_ITEM_KEY;
 766         key.offset = 0;
 767
 768         ret = btrfs_search_slot(NULL, local_root, &key, swarn->path, 0, 0);
 769         if (ret) {
 770                 btrfs_release_path(swarn->path);
 771                 goto err;
 772         }
 773
 774         eb = swarn->path->nodes[0];
 775         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 776                                         struct btrfs_inode_item);
 777         isize = btrfs_inode_size(eb, inode_item);
 778         nlink = btrfs_inode_nlink(eb, inode_item);
 779         btrfs_release_path(swarn->path);
 780
 781         /*
 782          * init_path might indirectly call vmalloc, or use GFP_KERNEL. Scrub
 783          * uses GFP_NOFS in this context, so we keep it consistent but it does
 784          * not seem to be strictly necessary.
 785          */
 786         nofs_flag = memalloc_nofs_save();
 787         ipath = init_ipath(4096, local_root, swarn->path);
 788         memalloc_nofs_restore(nofs_flag);
 789         if (IS_ERR(ipath)) {
 790                 ret = PTR_ERR(ipath);
 791                 ipath = NULL;
 792                 goto err;
 793         }
 794         ret = paths_from_inode(inum, ipath);
 795
 796         if (ret < 0)
 797                 goto err;
 798
 799         /*
 800          * we deliberately ignore the bit ipath might have been too small to
 801          * hold all of the paths here
 802          */
 803         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 804                 btrfs_warn_in_rcu(fs_info,
 805                                   "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu, length %llu, links %u (path: %s)",
 806                                   swarn->errstr, swarn->logical,
 807                                   rcu_str_deref(swarn->dev->name),
 808                                   (unsigned long long)swarn->sector,
 809                                   root, inum, offset,
 810                                   min(isize - offset, (u64)PAGE_SIZE), nlink,
 811                                   (char *)(unsigned long)ipath->fspath->val[i]);
 812
 813         free_ipath(ipath);
 814         return 0;
 815
 816 err:
 817         btrfs_warn_in_rcu(fs_info,
 818                           "%s at logical %llu on dev %s, sector %llu, root %llu, inode %llu, offset %llu: path resolving failed with ret=%d",
 819                           swarn->errstr, swarn->logical,
 820                           rcu_str_deref(swarn->dev->name),
 821                           (unsigned long long)swarn->sector,
 822                           root, inum, offset, ret);
 823
 824         free_ipath(ipath);
 825         return 0;
 826 }
 827
 828 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 829 {
 830         struct btrfs_device *dev;
 831         struct btrfs_fs_info *fs_info;
 832         struct btrfs_path *path;
 833         struct btrfs_key found_key;
 834         struct extent_buffer *eb;
 835         struct btrfs_extent_item *ei;
 836         struct scrub_warning swarn;
 837         unsigned long ptr = 0;
 838         u64 extent_item_pos;
 839         u64 flags = 0;
 840         u64 ref_root;
 841         u32 item_size;
 842         u8 ref_level = 0;
 843         int ret;
 844
 845         WARN_ON(sblock->page_count < 1);
 846         dev = sblock->pagev[0]->dev;
 847         fs_info = sblock->sctx->fs_info;
 848
 849         path = btrfs_alloc_path();
 850         if (!path)
 851                 return;
 852
 853         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 854         swarn.logical = sblock->pagev[0]->logical;
 855         swarn.errstr = errstr;
 856         swarn.dev = NULL;
 857
 858         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 859                                   &flags);
 860         if (ret < 0)
 861                 goto out;
 862
 863         extent_item_pos = swarn.logical - found_key.objectid;
 864         swarn.extent_item_size = found_key.offset;
 865
 866         eb = path->nodes[0];
 867         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 868         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 869
 870         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 871                 do {
 872                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 873                                                       item_size, &ref_root,
 874                                                       &ref_level);
 875                         btrfs_warn_in_rcu(fs_info,
 876                                 "%s at logical %llu on dev %s, sector %llu: metadata %s (level %d) in tree %llu",
 877                                 errstr, swarn.logical,
 878                                 rcu_str_deref(dev->name),
 879                                 (unsigned long long)swarn.sector,
 880                                 ref_level ? "node" : "leaf",
 881                                 ret < 0 ? -1 : ref_level,
 882                                 ret < 0 ? -1 : ref_root);
 883                 } while (ret != 1);
 884                 btrfs_release_path(path);
 885         } else {
 886                 btrfs_release_path(path);
 887                 swarn.path = path;
 888                 swarn.dev = dev;
 889                 iterate_extent_inodes(fs_info, found_key.objectid,
 890                                         extent_item_pos, 1,
 891                                         scrub_print_warning_inode, &swarn);
 892         }
 893
 894 out:
 895         btrfs_free_path(path);
 896 }
 897
 898 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 899 {
 900         struct page *page = NULL;
 901         unsigned long index;
 902         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 903         int ret;
 904         int corrected = 0;
 905         struct btrfs_key key;
 906         struct inode *inode = NULL;
 907         struct btrfs_fs_info *fs_info;
 908         u64 end = offset + PAGE_SIZE - 1;
 909         struct btrfs_root *local_root;
 910         int srcu_index;
 911
 912         key.objectid = root;
 913         key.type = BTRFS_ROOT_ITEM_KEY;
 914         key.offset = (u64)-1;
 915
 916         fs_info = fixup->root->fs_info;
 917         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 918
 919         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 920         if (IS_ERR(local_root)) {
 921                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 922                 return PTR_ERR(local_root);
 923         }
 924
 925         key.type = BTRFS_INODE_ITEM_KEY;
 926         key.objectid = inum;
 927         key.offset = 0;
 928         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 929         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 930         if (IS_ERR(inode))
 931                 return PTR_ERR(inode);
 932
 933         index = offset >> PAGE_SHIFT;
 934
 935         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 936         if (!page) {
 937                 ret = -ENOMEM;
 938                 goto out;
 939         }
 940
 941         if (PageUptodate(page)) {
 942                 if (PageDirty(page)) {
 943                         /*
 944                          * we need to write the data to the defect sector. the
 945                          * data that was in that sector is not in memory,
 946                          * because the page was modified. we must not write the
 947                          * modified page to that sector.
 948                          *
 949                          * TODO: what could be done here: wait for the delalloc
 950                          *       runner to write out that page (might involve
 951                          *       COW) and see whether the sector is still
 952                          *       referenced afterwards.
 953                          *
 954                          * For the meantime, we'll treat this error
 955                          * incorrectable, although there is a chance that a
 956                          * later scrub will find the bad sector again and that
 957                          * there's no dirty page in memory, then.
 958                          */
 959                         ret = -EIO;
 960                         goto out;
 961                 }
 962                 ret = repair_io_failure(fs_info, inum, offset, PAGE_SIZE,
 963                                         fixup->logical, page,
 964                                         offset - page_offset(page),
 965                                         fixup->mirror_num);
 966                 unlock_page(page);
 967                 corrected = !ret;
 968         } else {
 969                 /*
 970                  * we need to get good data first. the general readpage path
 971                  * will call repair_io_failure for us, we just have to make
 972                  * sure we read the bad mirror.
 973                  */
 974                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 975                                         EXTENT_DAMAGED);
 976                 if (ret) {
 977                         /* set_extent_bits should give proper error */
 978                         WARN_ON(ret > 0);
 979                         if (ret > 0)
 980                                 ret = -EFAULT;
 981                         goto out;
 982                 }
 983
 984                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 985                                                 btrfs_get_extent,
 986                                                 fixup->mirror_num);
 987                 wait_on_page_locked(page);
 988
 989                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 990                                                 end, EXTENT_DAMAGED, 0, NULL);
 991                 if (!corrected)
 992                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 993                                                 EXTENT_DAMAGED);
 994         }
 995
 996 out:
 997         if (page)
 998                 put_page(page);
 999
1000         iput(inode);
1001
1002         if (ret < 0)
1003                 return ret;
1004
1005         if (ret == 0 && corrected) {
1006                 /*
1007                  * we only need to call readpage for one of the inodes belonging
1008                  * to this extent. so make iterate_extent_inodes stop
1009                  */
1010                 return 1;
1011         }
1012
1013         return -EIO;
1014 }
1015
1016 static void scrub_fixup_nodatasum(struct btrfs_work *work)
1017 {
1018         struct btrfs_fs_info *fs_info;
1019         int ret;
1020         struct scrub_fixup_nodatasum *fixup;
1021         struct scrub_ctx *sctx;
1022         struct btrfs_trans_handle *trans = NULL;
1023         struct btrfs_path *path;
1024         int uncorrectable = 0;
1025
1026         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
1027         sctx = fixup->sctx;
1028         fs_info = fixup->root->fs_info;
1029
1030         path = btrfs_alloc_path();
1031         if (!path) {
1032                 spin_lock(&sctx->stat_lock);
1033                 ++sctx->stat.malloc_errors;
1034                 spin_unlock(&sctx->stat_lock);
1035                 uncorrectable = 1;
1036                 goto out;
1037         }
1038
1039         trans = btrfs_join_transaction(fixup->root);
1040         if (IS_ERR(trans)) {
1041                 uncorrectable = 1;
1042                 goto out;
1043         }
1044
1045         /*
1046          * the idea is to trigger a regular read through the standard path. we
1047          * read a page from the (failed) logical address by specifying the
1048          * corresponding copynum of the failed sector. thus, that readpage is
1049          * expected to fail.
1050          * that is the point where on-the-fly error correction will kick in
1051          * (once it's finished) and rewrite the failed sector if a good copy
1052          * can be found.
1053          */
1054         ret = iterate_inodes_from_logical(fixup->logical, fs_info, path,
1055                                           scrub_fixup_readpage, fixup);
1056         if (ret < 0) {
1057                 uncorrectable = 1;
1058                 goto out;
1059         }
1060         WARN_ON(ret != 1);
1061
1062         spin_lock(&sctx->stat_lock);
1063         ++sctx->stat.corrected_errors;
1064         spin_unlock(&sctx->stat_lock);
1065
1066 out:
1067         if (trans && !IS_ERR(trans))
1068                 btrfs_end_transaction(trans);
1069         if (uncorrectable) {
1070                 spin_lock(&sctx->stat_lock);
1071                 ++sctx->stat.uncorrectable_errors;
1072                 spin_unlock(&sctx->stat_lock);
1073                 btrfs_dev_replace_stats_inc(
1074                         &fs_info->dev_replace.num_uncorrectable_read_errors);
1075                 btrfs_err_rl_in_rcu(fs_info,
1076                     "unable to fixup (nodatasum) error at logical %llu on dev %s",
1077                         fixup->logical, rcu_str_deref(fixup->dev->name));
1078         }
1079
1080         btrfs_free_path(path);
1081         kfree(fixup);
1082
1083         scrub_pending_trans_workers_dec(sctx);
1084 }
1085
1086 static inline void scrub_get_recover(struct scrub_recover *recover)
1087 {
1088         refcount_inc(&recover->refs);
1089 }
1090
1091 static inline void scrub_put_recover(struct btrfs_fs_info *fs_info,
1092                                      struct scrub_recover *recover)
1093 {
1094         if (refcount_dec_and_test(&recover->refs)) {
1095                 btrfs_bio_counter_dec(fs_info);
1096                 btrfs_put_bbio(recover->bbio);
1097                 kfree(recover);
1098         }
1099 }
1100
1101 /*
1102  * scrub_handle_errored_block gets called when either verification of the
1103  * pages failed or the bio failed to read, e.g. with EIO. In the latter
1104  * case, this function handles all pages in the bio, even though only one
1105  * may be bad.
1106  * The goal of this function is to repair the errored block by using the
1107  * contents of one of the mirrors.
1108  */
1109 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
1110 {
1111         struct scrub_ctx *sctx = sblock_to_check->sctx;
1112         struct btrfs_device *dev;
1113         struct btrfs_fs_info *fs_info;
1114         u64 length;
1115         u64 logical;
1116         unsigned int failed_mirror_index;
1117         unsigned int is_metadata;
1118         unsigned int have_csum;
1119         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
1120         struct scrub_block *sblock_bad;
1121         int ret;
1122         int mirror_index;
1123         int page_num;
1124         int success;
1125         bool full_stripe_locked;
1126         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1127                                       DEFAULT_RATELIMIT_BURST);
1128
1129         BUG_ON(sblock_to_check->page_count < 1);
1130         fs_info = sctx->fs_info;
1131         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
1132                 /*
1133                  * if we find an error in a super block, we just report it.
1134                  * They will get written with the next transaction commit
1135                  * anyway
1136                  */
1137                 spin_lock(&sctx->stat_lock);
1138                 ++sctx->stat.super_errors;
1139                 spin_unlock(&sctx->stat_lock);
1140                 return 0;
1141         }
1142         length = sblock_to_check->page_count * PAGE_SIZE;
1143         logical = sblock_to_check->pagev[0]->logical;
1144         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
1145         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
1146         is_metadata = !(sblock_to_check->pagev[0]->flags &
1147                         BTRFS_EXTENT_FLAG_DATA);
1148         have_csum = sblock_to_check->pagev[0]->have_csum;
1149         dev = sblock_to_check->pagev[0]->dev;
1150
1151         /*
1152          * For RAID5/6, race can happen for a different device scrub thread.
1153          * For data corruption, Parity and Data threads will both try
1154          * to recovery the data.
1155          * Race can lead to doubly added csum error, or even unrecoverable
1156          * error.
1157          */
1158         ret = lock_full_stripe(fs_info, logical, &full_stripe_locked);
1159         if (ret < 0) {
1160                 spin_lock(&sctx->stat_lock);
1161                 if (ret == -ENOMEM)
1162                         sctx->stat.malloc_errors++;
1163                 sctx->stat.read_errors++;
1164                 sctx->stat.uncorrectable_errors++;
1165                 spin_unlock(&sctx->stat_lock);
1166                 return ret;
1167         }
1168
1169         /*
1170          * read all mirrors one after the other. This includes to
1171          * re-read the extent or metadata block that failed (that was
1172          * the cause that this fixup code is called) another time,
1173          * page by page this time in order to know which pages
1174          * caused I/O errors and which ones are good (for all mirrors).
1175          * It is the goal to handle the situation when more than one
1176          * mirror contains I/O errors, but the errors do not
1177          * overlap, i.e. the data can be repaired by selecting the
1178          * pages from those mirrors without I/O error on the
1179          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
1180          * would be that mirror #1 has an I/O error on the first page,
1181          * the second page is good, and mirror #2 has an I/O error on
1182          * the second page, but the first page is good.
1183          * Then the first page of the first mirror can be repaired by
1184          * taking the first page of the second mirror, and the
1185          * second page of the second mirror can be repaired by
1186          * copying the contents of the 2nd page of the 1st mirror.
1187          * One more note: if the pages of one mirror contain I/O
1188          * errors, the checksum cannot be verified. In order to get
1189          * the best data for repairing, the first attempt is to find
1190          * a mirror without I/O errors and with a validated checksum.
1191          * Only if this is not possible, the pages are picked from
1192          * mirrors with I/O errors without considering the checksum.
1193          * If the latter is the case, at the end, the checksum of the
1194          * repaired area is verified in order to correctly maintain
1195          * the statistics.
1196          */
1197
1198         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
1199                                       sizeof(*sblocks_for_recheck), GFP_NOFS);
1200         if (!sblocks_for_recheck) {
1201                 spin_lock(&sctx->stat_lock);
1202                 sctx->stat.malloc_errors++;
1203                 sctx->stat.read_errors++;
1204                 sctx->stat.uncorrectable_errors++;
1205                 spin_unlock(&sctx->stat_lock);
1206                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1207                 goto out;
1208         }
1209
1210         /* setup the context, map the logical blocks and alloc the pages */
1211         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
1212         if (ret) {
1213                 spin_lock(&sctx->stat_lock);
1214                 sctx->stat.read_errors++;
1215                 sctx->stat.uncorrectable_errors++;
1216                 spin_unlock(&sctx->stat_lock);
1217                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1218                 goto out;
1219         }
1220         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
1221         sblock_bad = sblocks_for_recheck + failed_mirror_index;
1222
1223         /* build and submit the bios for the failed mirror, check checksums */
1224         scrub_recheck_block(fs_info, sblock_bad, 1);
1225
1226         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
1227             sblock_bad->no_io_error_seen) {
1228                 /*
1229                  * the error disappeared after reading page by page, or
1230                  * the area was part of a huge bio and other parts of the
1231                  * bio caused I/O errors, or the block layer merged several
1232                  * read requests into one and the error is caused by a
1233                  * different bio (usually one of the two latter cases is
1234                  * the cause)
1235                  */
1236                 spin_lock(&sctx->stat_lock);
1237                 sctx->stat.unverified_errors++;
1238                 sblock_to_check->data_corrected = 1;
1239                 spin_unlock(&sctx->stat_lock);
1240
1241                 if (sctx->is_dev_replace)
1242                         scrub_write_block_to_dev_replace(sblock_bad);
1243                 goto out;
1244         }
1245
1246         if (!sblock_bad->no_io_error_seen) {
1247                 spin_lock(&sctx->stat_lock);
1248                 sctx->stat.read_errors++;
1249                 spin_unlock(&sctx->stat_lock);
1250                 if (__ratelimit(&_rs))
1251                         scrub_print_warning("i/o error", sblock_to_check);
1252                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1253         } else if (sblock_bad->checksum_error) {
1254                 spin_lock(&sctx->stat_lock);
1255                 sctx->stat.csum_errors++;
1256                 spin_unlock(&sctx->stat_lock);
1257                 if (__ratelimit(&_rs))
1258                         scrub_print_warning("checksum error", sblock_to_check);
1259                 btrfs_dev_stat_inc_and_print(dev,
1260                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
1261         } else if (sblock_bad->header_error) {
1262                 spin_lock(&sctx->stat_lock);
1263                 sctx->stat.verify_errors++;
1264                 spin_unlock(&sctx->stat_lock);
1265                 if (__ratelimit(&_rs))
1266                         scrub_print_warning("checksum/header error",
1267                                             sblock_to_check);
1268                 if (sblock_bad->generation_error)
1269                         btrfs_dev_stat_inc_and_print(dev,
1270                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1271                 else
1272                         btrfs_dev_stat_inc_and_print(dev,
1273                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1274         }
1275
1276         if (sctx->readonly) {
1277                 ASSERT(!sctx->is_dev_replace);
1278                 goto out;
1279         }
1280
1281         /*
1282          * NOTE: Even for nodatasum case, it's still possible that it's a
1283          * compressed data extent, thus scrub_fixup_nodatasum(), which write
1284          * inode page cache onto disk, could cause serious data corruption.
1285          *
1286          * So here we could only read from disk, and hope our recovery could
1287          * reach disk before the newer write.
1288          */
1289         if (0 && !is_metadata && !have_csum) {
1290                 struct scrub_fixup_nodatasum *fixup_nodatasum;
1291
1292                 WARN_ON(sctx->is_dev_replace);
1293
1294                 /*
1295                  * !is_metadata and !have_csum, this means that the data
1296                  * might not be COWed, that it might be modified
1297                  * concurrently. The general strategy to work on the
1298                  * commit root does not help in the case when COW is not
1299                  * used.
1300                  */
1301                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1302                 if (!fixup_nodatasum)
1303                         goto did_not_correct_error;
1304                 fixup_nodatasum->sctx = sctx;
1305                 fixup_nodatasum->dev = dev;
1306                 fixup_nodatasum->logical = logical;
1307                 fixup_nodatasum->root = fs_info->extent_root;
1308                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1309                 scrub_pending_trans_workers_inc(sctx);
1310                 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1311                                 scrub_fixup_nodatasum, NULL, NULL);
1312                 btrfs_queue_work(fs_info->scrub_workers,
1313                                  &fixup_nodatasum->work);
1314                 goto out;
1315         }
1316
1317         /*
1318          * now build and submit the bios for the other mirrors, check
1319          * checksums.
1320          * First try to pick the mirror which is completely without I/O
1321          * errors and also does not have a checksum error.
1322          * If one is found, and if a checksum is present, the full block
1323          * that is known to contain an error is rewritten. Afterwards
1324          * the block is known to be corrected.
1325          * If a mirror is found which is completely correct, and no
1326          * checksum is present, only those pages are rewritten that had
1327          * an I/O error in the block to be repaired, since it cannot be
1328          * determined, which copy of the other pages is better (and it
1329          * could happen otherwise that a correct page would be
1330          * overwritten by a bad one).
1331          */
1332         for (mirror_index = 0; ;mirror_index++) {
1333                 struct scrub_block *sblock_other;
1334
1335                 if (mirror_index == failed_mirror_index)
1336                         continue;
1337
1338                 /* raid56's mirror can be more than BTRFS_MAX_MIRRORS */
1339                 if (!scrub_is_page_on_raid56(sblock_bad->pagev[0])) {
1340                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1341                                 break;
1342                         if (!sblocks_for_recheck[mirror_index].page_count)
1343                                 break;
1344
1345                         sblock_other = sblocks_for_recheck + mirror_index;
1346                 } else {
1347                         struct scrub_recover *r = sblock_bad->pagev[0]->recover;
1348                         int max_allowed = r->bbio->num_stripes -
1349                                                 r->bbio->num_tgtdevs;
1350
1351                         if (mirror_index >= max_allowed)
1352                                 break;
1353                         if (!sblocks_for_recheck[1].page_count)
1354                                 break;
1355
1356                         ASSERT(failed_mirror_index == 0);
1357                         sblock_other = sblocks_for_recheck + 1;
1358                         sblock_other->pagev[0]->mirror_num = 1 + mirror_index;
1359                 }
1360
1361                 /* build and submit the bios, check checksums */
1362                 scrub_recheck_block(fs_info, sblock_other, 0);
1363
1364                 if (!sblock_other->header_error &&
1365                     !sblock_other->checksum_error &&
1366                     sblock_other->no_io_error_seen) {
1367                         if (sctx->is_dev_replace) {
1368                                 scrub_write_block_to_dev_replace(sblock_other);
1369                                 goto corrected_error;
1370                         } else {
1371                                 ret = scrub_repair_block_from_good_copy(
1372                                                 sblock_bad, sblock_other);
1373                                 if (!ret)
1374                                         goto corrected_error;
1375                         }
1376                 }
1377         }
1378
1379         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1380                 goto did_not_correct_error;
1381
1382         /*
1383          * In case of I/O errors in the area that is supposed to be
1384          * repaired, continue by picking good copies of those pages.
1385          * Select the good pages from mirrors to rewrite bad pages from
1386          * the area to fix. Afterwards verify the checksum of the block
1387          * that is supposed to be repaired. This verification step is
1388          * only done for the purpose of statistic counting and for the
1389          * final scrub report, whether errors remain.
1390          * A perfect algorithm could make use of the checksum and try
1391          * all possible combinations of pages from the different mirrors
1392          * until the checksum verification succeeds. For example, when
1393          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1394          * of mirror #2 is readable but the final checksum test fails,
1395          * then the 2nd page of mirror #3 could be tried, whether now
1396          * the final checksum succeeds. But this would be a rare
1397          * exception and is therefore not implemented. At least it is
1398          * avoided that the good copy is overwritten.
1399          * A more useful improvement would be to pick the sectors
1400          * without I/O error based on sector sizes (512 bytes on legacy
1401          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1402          * mirror could be repaired by taking 512 byte of a different
1403          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1404          * area are unreadable.
1405          */
1406         success = 1;
1407         for (page_num = 0; page_num < sblock_bad->page_count;
1408              page_num++) {
1409                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1410                 struct scrub_block *sblock_other = NULL;
1411
1412                 /* skip no-io-error page in scrub */
1413                 if (!page_bad->io_error && !sctx->is_dev_replace)
1414                         continue;
1415
1416                 /* try to find no-io-error page in mirrors */
1417                 if (page_bad->io_error) {
1418                         for (mirror_index = 0;
1419                              mirror_index < BTRFS_MAX_MIRRORS &&
1420                              sblocks_for_recheck[mirror_index].page_count > 0;
1421                              mirror_index++) {
1422                                 if (!sblocks_for_recheck[mirror_index].
1423                                     pagev[page_num]->io_error) {
1424                                         sblock_other = sblocks_for_recheck +
1425                                                        mirror_index;
1426                                         break;
1427                                 }
1428                         }
1429                         if (!sblock_other)
1430                                 success = 0;
1431                 }
1432
1433                 if (sctx->is_dev_replace) {
1434                         /*
1435                          * did not find a mirror to fetch the page
1436                          * from. scrub_write_page_to_dev_replace()
1437                          * handles this case (page->io_error), by
1438                          * filling the block with zeros before
1439                          * submitting the write request
1440                          */
1441                         if (!sblock_other)
1442                                 sblock_other = sblock_bad;
1443
1444                         if (scrub_write_page_to_dev_replace(sblock_other,
1445                                                             page_num) != 0) {
1446                                 btrfs_dev_replace_stats_inc(
1447                                         &fs_info->dev_replace.num_write_errors);
1448                                 success = 0;
1449                         }
1450                 } else if (sblock_other) {
1451                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1452                                                                sblock_other,
1453                                                                page_num, 0);
1454                         if (0 == ret)
1455                                 page_bad->io_error = 0;
1456                         else
1457                                 success = 0;
1458                 }
1459         }
1460
1461         if (success && !sctx->is_dev_replace) {
1462                 if (is_metadata || have_csum) {
1463                         /*
1464                          * need to verify the checksum now that all
1465                          * sectors on disk are repaired (the write
1466                          * request for data to be repaired is on its way).
1467                          * Just be lazy and use scrub_recheck_block()
1468                          * which re-reads the data before the checksum
1469                          * is verified, but most likely the data comes out
1470                          * of the page cache.
1471                          */
1472                         scrub_recheck_block(fs_info, sblock_bad, 1);
1473                         if (!sblock_bad->header_error &&
1474                             !sblock_bad->checksum_error &&
1475                             sblock_bad->no_io_error_seen)
1476                                 goto corrected_error;
1477                         else
1478                                 goto did_not_correct_error;
1479                 } else {
1480 corrected_error:
1481                         spin_lock(&sctx->stat_lock);
1482                         sctx->stat.corrected_errors++;
1483                         sblock_to_check->data_corrected = 1;
1484                         spin_unlock(&sctx->stat_lock);
1485                         btrfs_err_rl_in_rcu(fs_info,
1486                                 "fixed up error at logical %llu on dev %s",
1487                                 logical, rcu_str_deref(dev->name));
1488                 }
1489         } else {
1490 did_not_correct_error:
1491                 spin_lock(&sctx->stat_lock);
1492                 sctx->stat.uncorrectable_errors++;
1493                 spin_unlock(&sctx->stat_lock);
1494                 btrfs_err_rl_in_rcu(fs_info,
1495                         "unable to fixup (regular) error at logical %llu on dev %s",
1496                         logical, rcu_str_deref(dev->name));
1497         }
1498
1499 out:
1500         if (sblocks_for_recheck) {
1501                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1502                      mirror_index++) {
1503                         struct scrub_block *sblock = sblocks_for_recheck +
1504                                                      mirror_index;
1505                         struct scrub_recover *recover;
1506                         int page_index;
1507
1508                         for (page_index = 0; page_index < sblock->page_count;
1509                              page_index++) {
1510                                 sblock->pagev[page_index]->sblock = NULL;
1511                                 recover = sblock->pagev[page_index]->recover;
1512                                 if (recover) {
1513                                         scrub_put_recover(fs_info, recover);
1514                                         sblock->pagev[page_index]->recover =
1515                                                                         NULL;
1516                                 }
1517                                 scrub_page_put(sblock->pagev[page_index]);
1518                         }
1519                 }
1520                 kfree(sblocks_for_recheck);
1521         }
1522
1523         ret = unlock_full_stripe(fs_info, logical, full_stripe_locked);
1524         if (ret < 0)
1525                 return ret;
1526         return 0;
1527 }
1528
1529 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1530 {
1531         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1532                 return 2;
1533         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1534                 return 3;
1535         else
1536                 return (int)bbio->num_stripes;
1537 }
1538
1539 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1540                                                  u64 *raid_map,
1541                                                  u64 mapped_length,
1542                                                  int nstripes, int mirror,
1543                                                  int *stripe_index,
1544                                                  u64 *stripe_offset)
1545 {
1546         int i;
1547
1548         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1549                 /* RAID5/6 */
1550                 for (i = 0; i < nstripes; i++) {
1551                         if (raid_map[i] == RAID6_Q_STRIPE ||
1552                             raid_map[i] == RAID5_P_STRIPE)
1553                                 continue;
1554
1555                         if (logical >= raid_map[i] &&
1556                             logical < raid_map[i] + mapped_length)
1557                                 break;
1558                 }
1559
1560                 *stripe_index = i;
1561                 *stripe_offset = logical - raid_map[i];
1562         } else {
1563                 /* The other RAID type */
1564                 *stripe_index = mirror;
1565                 *stripe_offset = 0;
1566         }
1567 }
1568
1569 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1570                                      struct scrub_block *sblocks_for_recheck)
1571 {
1572         struct scrub_ctx *sctx = original_sblock->sctx;
1573         struct btrfs_fs_info *fs_info = sctx->fs_info;
1574         u64 length = original_sblock->page_count * PAGE_SIZE;
1575         u64 logical = original_sblock->pagev[0]->logical;
1576         u64 generation = original_sblock->pagev[0]->generation;
1577         u64 flags = original_sblock->pagev[0]->flags;
1578         u64 have_csum = original_sblock->pagev[0]->have_csum;
1579         struct scrub_recover *recover;
1580         struct btrfs_bio *bbio;
1581         u64 sublen;
1582         u64 mapped_length;
1583         u64 stripe_offset;
1584         int stripe_index;
1585         int page_index = 0;
1586         int mirror_index;
1587         int nmirrors;
1588         int ret;
1589
1590         /*
1591          * note: the two members refs and outstanding_pages
1592          * are not used (and not set) in the blocks that are used for
1593          * the recheck procedure
1594          */
1595
1596         while (length > 0) {
1597                 sublen = min_t(u64, length, PAGE_SIZE);
1598                 mapped_length = sublen;
1599                 bbio = NULL;
1600
1601                 /*
1602                  * with a length of PAGE_SIZE, each returned stripe
1603                  * represents one mirror
1604                  */
1605                 btrfs_bio_counter_inc_blocked(fs_info);
1606                 ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS,
1607                                 logical, &mapped_length, &bbio);
1608                 if (ret || !bbio || mapped_length < sublen) {
1609                         btrfs_put_bbio(bbio);
1610                         btrfs_bio_counter_dec(fs_info);
1611                         return -EIO;
1612                 }
1613
1614                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1615                 if (!recover) {
1616                         btrfs_put_bbio(bbio);
1617                         btrfs_bio_counter_dec(fs_info);
1618                         return -ENOMEM;
1619                 }
1620
1621                 refcount_set(&recover->refs, 1);
1622                 recover->bbio = bbio;
1623                 recover->map_length = mapped_length;
1624
1625                 BUG_ON(page_index >= SCRUB_MAX_PAGES_PER_BLOCK);
1626
1627                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1628
1629                 for (mirror_index = 0; mirror_index < nmirrors;
1630                      mirror_index++) {
1631                         struct scrub_block *sblock;
1632                         struct scrub_page *page;
1633
1634                         sblock = sblocks_for_recheck + mirror_index;
1635                         sblock->sctx = sctx;
1636
1637                         page = kzalloc(sizeof(*page), GFP_NOFS);
1638                         if (!page) {
1639 leave_nomem:
1640                                 spin_lock(&sctx->stat_lock);
1641                                 sctx->stat.malloc_errors++;
1642                                 spin_unlock(&sctx->stat_lock);
1643                                 scrub_put_recover(fs_info, recover);
1644                                 return -ENOMEM;
1645                         }
1646                         scrub_page_get(page);
1647                         sblock->pagev[page_index] = page;
1648                         page->sblock = sblock;
1649                         page->flags = flags;
1650                         page->generation = generation;
1651                         page->logical = logical;
1652                         page->have_csum = have_csum;
1653                         if (have_csum)
1654                                 memcpy(page->csum,
1655                                        original_sblock->pagev[0]->csum,
1656                                        sctx->csum_size);
1657
1658                         scrub_stripe_index_and_offset(logical,
1659                                                       bbio->map_type,
1660                                                       bbio->raid_map,
1661                                                       mapped_length,
1662                                                       bbio->num_stripes -
1663                                                       bbio->num_tgtdevs,
1664                                                       mirror_index,
1665                                                       &stripe_index,
1666                                                       &stripe_offset);
1667                         page->physical = bbio->stripes[stripe_index].physical +
1668                                          stripe_offset;
1669                         page->dev = bbio->stripes[stripe_index].dev;
1670
1671                         BUG_ON(page_index >= original_sblock->page_count);
1672                         page->physical_for_dev_replace =
1673                                 original_sblock->pagev[page_index]->
1674                                 physical_for_dev_replace;
1675                         /* for missing devices, dev->bdev is NULL */
1676                         page->mirror_num = mirror_index + 1;
1677                         sblock->page_count++;
1678                         page->page = alloc_page(GFP_NOFS);
1679                         if (!page->page)
1680                                 goto leave_nomem;
1681
1682                         scrub_get_recover(recover);
1683                         page->recover = recover;
1684                 }
1685                 scrub_put_recover(fs_info, recover);
1686                 length -= sublen;
1687                 logical += sublen;
1688                 page_index++;
1689         }
1690
1691         return 0;
1692 }
1693
1694 struct scrub_bio_ret {
1695         struct completion event;
1696         blk_status_t status;
1697 };
1698
1699 static void scrub_bio_wait_endio(struct bio *bio)
1700 {
1701         struct scrub_bio_ret *ret = bio->bi_private;
1702
1703         ret->status = bio->bi_status;
1704         complete(&ret->event);
1705 }
1706
1707 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1708                                         struct bio *bio,
1709                                         struct scrub_page *page)
1710 {
1711         struct scrub_bio_ret done;
1712         int ret;
1713         int mirror_num;
1714
1715         init_completion(&done.event);
1716         done.status = 0;
1717         bio->bi_iter.bi_sector = page->logical >> 9;
1718         bio->bi_private = &done;
1719         bio->bi_end_io = scrub_bio_wait_endio;
1720
1721         mirror_num = page->sblock->pagev[0]->mirror_num;
1722         ret = raid56_parity_recover(fs_info, bio, page->recover->bbio,
1723                                     page->recover->map_length,
1724                                     mirror_num, 0);
1725         if (ret)
1726                 return ret;
1727
1728         wait_for_completion_io(&done.event);
1729         if (done.status)
1730                 return -EIO;
1731
1732         return 0;
1733 }
1734
1735 /*
1736  * this function will check the on disk data for checksum errors, header
1737  * errors and read I/O errors. If any I/O errors happen, the exact pages
1738  * which are errored are marked as being bad. The goal is to enable scrub
1739  * to take those pages that are not errored from all the mirrors so that
1740  * the pages that are errored in the just handled mirror can be repaired.
1741  */
1742 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1743                                 struct scrub_block *sblock,
1744                                 int retry_failed_mirror)
1745 {
1746         int page_num;
1747
1748         sblock->no_io_error_seen = 1;
1749
1750         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1751                 struct bio *bio;
1752                 struct scrub_page *page = sblock->pagev[page_num];
1753
1754                 if (page->dev->bdev == NULL) {
1755                         page->io_error = 1;
1756                         sblock->no_io_error_seen = 0;
1757                         continue;
1758                 }
1759
1760                 WARN_ON(!page->page);
1761                 bio = btrfs_io_bio_alloc(1);
1762                 bio_set_dev(bio, page->dev->bdev);
1763
1764                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1765                 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1766                         if (scrub_submit_raid56_bio_wait(fs_info, bio, page)) {
1767                                 page->io_error = 1;
1768                                 sblock->no_io_error_seen = 0;
1769                         }
1770                 } else {
1771                         bio->bi_iter.bi_sector = page->physical >> 9;
1772                         bio_set_op_attrs(bio, REQ_OP_READ, 0);
1773
1774                         if (btrfsic_submit_bio_wait(bio)) {
1775                                 page->io_error = 1;
1776                                 sblock->no_io_error_seen = 0;
1777                         }
1778                 }
1779
1780                 bio_put(bio);
1781         }
1782
1783         if (sblock->no_io_error_seen)
1784                 scrub_recheck_block_checksum(sblock);
1785 }
1786
1787 static inline int scrub_check_fsid(u8 fsid[],
1788                                    struct scrub_page *spage)
1789 {
1790         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1791         int ret;
1792
1793         ret = memcmp(fsid, fs_devices->fsid, BTRFS_FSID_SIZE);
1794         return !ret;
1795 }
1796
1797 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1798 {
1799         sblock->header_error = 0;
1800         sblock->checksum_error = 0;
1801         sblock->generation_error = 0;
1802
1803         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1804                 scrub_checksum_data(sblock);
1805         else
1806                 scrub_checksum_tree_block(sblock);
1807 }
1808
1809 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1810                                              struct scrub_block *sblock_good)
1811 {
1812         int page_num;
1813         int ret = 0;
1814
1815         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1816                 int ret_sub;
1817
1818                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1819                                                            sblock_good,
1820                                                            page_num, 1);
1821                 if (ret_sub)
1822                         ret = ret_sub;
1823         }
1824
1825         return ret;
1826 }
1827
1828 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1829                                             struct scrub_block *sblock_good,
1830                                             int page_num, int force_write)
1831 {
1832         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1833         struct scrub_page *page_good = sblock_good->pagev[page_num];
1834         struct btrfs_fs_info *fs_info = sblock_bad->sctx->fs_info;
1835
1836         BUG_ON(page_bad->page == NULL);
1837         BUG_ON(page_good->page == NULL);
1838         if (force_write || sblock_bad->header_error ||
1839             sblock_bad->checksum_error || page_bad->io_error) {
1840                 struct bio *bio;
1841                 int ret;
1842
1843                 if (!page_bad->dev->bdev) {
1844                         btrfs_warn_rl(fs_info,
1845                                 "scrub_repair_page_from_good_copy(bdev == NULL) is unexpected");
1846                         return -EIO;
1847                 }
1848
1849                 bio = btrfs_io_bio_alloc(1);
1850                 bio_set_dev(bio, page_bad->dev->bdev);
1851                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1852                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1853
1854                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1855                 if (PAGE_SIZE != ret) {
1856                         bio_put(bio);
1857                         return -EIO;
1858                 }
1859
1860                 if (btrfsic_submit_bio_wait(bio)) {
1861                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1862                                 BTRFS_DEV_STAT_WRITE_ERRS);
1863                         btrfs_dev_replace_stats_inc(
1864                                 &fs_info->dev_replace.num_write_errors);
1865                         bio_put(bio);
1866                         return -EIO;
1867                 }
1868                 bio_put(bio);
1869         }
1870
1871         return 0;
1872 }
1873
1874 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1875 {
1876         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
1877         int page_num;
1878
1879         /*
1880          * This block is used for the check of the parity on the source device,
1881          * so the data needn't be written into the destination device.
1882          */
1883         if (sblock->sparity)
1884                 return;
1885
1886         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1887                 int ret;
1888
1889                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1890                 if (ret)
1891                         btrfs_dev_replace_stats_inc(
1892                                 &fs_info->dev_replace.num_write_errors);
1893         }
1894 }
1895
1896 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1897                                            int page_num)
1898 {
1899         struct scrub_page *spage = sblock->pagev[page_num];
1900
1901         BUG_ON(spage->page == NULL);
1902         if (spage->io_error) {
1903                 void *mapped_buffer = kmap_atomic(spage->page);
1904
1905                 clear_page(mapped_buffer);
1906                 flush_dcache_page(spage->page);
1907                 kunmap_atomic(mapped_buffer);
1908         }
1909         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1910 }
1911
1912 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1913                                     struct scrub_page *spage)
1914 {
1915         struct scrub_bio *sbio;
1916         int ret;
1917
1918         mutex_lock(&sctx->wr_lock);
1919 again:
1920         if (!sctx->wr_curr_bio) {
1921                 sctx->wr_curr_bio = kzalloc(sizeof(*sctx->wr_curr_bio),
1922                                               GFP_KERNEL);
1923                 if (!sctx->wr_curr_bio) {
1924                         mutex_unlock(&sctx->wr_lock);
1925                         return -ENOMEM;
1926                 }
1927                 sctx->wr_curr_bio->sctx = sctx;
1928                 sctx->wr_curr_bio->page_count = 0;
1929         }
1930         sbio = sctx->wr_curr_bio;
1931         if (sbio->page_count == 0) {
1932                 struct bio *bio;
1933
1934                 sbio->physical = spage->physical_for_dev_replace;
1935                 sbio->logical = spage->logical;
1936                 sbio->dev = sctx->wr_tgtdev;
1937                 bio = sbio->bio;
1938                 if (!bio) {
1939                         bio = btrfs_io_bio_alloc(sctx->pages_per_wr_bio);
1940                         sbio->bio = bio;
1941                 }
1942
1943                 bio->bi_private = sbio;
1944                 bio->bi_end_io = scrub_wr_bio_end_io;
1945                 bio_set_dev(bio, sbio->dev->bdev);
1946                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1947                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
1948                 sbio->status = 0;
1949         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1950                    spage->physical_for_dev_replace ||
1951                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1952                    spage->logical) {
1953                 scrub_wr_submit(sctx);
1954                 goto again;
1955         }
1956
1957         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1958         if (ret != PAGE_SIZE) {
1959                 if (sbio->page_count < 1) {
1960                         bio_put(sbio->bio);
1961                         sbio->bio = NULL;
1962                         mutex_unlock(&sctx->wr_lock);
1963                         return -EIO;
1964                 }
1965                 scrub_wr_submit(sctx);
1966                 goto again;
1967         }
1968
1969         sbio->pagev[sbio->page_count] = spage;
1970         scrub_page_get(spage);
1971         sbio->page_count++;
1972         if (sbio->page_count == sctx->pages_per_wr_bio)
1973                 scrub_wr_submit(sctx);
1974         mutex_unlock(&sctx->wr_lock);
1975
1976         return 0;
1977 }
1978
1979 static void scrub_wr_submit(struct scrub_ctx *sctx)
1980 {
1981         struct scrub_bio *sbio;
1982
1983         if (!sctx->wr_curr_bio)
1984                 return;
1985
1986         sbio = sctx->wr_curr_bio;
1987         sctx->wr_curr_bio = NULL;
1988         WARN_ON(!sbio->bio->bi_disk);
1989         scrub_pending_bio_inc(sctx);
1990         /* process all writes in a single worker thread. Then the block layer
1991          * orders the requests before sending them to the driver which
1992          * doubled the write performance on spinning disks when measured
1993          * with Linux 3.5 */
1994         btrfsic_submit_bio(sbio->bio);
1995 }
1996
1997 static void scrub_wr_bio_end_io(struct bio *bio)
1998 {
1999         struct scrub_bio *sbio = bio->bi_private;
2000         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2001
2002         sbio->status = bio->bi_status;
2003         sbio->bio = bio;
2004
2005         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
2006                          scrub_wr_bio_end_io_worker, NULL, NULL);
2007         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
2008 }
2009
2010 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
2011 {
2012         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2013         struct scrub_ctx *sctx = sbio->sctx;
2014         int i;
2015
2016         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
2017         if (sbio->status) {
2018                 struct btrfs_dev_replace *dev_replace =
2019                         &sbio->sctx->fs_info->dev_replace;
2020
2021                 for (i = 0; i < sbio->page_count; i++) {
2022                         struct scrub_page *spage = sbio->pagev[i];
2023
2024                         spage->io_error = 1;
2025                         btrfs_dev_replace_stats_inc(&dev_replace->
2026                                                     num_write_errors);
2027                 }
2028         }
2029
2030         for (i = 0; i < sbio->page_count; i++)
2031                 scrub_page_put(sbio->pagev[i]);
2032
2033         bio_put(sbio->bio);
2034         kfree(sbio);
2035         scrub_pending_bio_dec(sctx);
2036 }
2037
2038 static int scrub_checksum(struct scrub_block *sblock)
2039 {
2040         u64 flags;
2041         int ret;
2042
2043         /*
2044          * No need to initialize these stats currently,
2045          * because this function only use return value
2046          * instead of these stats value.
2047          *
2048          * Todo:
2049          * always use stats
2050          */
2051         sblock->header_error = 0;
2052         sblock->generation_error = 0;
2053         sblock->checksum_error = 0;
2054
2055         WARN_ON(sblock->page_count < 1);
2056         flags = sblock->pagev[0]->flags;
2057         ret = 0;
2058         if (flags & BTRFS_EXTENT_FLAG_DATA)
2059                 ret = scrub_checksum_data(sblock);
2060         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
2061                 ret = scrub_checksum_tree_block(sblock);
2062         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
2063                 (void)scrub_checksum_super(sblock);
2064         else
2065                 WARN_ON(1);
2066         if (ret)
2067                 scrub_handle_errored_block(sblock);
2068
2069         return ret;
2070 }
2071
2072 static int scrub_checksum_data(struct scrub_block *sblock)
2073 {
2074         struct scrub_ctx *sctx = sblock->sctx;
2075         u8 csum[BTRFS_CSUM_SIZE];
2076         u8 *on_disk_csum;
2077         struct page *page;
2078         void *buffer;
2079         u32 crc = ~(u32)0;
2080         u64 len;
2081         int index;
2082
2083         BUG_ON(sblock->page_count < 1);
2084         if (!sblock->pagev[0]->have_csum)
2085                 return 0;
2086
2087         on_disk_csum = sblock->pagev[0]->csum;
2088         page = sblock->pagev[0]->page;
2089         buffer = kmap_atomic(page);
2090
2091         len = sctx->fs_info->sectorsize;
2092         index = 0;
2093         for (;;) {
2094                 u64 l = min_t(u64, len, PAGE_SIZE);
2095
2096                 crc = btrfs_csum_data(buffer, crc, l);
2097                 kunmap_atomic(buffer);
2098                 len -= l;
2099                 if (len == 0)
2100                         break;
2101                 index++;
2102                 BUG_ON(index >= sblock->page_count);
2103                 BUG_ON(!sblock->pagev[index]->page);
2104                 page = sblock->pagev[index]->page;
2105                 buffer = kmap_atomic(page);
2106         }
2107
2108         btrfs_csum_final(crc, csum);
2109         if (memcmp(csum, on_disk_csum, sctx->csum_size))
2110                 sblock->checksum_error = 1;
2111
2112         return sblock->checksum_error;
2113 }
2114
2115 static int scrub_checksum_tree_block(struct scrub_block *sblock)
2116 {
2117         struct scrub_ctx *sctx = sblock->sctx;
2118         struct btrfs_header *h;
2119         struct btrfs_fs_info *fs_info = sctx->fs_info;
2120         u8 calculated_csum[BTRFS_CSUM_SIZE];
2121         u8 on_disk_csum[BTRFS_CSUM_SIZE];
2122         struct page *page;
2123         void *mapped_buffer;
2124         u64 mapped_size;
2125         void *p;
2126         u32 crc = ~(u32)0;
2127         u64 len;
2128         int index;
2129
2130         BUG_ON(sblock->page_count < 1);
2131         page = sblock->pagev[0]->page;
2132         mapped_buffer = kmap_atomic(page);
2133         h = (struct btrfs_header *)mapped_buffer;
2134         memcpy(on_disk_csum, h->csum, sctx->csum_size);
2135
2136         /*
2137          * we don't use the getter functions here, as we
2138          * a) don't have an extent buffer and
2139          * b) the page is already kmapped
2140          */
2141         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
2142                 sblock->header_error = 1;
2143
2144         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
2145                 sblock->header_error = 1;
2146                 sblock->generation_error = 1;
2147         }
2148
2149         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
2150                 sblock->header_error = 1;
2151
2152         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
2153                    BTRFS_UUID_SIZE))
2154                 sblock->header_error = 1;
2155
2156         len = sctx->fs_info->nodesize - BTRFS_CSUM_SIZE;
2157         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2158         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2159         index = 0;
2160         for (;;) {
2161                 u64 l = min_t(u64, len, mapped_size);
2162
2163                 crc = btrfs_csum_data(p, crc, l);
2164                 kunmap_atomic(mapped_buffer);
2165                 len -= l;
2166                 if (len == 0)
2167                         break;
2168                 index++;
2169                 BUG_ON(index >= sblock->page_count);
2170                 BUG_ON(!sblock->pagev[index]->page);
2171                 page = sblock->pagev[index]->page;
2172                 mapped_buffer = kmap_atomic(page);
2173                 mapped_size = PAGE_SIZE;
2174                 p = mapped_buffer;
2175         }
2176
2177         btrfs_csum_final(crc, calculated_csum);
2178         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2179                 sblock->checksum_error = 1;
2180
2181         return sblock->header_error || sblock->checksum_error;
2182 }
2183
2184 static int scrub_checksum_super(struct scrub_block *sblock)
2185 {
2186         struct btrfs_super_block *s;
2187         struct scrub_ctx *sctx = sblock->sctx;
2188         u8 calculated_csum[BTRFS_CSUM_SIZE];
2189         u8 on_disk_csum[BTRFS_CSUM_SIZE];
2190         struct page *page;
2191         void *mapped_buffer;
2192         u64 mapped_size;
2193         void *p;
2194         u32 crc = ~(u32)0;
2195         int fail_gen = 0;
2196         int fail_cor = 0;
2197         u64 len;
2198         int index;
2199
2200         BUG_ON(sblock->page_count < 1);
2201         page = sblock->pagev[0]->page;
2202         mapped_buffer = kmap_atomic(page);
2203         s = (struct btrfs_super_block *)mapped_buffer;
2204         memcpy(on_disk_csum, s->csum, sctx->csum_size);
2205
2206         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
2207                 ++fail_cor;
2208
2209         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
2210                 ++fail_gen;
2211
2212         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
2213                 ++fail_cor;
2214
2215         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
2216         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
2217         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
2218         index = 0;
2219         for (;;) {
2220                 u64 l = min_t(u64, len, mapped_size);
2221
2222                 crc = btrfs_csum_data(p, crc, l);
2223                 kunmap_atomic(mapped_buffer);
2224                 len -= l;
2225                 if (len == 0)
2226                         break;
2227                 index++;
2228                 BUG_ON(index >= sblock->page_count);
2229                 BUG_ON(!sblock->pagev[index]->page);
2230                 page = sblock->pagev[index]->page;
2231                 mapped_buffer = kmap_atomic(page);
2232                 mapped_size = PAGE_SIZE;
2233                 p = mapped_buffer;
2234         }
2235
2236         btrfs_csum_final(crc, calculated_csum);
2237         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
2238                 ++fail_cor;
2239
2240         if (fail_cor + fail_gen) {
2241                 /*
2242                  * if we find an error in a super block, we just report it.
2243                  * They will get written with the next transaction commit
2244                  * anyway
2245                  */
2246                 spin_lock(&sctx->stat_lock);
2247                 ++sctx->stat.super_errors;
2248                 spin_unlock(&sctx->stat_lock);
2249                 if (fail_cor)
2250                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2251                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
2252                 else
2253                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
2254                                 BTRFS_DEV_STAT_GENERATION_ERRS);
2255         }
2256
2257         return fail_cor + fail_gen;
2258 }
2259
2260 static void scrub_block_get(struct scrub_block *sblock)
2261 {
2262         refcount_inc(&sblock->refs);
2263 }
2264
2265 static void scrub_block_put(struct scrub_block *sblock)
2266 {
2267         if (refcount_dec_and_test(&sblock->refs)) {
2268                 int i;
2269
2270                 if (sblock->sparity)
2271                         scrub_parity_put(sblock->sparity);
2272
2273                 for (i = 0; i < sblock->page_count; i++)
2274                         scrub_page_put(sblock->pagev[i]);
2275                 kfree(sblock);
2276         }
2277 }
2278
2279 static void scrub_page_get(struct scrub_page *spage)
2280 {
2281         atomic_inc(&spage->refs);
2282 }
2283
2284 static void scrub_page_put(struct scrub_page *spage)
2285 {
2286         if (atomic_dec_and_test(&spage->refs)) {
2287                 if (spage->page)
2288                         __free_page(spage->page);
2289                 kfree(spage);
2290         }
2291 }
2292
2293 static void scrub_submit(struct scrub_ctx *sctx)
2294 {
2295         struct scrub_bio *sbio;
2296
2297         if (sctx->curr == -1)
2298                 return;
2299
2300         sbio = sctx->bios[sctx->curr];
2301         sctx->curr = -1;
2302         scrub_pending_bio_inc(sctx);
2303         btrfsic_submit_bio(sbio->bio);
2304 }
2305
2306 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2307                                     struct scrub_page *spage)
2308 {
2309         struct scrub_block *sblock = spage->sblock;
2310         struct scrub_bio *sbio;
2311         int ret;
2312
2313 again:
2314         /*
2315          * grab a fresh bio or wait for one to become available
2316          */
2317         while (sctx->curr == -1) {
2318                 spin_lock(&sctx->list_lock);
2319                 sctx->curr = sctx->first_free;
2320                 if (sctx->curr != -1) {
2321                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2322                         sctx->bios[sctx->curr]->next_free = -1;
2323                         sctx->bios[sctx->curr]->page_count = 0;
2324                         spin_unlock(&sctx->list_lock);
2325                 } else {
2326                         spin_unlock(&sctx->list_lock);
2327                         wait_event(sctx->list_wait, sctx->first_free != -1);
2328                 }
2329         }
2330         sbio = sctx->bios[sctx->curr];
2331         if (sbio->page_count == 0) {
2332                 struct bio *bio;
2333
2334                 sbio->physical = spage->physical;
2335                 sbio->logical = spage->logical;
2336                 sbio->dev = spage->dev;
2337                 bio = sbio->bio;
2338                 if (!bio) {
2339                         bio = btrfs_io_bio_alloc(sctx->pages_per_rd_bio);
2340                         sbio->bio = bio;
2341                 }
2342
2343                 bio->bi_private = sbio;
2344                 bio->bi_end_io = scrub_bio_end_io;
2345                 bio_set_dev(bio, sbio->dev->bdev);
2346                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2347                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2348                 sbio->status = 0;
2349         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2350                    spage->physical ||
2351                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2352                    spage->logical ||
2353                    sbio->dev != spage->dev) {
2354                 scrub_submit(sctx);
2355                 goto again;
2356         }
2357
2358         sbio->pagev[sbio->page_count] = spage;
2359         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2360         if (ret != PAGE_SIZE) {
2361                 if (sbio->page_count < 1) {
2362                         bio_put(sbio->bio);
2363                         sbio->bio = NULL;
2364                         return -EIO;
2365                 }
2366                 scrub_submit(sctx);
2367                 goto again;
2368         }
2369
2370         scrub_block_get(sblock); /* one for the page added to the bio */
2371         atomic_inc(&sblock->outstanding_pages);
2372         sbio->page_count++;
2373         if (sbio->page_count == sctx->pages_per_rd_bio)
2374                 scrub_submit(sctx);
2375
2376         return 0;
2377 }
2378
2379 static void scrub_missing_raid56_end_io(struct bio *bio)
2380 {
2381         struct scrub_block *sblock = bio->bi_private;
2382         struct btrfs_fs_info *fs_info = sblock->sctx->fs_info;
2383
2384         if (bio->bi_status)
2385                 sblock->no_io_error_seen = 0;
2386
2387         bio_put(bio);
2388
2389         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2390 }
2391
2392 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2393 {
2394         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2395         struct scrub_ctx *sctx = sblock->sctx;
2396         struct btrfs_fs_info *fs_info = sctx->fs_info;
2397         u64 logical;
2398         struct btrfs_device *dev;
2399
2400         logical = sblock->pagev[0]->logical;
2401         dev = sblock->pagev[0]->dev;
2402
2403         if (sblock->no_io_error_seen)
2404                 scrub_recheck_block_checksum(sblock);
2405
2406         if (!sblock->no_io_error_seen) {
2407                 spin_lock(&sctx->stat_lock);
2408                 sctx->stat.read_errors++;
2409                 spin_unlock(&sctx->stat_lock);
2410                 btrfs_err_rl_in_rcu(fs_info,
2411                         "IO error rebuilding logical %llu for dev %s",
2412                         logical, rcu_str_deref(dev->name));
2413         } else if (sblock->header_error || sblock->checksum_error) {
2414                 spin_lock(&sctx->stat_lock);
2415                 sctx->stat.uncorrectable_errors++;
2416                 spin_unlock(&sctx->stat_lock);
2417                 btrfs_err_rl_in_rcu(fs_info,
2418                         "failed to rebuild valid logical %llu for dev %s",
2419                         logical, rcu_str_deref(dev->name));
2420         } else {
2421                 scrub_write_block_to_dev_replace(sblock);
2422         }
2423
2424         scrub_block_put(sblock);
2425
2426         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2427                 mutex_lock(&sctx->wr_lock);
2428                 scrub_wr_submit(sctx);
2429                 mutex_unlock(&sctx->wr_lock);
2430         }
2431
2432         scrub_pending_bio_dec(sctx);
2433 }
2434
2435 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2436 {
2437         struct scrub_ctx *sctx = sblock->sctx;
2438         struct btrfs_fs_info *fs_info = sctx->fs_info;
2439         u64 length = sblock->page_count * PAGE_SIZE;
2440         u64 logical = sblock->pagev[0]->logical;
2441         struct btrfs_bio *bbio = NULL;
2442         struct bio *bio;
2443         struct btrfs_raid_bio *rbio;
2444         int ret;
2445         int i;
2446
2447         btrfs_bio_counter_inc_blocked(fs_info);
2448         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
2449                         &length, &bbio);
2450         if (ret || !bbio || !bbio->raid_map)
2451                 goto bbio_out;
2452
2453         if (WARN_ON(!sctx->is_dev_replace ||
2454                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2455                 /*
2456                  * We shouldn't be scrubbing a missing device. Even for dev
2457                  * replace, we should only get here for RAID 5/6. We either
2458                  * managed to mount something with no mirrors remaining or
2459                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2460                  */
2461                 goto bbio_out;
2462         }
2463
2464         bio = btrfs_io_bio_alloc(0);
2465         bio->bi_iter.bi_sector = logical >> 9;
2466         bio->bi_private = sblock;
2467         bio->bi_end_io = scrub_missing_raid56_end_io;
2468
2469         rbio = raid56_alloc_missing_rbio(fs_info, bio, bbio, length);
2470         if (!rbio)
2471                 goto rbio_out;
2472
2473         for (i = 0; i < sblock->page_count; i++) {
2474                 struct scrub_page *spage = sblock->pagev[i];
2475
2476                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2477         }
2478
2479         btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2480                         scrub_missing_raid56_worker, NULL, NULL);
2481         scrub_block_get(sblock);
2482         scrub_pending_bio_inc(sctx);
2483         raid56_submit_missing_rbio(rbio);
2484         return;
2485
2486 rbio_out:
2487         bio_put(bio);
2488 bbio_out:
2489         btrfs_bio_counter_dec(fs_info);
2490         btrfs_put_bbio(bbio);
2491         spin_lock(&sctx->stat_lock);
2492         sctx->stat.malloc_errors++;
2493         spin_unlock(&sctx->stat_lock);
2494 }
2495
2496 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2497                        u64 physical, struct btrfs_device *dev, u64 flags,
2498                        u64 gen, int mirror_num, u8 *csum, int force,
2499                        u64 physical_for_dev_replace)
2500 {
2501         struct scrub_block *sblock;
2502         int index;
2503
2504         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2505         if (!sblock) {
2506                 spin_lock(&sctx->stat_lock);
2507                 sctx->stat.malloc_errors++;
2508                 spin_unlock(&sctx->stat_lock);
2509                 return -ENOMEM;
2510         }
2511
2512         /* one ref inside this function, plus one for each page added to
2513          * a bio later on */
2514         refcount_set(&sblock->refs, 1);
2515         sblock->sctx = sctx;
2516         sblock->no_io_error_seen = 1;
2517
2518         for (index = 0; len > 0; index++) {
2519                 struct scrub_page *spage;
2520                 u64 l = min_t(u64, len, PAGE_SIZE);
2521
2522                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2523                 if (!spage) {
2524 leave_nomem:
2525                         spin_lock(&sctx->stat_lock);
2526                         sctx->stat.malloc_errors++;
2527                         spin_unlock(&sctx->stat_lock);
2528                         scrub_block_put(sblock);
2529                         return -ENOMEM;
2530                 }
2531                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2532                 scrub_page_get(spage);
2533                 sblock->pagev[index] = spage;
2534                 spage->sblock = sblock;
2535                 spage->dev = dev;
2536                 spage->flags = flags;
2537                 spage->generation = gen;
2538                 spage->logical = logical;
2539                 spage->physical = physical;
2540                 spage->physical_for_dev_replace = physical_for_dev_replace;
2541                 spage->mirror_num = mirror_num;
2542                 if (csum) {
2543                         spage->have_csum = 1;
2544                         memcpy(spage->csum, csum, sctx->csum_size);
2545                 } else {
2546                         spage->have_csum = 0;
2547                 }
2548                 sblock->page_count++;
2549                 spage->page = alloc_page(GFP_KERNEL);
2550                 if (!spage->page)
2551                         goto leave_nomem;
2552                 len -= l;
2553                 logical += l;
2554                 physical += l;
2555                 physical_for_dev_replace += l;
2556         }
2557
2558         WARN_ON(sblock->page_count == 0);
2559         if (dev->missing) {
2560                 /*
2561                  * This case should only be hit for RAID 5/6 device replace. See
2562                  * the comment in scrub_missing_raid56_pages() for details.
2563                  */
2564                 scrub_missing_raid56_pages(sblock);
2565         } else {
2566                 for (index = 0; index < sblock->page_count; index++) {
2567                         struct scrub_page *spage = sblock->pagev[index];
2568                         int ret;
2569
2570                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2571                         if (ret) {
2572                                 scrub_block_put(sblock);
2573                                 return ret;
2574                         }
2575                 }
2576
2577                 if (force)
2578                         scrub_submit(sctx);
2579         }
2580
2581         /* last one frees, either here or in bio completion for last page */
2582         scrub_block_put(sblock);
2583         return 0;
2584 }
2585
2586 static void scrub_bio_end_io(struct bio *bio)
2587 {
2588         struct scrub_bio *sbio = bio->bi_private;
2589         struct btrfs_fs_info *fs_info = sbio->dev->fs_info;
2590
2591         sbio->status = bio->bi_status;
2592         sbio->bio = bio;
2593
2594         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2595 }
2596
2597 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2598 {
2599         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2600         struct scrub_ctx *sctx = sbio->sctx;
2601         int i;
2602
2603         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2604         if (sbio->status) {
2605                 for (i = 0; i < sbio->page_count; i++) {
2606                         struct scrub_page *spage = sbio->pagev[i];
2607
2608                         spage->io_error = 1;
2609                         spage->sblock->no_io_error_seen = 0;
2610                 }
2611         }
2612
2613         /* now complete the scrub_block items that have all pages completed */
2614         for (i = 0; i < sbio->page_count; i++) {
2615                 struct scrub_page *spage = sbio->pagev[i];
2616                 struct scrub_block *sblock = spage->sblock;
2617
2618                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2619                         scrub_block_complete(sblock);
2620                 scrub_block_put(sblock);
2621         }
2622
2623         bio_put(sbio->bio);
2624         sbio->bio = NULL;
2625         spin_lock(&sctx->list_lock);
2626         sbio->next_free = sctx->first_free;
2627         sctx->first_free = sbio->index;
2628         spin_unlock(&sctx->list_lock);
2629
2630         if (sctx->is_dev_replace && sctx->flush_all_writes) {
2631                 mutex_lock(&sctx->wr_lock);
2632                 scrub_wr_submit(sctx);
2633                 mutex_unlock(&sctx->wr_lock);
2634         }
2635
2636         scrub_pending_bio_dec(sctx);
2637 }
2638
2639 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2640                                        unsigned long *bitmap,
2641                                        u64 start, u64 len)
2642 {
2643         u64 offset;
2644         u64 nsectors64;
2645         u32 nsectors;
2646         int sectorsize = sparity->sctx->fs_info->sectorsize;
2647
2648         if (len >= sparity->stripe_len) {
2649                 bitmap_set(bitmap, 0, sparity->nsectors);
2650                 return;
2651         }
2652
2653         start -= sparity->logic_start;
2654         start = div64_u64_rem(start, sparity->stripe_len, &offset);
2655         offset = div_u64(offset, sectorsize);
2656         nsectors64 = div_u64(len, sectorsize);
2657
2658         ASSERT(nsectors64 < UINT_MAX);
2659         nsectors = (u32)nsectors64;
2660
2661         if (offset + nsectors <= sparity->nsectors) {
2662                 bitmap_set(bitmap, offset, nsectors);
2663                 return;
2664         }
2665
2666         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2667         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2668 }
2669
2670 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2671                                                    u64 start, u64 len)
2672 {
2673         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2674 }
2675
2676 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2677                                                   u64 start, u64 len)
2678 {
2679         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2680 }
2681
2682 static void scrub_block_complete(struct scrub_block *sblock)
2683 {
2684         int corrupted = 0;
2685
2686         if (!sblock->no_io_error_seen) {
2687                 corrupted = 1;
2688                 scrub_handle_errored_block(sblock);
2689         } else {
2690                 /*
2691                  * if has checksum error, write via repair mechanism in
2692                  * dev replace case, otherwise write here in dev replace
2693                  * case.
2694                  */
2695                 corrupted = scrub_checksum(sblock);
2696                 if (!corrupted && sblock->sctx->is_dev_replace)
2697                         scrub_write_block_to_dev_replace(sblock);
2698         }
2699
2700         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2701                 u64 start = sblock->pagev[0]->logical;
2702                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2703                           PAGE_SIZE;
2704
2705                 scrub_parity_mark_sectors_error(sblock->sparity,
2706                                                 start, end - start);
2707         }
2708 }
2709
2710 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2711 {
2712         struct btrfs_ordered_sum *sum = NULL;
2713         unsigned long index;
2714         unsigned long num_sectors;
2715
2716         while (!list_empty(&sctx->csum_list)) {
2717                 sum = list_first_entry(&sctx->csum_list,
2718                                        struct btrfs_ordered_sum, list);
2719                 if (sum->bytenr > logical)
2720                         return 0;
2721                 if (sum->bytenr + sum->len > logical)
2722                         break;
2723
2724                 ++sctx->stat.csum_discards;
2725                 list_del(&sum->list);
2726                 kfree(sum);
2727                 sum = NULL;
2728         }
2729         if (!sum)
2730                 return 0;
2731
2732         index = div_u64(logical - sum->bytenr, sctx->fs_info->sectorsize);
2733         ASSERT(index < UINT_MAX);
2734
2735         num_sectors = sum->len / sctx->fs_info->sectorsize;
2736         memcpy(csum, sum->sums + index, sctx->csum_size);
2737         if (index == num_sectors - 1) {
2738                 list_del(&sum->list);
2739                 kfree(sum);
2740         }
2741         return 1;
2742 }
2743
2744 /* scrub extent tries to collect up to 64 kB for each bio */
2745 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2746                         u64 physical, struct btrfs_device *dev, u64 flags,
2747                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2748 {
2749         int ret;
2750         u8 csum[BTRFS_CSUM_SIZE];
2751         u32 blocksize;
2752
2753         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2754                 blocksize = sctx->fs_info->sectorsize;
2755                 spin_lock(&sctx->stat_lock);
2756                 sctx->stat.data_extents_scrubbed++;
2757                 sctx->stat.data_bytes_scrubbed += len;
2758                 spin_unlock(&sctx->stat_lock);
2759         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2760                 blocksize = sctx->fs_info->nodesize;
2761                 spin_lock(&sctx->stat_lock);
2762                 sctx->stat.tree_extents_scrubbed++;
2763                 sctx->stat.tree_bytes_scrubbed += len;
2764                 spin_unlock(&sctx->stat_lock);
2765         } else {
2766                 blocksize = sctx->fs_info->sectorsize;
2767                 WARN_ON(1);
2768         }
2769
2770         while (len) {
2771                 u64 l = min_t(u64, len, blocksize);
2772                 int have_csum = 0;
2773
2774                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2775                         /* push csums to sbio */
2776                         have_csum = scrub_find_csum(sctx, logical, csum);
2777                         if (have_csum == 0)
2778                                 ++sctx->stat.no_csum;
2779                         if (0 && sctx->is_dev_replace && !have_csum) {
2780                                 ret = copy_nocow_pages(sctx, logical, l,
2781                                                        mirror_num,
2782                                                       physical_for_dev_replace);
2783                                 goto behind_scrub_pages;
2784                         }
2785                 }
2786                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2787                                   mirror_num, have_csum ? csum : NULL, 0,
2788                                   physical_for_dev_replace);
2789 behind_scrub_pages:
2790                 if (ret)
2791                         return ret;
2792                 len -= l;
2793                 logical += l;
2794                 physical += l;
2795                 physical_for_dev_replace += l;
2796         }
2797         return 0;
2798 }
2799
2800 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2801                                   u64 logical, u64 len,
2802                                   u64 physical, struct btrfs_device *dev,
2803                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2804 {
2805         struct scrub_ctx *sctx = sparity->sctx;
2806         struct scrub_block *sblock;
2807         int index;
2808
2809         sblock = kzalloc(sizeof(*sblock), GFP_KERNEL);
2810         if (!sblock) {
2811                 spin_lock(&sctx->stat_lock);
2812                 sctx->stat.malloc_errors++;
2813                 spin_unlock(&sctx->stat_lock);
2814                 return -ENOMEM;
2815         }
2816
2817         /* one ref inside this function, plus one for each page added to
2818          * a bio later on */
2819         refcount_set(&sblock->refs, 1);
2820         sblock->sctx = sctx;
2821         sblock->no_io_error_seen = 1;
2822         sblock->sparity = sparity;
2823         scrub_parity_get(sparity);
2824
2825         for (index = 0; len > 0; index++) {
2826                 struct scrub_page *spage;
2827                 u64 l = min_t(u64, len, PAGE_SIZE);
2828
2829                 spage = kzalloc(sizeof(*spage), GFP_KERNEL);
2830                 if (!spage) {
2831 leave_nomem:
2832                         spin_lock(&sctx->stat_lock);
2833                         sctx->stat.malloc_errors++;
2834                         spin_unlock(&sctx->stat_lock);
2835                         scrub_block_put(sblock);
2836                         return -ENOMEM;
2837                 }
2838                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2839                 /* For scrub block */
2840                 scrub_page_get(spage);
2841                 sblock->pagev[index] = spage;
2842                 /* For scrub parity */
2843                 scrub_page_get(spage);
2844                 list_add_tail(&spage->list, &sparity->spages);
2845                 spage->sblock = sblock;
2846                 spage->dev = dev;
2847                 spage->flags = flags;
2848                 spage->generation = gen;
2849                 spage->logical = logical;
2850                 spage->physical = physical;
2851                 spage->mirror_num = mirror_num;
2852                 if (csum) {
2853                         spage->have_csum = 1;
2854                         memcpy(spage->csum, csum, sctx->csum_size);
2855                 } else {
2856                         spage->have_csum = 0;
2857                 }
2858                 sblock->page_count++;
2859                 spage->page = alloc_page(GFP_KERNEL);
2860                 if (!spage->page)
2861                         goto leave_nomem;
2862                 len -= l;
2863                 logical += l;
2864                 physical += l;
2865         }
2866
2867         WARN_ON(sblock->page_count == 0);
2868         for (index = 0; index < sblock->page_count; index++) {
2869                 struct scrub_page *spage = sblock->pagev[index];
2870                 int ret;
2871
2872                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2873                 if (ret) {
2874                         scrub_block_put(sblock);
2875                         return ret;
2876                 }
2877         }
2878
2879         /* last one frees, either here or in bio completion for last page */
2880         scrub_block_put(sblock);
2881         return 0;
2882 }
2883
2884 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2885                                    u64 logical, u64 len,
2886                                    u64 physical, struct btrfs_device *dev,
2887                                    u64 flags, u64 gen, int mirror_num)
2888 {
2889         struct scrub_ctx *sctx = sparity->sctx;
2890         int ret;
2891         u8 csum[BTRFS_CSUM_SIZE];
2892         u32 blocksize;
2893
2894         if (dev->missing) {
2895                 scrub_parity_mark_sectors_error(sparity, logical, len);
2896                 return 0;
2897         }
2898
2899         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2900                 blocksize = sctx->fs_info->sectorsize;
2901         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2902                 blocksize = sctx->fs_info->nodesize;
2903         } else {
2904                 blocksize = sctx->fs_info->sectorsize;
2905                 WARN_ON(1);
2906         }
2907
2908         while (len) {
2909                 u64 l = min_t(u64, len, blocksize);
2910                 int have_csum = 0;
2911
2912                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2913                         /* push csums to sbio */
2914                         have_csum = scrub_find_csum(sctx, logical, csum);
2915                         if (have_csum == 0)
2916                                 goto skip;
2917                 }
2918                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2919                                              flags, gen, mirror_num,
2920                                              have_csum ? csum : NULL);
2921                 if (ret)
2922                         return ret;
2923 skip:
2924                 len -= l;
2925                 logical += l;
2926                 physical += l;
2927         }
2928         return 0;
2929 }
2930
2931 /*
2932  * Given a physical address, this will calculate it's
2933  * logical offset. if this is a parity stripe, it will return
2934  * the most left data stripe's logical offset.
2935  *
2936  * return 0 if it is a data stripe, 1 means parity stripe.
2937  */
2938 static int get_raid56_logic_offset(u64 physical, int num,
2939                                    struct map_lookup *map, u64 *offset,
2940                                    u64 *stripe_start)
2941 {
2942         int i;
2943         int j = 0;
2944         u64 stripe_nr;
2945         u64 last_offset;
2946         u32 stripe_index;
2947         u32 rot;
2948
2949         last_offset = (physical - map->stripes[num].physical) *
2950                       nr_data_stripes(map);
2951         if (stripe_start)
2952                 *stripe_start = last_offset;
2953
2954         *offset = last_offset;
2955         for (i = 0; i < nr_data_stripes(map); i++) {
2956                 *offset = last_offset + i * map->stripe_len;
2957
2958                 stripe_nr = div64_u64(*offset, map->stripe_len);
2959                 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2960
2961                 /* Work out the disk rotation on this stripe-set */
2962                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2963                 /* calculate which stripe this data locates */
2964                 rot += i;
2965                 stripe_index = rot % map->num_stripes;
2966                 if (stripe_index == num)
2967                         return 0;
2968                 if (stripe_index < num)
2969                         j++;
2970         }
2971         *offset = last_offset + j * map->stripe_len;
2972         return 1;
2973 }
2974
2975 static void scrub_free_parity(struct scrub_parity *sparity)
2976 {
2977         struct scrub_ctx *sctx = sparity->sctx;
2978         struct scrub_page *curr, *next;
2979         int nbits;
2980
2981         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2982         if (nbits) {
2983                 spin_lock(&sctx->stat_lock);
2984                 sctx->stat.read_errors += nbits;
2985                 sctx->stat.uncorrectable_errors += nbits;
2986                 spin_unlock(&sctx->stat_lock);
2987         }
2988
2989         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2990                 list_del_init(&curr->list);
2991                 scrub_page_put(curr);
2992         }
2993
2994         kfree(sparity);
2995 }
2996
2997 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2998 {
2999         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
3000                                                     work);
3001         struct scrub_ctx *sctx = sparity->sctx;
3002
3003         scrub_free_parity(sparity);
3004         scrub_pending_bio_dec(sctx);
3005 }
3006
3007 static void scrub_parity_bio_endio(struct bio *bio)
3008 {
3009         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
3010         struct btrfs_fs_info *fs_info = sparity->sctx->fs_info;
3011
3012         if (bio->bi_status)
3013                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3014                           sparity->nsectors);
3015
3016         bio_put(bio);
3017
3018         btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
3019                         scrub_parity_bio_endio_worker, NULL, NULL);
3020         btrfs_queue_work(fs_info->scrub_parity_workers, &sparity->work);
3021 }
3022
3023 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
3024 {
3025         struct scrub_ctx *sctx = sparity->sctx;
3026         struct btrfs_fs_info *fs_info = sctx->fs_info;
3027         struct bio *bio;
3028         struct btrfs_raid_bio *rbio;
3029         struct btrfs_bio *bbio = NULL;
3030         u64 length;
3031         int ret;
3032
3033         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
3034                            sparity->nsectors))
3035                 goto out;
3036
3037         length = sparity->logic_end - sparity->logic_start;
3038
3039         btrfs_bio_counter_inc_blocked(fs_info);
3040         ret = btrfs_map_sblock(fs_info, BTRFS_MAP_WRITE, sparity->logic_start,
3041                                &length, &bbio);
3042         if (ret || !bbio || !bbio->raid_map)
3043                 goto bbio_out;
3044
3045         bio = btrfs_io_bio_alloc(0);
3046         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
3047         bio->bi_private = sparity;
3048         bio->bi_end_io = scrub_parity_bio_endio;
3049
3050         rbio = raid56_parity_alloc_scrub_rbio(fs_info, bio, bbio,
3051                                               length, sparity->scrub_dev,
3052                                               sparity->dbitmap,
3053                                               sparity->nsectors);
3054         if (!rbio)
3055                 goto rbio_out;
3056
3057         scrub_pending_bio_inc(sctx);
3058         raid56_parity_submit_scrub_rbio(rbio);
3059         return;
3060
3061 rbio_out:
3062         bio_put(bio);
3063 bbio_out:
3064         btrfs_bio_counter_dec(fs_info);
3065         btrfs_put_bbio(bbio);
3066         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
3067                   sparity->nsectors);
3068         spin_lock(&sctx->stat_lock);
3069         sctx->stat.malloc_errors++;
3070         spin_unlock(&sctx->stat_lock);
3071 out:
3072         scrub_free_parity(sparity);
3073 }
3074
3075 static inline int scrub_calc_parity_bitmap_len(int nsectors)
3076 {
3077         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * sizeof(long);
3078 }
3079
3080 static void scrub_parity_get(struct scrub_parity *sparity)
3081 {
3082         refcount_inc(&sparity->refs);
3083 }
3084
3085 static void scrub_parity_put(struct scrub_parity *sparity)
3086 {
3087         if (!refcount_dec_and_test(&sparity->refs))
3088                 return;
3089
3090         scrub_parity_check_and_repair(sparity);
3091 }
3092
3093 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
3094                                                   struct map_lookup *map,
3095                                                   struct btrfs_device *sdev,
3096                                                   struct btrfs_path *path,
3097                                                   u64 logic_start,
3098                                                   u64 logic_end)
3099 {
3100         struct btrfs_fs_info *fs_info = sctx->fs_info;
3101         struct btrfs_root *root = fs_info->extent_root;
3102         struct btrfs_root *csum_root = fs_info->csum_root;
3103         struct btrfs_extent_item *extent;
3104         struct btrfs_bio *bbio = NULL;
3105         u64 flags;
3106         int ret;
3107         int slot;
3108         struct extent_buffer *l;
3109         struct btrfs_key key;
3110         u64 generation;
3111         u64 extent_logical;
3112         u64 extent_physical;
3113         u64 extent_len;
3114         u64 mapped_length;
3115         struct btrfs_device *extent_dev;
3116         struct scrub_parity *sparity;
3117         int nsectors;
3118         int bitmap_len;
3119         int extent_mirror_num;
3120         int stop_loop = 0;
3121
3122         nsectors = div_u64(map->stripe_len, fs_info->sectorsize);
3123         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
3124         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
3125                           GFP_NOFS);
3126         if (!sparity) {
3127                 spin_lock(&sctx->stat_lock);
3128                 sctx->stat.malloc_errors++;
3129                 spin_unlock(&sctx->stat_lock);
3130                 return -ENOMEM;
3131         }
3132
3133         sparity->stripe_len = map->stripe_len;
3134         sparity->nsectors = nsectors;
3135         sparity->sctx = sctx;
3136         sparity->scrub_dev = sdev;
3137         sparity->logic_start = logic_start;
3138         sparity->logic_end = logic_end;
3139         refcount_set(&sparity->refs, 1);
3140         INIT_LIST_HEAD(&sparity->spages);
3141         sparity->dbitmap = sparity->bitmap;
3142         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
3143
3144         ret = 0;
3145         while (logic_start < logic_end) {
3146                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3147                         key.type = BTRFS_METADATA_ITEM_KEY;
3148                 else
3149                         key.type = BTRFS_EXTENT_ITEM_KEY;
3150                 key.objectid = logic_start;
3151                 key.offset = (u64)-1;
3152
3153                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3154                 if (ret < 0)
3155                         goto out;
3156
3157                 if (ret > 0) {
3158                         ret = btrfs_previous_extent_item(root, path, 0);
3159                         if (ret < 0)
3160                                 goto out;
3161                         if (ret > 0) {
3162                                 btrfs_release_path(path);
3163                                 ret = btrfs_search_slot(NULL, root, &key,
3164                                                         path, 0, 0);
3165                                 if (ret < 0)
3166                                         goto out;
3167                         }
3168                 }
3169
3170                 stop_loop = 0;
3171                 while (1) {
3172                         u64 bytes;
3173
3174                         l = path->nodes[0];
3175                         slot = path->slots[0];
3176                         if (slot >= btrfs_header_nritems(l)) {
3177                                 ret = btrfs_next_leaf(root, path);
3178                                 if (ret == 0)
3179                                         continue;
3180                                 if (ret < 0)
3181                                         goto out;
3182
3183                                 stop_loop = 1;
3184                                 break;
3185                         }
3186                         btrfs_item_key_to_cpu(l, &key, slot);
3187
3188                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3189                             key.type != BTRFS_METADATA_ITEM_KEY)
3190                                 goto next;
3191
3192                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3193                                 bytes = fs_info->nodesize;
3194                         else
3195                                 bytes = key.offset;
3196
3197                         if (key.objectid + bytes <= logic_start)
3198                                 goto next;
3199
3200                         if (key.objectid >= logic_end) {
3201                                 stop_loop = 1;
3202                                 break;
3203                         }
3204
3205                         while (key.objectid >= logic_start + map->stripe_len)
3206                                 logic_start += map->stripe_len;
3207
3208                         extent = btrfs_item_ptr(l, slot,
3209                                                 struct btrfs_extent_item);
3210                         flags = btrfs_extent_flags(l, extent);
3211                         generation = btrfs_extent_generation(l, extent);
3212
3213                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3214                             (key.objectid < logic_start ||
3215                              key.objectid + bytes >
3216                              logic_start + map->stripe_len)) {
3217                                 btrfs_err(fs_info,
3218                                           "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3219                                           key.objectid, logic_start);
3220                                 spin_lock(&sctx->stat_lock);
3221                                 sctx->stat.uncorrectable_errors++;
3222                                 spin_unlock(&sctx->stat_lock);
3223                                 goto next;
3224                         }
3225 again:
3226                         extent_logical = key.objectid;
3227                         extent_len = bytes;
3228
3229                         if (extent_logical < logic_start) {
3230                                 extent_len -= logic_start - extent_logical;
3231                                 extent_logical = logic_start;
3232                         }
3233
3234                         if (extent_logical + extent_len >
3235                             logic_start + map->stripe_len)
3236                                 extent_len = logic_start + map->stripe_len -
3237                                              extent_logical;
3238
3239                         scrub_parity_mark_sectors_data(sparity, extent_logical,
3240                                                        extent_len);
3241
3242                         mapped_length = extent_len;
3243                         bbio = NULL;
3244                         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ,
3245                                         extent_logical, &mapped_length, &bbio,
3246                                         0);
3247                         if (!ret) {
3248                                 if (!bbio || mapped_length < extent_len)
3249                                         ret = -EIO;
3250                         }
3251                         if (ret) {
3252                                 btrfs_put_bbio(bbio);
3253                                 goto out;
3254                         }
3255                         extent_physical = bbio->stripes[0].physical;
3256                         extent_mirror_num = bbio->mirror_num;
3257                         extent_dev = bbio->stripes[0].dev;
3258                         btrfs_put_bbio(bbio);
3259
3260                         ret = btrfs_lookup_csums_range(csum_root,
3261                                                 extent_logical,
3262                                                 extent_logical + extent_len - 1,
3263                                                 &sctx->csum_list, 1);
3264                         if (ret)
3265                                 goto out;
3266
3267                         ret = scrub_extent_for_parity(sparity, extent_logical,
3268                                                       extent_len,
3269                                                       extent_physical,
3270                                                       extent_dev, flags,
3271                                                       generation,
3272                                                       extent_mirror_num);
3273
3274                         scrub_free_csums(sctx);
3275
3276                         if (ret)
3277                                 goto out;
3278
3279                         if (extent_logical + extent_len <
3280                             key.objectid + bytes) {
3281                                 logic_start += map->stripe_len;
3282
3283                                 if (logic_start >= logic_end) {
3284                                         stop_loop = 1;
3285                                         break;
3286                                 }
3287
3288                                 if (logic_start < key.objectid + bytes) {
3289                                         cond_resched();
3290                                         goto again;
3291                                 }
3292                         }
3293 next:
3294                         path->slots[0]++;
3295                 }
3296
3297                 btrfs_release_path(path);
3298
3299                 if (stop_loop)
3300                         break;
3301
3302                 logic_start += map->stripe_len;
3303         }
3304 out:
3305         if (ret < 0)
3306                 scrub_parity_mark_sectors_error(sparity, logic_start,
3307                                                 logic_end - logic_start);
3308         scrub_parity_put(sparity);
3309         scrub_submit(sctx);
3310         mutex_lock(&sctx->wr_lock);
3311         scrub_wr_submit(sctx);
3312         mutex_unlock(&sctx->wr_lock);
3313
3314         btrfs_release_path(path);
3315         return ret < 0 ? ret : 0;
3316 }
3317
3318 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3319                                            struct map_lookup *map,
3320                                            struct btrfs_device *scrub_dev,
3321                                            int num, u64 base, u64 length,
3322                                            int is_dev_replace)
3323 {
3324         struct btrfs_path *path, *ppath;
3325         struct btrfs_fs_info *fs_info = sctx->fs_info;
3326         struct btrfs_root *root = fs_info->extent_root;
3327         struct btrfs_root *csum_root = fs_info->csum_root;
3328         struct btrfs_extent_item *extent;
3329         struct blk_plug plug;
3330         u64 flags;
3331         int ret;
3332         int slot;
3333         u64 nstripes;
3334         struct extent_buffer *l;
3335         u64 physical;
3336         u64 logical;
3337         u64 logic_end;
3338         u64 physical_end;
3339         u64 generation;
3340         int mirror_num;
3341         struct reada_control *reada1;
3342         struct reada_control *reada2;
3343         struct btrfs_key key;
3344         struct btrfs_key key_end;
3345         u64 increment = map->stripe_len;
3346         u64 offset;
3347         u64 extent_logical;
3348         u64 extent_physical;
3349         u64 extent_len;
3350         u64 stripe_logical;
3351         u64 stripe_end;
3352         struct btrfs_device *extent_dev;
3353         int extent_mirror_num;
3354         int stop_loop = 0;
3355
3356         physical = map->stripes[num].physical;
3357         offset = 0;
3358         nstripes = div64_u64(length, map->stripe_len);
3359         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3360                 offset = map->stripe_len * num;
3361                 increment = map->stripe_len * map->num_stripes;
3362                 mirror_num = 1;
3363         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3364                 int factor = map->num_stripes / map->sub_stripes;
3365                 offset = map->stripe_len * (num / map->sub_stripes);
3366                 increment = map->stripe_len * factor;
3367                 mirror_num = num % map->sub_stripes + 1;
3368         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3369                 increment = map->stripe_len;
3370                 mirror_num = num % map->num_stripes + 1;
3371         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3372                 increment = map->stripe_len;
3373                 mirror_num = num % map->num_stripes + 1;
3374         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3375                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3376                 increment = map->stripe_len * nr_data_stripes(map);
3377                 mirror_num = 1;
3378         } else {
3379                 increment = map->stripe_len;
3380                 mirror_num = 1;
3381         }
3382
3383         path = btrfs_alloc_path();
3384         if (!path)
3385                 return -ENOMEM;
3386
3387         ppath = btrfs_alloc_path();
3388         if (!ppath) {
3389                 btrfs_free_path(path);
3390                 return -ENOMEM;
3391         }
3392
3393         /*
3394          * work on commit root. The related disk blocks are static as
3395          * long as COW is applied. This means, it is save to rewrite
3396          * them to repair disk errors without any race conditions
3397          */
3398         path->search_commit_root = 1;
3399         path->skip_locking = 1;
3400
3401         ppath->search_commit_root = 1;
3402         ppath->skip_locking = 1;
3403         /*
3404          * trigger the readahead for extent tree csum tree and wait for
3405          * completion. During readahead, the scrub is officially paused
3406          * to not hold off transaction commits
3407          */
3408         logical = base + offset;
3409         physical_end = physical + nstripes * map->stripe_len;
3410         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3411                 get_raid56_logic_offset(physical_end, num,
3412                                         map, &logic_end, NULL);
3413                 logic_end += base;
3414         } else {
3415                 logic_end = logical + increment * nstripes;
3416         }
3417         wait_event(sctx->list_wait,
3418                    atomic_read(&sctx->bios_in_flight) == 0);
3419         scrub_blocked_if_needed(fs_info);
3420
3421         /* FIXME it might be better to start readahead at commit root */
3422         key.objectid = logical;
3423         key.type = BTRFS_EXTENT_ITEM_KEY;
3424         key.offset = (u64)0;
3425         key_end.objectid = logic_end;
3426         key_end.type = BTRFS_METADATA_ITEM_KEY;
3427         key_end.offset = (u64)-1;
3428         reada1 = btrfs_reada_add(root, &key, &key_end);
3429
3430         key.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3431         key.type = BTRFS_EXTENT_CSUM_KEY;
3432         key.offset = logical;
3433         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3434         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3435         key_end.offset = logic_end;
3436         reada2 = btrfs_reada_add(csum_root, &key, &key_end);
3437
3438         if (!IS_ERR(reada1))
3439                 btrfs_reada_wait(reada1);
3440         if (!IS_ERR(reada2))
3441                 btrfs_reada_wait(reada2);
3442
3443
3444         /*
3445          * collect all data csums for the stripe to avoid seeking during
3446          * the scrub. This might currently (crc32) end up to be about 1MB
3447          */
3448         blk_start_plug(&plug);
3449
3450         /*
3451          * now find all extents for each stripe and scrub them
3452          */
3453         ret = 0;
3454         while (physical < physical_end) {
3455                 /*
3456                  * canceled?
3457                  */
3458                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3459                     atomic_read(&sctx->cancel_req)) {
3460                         ret = -ECANCELED;
3461                         goto out;
3462                 }
3463                 /*
3464                  * check to see if we have to pause
3465                  */
3466                 if (atomic_read(&fs_info->scrub_pause_req)) {
3467                         /* push queued extents */
3468                         sctx->flush_all_writes = true;
3469                         scrub_submit(sctx);
3470                         mutex_lock(&sctx->wr_lock);
3471                         scrub_wr_submit(sctx);
3472                         mutex_unlock(&sctx->wr_lock);
3473                         wait_event(sctx->list_wait,
3474                                    atomic_read(&sctx->bios_in_flight) == 0);
3475                         sctx->flush_all_writes = false;
3476                         scrub_blocked_if_needed(fs_info);
3477                 }
3478
3479                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3480                         ret = get_raid56_logic_offset(physical, num, map,
3481                                                       &logical,
3482                                                       &stripe_logical);
3483                         logical += base;
3484                         if (ret) {
3485                                 /* it is parity strip */
3486                                 stripe_logical += base;
3487                                 stripe_end = stripe_logical + increment;
3488                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3489                                                           ppath, stripe_logical,
3490                                                           stripe_end);
3491                                 if (ret)
3492                                         goto out;
3493                                 goto skip;
3494                         }
3495                 }
3496
3497                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3498                         key.type = BTRFS_METADATA_ITEM_KEY;
3499                 else
3500                         key.type = BTRFS_EXTENT_ITEM_KEY;
3501                 key.objectid = logical;
3502                 key.offset = (u64)-1;
3503
3504                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3505                 if (ret < 0)
3506                         goto out;
3507
3508                 if (ret > 0) {
3509                         ret = btrfs_previous_extent_item(root, path, 0);
3510                         if (ret < 0)
3511                                 goto out;
3512                         if (ret > 0) {
3513                                 /* there's no smaller item, so stick with the
3514                                  * larger one */
3515                                 btrfs_release_path(path);
3516                                 ret = btrfs_search_slot(NULL, root, &key,
3517                                                         path, 0, 0);
3518                                 if (ret < 0)
3519                                         goto out;
3520                         }
3521                 }
3522
3523                 stop_loop = 0;
3524                 while (1) {
3525                         u64 bytes;
3526
3527                         l = path->nodes[0];
3528                         slot = path->slots[0];
3529                         if (slot >= btrfs_header_nritems(l)) {
3530                                 ret = btrfs_next_leaf(root, path);
3531                                 if (ret == 0)
3532                                         continue;
3533                                 if (ret < 0)
3534                                         goto out;
3535
3536                                 stop_loop = 1;
3537                                 break;
3538                         }
3539                         btrfs_item_key_to_cpu(l, &key, slot);
3540
3541                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3542                             key.type != BTRFS_METADATA_ITEM_KEY)
3543                                 goto next;
3544
3545                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3546                                 bytes = fs_info->nodesize;
3547                         else
3548                                 bytes = key.offset;
3549
3550                         if (key.objectid + bytes <= logical)
3551                                 goto next;
3552
3553                         if (key.objectid >= logical + map->stripe_len) {
3554                                 /* out of this device extent */
3555                                 if (key.objectid >= logic_end)
3556                                         stop_loop = 1;
3557                                 break;
3558                         }
3559
3560                         extent = btrfs_item_ptr(l, slot,
3561                                                 struct btrfs_extent_item);
3562                         flags = btrfs_extent_flags(l, extent);
3563                         generation = btrfs_extent_generation(l, extent);
3564
3565                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3566                             (key.objectid < logical ||
3567                              key.objectid + bytes >
3568                              logical + map->stripe_len)) {
3569                                 btrfs_err(fs_info,
3570                                            "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
3571                                        key.objectid, logical);
3572                                 spin_lock(&sctx->stat_lock);
3573                                 sctx->stat.uncorrectable_errors++;
3574                                 spin_unlock(&sctx->stat_lock);
3575                                 goto next;
3576                         }
3577
3578 again:
3579                         extent_logical = key.objectid;
3580                         extent_len = bytes;
3581
3582                         /*
3583                          * trim extent to this stripe
3584                          */
3585                         if (extent_logical < logical) {
3586                                 extent_len -= logical - extent_logical;
3587                                 extent_logical = logical;
3588                         }
3589                         if (extent_logical + extent_len >
3590                             logical + map->stripe_len) {
3591                                 extent_len = logical + map->stripe_len -
3592                                              extent_logical;
3593                         }
3594
3595                         extent_physical = extent_logical - logical + physical;
3596                         extent_dev = scrub_dev;
3597                         extent_mirror_num = mirror_num;
3598                         if (is_dev_replace)
3599                                 scrub_remap_extent(fs_info, extent_logical,
3600                                                    extent_len, &extent_physical,
3601                                                    &extent_dev,
3602                                                    &extent_mirror_num);
3603
3604                         ret = btrfs_lookup_csums_range(csum_root,
3605                                                        extent_logical,
3606                                                        extent_logical +
3607                                                        extent_len - 1,
3608                                                        &sctx->csum_list, 1);
3609                         if (ret)
3610                                 goto out;
3611
3612                         ret = scrub_extent(sctx, extent_logical, extent_len,
3613                                            extent_physical, extent_dev, flags,
3614                                            generation, extent_mirror_num,
3615                                            extent_logical - logical + physical);
3616
3617                         scrub_free_csums(sctx);
3618
3619                         if (ret)
3620                                 goto out;
3621
3622                         if (extent_logical + extent_len <
3623                             key.objectid + bytes) {
3624                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3625                                         /*
3626                                          * loop until we find next data stripe
3627                                          * or we have finished all stripes.
3628                                          */
3629 loop:
3630                                         physical += map->stripe_len;
3631                                         ret = get_raid56_logic_offset(physical,
3632                                                         num, map, &logical,
3633                                                         &stripe_logical);
3634                                         logical += base;
3635
3636                                         if (ret && physical < physical_end) {
3637                                                 stripe_logical += base;
3638                                                 stripe_end = stripe_logical +
3639                                                                 increment;
3640                                                 ret = scrub_raid56_parity(sctx,
3641                                                         map, scrub_dev, ppath,
3642                                                         stripe_logical,
3643                                                         stripe_end);
3644                                                 if (ret)
3645                                                         goto out;
3646                                                 goto loop;
3647                                         }
3648                                 } else {
3649                                         physical += map->stripe_len;
3650                                         logical += increment;
3651                                 }
3652                                 if (logical < key.objectid + bytes) {
3653                                         cond_resched();
3654                                         goto again;
3655                                 }
3656
3657                                 if (physical >= physical_end) {
3658                                         stop_loop = 1;
3659                                         break;
3660                                 }
3661                         }
3662 next:
3663                         path->slots[0]++;
3664                 }
3665                 btrfs_release_path(path);
3666 skip:
3667                 logical += increment;
3668                 physical += map->stripe_len;
3669                 spin_lock(&sctx->stat_lock);
3670                 if (stop_loop)
3671                         sctx->stat.last_physical = map->stripes[num].physical +
3672                                                    length;
3673                 else
3674                         sctx->stat.last_physical = physical;
3675                 spin_unlock(&sctx->stat_lock);
3676                 if (stop_loop)
3677                         break;
3678         }
3679 out:
3680         /* push queued extents */
3681         scrub_submit(sctx);
3682         mutex_lock(&sctx->wr_lock);
3683         scrub_wr_submit(sctx);
3684         mutex_unlock(&sctx->wr_lock);
3685
3686         blk_finish_plug(&plug);
3687         btrfs_free_path(path);
3688         btrfs_free_path(ppath);
3689         return ret < 0 ? ret : 0;
3690 }
3691
3692 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3693                                           struct btrfs_device *scrub_dev,
3694                                           u64 chunk_offset, u64 length,
3695                                           u64 dev_offset,
3696                                           struct btrfs_block_group_cache *cache,
3697                                           int is_dev_replace)
3698 {
3699         struct btrfs_fs_info *fs_info = sctx->fs_info;
3700         struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
3701         struct map_lookup *map;
3702         struct extent_map *em;
3703         int i;
3704         int ret = 0;
3705
3706         read_lock(&map_tree->map_tree.lock);
3707         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3708         read_unlock(&map_tree->map_tree.lock);
3709
3710         if (!em) {
3711                 /*
3712                  * Might have been an unused block group deleted by the cleaner
3713                  * kthread or relocation.
3714                  */
3715                 spin_lock(&cache->lock);
3716                 if (!cache->removed)
3717                         ret = -EINVAL;
3718                 spin_unlock(&cache->lock);
3719
3720                 return ret;
3721         }
3722
3723         map = em->map_lookup;
3724         if (em->start != chunk_offset)
3725                 goto out;
3726
3727         if (em->len < length)
3728                 goto out;
3729
3730         for (i = 0; i < map->num_stripes; ++i) {
3731                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3732                     map->stripes[i].physical == dev_offset) {
3733                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3734                                            chunk_offset, length,
3735                                            is_dev_replace);
3736                         if (ret)
3737                                 goto out;
3738                 }
3739         }
3740 out:
3741         free_extent_map(em);
3742
3743         return ret;
3744 }
3745
3746 static noinline_for_stack
3747 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3748                            struct btrfs_device *scrub_dev, u64 start, u64 end,
3749                            int is_dev_replace)
3750 {
3751         struct btrfs_dev_extent *dev_extent = NULL;
3752         struct btrfs_path *path;
3753         struct btrfs_fs_info *fs_info = sctx->fs_info;
3754         struct btrfs_root *root = fs_info->dev_root;
3755         u64 length;
3756         u64 chunk_offset;
3757         int ret = 0;
3758         int ro_set;
3759         int slot;
3760         struct extent_buffer *l;
3761         struct btrfs_key key;
3762         struct btrfs_key found_key;
3763         struct btrfs_block_group_cache *cache;
3764         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3765
3766         path = btrfs_alloc_path();
3767         if (!path)
3768                 return -ENOMEM;
3769
3770         path->reada = READA_FORWARD;
3771         path->search_commit_root = 1;
3772         path->skip_locking = 1;
3773
3774         key.objectid = scrub_dev->devid;
3775         key.offset = 0ull;
3776         key.type = BTRFS_DEV_EXTENT_KEY;
3777
3778         while (1) {
3779                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3780                 if (ret < 0)
3781                         break;
3782                 if (ret > 0) {
3783                         if (path->slots[0] >=
3784                             btrfs_header_nritems(path->nodes[0])) {
3785                                 ret = btrfs_next_leaf(root, path);
3786                                 if (ret < 0)
3787                                         break;
3788                                 if (ret > 0) {
3789                                         ret = 0;
3790                                         break;
3791                                 }
3792                         } else {
3793                                 ret = 0;
3794                         }
3795                 }
3796
3797                 l = path->nodes[0];
3798                 slot = path->slots[0];
3799
3800                 btrfs_item_key_to_cpu(l, &found_key, slot);
3801
3802                 if (found_key.objectid != scrub_dev->devid)
3803                         break;
3804
3805                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3806                         break;
3807
3808                 if (found_key.offset >= end)
3809                         break;
3810
3811                 if (found_key.offset < key.offset)
3812                         break;
3813
3814                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3815                 length = btrfs_dev_extent_length(l, dev_extent);
3816
3817                 if (found_key.offset + length <= start)
3818                         goto skip;
3819
3820                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3821
3822                 /*
3823                  * get a reference on the corresponding block group to prevent
3824                  * the chunk from going away while we scrub it
3825                  */
3826                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3827
3828                 /* some chunks are removed but not committed to disk yet,
3829                  * continue scrubbing */
3830                 if (!cache)
3831                         goto skip;
3832
3833                 /*
3834                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3835                  * to avoid deadlock caused by:
3836                  * btrfs_inc_block_group_ro()
3837                  * -> btrfs_wait_for_commit()
3838                  * -> btrfs_commit_transaction()
3839                  * -> btrfs_scrub_pause()
3840                  */
3841                 scrub_pause_on(fs_info);
3842                 ret = btrfs_inc_block_group_ro(fs_info, cache);
3843                 if (!ret && is_dev_replace) {
3844                         /*
3845                          * If we are doing a device replace wait for any tasks
3846                          * that started dellaloc right before we set the block
3847                          * group to RO mode, as they might have just allocated
3848                          * an extent from it or decided they could do a nocow
3849                          * write. And if any such tasks did that, wait for their
3850                          * ordered extents to complete and then commit the
3851                          * current transaction, so that we can later see the new
3852                          * extent items in the extent tree - the ordered extents
3853                          * create delayed data references (for cow writes) when
3854                          * they complete, which will be run and insert the
3855                          * corresponding extent items into the extent tree when
3856                          * we commit the transaction they used when running
3857                          * inode.c:btrfs_finish_ordered_io(). We later use
3858                          * the commit root of the extent tree to find extents
3859                          * to copy from the srcdev into the tgtdev, and we don't
3860                          * want to miss any new extents.
3861                          */
3862                         btrfs_wait_block_group_reservations(cache);
3863                         btrfs_wait_nocow_writers(cache);
3864                         ret = btrfs_wait_ordered_roots(fs_info, U64_MAX,
3865                                                        cache->key.objectid,
3866                                                        cache->key.offset);
3867                         if (ret > 0) {
3868                                 struct btrfs_trans_handle *trans;
3869
3870                                 trans = btrfs_join_transaction(root);
3871                                 if (IS_ERR(trans))
3872                                         ret = PTR_ERR(trans);
3873                                 else
3874                                         ret = btrfs_commit_transaction(trans);
3875                                 if (ret) {
3876                                         scrub_pause_off(fs_info);
3877                                         btrfs_put_block_group(cache);
3878                                         break;
3879                                 }
3880                         }
3881                 }
3882                 scrub_pause_off(fs_info);
3883
3884                 if (ret == 0) {
3885                         ro_set = 1;
3886                 } else if (ret == -ENOSPC) {
3887                         /*
3888                          * btrfs_inc_block_group_ro return -ENOSPC when it
3889                          * failed in creating new chunk for metadata.
3890                          * It is not a problem for scrub/replace, because
3891                          * metadata are always cowed, and our scrub paused
3892                          * commit_transactions.
3893                          */
3894                         ro_set = 0;
3895                 } else {
3896                         btrfs_warn(fs_info,
3897                                    "failed setting block group ro: %d", ret);
3898                         btrfs_put_block_group(cache);
3899                         break;
3900                 }
3901
3902                 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3903                 dev_replace->cursor_right = found_key.offset + length;
3904                 dev_replace->cursor_left = found_key.offset;
3905                 dev_replace->item_needs_writeback = 1;
3906                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3907                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3908                                   found_key.offset, cache, is_dev_replace);
3909
3910                 /*
3911                  * flush, submit all pending read and write bios, afterwards
3912                  * wait for them.
3913                  * Note that in the dev replace case, a read request causes
3914                  * write requests that are submitted in the read completion
3915                  * worker. Therefore in the current situation, it is required
3916                  * that all write requests are flushed, so that all read and
3917                  * write requests are really completed when bios_in_flight
3918                  * changes to 0.
3919                  */
3920                 sctx->flush_all_writes = true;
3921                 scrub_submit(sctx);
3922                 mutex_lock(&sctx->wr_lock);
3923                 scrub_wr_submit(sctx);
3924                 mutex_unlock(&sctx->wr_lock);
3925
3926                 wait_event(sctx->list_wait,
3927                            atomic_read(&sctx->bios_in_flight) == 0);
3928
3929                 scrub_pause_on(fs_info);
3930
3931                 /*
3932                  * must be called before we decrease @scrub_paused.
3933                  * make sure we don't block transaction commit while
3934                  * we are waiting pending workers finished.
3935                  */
3936                 wait_event(sctx->list_wait,
3937                            atomic_read(&sctx->workers_pending) == 0);
3938                 sctx->flush_all_writes = false;
3939
3940                 scrub_pause_off(fs_info);
3941
3942                 btrfs_dev_replace_lock(&fs_info->dev_replace, 1);
3943                 dev_replace->cursor_left = dev_replace->cursor_right;
3944                 dev_replace->item_needs_writeback = 1;
3945                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 1);
3946
3947                 if (ro_set)
3948                         btrfs_dec_block_group_ro(cache);
3949
3950                 /*
3951                  * We might have prevented the cleaner kthread from deleting
3952                  * this block group if it was already unused because we raced
3953                  * and set it to RO mode first. So add it back to the unused
3954                  * list, otherwise it might not ever be deleted unless a manual
3955                  * balance is triggered or it becomes used and unused again.
3956                  */
3957                 spin_lock(&cache->lock);
3958                 if (!cache->removed && !cache->ro && cache->reserved == 0 &&
3959                     btrfs_block_group_used(&cache->item) == 0) {
3960                         spin_unlock(&cache->lock);
3961                         spin_lock(&fs_info->unused_bgs_lock);
3962                         if (list_empty(&cache->bg_list)) {
3963                                 btrfs_get_block_group(cache);
3964                                 list_add_tail(&cache->bg_list,
3965                                               &fs_info->unused_bgs);
3966                         }
3967                         spin_unlock(&fs_info->unused_bgs_lock);
3968                 } else {
3969                         spin_unlock(&cache->lock);
3970                 }
3971
3972                 btrfs_put_block_group(cache);
3973                 if (ret)
3974                         break;
3975                 if (is_dev_replace &&
3976                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3977                         ret = -EIO;
3978                         break;
3979                 }
3980                 if (sctx->stat.malloc_errors > 0) {
3981                         ret = -ENOMEM;
3982                         break;
3983                 }
3984 skip:
3985                 key.offset = found_key.offset + length;
3986                 btrfs_release_path(path);
3987         }
3988
3989         btrfs_free_path(path);
3990
3991         return ret;
3992 }
3993
3994 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3995                                            struct btrfs_device *scrub_dev)
3996 {
3997         int     i;
3998         u64     bytenr;
3999         u64     gen;
4000         int     ret;
4001         struct btrfs_fs_info *fs_info = sctx->fs_info;
4002
4003         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4004                 return -EIO;
4005
4006         /* Seed devices of a new filesystem has their own generation. */
4007         if (scrub_dev->fs_devices != fs_info->fs_devices)
4008                 gen = scrub_dev->generation;
4009         else
4010                 gen = fs_info->last_trans_committed;
4011
4012         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
4013                 bytenr = btrfs_sb_offset(i);
4014                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
4015                     scrub_dev->commit_total_bytes)
4016                         break;
4017
4018                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
4019                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
4020                                   NULL, 1, bytenr);
4021                 if (ret)
4022                         return ret;
4023         }
4024         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4025
4026         return 0;
4027 }
4028
4029 /*
4030  * get a reference count on fs_info->scrub_workers. start worker if necessary
4031  */
4032 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
4033                                                 int is_dev_replace)
4034 {
4035         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
4036         int max_active = fs_info->thread_pool_size;
4037
4038         if (fs_info->scrub_workers_refcnt == 0) {
4039                 fs_info->scrub_workers = btrfs_alloc_workqueue(fs_info, "scrub",
4040                                 flags, is_dev_replace ? 1 : max_active, 4);
4041                 if (!fs_info->scrub_workers)
4042                         goto fail_scrub_workers;
4043
4044                 fs_info->scrub_wr_completion_workers =
4045                         btrfs_alloc_workqueue(fs_info, "scrubwrc", flags,
4046                                               max_active, 2);
4047                 if (!fs_info->scrub_wr_completion_workers)
4048                         goto fail_scrub_wr_completion_workers;
4049
4050                 fs_info->scrub_nocow_workers =
4051                         btrfs_alloc_workqueue(fs_info, "scrubnc", flags, 1, 0);
4052                 if (!fs_info->scrub_nocow_workers)
4053                         goto fail_scrub_nocow_workers;
4054                 fs_info->scrub_parity_workers =
4055                         btrfs_alloc_workqueue(fs_info, "scrubparity", flags,
4056                                               max_active, 2);
4057                 if (!fs_info->scrub_parity_workers)
4058                         goto fail_scrub_parity_workers;
4059         }
4060         ++fs_info->scrub_workers_refcnt;
4061         return 0;
4062
4063 fail_scrub_parity_workers:
4064         btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4065 fail_scrub_nocow_workers:
4066         btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4067 fail_scrub_wr_completion_workers:
4068         btrfs_destroy_workqueue(fs_info->scrub_workers);
4069 fail_scrub_workers:
4070         return -ENOMEM;
4071 }
4072
4073 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
4074 {
4075         if (--fs_info->scrub_workers_refcnt == 0) {
4076                 btrfs_destroy_workqueue(fs_info->scrub_workers);
4077                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
4078                 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
4079                 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
4080         }
4081         WARN_ON(fs_info->scrub_workers_refcnt < 0);
4082 }
4083
4084 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
4085                     u64 end, struct btrfs_scrub_progress *progress,
4086                     int readonly, int is_dev_replace)
4087 {
4088         struct scrub_ctx *sctx;
4089         int ret;
4090         struct btrfs_device *dev;
4091         struct rcu_string *name;
4092
4093         if (btrfs_fs_closing(fs_info))
4094                 return -EINVAL;
4095
4096         if (fs_info->nodesize > BTRFS_STRIPE_LEN) {
4097                 /*
4098                  * in this case scrub is unable to calculate the checksum
4099                  * the way scrub is implemented. Do not handle this
4100                  * situation at all because it won't ever happen.
4101                  */
4102                 btrfs_err(fs_info,
4103                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
4104                        fs_info->nodesize,
4105                        BTRFS_STRIPE_LEN);
4106                 return -EINVAL;
4107         }
4108
4109         if (fs_info->sectorsize != PAGE_SIZE) {
4110                 /* not supported for data w/o checksums */
4111                 btrfs_err_rl(fs_info,
4112                            "scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails",
4113                        fs_info->sectorsize, PAGE_SIZE);
4114                 return -EINVAL;
4115         }
4116
4117         if (fs_info->nodesize >
4118             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
4119             fs_info->sectorsize > PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
4120                 /*
4121                  * would exhaust the array bounds of pagev member in
4122                  * struct scrub_block
4123                  */
4124                 btrfs_err(fs_info,
4125                           "scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
4126                        fs_info->nodesize,
4127                        SCRUB_MAX_PAGES_PER_BLOCK,
4128                        fs_info->sectorsize,
4129                        SCRUB_MAX_PAGES_PER_BLOCK);
4130                 return -EINVAL;
4131         }
4132
4133
4134         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4135         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4136         if (!dev || (dev->missing && !is_dev_replace)) {
4137                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4138                 return -ENODEV;
4139         }
4140
4141         if (!is_dev_replace && !readonly && !dev->writeable) {
4142                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4143                 rcu_read_lock();
4144                 name = rcu_dereference(dev->name);
4145                 btrfs_err(fs_info, "scrub: device %s is not writable",
4146                           name->str);
4147                 rcu_read_unlock();
4148                 return -EROFS;
4149         }
4150
4151         mutex_lock(&fs_info->scrub_lock);
4152         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
4153                 mutex_unlock(&fs_info->scrub_lock);
4154                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4155                 return -EIO;
4156         }
4157
4158         btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
4159         if (dev->scrub_device ||
4160             (!is_dev_replace &&
4161              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
4162                 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
4163                 mutex_unlock(&fs_info->scrub_lock);
4164                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4165                 return -EINPROGRESS;
4166         }
4167         btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
4168
4169         ret = scrub_workers_get(fs_info, is_dev_replace);
4170         if (ret) {
4171                 mutex_unlock(&fs_info->scrub_lock);
4172                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4173                 return ret;
4174         }
4175
4176         sctx = scrub_setup_ctx(dev, is_dev_replace);
4177         if (IS_ERR(sctx)) {
4178                 mutex_unlock(&fs_info->scrub_lock);
4179                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4180                 scrub_workers_put(fs_info);
4181                 return PTR_ERR(sctx);
4182         }
4183         sctx->readonly = readonly;
4184         dev->scrub_device = sctx;
4185         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4186
4187         /*
4188          * checking @scrub_pause_req here, we can avoid
4189          * race between committing transaction and scrubbing.
4190          */
4191         __scrub_blocked_if_needed(fs_info);
4192         atomic_inc(&fs_info->scrubs_running);
4193         mutex_unlock(&fs_info->scrub_lock);
4194
4195         if (!is_dev_replace) {
4196                 /*
4197                  * by holding device list mutex, we can
4198                  * kick off writing super in log tree sync.
4199                  */
4200                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
4201                 ret = scrub_supers(sctx, dev);
4202                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4203         }
4204
4205         if (!ret)
4206                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
4207                                              is_dev_replace);
4208
4209         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
4210         atomic_dec(&fs_info->scrubs_running);
4211         wake_up(&fs_info->scrub_pause_wait);
4212
4213         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
4214
4215         if (progress)
4216                 memcpy(progress, &sctx->stat, sizeof(*progress));
4217
4218         mutex_lock(&fs_info->scrub_lock);
4219         dev->scrub_device = NULL;
4220         scrub_workers_put(fs_info);
4221         mutex_unlock(&fs_info->scrub_lock);
4222
4223         scrub_put_ctx(sctx);
4224
4225         return ret;
4226 }
4227
4228 void btrfs_scrub_pause(struct btrfs_fs_info *fs_info)
4229 {
4230         mutex_lock(&fs_info->scrub_lock);
4231         atomic_inc(&fs_info->scrub_pause_req);
4232         while (atomic_read(&fs_info->scrubs_paused) !=
4233                atomic_read(&fs_info->scrubs_running)) {
4234                 mutex_unlock(&fs_info->scrub_lock);
4235                 wait_event(fs_info->scrub_pause_wait,
4236                            atomic_read(&fs_info->scrubs_paused) ==
4237                            atomic_read(&fs_info->scrubs_running));
4238                 mutex_lock(&fs_info->scrub_lock);
4239         }
4240         mutex_unlock(&fs_info->scrub_lock);
4241 }
4242
4243 void btrfs_scrub_continue(struct btrfs_fs_info *fs_info)
4244 {
4245         atomic_dec(&fs_info->scrub_pause_req);
4246         wake_up(&fs_info->scrub_pause_wait);
4247 }
4248
4249 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
4250 {
4251         mutex_lock(&fs_info->scrub_lock);
4252         if (!atomic_read(&fs_info->scrubs_running)) {
4253                 mutex_unlock(&fs_info->scrub_lock);
4254                 return -ENOTCONN;
4255         }
4256
4257         atomic_inc(&fs_info->scrub_cancel_req);
4258         while (atomic_read(&fs_info->scrubs_running)) {
4259                 mutex_unlock(&fs_info->scrub_lock);
4260                 wait_event(fs_info->scrub_pause_wait,
4261                            atomic_read(&fs_info->scrubs_running) == 0);
4262                 mutex_lock(&fs_info->scrub_lock);
4263         }
4264         atomic_dec(&fs_info->scrub_cancel_req);
4265         mutex_unlock(&fs_info->scrub_lock);
4266
4267         return 0;
4268 }
4269
4270 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
4271                            struct btrfs_device *dev)
4272 {
4273         struct scrub_ctx *sctx;
4274
4275         mutex_lock(&fs_info->scrub_lock);
4276         sctx = dev->scrub_device;
4277         if (!sctx) {
4278                 mutex_unlock(&fs_info->scrub_lock);
4279                 return -ENOTCONN;
4280         }
4281         atomic_inc(&sctx->cancel_req);
4282         while (dev->scrub_device) {
4283                 mutex_unlock(&fs_info->scrub_lock);
4284                 wait_event(fs_info->scrub_pause_wait,
4285                            dev->scrub_device == NULL);
4286                 mutex_lock(&fs_info->scrub_lock);
4287         }
4288         mutex_unlock(&fs_info->scrub_lock);
4289
4290         return 0;
4291 }
4292
4293 int btrfs_scrub_progress(struct btrfs_fs_info *fs_info, u64 devid,
4294                          struct btrfs_scrub_progress *progress)
4295 {
4296         struct btrfs_device *dev;
4297         struct scrub_ctx *sctx = NULL;
4298
4299         mutex_lock(&fs_info->fs_devices->device_list_mutex);
4300         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
4301         if (dev)
4302                 sctx = dev->scrub_device;
4303         if (sctx)
4304                 memcpy(progress, &sctx->stat, sizeof(*progress));
4305         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
4306
4307         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
4308 }
4309
4310 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
4311                                u64 extent_logical, u64 extent_len,
4312                                u64 *extent_physical,
4313                                struct btrfs_device **extent_dev,
4314                                int *extent_mirror_num)
4315 {
4316         u64 mapped_length;
4317         struct btrfs_bio *bbio = NULL;
4318         int ret;
4319
4320         mapped_length = extent_len;
4321         ret = btrfs_map_block(fs_info, BTRFS_MAP_READ, extent_logical,
4322                               &mapped_length, &bbio, 0);
4323         if (ret || !bbio || mapped_length < extent_len ||
4324             !bbio->stripes[0].dev->bdev) {
4325                 btrfs_put_bbio(bbio);
4326                 return;
4327         }
4328
4329         *extent_physical = bbio->stripes[0].physical;
4330         *extent_mirror_num = bbio->mirror_num;
4331         *extent_dev = bbio->stripes[0].dev;
4332         btrfs_put_bbio(bbio);
4333 }
4334
4335 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4336                             int mirror_num, u64 physical_for_dev_replace)
4337 {
4338         struct scrub_copy_nocow_ctx *nocow_ctx;
4339         struct btrfs_fs_info *fs_info = sctx->fs_info;
4340
4341         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4342         if (!nocow_ctx) {
4343                 spin_lock(&sctx->stat_lock);
4344                 sctx->stat.malloc_errors++;
4345                 spin_unlock(&sctx->stat_lock);
4346                 return -ENOMEM;
4347         }
4348
4349         scrub_pending_trans_workers_inc(sctx);
4350
4351         nocow_ctx->sctx = sctx;
4352         nocow_ctx->logical = logical;
4353         nocow_ctx->len = len;
4354         nocow_ctx->mirror_num = mirror_num;
4355         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4356         btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4357                         copy_nocow_pages_worker, NULL, NULL);
4358         INIT_LIST_HEAD(&nocow_ctx->inodes);
4359         btrfs_queue_work(fs_info->scrub_nocow_workers,
4360                          &nocow_ctx->work);
4361
4362         return 0;
4363 }
4364
4365 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4366 {
4367         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4368         struct scrub_nocow_inode *nocow_inode;
4369
4370         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4371         if (!nocow_inode)
4372                 return -ENOMEM;
4373         nocow_inode->inum = inum;
4374         nocow_inode->offset = offset;
4375         nocow_inode->root = root;
4376         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4377         return 0;
4378 }
4379
4380 #define COPY_COMPLETE 1
4381
4382 static void copy_nocow_pages_worker(struct btrfs_work *work)
4383 {
4384         struct scrub_copy_nocow_ctx *nocow_ctx =
4385                 container_of(work, struct scrub_copy_nocow_ctx, work);
4386         struct scrub_ctx *sctx = nocow_ctx->sctx;
4387         struct btrfs_fs_info *fs_info = sctx->fs_info;
4388         struct btrfs_root *root = fs_info->extent_root;
4389         u64 logical = nocow_ctx->logical;
4390         u64 len = nocow_ctx->len;
4391         int mirror_num = nocow_ctx->mirror_num;
4392         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4393         int ret;
4394         struct btrfs_trans_handle *trans = NULL;
4395         struct btrfs_path *path;
4396         int not_written = 0;
4397
4398         path = btrfs_alloc_path();
4399         if (!path) {
4400                 spin_lock(&sctx->stat_lock);
4401                 sctx->stat.malloc_errors++;
4402                 spin_unlock(&sctx->stat_lock);
4403                 not_written = 1;
4404                 goto out;
4405         }
4406
4407         trans = btrfs_join_transaction(root);
4408         if (IS_ERR(trans)) {
4409                 not_written = 1;
4410                 goto out;
4411         }
4412
4413         ret = iterate_inodes_from_logical(logical, fs_info, path,
4414                                           record_inode_for_nocow, nocow_ctx);
4415         if (ret != 0 && ret != -ENOENT) {
4416                 btrfs_warn(fs_info,
4417                            "iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d",
4418                            logical, physical_for_dev_replace, len, mirror_num,
4419                            ret);
4420                 not_written = 1;
4421                 goto out;
4422         }
4423
4424         btrfs_end_transaction(trans);
4425         trans = NULL;
4426         while (!list_empty(&nocow_ctx->inodes)) {
4427                 struct scrub_nocow_inode *entry;
4428                 entry = list_first_entry(&nocow_ctx->inodes,
4429                                          struct scrub_nocow_inode,
4430                                          list);
4431                 list_del_init(&entry->list);
4432                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4433                                                  entry->root, nocow_ctx);
4434                 kfree(entry);
4435                 if (ret == COPY_COMPLETE) {
4436                         ret = 0;
4437                         break;
4438                 } else if (ret) {
4439                         break;
4440                 }
4441         }
4442 out:
4443         while (!list_empty(&nocow_ctx->inodes)) {
4444                 struct scrub_nocow_inode *entry;
4445                 entry = list_first_entry(&nocow_ctx->inodes,
4446                                          struct scrub_nocow_inode,
4447                                          list);
4448                 list_del_init(&entry->list);
4449                 kfree(entry);
4450         }
4451         if (trans && !IS_ERR(trans))
4452                 btrfs_end_transaction(trans);
4453         if (not_written)
4454                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4455                                             num_uncorrectable_read_errors);
4456
4457         btrfs_free_path(path);
4458         kfree(nocow_ctx);
4459
4460         scrub_pending_trans_workers_dec(sctx);
4461 }
4462
4463 static int check_extent_to_block(struct btrfs_inode *inode, u64 start, u64 len,
4464                                  u64 logical)
4465 {
4466         struct extent_state *cached_state = NULL;
4467         struct btrfs_ordered_extent *ordered;
4468         struct extent_io_tree *io_tree;
4469         struct extent_map *em;
4470         u64 lockstart = start, lockend = start + len - 1;
4471         int ret = 0;
4472
4473         io_tree = &inode->io_tree;
4474
4475         lock_extent_bits(io_tree, lockstart, lockend, &cached_state);
4476         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4477         if (ordered) {
4478                 btrfs_put_ordered_extent(ordered);
4479                 ret = 1;
4480                 goto out_unlock;
4481         }
4482
4483         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4484         if (IS_ERR(em)) {
4485                 ret = PTR_ERR(em);
4486                 goto out_unlock;
4487         }
4488
4489         /*
4490          * This extent does not actually cover the logical extent anymore,
4491          * move on to the next inode.
4492          */
4493         if (em->block_start > logical ||
4494             em->block_start + em->block_len < logical + len) {
4495                 free_extent_map(em);
4496                 ret = 1;
4497                 goto out_unlock;
4498         }
4499         free_extent_map(em);
4500
4501 out_unlock:
4502         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4503                              GFP_NOFS);
4504         return ret;
4505 }
4506
4507 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4508                                       struct scrub_copy_nocow_ctx *nocow_ctx)
4509 {
4510         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->fs_info;
4511         struct btrfs_key key;
4512         struct inode *inode;
4513         struct page *page;
4514         struct btrfs_root *local_root;
4515         struct extent_io_tree *io_tree;
4516         u64 physical_for_dev_replace;
4517         u64 nocow_ctx_logical;
4518         u64 len = nocow_ctx->len;
4519         unsigned long index;
4520         int srcu_index;
4521         int ret = 0;
4522         int err = 0;
4523
4524         key.objectid = root;
4525         key.type = BTRFS_ROOT_ITEM_KEY;
4526         key.offset = (u64)-1;
4527
4528         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4529
4530         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4531         if (IS_ERR(local_root)) {
4532                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4533                 return PTR_ERR(local_root);
4534         }
4535
4536         key.type = BTRFS_INODE_ITEM_KEY;
4537         key.objectid = inum;
4538         key.offset = 0;
4539         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4540         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4541         if (IS_ERR(inode))
4542                 return PTR_ERR(inode);
4543
4544         /* Avoid truncate/dio/punch hole.. */
4545         inode_lock(inode);
4546         inode_dio_wait(inode);
4547
4548         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4549         io_tree = &BTRFS_I(inode)->io_tree;
4550         nocow_ctx_logical = nocow_ctx->logical;
4551
4552         ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4553                         nocow_ctx_logical);
4554         if (ret) {
4555                 ret = ret > 0 ? 0 : ret;
4556                 goto out;
4557         }
4558
4559         while (len >= PAGE_SIZE) {
4560                 index = offset >> PAGE_SHIFT;
4561 again:
4562                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4563                 if (!page) {
4564                         btrfs_err(fs_info, "find_or_create_page() failed");
4565                         ret = -ENOMEM;
4566                         goto out;
4567                 }
4568
4569                 if (PageUptodate(page)) {
4570                         if (PageDirty(page))
4571                                 goto next_page;
4572                 } else {
4573                         ClearPageError(page);
4574                         err = extent_read_full_page(io_tree, page,
4575                                                            btrfs_get_extent,
4576                                                            nocow_ctx->mirror_num);
4577                         if (err) {
4578                                 ret = err;
4579                                 goto next_page;
4580                         }
4581
4582                         lock_page(page);
4583                         /*
4584                          * If the page has been remove from the page cache,
4585                          * the data on it is meaningless, because it may be
4586                          * old one, the new data may be written into the new
4587                          * page in the page cache.
4588                          */
4589                         if (page->mapping != inode->i_mapping) {
4590                                 unlock_page(page);
4591                                 put_page(page);
4592                                 goto again;
4593                         }
4594                         if (!PageUptodate(page)) {
4595                                 ret = -EIO;
4596                                 goto next_page;
4597                         }
4598                 }
4599
4600                 ret = check_extent_to_block(BTRFS_I(inode), offset, len,
4601                                             nocow_ctx_logical);
4602                 if (ret) {
4603                         ret = ret > 0 ? 0 : ret;
4604                         goto next_page;
4605                 }
4606
4607                 err = write_page_nocow(nocow_ctx->sctx,
4608                                        physical_for_dev_replace, page);
4609                 if (err)
4610                         ret = err;
4611 next_page:
4612                 unlock_page(page);
4613                 put_page(page);
4614
4615                 if (ret)
4616                         break;
4617
4618                 offset += PAGE_SIZE;
4619                 physical_for_dev_replace += PAGE_SIZE;
4620                 nocow_ctx_logical += PAGE_SIZE;
4621                 len -= PAGE_SIZE;
4622         }
4623         ret = COPY_COMPLETE;
4624 out:
4625         inode_unlock(inode);
4626         iput(inode);
4627         return ret;
4628 }
4629
4630 static int write_page_nocow(struct scrub_ctx *sctx,
4631                             u64 physical_for_dev_replace, struct page *page)
4632 {
4633         struct bio *bio;
4634         struct btrfs_device *dev;
4635         int ret;
4636
4637         dev = sctx->wr_tgtdev;
4638         if (!dev)
4639                 return -EIO;
4640         if (!dev->bdev) {
4641                 btrfs_warn_rl(dev->fs_info,
4642                         "scrub write_page_nocow(bdev == NULL) is unexpected");
4643                 return -EIO;
4644         }
4645         bio = btrfs_io_bio_alloc(1);
4646         bio->bi_iter.bi_size = 0;
4647         bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4648         bio_set_dev(bio, dev->bdev);
4649         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC;
4650         ret = bio_add_page(bio, page, PAGE_SIZE, 0);
4651         if (ret != PAGE_SIZE) {
4652 leave_with_eio:
4653                 bio_put(bio);
4654                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4655                 return -EIO;
4656         }
4657
4658         if (btrfsic_submit_bio_wait(bio))
4659                 goto leave_with_eio;
4660
4661         bio_put(bio);
4662         return 0;
4663 }