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