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