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         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 922                 sblocks_for_recheck = NULL;
 923                 goto nodatasum_case;
 924         }
 925
 926         /*
 927          * read all mirrors one after the other. This includes to
 928          * re-read the extent or metadata block that failed (that was
 929          * the cause that this fixup code is called) another time,
 930          * page by page this time in order to know which pages
 931          * caused I/O errors and which ones are good (for all mirrors).
 932          * It is the goal to handle the situation when more than one
 933          * mirror contains I/O errors, but the errors do not
 934          * overlap, i.e. the data can be repaired by selecting the
 935          * pages from those mirrors without I/O error on the
 936          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 937          * would be that mirror #1 has an I/O error on the first page,
 938          * the second page is good, and mirror #2 has an I/O error on
 939          * the second page, but the first page is good.
 940          * Then the first page of the first mirror can be repaired by
 941          * taking the first page of the second mirror, and the
 942          * second page of the second mirror can be repaired by
 943          * copying the contents of the 2nd page of the 1st mirror.
 944          * One more note: if the pages of one mirror contain I/O
 945          * errors, the checksum cannot be verified. In order to get
 946          * the best data for repairing, the first attempt is to find
 947          * a mirror without I/O errors and with a validated checksum.
 948          * Only if this is not possible, the pages are picked from
 949          * mirrors with I/O errors without considering the checksum.
 950          * If the latter is the case, at the end, the checksum of the
 951          * repaired area is verified in order to correctly maintain
 952          * the statistics.
 953          */
 954
 955         sblocks_for_recheck = kcalloc(BTRFS_MAX_MIRRORS,
 956                                       sizeof(*sblocks_for_recheck), GFP_NOFS);
 957         if (!sblocks_for_recheck) {
 958                 spin_lock(&sctx->stat_lock);
 959                 sctx->stat.malloc_errors++;
 960                 sctx->stat.read_errors++;
 961                 sctx->stat.uncorrectable_errors++;
 962                 spin_unlock(&sctx->stat_lock);
 963                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 964                 goto out;
 965         }
 966
 967         /* setup the context, map the logical blocks and alloc the pages */
 968         ret = scrub_setup_recheck_block(sblock_to_check, sblocks_for_recheck);
 969         if (ret) {
 970                 spin_lock(&sctx->stat_lock);
 971                 sctx->stat.read_errors++;
 972                 sctx->stat.uncorrectable_errors++;
 973                 spin_unlock(&sctx->stat_lock);
 974                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 975                 goto out;
 976         }
 977         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 978         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 979
 980         /* build and submit the bios for the failed mirror, check checksums */
 981         scrub_recheck_block(fs_info, sblock_bad, 1);
 982
 983         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 984             sblock_bad->no_io_error_seen) {
 985                 /*
 986                  * the error disappeared after reading page by page, or
 987                  * the area was part of a huge bio and other parts of the
 988                  * bio caused I/O errors, or the block layer merged several
 989                  * read requests into one and the error is caused by a
 990                  * different bio (usually one of the two latter cases is
 991                  * the cause)
 992                  */
 993                 spin_lock(&sctx->stat_lock);
 994                 sctx->stat.unverified_errors++;
 995                 sblock_to_check->data_corrected = 1;
 996                 spin_unlock(&sctx->stat_lock);
 997
 998                 if (sctx->is_dev_replace)
 999                         scrub_write_block_to_dev_replace(sblock_bad);
1000                 goto out;
1001         }
1002
1003         if (!sblock_bad->no_io_error_seen) {
1004                 spin_lock(&sctx->stat_lock);
1005                 sctx->stat.read_errors++;
1006                 spin_unlock(&sctx->stat_lock);
1007                 if (__ratelimit(&_rs))
1008                         scrub_print_warning("i/o error", sblock_to_check);
1009                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
1010         } else if (sblock_bad->checksum_error) {
1011                 spin_lock(&sctx->stat_lock);
1012                 sctx->stat.csum_errors++;
1013                 spin_unlock(&sctx->stat_lock);
1014                 if (__ratelimit(&_rs))
1015                         scrub_print_warning("checksum error", sblock_to_check);
1016                 btrfs_dev_stat_inc_and_print(dev,
1017                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
1018         } else if (sblock_bad->header_error) {
1019                 spin_lock(&sctx->stat_lock);
1020                 sctx->stat.verify_errors++;
1021                 spin_unlock(&sctx->stat_lock);
1022                 if (__ratelimit(&_rs))
1023                         scrub_print_warning("checksum/header error",
1024                                             sblock_to_check);
1025                 if (sblock_bad->generation_error)
1026                         btrfs_dev_stat_inc_and_print(dev,
1027                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1028                 else
1029                         btrfs_dev_stat_inc_and_print(dev,
1030                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1031         }
1032
1033         if (sctx->readonly) {
1034                 ASSERT(!sctx->is_dev_replace);
1035                 goto out;
1036         }
1037
1038         if (!is_metadata && !have_csum) {
1039                 struct scrub_fixup_nodatasum *fixup_nodatasum;
1040
1041                 WARN_ON(sctx->is_dev_replace);
1042
1043 nodatasum_case:
1044
1045                 /*
1046                  * !is_metadata and !have_csum, this means that the data
1047                  * might not be COW'ed, that it might be modified
1048                  * concurrently. The general strategy to work on the
1049                  * commit root does not help in the case when COW is not
1050                  * used.
1051                  */
1052                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
1053                 if (!fixup_nodatasum)
1054                         goto did_not_correct_error;
1055                 fixup_nodatasum->sctx = sctx;
1056                 fixup_nodatasum->dev = dev;
1057                 fixup_nodatasum->logical = logical;
1058                 fixup_nodatasum->root = fs_info->extent_root;
1059                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1060                 scrub_pending_trans_workers_inc(sctx);
1061                 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1062                                 scrub_fixup_nodatasum, NULL, NULL);
1063                 btrfs_queue_work(fs_info->scrub_workers,
1064                                  &fixup_nodatasum->work);
1065                 goto out;
1066         }
1067
1068         /*
1069          * now build and submit the bios for the other mirrors, check
1070          * checksums.
1071          * First try to pick the mirror which is completely without I/O
1072          * errors and also does not have a checksum error.
1073          * If one is found, and if a checksum is present, the full block
1074          * that is known to contain an error is rewritten. Afterwards
1075          * the block is known to be corrected.
1076          * If a mirror is found which is completely correct, and no
1077          * checksum is present, only those pages are rewritten that had
1078          * an I/O error in the block to be repaired, since it cannot be
1079          * determined, which copy of the other pages is better (and it
1080          * could happen otherwise that a correct page would be
1081          * overwritten by a bad one).
1082          */
1083         for (mirror_index = 0;
1084              mirror_index < BTRFS_MAX_MIRRORS &&
1085              sblocks_for_recheck[mirror_index].page_count > 0;
1086              mirror_index++) {
1087                 struct scrub_block *sblock_other;
1088
1089                 if (mirror_index == failed_mirror_index)
1090                         continue;
1091                 sblock_other = sblocks_for_recheck + mirror_index;
1092
1093                 /* build and submit the bios, check checksums */
1094                 scrub_recheck_block(fs_info, sblock_other, 0);
1095
1096                 if (!sblock_other->header_error &&
1097                     !sblock_other->checksum_error &&
1098                     sblock_other->no_io_error_seen) {
1099                         if (sctx->is_dev_replace) {
1100                                 scrub_write_block_to_dev_replace(sblock_other);
1101                                 goto corrected_error;
1102                         } else {
1103                                 ret = scrub_repair_block_from_good_copy(
1104                                                 sblock_bad, sblock_other);
1105                                 if (!ret)
1106                                         goto corrected_error;
1107                         }
1108                 }
1109         }
1110
1111         if (sblock_bad->no_io_error_seen && !sctx->is_dev_replace)
1112                 goto did_not_correct_error;
1113
1114         /*
1115          * In case of I/O errors in the area that is supposed to be
1116          * repaired, continue by picking good copies of those pages.
1117          * Select the good pages from mirrors to rewrite bad pages from
1118          * the area to fix. Afterwards verify the checksum of the block
1119          * that is supposed to be repaired. This verification step is
1120          * only done for the purpose of statistic counting and for the
1121          * final scrub report, whether errors remain.
1122          * A perfect algorithm could make use of the checksum and try
1123          * all possible combinations of pages from the different mirrors
1124          * until the checksum verification succeeds. For example, when
1125          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1126          * of mirror #2 is readable but the final checksum test fails,
1127          * then the 2nd page of mirror #3 could be tried, whether now
1128          * the final checksum succeedes. But this would be a rare
1129          * exception and is therefore not implemented. At least it is
1130          * avoided that the good copy is overwritten.
1131          * A more useful improvement would be to pick the sectors
1132          * without I/O error based on sector sizes (512 bytes on legacy
1133          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1134          * mirror could be repaired by taking 512 byte of a different
1135          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1136          * area are unreadable.
1137          */
1138         success = 1;
1139         for (page_num = 0; page_num < sblock_bad->page_count;
1140              page_num++) {
1141                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1142                 struct scrub_block *sblock_other = NULL;
1143
1144                 /* skip no-io-error page in scrub */
1145                 if (!page_bad->io_error && !sctx->is_dev_replace)
1146                         continue;
1147
1148                 /* try to find no-io-error page in mirrors */
1149                 if (page_bad->io_error) {
1150                         for (mirror_index = 0;
1151                              mirror_index < BTRFS_MAX_MIRRORS &&
1152                              sblocks_for_recheck[mirror_index].page_count > 0;
1153                              mirror_index++) {
1154                                 if (!sblocks_for_recheck[mirror_index].
1155                                     pagev[page_num]->io_error) {
1156                                         sblock_other = sblocks_for_recheck +
1157                                                        mirror_index;
1158                                         break;
1159                                 }
1160                         }
1161                         if (!sblock_other)
1162                                 success = 0;
1163                 }
1164
1165                 if (sctx->is_dev_replace) {
1166                         /*
1167                          * did not find a mirror to fetch the page
1168                          * from. scrub_write_page_to_dev_replace()
1169                          * handles this case (page->io_error), by
1170                          * filling the block with zeros before
1171                          * submitting the write request
1172                          */
1173                         if (!sblock_other)
1174                                 sblock_other = sblock_bad;
1175
1176                         if (scrub_write_page_to_dev_replace(sblock_other,
1177                                                             page_num) != 0) {
1178                                 btrfs_dev_replace_stats_inc(
1179                                         &sctx->dev_root->
1180                                         fs_info->dev_replace.
1181                                         num_write_errors);
1182                                 success = 0;
1183                         }
1184                 } else if (sblock_other) {
1185                         ret = scrub_repair_page_from_good_copy(sblock_bad,
1186                                                                sblock_other,
1187                                                                page_num, 0);
1188                         if (0 == ret)
1189                                 page_bad->io_error = 0;
1190                         else
1191                                 success = 0;
1192                 }
1193         }
1194
1195         if (success && !sctx->is_dev_replace) {
1196                 if (is_metadata || have_csum) {
1197                         /*
1198                          * need to verify the checksum now that all
1199                          * sectors on disk are repaired (the write
1200                          * request for data to be repaired is on its way).
1201                          * Just be lazy and use scrub_recheck_block()
1202                          * which re-reads the data before the checksum
1203                          * is verified, but most likely the data comes out
1204                          * of the page cache.
1205                          */
1206                         scrub_recheck_block(fs_info, sblock_bad, 1);
1207                         if (!sblock_bad->header_error &&
1208                             !sblock_bad->checksum_error &&
1209                             sblock_bad->no_io_error_seen)
1210                                 goto corrected_error;
1211                         else
1212                                 goto did_not_correct_error;
1213                 } else {
1214 corrected_error:
1215                         spin_lock(&sctx->stat_lock);
1216                         sctx->stat.corrected_errors++;
1217                         sblock_to_check->data_corrected = 1;
1218                         spin_unlock(&sctx->stat_lock);
1219                         btrfs_err_rl_in_rcu(fs_info,
1220                                 "fixed up error at logical %llu on dev %s",
1221                                 logical, rcu_str_deref(dev->name));
1222                 }
1223         } else {
1224 did_not_correct_error:
1225                 spin_lock(&sctx->stat_lock);
1226                 sctx->stat.uncorrectable_errors++;
1227                 spin_unlock(&sctx->stat_lock);
1228                 btrfs_err_rl_in_rcu(fs_info,
1229                         "unable to fixup (regular) error at logical %llu on dev %s",
1230                         logical, rcu_str_deref(dev->name));
1231         }
1232
1233 out:
1234         if (sblocks_for_recheck) {
1235                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1236                      mirror_index++) {
1237                         struct scrub_block *sblock = sblocks_for_recheck +
1238                                                      mirror_index;
1239                         struct scrub_recover *recover;
1240                         int page_index;
1241
1242                         for (page_index = 0; page_index < sblock->page_count;
1243                              page_index++) {
1244                                 sblock->pagev[page_index]->sblock = NULL;
1245                                 recover = sblock->pagev[page_index]->recover;
1246                                 if (recover) {
1247                                         scrub_put_recover(recover);
1248                                         sblock->pagev[page_index]->recover =
1249                                                                         NULL;
1250                                 }
1251                                 scrub_page_put(sblock->pagev[page_index]);
1252                         }
1253                 }
1254                 kfree(sblocks_for_recheck);
1255         }
1256
1257         return 0;
1258 }
1259
1260 static inline int scrub_nr_raid_mirrors(struct btrfs_bio *bbio)
1261 {
1262         if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID5)
1263                 return 2;
1264         else if (bbio->map_type & BTRFS_BLOCK_GROUP_RAID6)
1265                 return 3;
1266         else
1267                 return (int)bbio->num_stripes;
1268 }
1269
1270 static inline void scrub_stripe_index_and_offset(u64 logical, u64 map_type,
1271                                                  u64 *raid_map,
1272                                                  u64 mapped_length,
1273                                                  int nstripes, int mirror,
1274                                                  int *stripe_index,
1275                                                  u64 *stripe_offset)
1276 {
1277         int i;
1278
1279         if (map_type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
1280                 /* RAID5/6 */
1281                 for (i = 0; i < nstripes; i++) {
1282                         if (raid_map[i] == RAID6_Q_STRIPE ||
1283                             raid_map[i] == RAID5_P_STRIPE)
1284                                 continue;
1285
1286                         if (logical >= raid_map[i] &&
1287                             logical < raid_map[i] + mapped_length)
1288                                 break;
1289                 }
1290
1291                 *stripe_index = i;
1292                 *stripe_offset = logical - raid_map[i];
1293         } else {
1294                 /* The other RAID type */
1295                 *stripe_index = mirror;
1296                 *stripe_offset = 0;
1297         }
1298 }
1299
1300 static int scrub_setup_recheck_block(struct scrub_block *original_sblock,
1301                                      struct scrub_block *sblocks_for_recheck)
1302 {
1303         struct scrub_ctx *sctx = original_sblock->sctx;
1304         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
1305         u64 length = original_sblock->page_count * PAGE_SIZE;
1306         u64 logical = original_sblock->pagev[0]->logical;
1307         u64 generation = original_sblock->pagev[0]->generation;
1308         u64 flags = original_sblock->pagev[0]->flags;
1309         u64 have_csum = original_sblock->pagev[0]->have_csum;
1310         struct scrub_recover *recover;
1311         struct btrfs_bio *bbio;
1312         u64 sublen;
1313         u64 mapped_length;
1314         u64 stripe_offset;
1315         int stripe_index;
1316         int page_index = 0;
1317         int mirror_index;
1318         int nmirrors;
1319         int ret;
1320
1321         /*
1322          * note: the two members refs and outstanding_pages
1323          * are not used (and not set) in the blocks that are used for
1324          * the recheck procedure
1325          */
1326
1327         while (length > 0) {
1328                 sublen = min_t(u64, length, PAGE_SIZE);
1329                 mapped_length = sublen;
1330                 bbio = NULL;
1331
1332                 /*
1333                  * with a length of PAGE_SIZE, each returned stripe
1334                  * represents one mirror
1335                  */
1336                 ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical,
1337                                        &mapped_length, &bbio, 0, 1);
1338                 if (ret || !bbio || mapped_length < sublen) {
1339                         btrfs_put_bbio(bbio);
1340                         return -EIO;
1341                 }
1342
1343                 recover = kzalloc(sizeof(struct scrub_recover), GFP_NOFS);
1344                 if (!recover) {
1345                         btrfs_put_bbio(bbio);
1346                         return -ENOMEM;
1347                 }
1348
1349                 atomic_set(&recover->refs, 1);
1350                 recover->bbio = bbio;
1351                 recover->map_length = mapped_length;
1352
1353                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1354
1355                 nmirrors = min(scrub_nr_raid_mirrors(bbio), BTRFS_MAX_MIRRORS);
1356
1357                 for (mirror_index = 0; mirror_index < nmirrors;
1358                      mirror_index++) {
1359                         struct scrub_block *sblock;
1360                         struct scrub_page *page;
1361
1362                         sblock = sblocks_for_recheck + mirror_index;
1363                         sblock->sctx = sctx;
1364
1365                         page = kzalloc(sizeof(*page), GFP_NOFS);
1366                         if (!page) {
1367 leave_nomem:
1368                                 spin_lock(&sctx->stat_lock);
1369                                 sctx->stat.malloc_errors++;
1370                                 spin_unlock(&sctx->stat_lock);
1371                                 scrub_put_recover(recover);
1372                                 return -ENOMEM;
1373                         }
1374                         scrub_page_get(page);
1375                         sblock->pagev[page_index] = page;
1376                         page->sblock = sblock;
1377                         page->flags = flags;
1378                         page->generation = generation;
1379                         page->logical = logical;
1380                         page->have_csum = have_csum;
1381                         if (have_csum)
1382                                 memcpy(page->csum,
1383                                        original_sblock->pagev[0]->csum,
1384                                        sctx->csum_size);
1385
1386                         scrub_stripe_index_and_offset(logical,
1387                                                       bbio->map_type,
1388                                                       bbio->raid_map,
1389                                                       mapped_length,
1390                                                       bbio->num_stripes -
1391                                                       bbio->num_tgtdevs,
1392                                                       mirror_index,
1393                                                       &stripe_index,
1394                                                       &stripe_offset);
1395                         page->physical = bbio->stripes[stripe_index].physical +
1396                                          stripe_offset;
1397                         page->dev = bbio->stripes[stripe_index].dev;
1398
1399                         BUG_ON(page_index >= original_sblock->page_count);
1400                         page->physical_for_dev_replace =
1401                                 original_sblock->pagev[page_index]->
1402                                 physical_for_dev_replace;
1403                         /* for missing devices, dev->bdev is NULL */
1404                         page->mirror_num = mirror_index + 1;
1405                         sblock->page_count++;
1406                         page->page = alloc_page(GFP_NOFS);
1407                         if (!page->page)
1408                                 goto leave_nomem;
1409
1410                         scrub_get_recover(recover);
1411                         page->recover = recover;
1412                 }
1413                 scrub_put_recover(recover);
1414                 length -= sublen;
1415                 logical += sublen;
1416                 page_index++;
1417         }
1418
1419         return 0;
1420 }
1421
1422 struct scrub_bio_ret {
1423         struct completion event;
1424         int error;
1425 };
1426
1427 static void scrub_bio_wait_endio(struct bio *bio)
1428 {
1429         struct scrub_bio_ret *ret = bio->bi_private;
1430
1431         ret->error = bio->bi_error;
1432         complete(&ret->event);
1433 }
1434
1435 static inline int scrub_is_page_on_raid56(struct scrub_page *page)
1436 {
1437         return page->recover &&
1438                (page->recover->bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK);
1439 }
1440
1441 static int scrub_submit_raid56_bio_wait(struct btrfs_fs_info *fs_info,
1442                                         struct bio *bio,
1443                                         struct scrub_page *page)
1444 {
1445         struct scrub_bio_ret done;
1446         int ret;
1447
1448         init_completion(&done.event);
1449         done.error = 0;
1450         bio->bi_iter.bi_sector = page->logical >> 9;
1451         bio->bi_private = &done;
1452         bio->bi_end_io = scrub_bio_wait_endio;
1453
1454         ret = raid56_parity_recover(fs_info->fs_root, bio, page->recover->bbio,
1455                                     page->recover->map_length,
1456                                     page->mirror_num, 0);
1457         if (ret)
1458                 return ret;
1459
1460         wait_for_completion(&done.event);
1461         if (done.error)
1462                 return -EIO;
1463
1464         return 0;
1465 }
1466
1467 /*
1468  * this function will check the on disk data for checksum errors, header
1469  * errors and read I/O errors. If any I/O errors happen, the exact pages
1470  * which are errored are marked as being bad. The goal is to enable scrub
1471  * to take those pages that are not errored from all the mirrors so that
1472  * the pages that are errored in the just handled mirror can be repaired.
1473  */
1474 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1475                                 struct scrub_block *sblock,
1476                                 int retry_failed_mirror)
1477 {
1478         int page_num;
1479
1480         sblock->no_io_error_seen = 1;
1481
1482         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1483                 struct bio *bio;
1484                 struct scrub_page *page = sblock->pagev[page_num];
1485
1486                 if (page->dev->bdev == NULL) {
1487                         page->io_error = 1;
1488                         sblock->no_io_error_seen = 0;
1489                         continue;
1490                 }
1491
1492                 WARN_ON(!page->page);
1493                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1494                 if (!bio) {
1495                         page->io_error = 1;
1496                         sblock->no_io_error_seen = 0;
1497                         continue;
1498                 }
1499                 bio->bi_bdev = page->dev->bdev;
1500
1501                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1502                 if (!retry_failed_mirror && scrub_is_page_on_raid56(page)) {
1503                         if (scrub_submit_raid56_bio_wait(fs_info, bio, page))
1504                                 sblock->no_io_error_seen = 0;
1505                 } else {
1506                         bio->bi_iter.bi_sector = page->physical >> 9;
1507
1508                         if (btrfsic_submit_bio_wait(READ, bio))
1509                                 sblock->no_io_error_seen = 0;
1510                 }
1511
1512                 bio_put(bio);
1513         }
1514
1515         if (sblock->no_io_error_seen)
1516                 scrub_recheck_block_checksum(sblock);
1517
1518         return;
1519 }
1520
1521 static inline int scrub_check_fsid(u8 fsid[],
1522                                    struct scrub_page *spage)
1523 {
1524         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1525         int ret;
1526
1527         ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1528         return !ret;
1529 }
1530
1531 static void scrub_recheck_block_checksum(struct scrub_block *sblock)
1532 {
1533         sblock->header_error = 0;
1534         sblock->checksum_error = 0;
1535         sblock->generation_error = 0;
1536
1537         if (sblock->pagev[0]->flags & BTRFS_EXTENT_FLAG_DATA)
1538                 scrub_checksum_data(sblock);
1539         else
1540                 scrub_checksum_tree_block(sblock);
1541 }
1542
1543 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1544                                              struct scrub_block *sblock_good)
1545 {
1546         int page_num;
1547         int ret = 0;
1548
1549         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1550                 int ret_sub;
1551
1552                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1553                                                            sblock_good,
1554                                                            page_num, 1);
1555                 if (ret_sub)
1556                         ret = ret_sub;
1557         }
1558
1559         return ret;
1560 }
1561
1562 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1563                                             struct scrub_block *sblock_good,
1564                                             int page_num, int force_write)
1565 {
1566         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1567         struct scrub_page *page_good = sblock_good->pagev[page_num];
1568
1569         BUG_ON(page_bad->page == NULL);
1570         BUG_ON(page_good->page == NULL);
1571         if (force_write || sblock_bad->header_error ||
1572             sblock_bad->checksum_error || page_bad->io_error) {
1573                 struct bio *bio;
1574                 int ret;
1575
1576                 if (!page_bad->dev->bdev) {
1577                         btrfs_warn_rl(sblock_bad->sctx->dev_root->fs_info,
1578                                 "scrub_repair_page_from_good_copy(bdev == NULL) "
1579                                 "is unexpected");
1580                         return -EIO;
1581                 }
1582
1583                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1584                 if (!bio)
1585                         return -EIO;
1586                 bio->bi_bdev = page_bad->dev->bdev;
1587                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1588
1589                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1590                 if (PAGE_SIZE != ret) {
1591                         bio_put(bio);
1592                         return -EIO;
1593                 }
1594
1595                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1596                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1597                                 BTRFS_DEV_STAT_WRITE_ERRS);
1598                         btrfs_dev_replace_stats_inc(
1599                                 &sblock_bad->sctx->dev_root->fs_info->
1600                                 dev_replace.num_write_errors);
1601                         bio_put(bio);
1602                         return -EIO;
1603                 }
1604                 bio_put(bio);
1605         }
1606
1607         return 0;
1608 }
1609
1610 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1611 {
1612         int page_num;
1613
1614         /*
1615          * This block is used for the check of the parity on the source device,
1616          * so the data needn't be written into the destination device.
1617          */
1618         if (sblock->sparity)
1619                 return;
1620
1621         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1622                 int ret;
1623
1624                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1625                 if (ret)
1626                         btrfs_dev_replace_stats_inc(
1627                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1628                                 num_write_errors);
1629         }
1630 }
1631
1632 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1633                                            int page_num)
1634 {
1635         struct scrub_page *spage = sblock->pagev[page_num];
1636
1637         BUG_ON(spage->page == NULL);
1638         if (spage->io_error) {
1639                 void *mapped_buffer = kmap_atomic(spage->page);
1640
1641                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1642                 flush_dcache_page(spage->page);
1643                 kunmap_atomic(mapped_buffer);
1644         }
1645         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1646 }
1647
1648 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1649                                     struct scrub_page *spage)
1650 {
1651         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1652         struct scrub_bio *sbio;
1653         int ret;
1654
1655         mutex_lock(&wr_ctx->wr_lock);
1656 again:
1657         if (!wr_ctx->wr_curr_bio) {
1658                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1659                                               GFP_NOFS);
1660                 if (!wr_ctx->wr_curr_bio) {
1661                         mutex_unlock(&wr_ctx->wr_lock);
1662                         return -ENOMEM;
1663                 }
1664                 wr_ctx->wr_curr_bio->sctx = sctx;
1665                 wr_ctx->wr_curr_bio->page_count = 0;
1666         }
1667         sbio = wr_ctx->wr_curr_bio;
1668         if (sbio->page_count == 0) {
1669                 struct bio *bio;
1670
1671                 sbio->physical = spage->physical_for_dev_replace;
1672                 sbio->logical = spage->logical;
1673                 sbio->dev = wr_ctx->tgtdev;
1674                 bio = sbio->bio;
1675                 if (!bio) {
1676                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1677                         if (!bio) {
1678                                 mutex_unlock(&wr_ctx->wr_lock);
1679                                 return -ENOMEM;
1680                         }
1681                         sbio->bio = bio;
1682                 }
1683
1684                 bio->bi_private = sbio;
1685                 bio->bi_end_io = scrub_wr_bio_end_io;
1686                 bio->bi_bdev = sbio->dev->bdev;
1687                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1688                 sbio->err = 0;
1689         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1690                    spage->physical_for_dev_replace ||
1691                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1692                    spage->logical) {
1693                 scrub_wr_submit(sctx);
1694                 goto again;
1695         }
1696
1697         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1698         if (ret != PAGE_SIZE) {
1699                 if (sbio->page_count < 1) {
1700                         bio_put(sbio->bio);
1701                         sbio->bio = NULL;
1702                         mutex_unlock(&wr_ctx->wr_lock);
1703                         return -EIO;
1704                 }
1705                 scrub_wr_submit(sctx);
1706                 goto again;
1707         }
1708
1709         sbio->pagev[sbio->page_count] = spage;
1710         scrub_page_get(spage);
1711         sbio->page_count++;
1712         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1713                 scrub_wr_submit(sctx);
1714         mutex_unlock(&wr_ctx->wr_lock);
1715
1716         return 0;
1717 }
1718
1719 static void scrub_wr_submit(struct scrub_ctx *sctx)
1720 {
1721         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1722         struct scrub_bio *sbio;
1723
1724         if (!wr_ctx->wr_curr_bio)
1725                 return;
1726
1727         sbio = wr_ctx->wr_curr_bio;
1728         wr_ctx->wr_curr_bio = NULL;
1729         WARN_ON(!sbio->bio->bi_bdev);
1730         scrub_pending_bio_inc(sctx);
1731         /* process all writes in a single worker thread. Then the block layer
1732          * orders the requests before sending them to the driver which
1733          * doubled the write performance on spinning disks when measured
1734          * with Linux 3.5 */
1735         btrfsic_submit_bio(WRITE, sbio->bio);
1736 }
1737
1738 static void scrub_wr_bio_end_io(struct bio *bio)
1739 {
1740         struct scrub_bio *sbio = bio->bi_private;
1741         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1742
1743         sbio->err = bio->bi_error;
1744         sbio->bio = bio;
1745
1746         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1747                          scrub_wr_bio_end_io_worker, NULL, NULL);
1748         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1749 }
1750
1751 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1752 {
1753         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1754         struct scrub_ctx *sctx = sbio->sctx;
1755         int i;
1756
1757         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1758         if (sbio->err) {
1759                 struct btrfs_dev_replace *dev_replace =
1760                         &sbio->sctx->dev_root->fs_info->dev_replace;
1761
1762                 for (i = 0; i < sbio->page_count; i++) {
1763                         struct scrub_page *spage = sbio->pagev[i];
1764
1765                         spage->io_error = 1;
1766                         btrfs_dev_replace_stats_inc(&dev_replace->
1767                                                     num_write_errors);
1768                 }
1769         }
1770
1771         for (i = 0; i < sbio->page_count; i++)
1772                 scrub_page_put(sbio->pagev[i]);
1773
1774         bio_put(sbio->bio);
1775         kfree(sbio);
1776         scrub_pending_bio_dec(sctx);
1777 }
1778
1779 static int scrub_checksum(struct scrub_block *sblock)
1780 {
1781         u64 flags;
1782         int ret;
1783
1784         /*
1785          * No need to initialize these stats currently,
1786          * because this function only use return value
1787          * instead of these stats value.
1788          *
1789          * Todo:
1790          * always use stats
1791          */
1792         sblock->header_error = 0;
1793         sblock->generation_error = 0;
1794         sblock->checksum_error = 0;
1795
1796         WARN_ON(sblock->page_count < 1);
1797         flags = sblock->pagev[0]->flags;
1798         ret = 0;
1799         if (flags & BTRFS_EXTENT_FLAG_DATA)
1800                 ret = scrub_checksum_data(sblock);
1801         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1802                 ret = scrub_checksum_tree_block(sblock);
1803         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1804                 (void)scrub_checksum_super(sblock);
1805         else
1806                 WARN_ON(1);
1807         if (ret)
1808                 scrub_handle_errored_block(sblock);
1809
1810         return ret;
1811 }
1812
1813 static int scrub_checksum_data(struct scrub_block *sblock)
1814 {
1815         struct scrub_ctx *sctx = sblock->sctx;
1816         u8 csum[BTRFS_CSUM_SIZE];
1817         u8 *on_disk_csum;
1818         struct page *page;
1819         void *buffer;
1820         u32 crc = ~(u32)0;
1821         u64 len;
1822         int index;
1823
1824         BUG_ON(sblock->page_count < 1);
1825         if (!sblock->pagev[0]->have_csum)
1826                 return 0;
1827
1828         on_disk_csum = sblock->pagev[0]->csum;
1829         page = sblock->pagev[0]->page;
1830         buffer = kmap_atomic(page);
1831
1832         len = sctx->sectorsize;
1833         index = 0;
1834         for (;;) {
1835                 u64 l = min_t(u64, len, PAGE_SIZE);
1836
1837                 crc = btrfs_csum_data(buffer, crc, l);
1838                 kunmap_atomic(buffer);
1839                 len -= l;
1840                 if (len == 0)
1841                         break;
1842                 index++;
1843                 BUG_ON(index >= sblock->page_count);
1844                 BUG_ON(!sblock->pagev[index]->page);
1845                 page = sblock->pagev[index]->page;
1846                 buffer = kmap_atomic(page);
1847         }
1848
1849         btrfs_csum_final(crc, csum);
1850         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1851                 sblock->checksum_error = 1;
1852
1853         return sblock->checksum_error;
1854 }
1855
1856 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1857 {
1858         struct scrub_ctx *sctx = sblock->sctx;
1859         struct btrfs_header *h;
1860         struct btrfs_root *root = sctx->dev_root;
1861         struct btrfs_fs_info *fs_info = root->fs_info;
1862         u8 calculated_csum[BTRFS_CSUM_SIZE];
1863         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1864         struct page *page;
1865         void *mapped_buffer;
1866         u64 mapped_size;
1867         void *p;
1868         u32 crc = ~(u32)0;
1869         u64 len;
1870         int index;
1871
1872         BUG_ON(sblock->page_count < 1);
1873         page = sblock->pagev[0]->page;
1874         mapped_buffer = kmap_atomic(page);
1875         h = (struct btrfs_header *)mapped_buffer;
1876         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1877
1878         /*
1879          * we don't use the getter functions here, as we
1880          * a) don't have an extent buffer and
1881          * b) the page is already kmapped
1882          */
1883         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1884                 sblock->header_error = 1;
1885
1886         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h)) {
1887                 sblock->header_error = 1;
1888                 sblock->generation_error = 1;
1889         }
1890
1891         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1892                 sblock->header_error = 1;
1893
1894         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1895                    BTRFS_UUID_SIZE))
1896                 sblock->header_error = 1;
1897
1898         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1899         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1900         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1901         index = 0;
1902         for (;;) {
1903                 u64 l = min_t(u64, len, mapped_size);
1904
1905                 crc = btrfs_csum_data(p, crc, l);
1906                 kunmap_atomic(mapped_buffer);
1907                 len -= l;
1908                 if (len == 0)
1909                         break;
1910                 index++;
1911                 BUG_ON(index >= sblock->page_count);
1912                 BUG_ON(!sblock->pagev[index]->page);
1913                 page = sblock->pagev[index]->page;
1914                 mapped_buffer = kmap_atomic(page);
1915                 mapped_size = PAGE_SIZE;
1916                 p = mapped_buffer;
1917         }
1918
1919         btrfs_csum_final(crc, calculated_csum);
1920         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1921                 sblock->checksum_error = 1;
1922
1923         return sblock->header_error || sblock->checksum_error;
1924 }
1925
1926 static int scrub_checksum_super(struct scrub_block *sblock)
1927 {
1928         struct btrfs_super_block *s;
1929         struct scrub_ctx *sctx = sblock->sctx;
1930         u8 calculated_csum[BTRFS_CSUM_SIZE];
1931         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1932         struct page *page;
1933         void *mapped_buffer;
1934         u64 mapped_size;
1935         void *p;
1936         u32 crc = ~(u32)0;
1937         int fail_gen = 0;
1938         int fail_cor = 0;
1939         u64 len;
1940         int index;
1941
1942         BUG_ON(sblock->page_count < 1);
1943         page = sblock->pagev[0]->page;
1944         mapped_buffer = kmap_atomic(page);
1945         s = (struct btrfs_super_block *)mapped_buffer;
1946         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1947
1948         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1949                 ++fail_cor;
1950
1951         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1952                 ++fail_gen;
1953
1954         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1955                 ++fail_cor;
1956
1957         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1958         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1959         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1960         index = 0;
1961         for (;;) {
1962                 u64 l = min_t(u64, len, mapped_size);
1963
1964                 crc = btrfs_csum_data(p, crc, l);
1965                 kunmap_atomic(mapped_buffer);
1966                 len -= l;
1967                 if (len == 0)
1968                         break;
1969                 index++;
1970                 BUG_ON(index >= sblock->page_count);
1971                 BUG_ON(!sblock->pagev[index]->page);
1972                 page = sblock->pagev[index]->page;
1973                 mapped_buffer = kmap_atomic(page);
1974                 mapped_size = PAGE_SIZE;
1975                 p = mapped_buffer;
1976         }
1977
1978         btrfs_csum_final(crc, calculated_csum);
1979         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1980                 ++fail_cor;
1981
1982         if (fail_cor + fail_gen) {
1983                 /*
1984                  * if we find an error in a super block, we just report it.
1985                  * They will get written with the next transaction commit
1986                  * anyway
1987                  */
1988                 spin_lock(&sctx->stat_lock);
1989                 ++sctx->stat.super_errors;
1990                 spin_unlock(&sctx->stat_lock);
1991                 if (fail_cor)
1992                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1993                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1994                 else
1995                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1996                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1997         }
1998
1999         return fail_cor + fail_gen;
2000 }
2001
2002 static void scrub_block_get(struct scrub_block *sblock)
2003 {
2004         atomic_inc(&sblock->refs);
2005 }
2006
2007 static void scrub_block_put(struct scrub_block *sblock)
2008 {
2009         if (atomic_dec_and_test(&sblock->refs)) {
2010                 int i;
2011
2012                 if (sblock->sparity)
2013                         scrub_parity_put(sblock->sparity);
2014
2015                 for (i = 0; i < sblock->page_count; i++)
2016                         scrub_page_put(sblock->pagev[i]);
2017                 kfree(sblock);
2018         }
2019 }
2020
2021 static void scrub_page_get(struct scrub_page *spage)
2022 {
2023         atomic_inc(&spage->refs);
2024 }
2025
2026 static void scrub_page_put(struct scrub_page *spage)
2027 {
2028         if (atomic_dec_and_test(&spage->refs)) {
2029                 if (spage->page)
2030                         __free_page(spage->page);
2031                 kfree(spage);
2032         }
2033 }
2034
2035 static void scrub_submit(struct scrub_ctx *sctx)
2036 {
2037         struct scrub_bio *sbio;
2038
2039         if (sctx->curr == -1)
2040                 return;
2041
2042         sbio = sctx->bios[sctx->curr];
2043         sctx->curr = -1;
2044         scrub_pending_bio_inc(sctx);
2045         btrfsic_submit_bio(READ, sbio->bio);
2046 }
2047
2048 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
2049                                     struct scrub_page *spage)
2050 {
2051         struct scrub_block *sblock = spage->sblock;
2052         struct scrub_bio *sbio;
2053         int ret;
2054
2055 again:
2056         /*
2057          * grab a fresh bio or wait for one to become available
2058          */
2059         while (sctx->curr == -1) {
2060                 spin_lock(&sctx->list_lock);
2061                 sctx->curr = sctx->first_free;
2062                 if (sctx->curr != -1) {
2063                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
2064                         sctx->bios[sctx->curr]->next_free = -1;
2065                         sctx->bios[sctx->curr]->page_count = 0;
2066                         spin_unlock(&sctx->list_lock);
2067                 } else {
2068                         spin_unlock(&sctx->list_lock);
2069                         wait_event(sctx->list_wait, sctx->first_free != -1);
2070                 }
2071         }
2072         sbio = sctx->bios[sctx->curr];
2073         if (sbio->page_count == 0) {
2074                 struct bio *bio;
2075
2076                 sbio->physical = spage->physical;
2077                 sbio->logical = spage->logical;
2078                 sbio->dev = spage->dev;
2079                 bio = sbio->bio;
2080                 if (!bio) {
2081                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
2082                         if (!bio)
2083                                 return -ENOMEM;
2084                         sbio->bio = bio;
2085                 }
2086
2087                 bio->bi_private = sbio;
2088                 bio->bi_end_io = scrub_bio_end_io;
2089                 bio->bi_bdev = sbio->dev->bdev;
2090                 bio->bi_iter.bi_sector = sbio->physical >> 9;
2091                 sbio->err = 0;
2092         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
2093                    spage->physical ||
2094                    sbio->logical + sbio->page_count * PAGE_SIZE !=
2095                    spage->logical ||
2096                    sbio->dev != spage->dev) {
2097                 scrub_submit(sctx);
2098                 goto again;
2099         }
2100
2101         sbio->pagev[sbio->page_count] = spage;
2102         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
2103         if (ret != PAGE_SIZE) {
2104                 if (sbio->page_count < 1) {
2105                         bio_put(sbio->bio);
2106                         sbio->bio = NULL;
2107                         return -EIO;
2108                 }
2109                 scrub_submit(sctx);
2110                 goto again;
2111         }
2112
2113         scrub_block_get(sblock); /* one for the page added to the bio */
2114         atomic_inc(&sblock->outstanding_pages);
2115         sbio->page_count++;
2116         if (sbio->page_count == sctx->pages_per_rd_bio)
2117                 scrub_submit(sctx);
2118
2119         return 0;
2120 }
2121
2122 static void scrub_missing_raid56_end_io(struct bio *bio)
2123 {
2124         struct scrub_block *sblock = bio->bi_private;
2125         struct btrfs_fs_info *fs_info = sblock->sctx->dev_root->fs_info;
2126
2127         if (bio->bi_error)
2128                 sblock->no_io_error_seen = 0;
2129
2130         btrfs_queue_work(fs_info->scrub_workers, &sblock->work);
2131 }
2132
2133 static void scrub_missing_raid56_worker(struct btrfs_work *work)
2134 {
2135         struct scrub_block *sblock = container_of(work, struct scrub_block, work);
2136         struct scrub_ctx *sctx = sblock->sctx;
2137         u64 logical;
2138         struct btrfs_device *dev;
2139
2140         logical = sblock->pagev[0]->logical;
2141         dev = sblock->pagev[0]->dev;
2142
2143         if (sblock->no_io_error_seen)
2144                 scrub_recheck_block_checksum(sblock);
2145
2146         if (!sblock->no_io_error_seen) {
2147                 spin_lock(&sctx->stat_lock);
2148                 sctx->stat.read_errors++;
2149                 spin_unlock(&sctx->stat_lock);
2150                 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2151                         "IO error rebuilding logical %llu for dev %s",
2152                         logical, rcu_str_deref(dev->name));
2153         } else if (sblock->header_error || sblock->checksum_error) {
2154                 spin_lock(&sctx->stat_lock);
2155                 sctx->stat.uncorrectable_errors++;
2156                 spin_unlock(&sctx->stat_lock);
2157                 btrfs_err_rl_in_rcu(sctx->dev_root->fs_info,
2158                         "failed to rebuild valid logical %llu for dev %s",
2159                         logical, rcu_str_deref(dev->name));
2160         } else {
2161                 scrub_write_block_to_dev_replace(sblock);
2162         }
2163
2164         scrub_block_put(sblock);
2165
2166         if (sctx->is_dev_replace &&
2167             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2168                 mutex_lock(&sctx->wr_ctx.wr_lock);
2169                 scrub_wr_submit(sctx);
2170                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2171         }
2172
2173         scrub_pending_bio_dec(sctx);
2174 }
2175
2176 static void scrub_missing_raid56_pages(struct scrub_block *sblock)
2177 {
2178         struct scrub_ctx *sctx = sblock->sctx;
2179         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2180         u64 length = sblock->page_count * PAGE_SIZE;
2181         u64 logical = sblock->pagev[0]->logical;
2182         struct btrfs_bio *bbio;
2183         struct bio *bio;
2184         struct btrfs_raid_bio *rbio;
2185         int ret;
2186         int i;
2187
2188         ret = btrfs_map_sblock(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
2189                                &bbio, 0, 1);
2190         if (ret || !bbio || !bbio->raid_map)
2191                 goto bbio_out;
2192
2193         if (WARN_ON(!sctx->is_dev_replace ||
2194                     !(bbio->map_type & BTRFS_BLOCK_GROUP_RAID56_MASK))) {
2195                 /*
2196                  * We shouldn't be scrubbing a missing device. Even for dev
2197                  * replace, we should only get here for RAID 5/6. We either
2198                  * managed to mount something with no mirrors remaining or
2199                  * there's a bug in scrub_remap_extent()/btrfs_map_block().
2200                  */
2201                 goto bbio_out;
2202         }
2203
2204         bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2205         if (!bio)
2206                 goto bbio_out;
2207
2208         bio->bi_iter.bi_sector = logical >> 9;
2209         bio->bi_private = sblock;
2210         bio->bi_end_io = scrub_missing_raid56_end_io;
2211
2212         rbio = raid56_alloc_missing_rbio(sctx->dev_root, bio, bbio, length);
2213         if (!rbio)
2214                 goto rbio_out;
2215
2216         for (i = 0; i < sblock->page_count; i++) {
2217                 struct scrub_page *spage = sblock->pagev[i];
2218
2219                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2220         }
2221
2222         btrfs_init_work(&sblock->work, btrfs_scrub_helper,
2223                         scrub_missing_raid56_worker, NULL, NULL);
2224         scrub_block_get(sblock);
2225         scrub_pending_bio_inc(sctx);
2226         raid56_submit_missing_rbio(rbio);
2227         return;
2228
2229 rbio_out:
2230         bio_put(bio);
2231 bbio_out:
2232         btrfs_put_bbio(bbio);
2233         spin_lock(&sctx->stat_lock);
2234         sctx->stat.malloc_errors++;
2235         spin_unlock(&sctx->stat_lock);
2236 }
2237
2238 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2239                        u64 physical, struct btrfs_device *dev, u64 flags,
2240                        u64 gen, int mirror_num, u8 *csum, int force,
2241                        u64 physical_for_dev_replace)
2242 {
2243         struct scrub_block *sblock;
2244         int index;
2245
2246         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2247         if (!sblock) {
2248                 spin_lock(&sctx->stat_lock);
2249                 sctx->stat.malloc_errors++;
2250                 spin_unlock(&sctx->stat_lock);
2251                 return -ENOMEM;
2252         }
2253
2254         /* one ref inside this function, plus one for each page added to
2255          * a bio later on */
2256         atomic_set(&sblock->refs, 1);
2257         sblock->sctx = sctx;
2258         sblock->no_io_error_seen = 1;
2259
2260         for (index = 0; len > 0; index++) {
2261                 struct scrub_page *spage;
2262                 u64 l = min_t(u64, len, PAGE_SIZE);
2263
2264                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2265                 if (!spage) {
2266 leave_nomem:
2267                         spin_lock(&sctx->stat_lock);
2268                         sctx->stat.malloc_errors++;
2269                         spin_unlock(&sctx->stat_lock);
2270                         scrub_block_put(sblock);
2271                         return -ENOMEM;
2272                 }
2273                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2274                 scrub_page_get(spage);
2275                 sblock->pagev[index] = spage;
2276                 spage->sblock = sblock;
2277                 spage->dev = dev;
2278                 spage->flags = flags;
2279                 spage->generation = gen;
2280                 spage->logical = logical;
2281                 spage->physical = physical;
2282                 spage->physical_for_dev_replace = physical_for_dev_replace;
2283                 spage->mirror_num = mirror_num;
2284                 if (csum) {
2285                         spage->have_csum = 1;
2286                         memcpy(spage->csum, csum, sctx->csum_size);
2287                 } else {
2288                         spage->have_csum = 0;
2289                 }
2290                 sblock->page_count++;
2291                 spage->page = alloc_page(GFP_NOFS);
2292                 if (!spage->page)
2293                         goto leave_nomem;
2294                 len -= l;
2295                 logical += l;
2296                 physical += l;
2297                 physical_for_dev_replace += l;
2298         }
2299
2300         WARN_ON(sblock->page_count == 0);
2301         if (dev->missing) {
2302                 /*
2303                  * This case should only be hit for RAID 5/6 device replace. See
2304                  * the comment in scrub_missing_raid56_pages() for details.
2305                  */
2306                 scrub_missing_raid56_pages(sblock);
2307         } else {
2308                 for (index = 0; index < sblock->page_count; index++) {
2309                         struct scrub_page *spage = sblock->pagev[index];
2310                         int ret;
2311
2312                         ret = scrub_add_page_to_rd_bio(sctx, spage);
2313                         if (ret) {
2314                                 scrub_block_put(sblock);
2315                                 return ret;
2316                         }
2317                 }
2318
2319                 if (force)
2320                         scrub_submit(sctx);
2321         }
2322
2323         /* last one frees, either here or in bio completion for last page */
2324         scrub_block_put(sblock);
2325         return 0;
2326 }
2327
2328 static void scrub_bio_end_io(struct bio *bio)
2329 {
2330         struct scrub_bio *sbio = bio->bi_private;
2331         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2332
2333         sbio->err = bio->bi_error;
2334         sbio->bio = bio;
2335
2336         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2337 }
2338
2339 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2340 {
2341         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2342         struct scrub_ctx *sctx = sbio->sctx;
2343         int i;
2344
2345         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2346         if (sbio->err) {
2347                 for (i = 0; i < sbio->page_count; i++) {
2348                         struct scrub_page *spage = sbio->pagev[i];
2349
2350                         spage->io_error = 1;
2351                         spage->sblock->no_io_error_seen = 0;
2352                 }
2353         }
2354
2355         /* now complete the scrub_block items that have all pages completed */
2356         for (i = 0; i < sbio->page_count; i++) {
2357                 struct scrub_page *spage = sbio->pagev[i];
2358                 struct scrub_block *sblock = spage->sblock;
2359
2360                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2361                         scrub_block_complete(sblock);
2362                 scrub_block_put(sblock);
2363         }
2364
2365         bio_put(sbio->bio);
2366         sbio->bio = NULL;
2367         spin_lock(&sctx->list_lock);
2368         sbio->next_free = sctx->first_free;
2369         sctx->first_free = sbio->index;
2370         spin_unlock(&sctx->list_lock);
2371
2372         if (sctx->is_dev_replace &&
2373             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2374                 mutex_lock(&sctx->wr_ctx.wr_lock);
2375                 scrub_wr_submit(sctx);
2376                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2377         }
2378
2379         scrub_pending_bio_dec(sctx);
2380 }
2381
2382 static inline void __scrub_mark_bitmap(struct scrub_parity *sparity,
2383                                        unsigned long *bitmap,
2384                                        u64 start, u64 len)
2385 {
2386         u32 offset;
2387         int nsectors;
2388         int sectorsize = sparity->sctx->dev_root->sectorsize;
2389
2390         if (len >= sparity->stripe_len) {
2391                 bitmap_set(bitmap, 0, sparity->nsectors);
2392                 return;
2393         }
2394
2395         start -= sparity->logic_start;
2396         start = div_u64_rem(start, sparity->stripe_len, &offset);
2397         offset /= sectorsize;
2398         nsectors = (int)len / sectorsize;
2399
2400         if (offset + nsectors <= sparity->nsectors) {
2401                 bitmap_set(bitmap, offset, nsectors);
2402                 return;
2403         }
2404
2405         bitmap_set(bitmap, offset, sparity->nsectors - offset);
2406         bitmap_set(bitmap, 0, nsectors - (sparity->nsectors - offset));
2407 }
2408
2409 static inline void scrub_parity_mark_sectors_error(struct scrub_parity *sparity,
2410                                                    u64 start, u64 len)
2411 {
2412         __scrub_mark_bitmap(sparity, sparity->ebitmap, start, len);
2413 }
2414
2415 static inline void scrub_parity_mark_sectors_data(struct scrub_parity *sparity,
2416                                                   u64 start, u64 len)
2417 {
2418         __scrub_mark_bitmap(sparity, sparity->dbitmap, start, len);
2419 }
2420
2421 static void scrub_block_complete(struct scrub_block *sblock)
2422 {
2423         int corrupted = 0;
2424
2425         if (!sblock->no_io_error_seen) {
2426                 corrupted = 1;
2427                 scrub_handle_errored_block(sblock);
2428         } else {
2429                 /*
2430                  * if has checksum error, write via repair mechanism in
2431                  * dev replace case, otherwise write here in dev replace
2432                  * case.
2433                  */
2434                 corrupted = scrub_checksum(sblock);
2435                 if (!corrupted && sblock->sctx->is_dev_replace)
2436                         scrub_write_block_to_dev_replace(sblock);
2437         }
2438
2439         if (sblock->sparity && corrupted && !sblock->data_corrected) {
2440                 u64 start = sblock->pagev[0]->logical;
2441                 u64 end = sblock->pagev[sblock->page_count - 1]->logical +
2442                           PAGE_SIZE;
2443
2444                 scrub_parity_mark_sectors_error(sblock->sparity,
2445                                                 start, end - start);
2446         }
2447 }
2448
2449 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u8 *csum)
2450 {
2451         struct btrfs_ordered_sum *sum = NULL;
2452         unsigned long index;
2453         unsigned long num_sectors;
2454
2455         while (!list_empty(&sctx->csum_list)) {
2456                 sum = list_first_entry(&sctx->csum_list,
2457                                        struct btrfs_ordered_sum, list);
2458                 if (sum->bytenr > logical)
2459                         return 0;
2460                 if (sum->bytenr + sum->len > logical)
2461                         break;
2462
2463                 ++sctx->stat.csum_discards;
2464                 list_del(&sum->list);
2465                 kfree(sum);
2466                 sum = NULL;
2467         }
2468         if (!sum)
2469                 return 0;
2470
2471         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2472         num_sectors = sum->len / sctx->sectorsize;
2473         memcpy(csum, sum->sums + index, sctx->csum_size);
2474         if (index == num_sectors - 1) {
2475                 list_del(&sum->list);
2476                 kfree(sum);
2477         }
2478         return 1;
2479 }
2480
2481 /* scrub extent tries to collect up to 64 kB for each bio */
2482 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2483                         u64 physical, struct btrfs_device *dev, u64 flags,
2484                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2485 {
2486         int ret;
2487         u8 csum[BTRFS_CSUM_SIZE];
2488         u32 blocksize;
2489
2490         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2491                 blocksize = sctx->sectorsize;
2492                 spin_lock(&sctx->stat_lock);
2493                 sctx->stat.data_extents_scrubbed++;
2494                 sctx->stat.data_bytes_scrubbed += len;
2495                 spin_unlock(&sctx->stat_lock);
2496         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2497                 blocksize = sctx->nodesize;
2498                 spin_lock(&sctx->stat_lock);
2499                 sctx->stat.tree_extents_scrubbed++;
2500                 sctx->stat.tree_bytes_scrubbed += len;
2501                 spin_unlock(&sctx->stat_lock);
2502         } else {
2503                 blocksize = sctx->sectorsize;
2504                 WARN_ON(1);
2505         }
2506
2507         while (len) {
2508                 u64 l = min_t(u64, len, blocksize);
2509                 int have_csum = 0;
2510
2511                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2512                         /* push csums to sbio */
2513                         have_csum = scrub_find_csum(sctx, logical, csum);
2514                         if (have_csum == 0)
2515                                 ++sctx->stat.no_csum;
2516                         if (sctx->is_dev_replace && !have_csum) {
2517                                 ret = copy_nocow_pages(sctx, logical, l,
2518                                                        mirror_num,
2519                                                       physical_for_dev_replace);
2520                                 goto behind_scrub_pages;
2521                         }
2522                 }
2523                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2524                                   mirror_num, have_csum ? csum : NULL, 0,
2525                                   physical_for_dev_replace);
2526 behind_scrub_pages:
2527                 if (ret)
2528                         return ret;
2529                 len -= l;
2530                 logical += l;
2531                 physical += l;
2532                 physical_for_dev_replace += l;
2533         }
2534         return 0;
2535 }
2536
2537 static int scrub_pages_for_parity(struct scrub_parity *sparity,
2538                                   u64 logical, u64 len,
2539                                   u64 physical, struct btrfs_device *dev,
2540                                   u64 flags, u64 gen, int mirror_num, u8 *csum)
2541 {
2542         struct scrub_ctx *sctx = sparity->sctx;
2543         struct scrub_block *sblock;
2544         int index;
2545
2546         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2547         if (!sblock) {
2548                 spin_lock(&sctx->stat_lock);
2549                 sctx->stat.malloc_errors++;
2550                 spin_unlock(&sctx->stat_lock);
2551                 return -ENOMEM;
2552         }
2553
2554         /* one ref inside this function, plus one for each page added to
2555          * a bio later on */
2556         atomic_set(&sblock->refs, 1);
2557         sblock->sctx = sctx;
2558         sblock->no_io_error_seen = 1;
2559         sblock->sparity = sparity;
2560         scrub_parity_get(sparity);
2561
2562         for (index = 0; len > 0; index++) {
2563                 struct scrub_page *spage;
2564                 u64 l = min_t(u64, len, PAGE_SIZE);
2565
2566                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2567                 if (!spage) {
2568 leave_nomem:
2569                         spin_lock(&sctx->stat_lock);
2570                         sctx->stat.malloc_errors++;
2571                         spin_unlock(&sctx->stat_lock);
2572                         scrub_block_put(sblock);
2573                         return -ENOMEM;
2574                 }
2575                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2576                 /* For scrub block */
2577                 scrub_page_get(spage);
2578                 sblock->pagev[index] = spage;
2579                 /* For scrub parity */
2580                 scrub_page_get(spage);
2581                 list_add_tail(&spage->list, &sparity->spages);
2582                 spage->sblock = sblock;
2583                 spage->dev = dev;
2584                 spage->flags = flags;
2585                 spage->generation = gen;
2586                 spage->logical = logical;
2587                 spage->physical = physical;
2588                 spage->mirror_num = mirror_num;
2589                 if (csum) {
2590                         spage->have_csum = 1;
2591                         memcpy(spage->csum, csum, sctx->csum_size);
2592                 } else {
2593                         spage->have_csum = 0;
2594                 }
2595                 sblock->page_count++;
2596                 spage->page = alloc_page(GFP_NOFS);
2597                 if (!spage->page)
2598                         goto leave_nomem;
2599                 len -= l;
2600                 logical += l;
2601                 physical += l;
2602         }
2603
2604         WARN_ON(sblock->page_count == 0);
2605         for (index = 0; index < sblock->page_count; index++) {
2606                 struct scrub_page *spage = sblock->pagev[index];
2607                 int ret;
2608
2609                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2610                 if (ret) {
2611                         scrub_block_put(sblock);
2612                         return ret;
2613                 }
2614         }
2615
2616         /* last one frees, either here or in bio completion for last page */
2617         scrub_block_put(sblock);
2618         return 0;
2619 }
2620
2621 static int scrub_extent_for_parity(struct scrub_parity *sparity,
2622                                    u64 logical, u64 len,
2623                                    u64 physical, struct btrfs_device *dev,
2624                                    u64 flags, u64 gen, int mirror_num)
2625 {
2626         struct scrub_ctx *sctx = sparity->sctx;
2627         int ret;
2628         u8 csum[BTRFS_CSUM_SIZE];
2629         u32 blocksize;
2630
2631         if (dev->missing) {
2632                 scrub_parity_mark_sectors_error(sparity, logical, len);
2633                 return 0;
2634         }
2635
2636         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2637                 blocksize = sctx->sectorsize;
2638         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2639                 blocksize = sctx->nodesize;
2640         } else {
2641                 blocksize = sctx->sectorsize;
2642                 WARN_ON(1);
2643         }
2644
2645         while (len) {
2646                 u64 l = min_t(u64, len, blocksize);
2647                 int have_csum = 0;
2648
2649                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2650                         /* push csums to sbio */
2651                         have_csum = scrub_find_csum(sctx, logical, csum);
2652                         if (have_csum == 0)
2653                                 goto skip;
2654                 }
2655                 ret = scrub_pages_for_parity(sparity, logical, l, physical, dev,
2656                                              flags, gen, mirror_num,
2657                                              have_csum ? csum : NULL);
2658                 if (ret)
2659                         return ret;
2660 skip:
2661                 len -= l;
2662                 logical += l;
2663                 physical += l;
2664         }
2665         return 0;
2666 }
2667
2668 /*
2669  * Given a physical address, this will calculate it's
2670  * logical offset. if this is a parity stripe, it will return
2671  * the most left data stripe's logical offset.
2672  *
2673  * return 0 if it is a data stripe, 1 means parity stripe.
2674  */
2675 static int get_raid56_logic_offset(u64 physical, int num,
2676                                    struct map_lookup *map, u64 *offset,
2677                                    u64 *stripe_start)
2678 {
2679         int i;
2680         int j = 0;
2681         u64 stripe_nr;
2682         u64 last_offset;
2683         u32 stripe_index;
2684         u32 rot;
2685
2686         last_offset = (physical - map->stripes[num].physical) *
2687                       nr_data_stripes(map);
2688         if (stripe_start)
2689                 *stripe_start = last_offset;
2690
2691         *offset = last_offset;
2692         for (i = 0; i < nr_data_stripes(map); i++) {
2693                 *offset = last_offset + i * map->stripe_len;
2694
2695                 stripe_nr = div_u64(*offset, map->stripe_len);
2696                 stripe_nr = div_u64(stripe_nr, nr_data_stripes(map));
2697
2698                 /* Work out the disk rotation on this stripe-set */
2699                 stripe_nr = div_u64_rem(stripe_nr, map->num_stripes, &rot);
2700                 /* calculate which stripe this data locates */
2701                 rot += i;
2702                 stripe_index = rot % map->num_stripes;
2703                 if (stripe_index == num)
2704                         return 0;
2705                 if (stripe_index < num)
2706                         j++;
2707         }
2708         *offset = last_offset + j * map->stripe_len;
2709         return 1;
2710 }
2711
2712 static void scrub_free_parity(struct scrub_parity *sparity)
2713 {
2714         struct scrub_ctx *sctx = sparity->sctx;
2715         struct scrub_page *curr, *next;
2716         int nbits;
2717
2718         nbits = bitmap_weight(sparity->ebitmap, sparity->nsectors);
2719         if (nbits) {
2720                 spin_lock(&sctx->stat_lock);
2721                 sctx->stat.read_errors += nbits;
2722                 sctx->stat.uncorrectable_errors += nbits;
2723                 spin_unlock(&sctx->stat_lock);
2724         }
2725
2726         list_for_each_entry_safe(curr, next, &sparity->spages, list) {
2727                 list_del_init(&curr->list);
2728                 scrub_page_put(curr);
2729         }
2730
2731         kfree(sparity);
2732 }
2733
2734 static void scrub_parity_bio_endio_worker(struct btrfs_work *work)
2735 {
2736         struct scrub_parity *sparity = container_of(work, struct scrub_parity,
2737                                                     work);
2738         struct scrub_ctx *sctx = sparity->sctx;
2739
2740         scrub_free_parity(sparity);
2741         scrub_pending_bio_dec(sctx);
2742 }
2743
2744 static void scrub_parity_bio_endio(struct bio *bio)
2745 {
2746         struct scrub_parity *sparity = (struct scrub_parity *)bio->bi_private;
2747
2748         if (bio->bi_error)
2749                 bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2750                           sparity->nsectors);
2751
2752         bio_put(bio);
2753
2754         btrfs_init_work(&sparity->work, btrfs_scrubparity_helper,
2755                         scrub_parity_bio_endio_worker, NULL, NULL);
2756         btrfs_queue_work(sparity->sctx->dev_root->fs_info->scrub_parity_workers,
2757                          &sparity->work);
2758 }
2759
2760 static void scrub_parity_check_and_repair(struct scrub_parity *sparity)
2761 {
2762         struct scrub_ctx *sctx = sparity->sctx;
2763         struct bio *bio;
2764         struct btrfs_raid_bio *rbio;
2765         struct scrub_page *spage;
2766         struct btrfs_bio *bbio = NULL;
2767         u64 length;
2768         int ret;
2769
2770         if (!bitmap_andnot(sparity->dbitmap, sparity->dbitmap, sparity->ebitmap,
2771                            sparity->nsectors))
2772                 goto out;
2773
2774         length = sparity->logic_end - sparity->logic_start;
2775         ret = btrfs_map_sblock(sctx->dev_root->fs_info, WRITE,
2776                                sparity->logic_start,
2777                                &length, &bbio, 0, 1);
2778         if (ret || !bbio || !bbio->raid_map)
2779                 goto bbio_out;
2780
2781         bio = btrfs_io_bio_alloc(GFP_NOFS, 0);
2782         if (!bio)
2783                 goto bbio_out;
2784
2785         bio->bi_iter.bi_sector = sparity->logic_start >> 9;
2786         bio->bi_private = sparity;
2787         bio->bi_end_io = scrub_parity_bio_endio;
2788
2789         rbio = raid56_parity_alloc_scrub_rbio(sctx->dev_root, bio, bbio,
2790                                               length, sparity->scrub_dev,
2791                                               sparity->dbitmap,
2792                                               sparity->nsectors);
2793         if (!rbio)
2794                 goto rbio_out;
2795
2796         list_for_each_entry(spage, &sparity->spages, list)
2797                 raid56_add_scrub_pages(rbio, spage->page, spage->logical);
2798
2799         scrub_pending_bio_inc(sctx);
2800         raid56_parity_submit_scrub_rbio(rbio);
2801         return;
2802
2803 rbio_out:
2804         bio_put(bio);
2805 bbio_out:
2806         btrfs_put_bbio(bbio);
2807         bitmap_or(sparity->ebitmap, sparity->ebitmap, sparity->dbitmap,
2808                   sparity->nsectors);
2809         spin_lock(&sctx->stat_lock);
2810         sctx->stat.malloc_errors++;
2811         spin_unlock(&sctx->stat_lock);
2812 out:
2813         scrub_free_parity(sparity);
2814 }
2815
2816 static inline int scrub_calc_parity_bitmap_len(int nsectors)
2817 {
2818         return DIV_ROUND_UP(nsectors, BITS_PER_LONG) * (BITS_PER_LONG / 8);
2819 }
2820
2821 static void scrub_parity_get(struct scrub_parity *sparity)
2822 {
2823         atomic_inc(&sparity->refs);
2824 }
2825
2826 static void scrub_parity_put(struct scrub_parity *sparity)
2827 {
2828         if (!atomic_dec_and_test(&sparity->refs))
2829                 return;
2830
2831         scrub_parity_check_and_repair(sparity);
2832 }
2833
2834 static noinline_for_stack int scrub_raid56_parity(struct scrub_ctx *sctx,
2835                                                   struct map_lookup *map,
2836                                                   struct btrfs_device *sdev,
2837                                                   struct btrfs_path *path,
2838                                                   u64 logic_start,
2839                                                   u64 logic_end)
2840 {
2841         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2842         struct btrfs_root *root = fs_info->extent_root;
2843         struct btrfs_root *csum_root = fs_info->csum_root;
2844         struct btrfs_extent_item *extent;
2845         struct btrfs_bio *bbio = NULL;
2846         u64 flags;
2847         int ret;
2848         int slot;
2849         struct extent_buffer *l;
2850         struct btrfs_key key;
2851         u64 generation;
2852         u64 extent_logical;
2853         u64 extent_physical;
2854         u64 extent_len;
2855         u64 mapped_length;
2856         struct btrfs_device *extent_dev;
2857         struct scrub_parity *sparity;
2858         int nsectors;
2859         int bitmap_len;
2860         int extent_mirror_num;
2861         int stop_loop = 0;
2862
2863         nsectors = map->stripe_len / root->sectorsize;
2864         bitmap_len = scrub_calc_parity_bitmap_len(nsectors);
2865         sparity = kzalloc(sizeof(struct scrub_parity) + 2 * bitmap_len,
2866                           GFP_NOFS);
2867         if (!sparity) {
2868                 spin_lock(&sctx->stat_lock);
2869                 sctx->stat.malloc_errors++;
2870                 spin_unlock(&sctx->stat_lock);
2871                 return -ENOMEM;
2872         }
2873
2874         sparity->stripe_len = map->stripe_len;
2875         sparity->nsectors = nsectors;
2876         sparity->sctx = sctx;
2877         sparity->scrub_dev = sdev;
2878         sparity->logic_start = logic_start;
2879         sparity->logic_end = logic_end;
2880         atomic_set(&sparity->refs, 1);
2881         INIT_LIST_HEAD(&sparity->spages);
2882         sparity->dbitmap = sparity->bitmap;
2883         sparity->ebitmap = (void *)sparity->bitmap + bitmap_len;
2884
2885         ret = 0;
2886         while (logic_start < logic_end) {
2887                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2888                         key.type = BTRFS_METADATA_ITEM_KEY;
2889                 else
2890                         key.type = BTRFS_EXTENT_ITEM_KEY;
2891                 key.objectid = logic_start;
2892                 key.offset = (u64)-1;
2893
2894                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2895                 if (ret < 0)
2896                         goto out;
2897
2898                 if (ret > 0) {
2899                         ret = btrfs_previous_extent_item(root, path, 0);
2900                         if (ret < 0)
2901                                 goto out;
2902                         if (ret > 0) {
2903                                 btrfs_release_path(path);
2904                                 ret = btrfs_search_slot(NULL, root, &key,
2905                                                         path, 0, 0);
2906                                 if (ret < 0)
2907                                         goto out;
2908                         }
2909                 }
2910
2911                 stop_loop = 0;
2912                 while (1) {
2913                         u64 bytes;
2914
2915                         l = path->nodes[0];
2916                         slot = path->slots[0];
2917                         if (slot >= btrfs_header_nritems(l)) {
2918                                 ret = btrfs_next_leaf(root, path);
2919                                 if (ret == 0)
2920                                         continue;
2921                                 if (ret < 0)
2922                                         goto out;
2923
2924                                 stop_loop = 1;
2925                                 break;
2926                         }
2927                         btrfs_item_key_to_cpu(l, &key, slot);
2928
2929                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2930                             key.type != BTRFS_METADATA_ITEM_KEY)
2931                                 goto next;
2932
2933                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2934                                 bytes = root->nodesize;
2935                         else
2936                                 bytes = key.offset;
2937
2938                         if (key.objectid + bytes <= logic_start)
2939                                 goto next;
2940
2941                         if (key.objectid >= logic_end) {
2942                                 stop_loop = 1;
2943                                 break;
2944                         }
2945
2946                         while (key.objectid >= logic_start + map->stripe_len)
2947                                 logic_start += map->stripe_len;
2948
2949                         extent = btrfs_item_ptr(l, slot,
2950                                                 struct btrfs_extent_item);
2951                         flags = btrfs_extent_flags(l, extent);
2952                         generation = btrfs_extent_generation(l, extent);
2953
2954                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
2955                             (key.objectid < logic_start ||
2956                              key.objectid + bytes >
2957                              logic_start + map->stripe_len)) {
2958                                 btrfs_err(fs_info, "scrub: tree block %llu spanning stripes, ignored. logical=%llu",
2959                                           key.objectid, logic_start);
2960                                 spin_lock(&sctx->stat_lock);
2961                                 sctx->stat.uncorrectable_errors++;
2962                                 spin_unlock(&sctx->stat_lock);
2963                                 goto next;
2964                         }
2965 again:
2966                         extent_logical = key.objectid;
2967                         extent_len = bytes;
2968
2969                         if (extent_logical < logic_start) {
2970                                 extent_len -= logic_start - extent_logical;
2971                                 extent_logical = logic_start;
2972                         }
2973
2974                         if (extent_logical + extent_len >
2975                             logic_start + map->stripe_len)
2976                                 extent_len = logic_start + map->stripe_len -
2977                                              extent_logical;
2978
2979                         scrub_parity_mark_sectors_data(sparity, extent_logical,
2980                                                        extent_len);
2981
2982                         mapped_length = extent_len;
2983                         ret = btrfs_map_block(fs_info, READ, extent_logical,
2984                                               &mapped_length, &bbio, 0);
2985                         if (!ret) {
2986                                 if (!bbio || mapped_length < extent_len)
2987                                         ret = -EIO;
2988                         }
2989                         if (ret) {
2990                                 btrfs_put_bbio(bbio);
2991                                 goto out;
2992                         }
2993                         extent_physical = bbio->stripes[0].physical;
2994                         extent_mirror_num = bbio->mirror_num;
2995                         extent_dev = bbio->stripes[0].dev;
2996                         btrfs_put_bbio(bbio);
2997
2998                         ret = btrfs_lookup_csums_range(csum_root,
2999                                                 extent_logical,
3000                                                 extent_logical + extent_len - 1,
3001                                                 &sctx->csum_list, 1);
3002                         if (ret)
3003                                 goto out;
3004
3005                         ret = scrub_extent_for_parity(sparity, extent_logical,
3006                                                       extent_len,
3007                                                       extent_physical,
3008                                                       extent_dev, flags,
3009                                                       generation,
3010                                                       extent_mirror_num);
3011
3012                         scrub_free_csums(sctx);
3013
3014                         if (ret)
3015                                 goto out;
3016
3017                         if (extent_logical + extent_len <
3018                             key.objectid + bytes) {
3019                                 logic_start += map->stripe_len;
3020
3021                                 if (logic_start >= logic_end) {
3022                                         stop_loop = 1;
3023                                         break;
3024                                 }
3025
3026                                 if (logic_start < key.objectid + bytes) {
3027                                         cond_resched();
3028                                         goto again;
3029                                 }
3030                         }
3031 next:
3032                         path->slots[0]++;
3033                 }
3034
3035                 btrfs_release_path(path);
3036
3037                 if (stop_loop)
3038                         break;
3039
3040                 logic_start += map->stripe_len;
3041         }
3042 out:
3043         if (ret < 0)
3044                 scrub_parity_mark_sectors_error(sparity, logic_start,
3045                                                 logic_end - logic_start);
3046         scrub_parity_put(sparity);
3047         scrub_submit(sctx);
3048         mutex_lock(&sctx->wr_ctx.wr_lock);
3049         scrub_wr_submit(sctx);
3050         mutex_unlock(&sctx->wr_ctx.wr_lock);
3051
3052         btrfs_release_path(path);
3053         return ret < 0 ? ret : 0;
3054 }
3055
3056 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
3057                                            struct map_lookup *map,
3058                                            struct btrfs_device *scrub_dev,
3059                                            int num, u64 base, u64 length,
3060                                            int is_dev_replace)
3061 {
3062         struct btrfs_path *path, *ppath;
3063         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3064         struct btrfs_root *root = fs_info->extent_root;
3065         struct btrfs_root *csum_root = fs_info->csum_root;
3066         struct btrfs_extent_item *extent;
3067         struct blk_plug plug;
3068         u64 flags;
3069         int ret;
3070         int slot;
3071         u64 nstripes;
3072         struct extent_buffer *l;
3073         struct btrfs_key key;
3074         u64 physical;
3075         u64 logical;
3076         u64 logic_end;
3077         u64 physical_end;
3078         u64 generation;
3079         int mirror_num;
3080         struct reada_control *reada1;
3081         struct reada_control *reada2;
3082         struct btrfs_key key_start;
3083         struct btrfs_key key_end;
3084         u64 increment = map->stripe_len;
3085         u64 offset;
3086         u64 extent_logical;
3087         u64 extent_physical;
3088         u64 extent_len;
3089         u64 stripe_logical;
3090         u64 stripe_end;
3091         struct btrfs_device *extent_dev;
3092         int extent_mirror_num;
3093         int stop_loop = 0;
3094
3095         physical = map->stripes[num].physical;
3096         offset = 0;
3097         nstripes = div_u64(length, map->stripe_len);
3098         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
3099                 offset = map->stripe_len * num;
3100                 increment = map->stripe_len * map->num_stripes;
3101                 mirror_num = 1;
3102         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
3103                 int factor = map->num_stripes / map->sub_stripes;
3104                 offset = map->stripe_len * (num / map->sub_stripes);
3105                 increment = map->stripe_len * factor;
3106                 mirror_num = num % map->sub_stripes + 1;
3107         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
3108                 increment = map->stripe_len;
3109                 mirror_num = num % map->num_stripes + 1;
3110         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
3111                 increment = map->stripe_len;
3112                 mirror_num = num % map->num_stripes + 1;
3113         } else if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3114                 get_raid56_logic_offset(physical, num, map, &offset, NULL);
3115                 increment = map->stripe_len * nr_data_stripes(map);
3116                 mirror_num = 1;
3117         } else {
3118                 increment = map->stripe_len;
3119                 mirror_num = 1;
3120         }
3121
3122         path = btrfs_alloc_path();
3123         if (!path)
3124                 return -ENOMEM;
3125
3126         ppath = btrfs_alloc_path();
3127         if (!ppath) {
3128                 btrfs_free_path(path);
3129                 return -ENOMEM;
3130         }
3131
3132         /*
3133          * work on commit root. The related disk blocks are static as
3134          * long as COW is applied. This means, it is save to rewrite
3135          * them to repair disk errors without any race conditions
3136          */
3137         path->search_commit_root = 1;
3138         path->skip_locking = 1;
3139
3140         ppath->search_commit_root = 1;
3141         ppath->skip_locking = 1;
3142         /*
3143          * trigger the readahead for extent tree csum tree and wait for
3144          * completion. During readahead, the scrub is officially paused
3145          * to not hold off transaction commits
3146          */
3147         logical = base + offset;
3148         physical_end = physical + nstripes * map->stripe_len;
3149         if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3150                 get_raid56_logic_offset(physical_end, num,
3151                                         map, &logic_end, NULL);
3152                 logic_end += base;
3153         } else {
3154                 logic_end = logical + increment * nstripes;
3155         }
3156         wait_event(sctx->list_wait,
3157                    atomic_read(&sctx->bios_in_flight) == 0);
3158         scrub_blocked_if_needed(fs_info);
3159
3160         /* FIXME it might be better to start readahead at commit root */
3161         key_start.objectid = logical;
3162         key_start.type = BTRFS_EXTENT_ITEM_KEY;
3163         key_start.offset = (u64)0;
3164         key_end.objectid = logic_end;
3165         key_end.type = BTRFS_METADATA_ITEM_KEY;
3166         key_end.offset = (u64)-1;
3167         reada1 = btrfs_reada_add(root, &key_start, &key_end);
3168
3169         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3170         key_start.type = BTRFS_EXTENT_CSUM_KEY;
3171         key_start.offset = logical;
3172         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
3173         key_end.type = BTRFS_EXTENT_CSUM_KEY;
3174         key_end.offset = logic_end;
3175         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
3176
3177         if (!IS_ERR(reada1))
3178                 btrfs_reada_wait(reada1);
3179         if (!IS_ERR(reada2))
3180                 btrfs_reada_wait(reada2);
3181
3182
3183         /*
3184          * collect all data csums for the stripe to avoid seeking during
3185          * the scrub. This might currently (crc32) end up to be about 1MB
3186          */
3187         blk_start_plug(&plug);
3188
3189         /*
3190          * now find all extents for each stripe and scrub them
3191          */
3192         ret = 0;
3193         while (physical < physical_end) {
3194                 /*
3195                  * canceled?
3196                  */
3197                 if (atomic_read(&fs_info->scrub_cancel_req) ||
3198                     atomic_read(&sctx->cancel_req)) {
3199                         ret = -ECANCELED;
3200                         goto out;
3201                 }
3202                 /*
3203                  * check to see if we have to pause
3204                  */
3205                 if (atomic_read(&fs_info->scrub_pause_req)) {
3206                         /* push queued extents */
3207                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3208                         scrub_submit(sctx);
3209                         mutex_lock(&sctx->wr_ctx.wr_lock);
3210                         scrub_wr_submit(sctx);
3211                         mutex_unlock(&sctx->wr_ctx.wr_lock);
3212                         wait_event(sctx->list_wait,
3213                                    atomic_read(&sctx->bios_in_flight) == 0);
3214                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3215                         scrub_blocked_if_needed(fs_info);
3216                 }
3217
3218                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3219                         ret = get_raid56_logic_offset(physical, num, map,
3220                                                       &logical,
3221                                                       &stripe_logical);
3222                         logical += base;
3223                         if (ret) {
3224                                 /* it is parity strip */
3225                                 stripe_logical += base;
3226                                 stripe_end = stripe_logical + increment;
3227                                 ret = scrub_raid56_parity(sctx, map, scrub_dev,
3228                                                           ppath, stripe_logical,
3229                                                           stripe_end);
3230                                 if (ret)
3231                                         goto out;
3232                                 goto skip;
3233                         }
3234                 }
3235
3236                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
3237                         key.type = BTRFS_METADATA_ITEM_KEY;
3238                 else
3239                         key.type = BTRFS_EXTENT_ITEM_KEY;
3240                 key.objectid = logical;
3241                 key.offset = (u64)-1;
3242
3243                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3244                 if (ret < 0)
3245                         goto out;
3246
3247                 if (ret > 0) {
3248                         ret = btrfs_previous_extent_item(root, path, 0);
3249                         if (ret < 0)
3250                                 goto out;
3251                         if (ret > 0) {
3252                                 /* there's no smaller item, so stick with the
3253                                  * larger one */
3254                                 btrfs_release_path(path);
3255                                 ret = btrfs_search_slot(NULL, root, &key,
3256                                                         path, 0, 0);
3257                                 if (ret < 0)
3258                                         goto out;
3259                         }
3260                 }
3261
3262                 stop_loop = 0;
3263                 while (1) {
3264                         u64 bytes;
3265
3266                         l = path->nodes[0];
3267                         slot = path->slots[0];
3268                         if (slot >= btrfs_header_nritems(l)) {
3269                                 ret = btrfs_next_leaf(root, path);
3270                                 if (ret == 0)
3271                                         continue;
3272                                 if (ret < 0)
3273                                         goto out;
3274
3275                                 stop_loop = 1;
3276                                 break;
3277                         }
3278                         btrfs_item_key_to_cpu(l, &key, slot);
3279
3280                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
3281                             key.type != BTRFS_METADATA_ITEM_KEY)
3282                                 goto next;
3283
3284                         if (key.type == BTRFS_METADATA_ITEM_KEY)
3285                                 bytes = root->nodesize;
3286                         else
3287                                 bytes = key.offset;
3288
3289                         if (key.objectid + bytes <= logical)
3290                                 goto next;
3291
3292                         if (key.objectid >= logical + map->stripe_len) {
3293                                 /* out of this device extent */
3294                                 if (key.objectid >= logic_end)
3295                                         stop_loop = 1;
3296                                 break;
3297                         }
3298
3299                         extent = btrfs_item_ptr(l, slot,
3300                                                 struct btrfs_extent_item);
3301                         flags = btrfs_extent_flags(l, extent);
3302                         generation = btrfs_extent_generation(l, extent);
3303
3304                         if ((flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) &&
3305                             (key.objectid < logical ||
3306                              key.objectid + bytes >
3307                              logical + map->stripe_len)) {
3308                                 btrfs_err(fs_info,
3309                                            "scrub: tree block %llu spanning "
3310                                            "stripes, ignored. logical=%llu",
3311                                        key.objectid, logical);
3312                                 spin_lock(&sctx->stat_lock);
3313                                 sctx->stat.uncorrectable_errors++;
3314                                 spin_unlock(&sctx->stat_lock);
3315                                 goto next;
3316                         }
3317
3318 again:
3319                         extent_logical = key.objectid;
3320                         extent_len = bytes;
3321
3322                         /*
3323                          * trim extent to this stripe
3324                          */
3325                         if (extent_logical < logical) {
3326                                 extent_len -= logical - extent_logical;
3327                                 extent_logical = logical;
3328                         }
3329                         if (extent_logical + extent_len >
3330                             logical + map->stripe_len) {
3331                                 extent_len = logical + map->stripe_len -
3332                                              extent_logical;
3333                         }
3334
3335                         extent_physical = extent_logical - logical + physical;
3336                         extent_dev = scrub_dev;
3337                         extent_mirror_num = mirror_num;
3338                         if (is_dev_replace)
3339                                 scrub_remap_extent(fs_info, extent_logical,
3340                                                    extent_len, &extent_physical,
3341                                                    &extent_dev,
3342                                                    &extent_mirror_num);
3343
3344                         ret = btrfs_lookup_csums_range(csum_root,
3345                                                        extent_logical,
3346                                                        extent_logical +
3347                                                        extent_len - 1,
3348                                                        &sctx->csum_list, 1);
3349                         if (ret)
3350                                 goto out;
3351
3352                         ret = scrub_extent(sctx, extent_logical, extent_len,
3353                                            extent_physical, extent_dev, flags,
3354                                            generation, extent_mirror_num,
3355                                            extent_logical - logical + physical);
3356
3357                         scrub_free_csums(sctx);
3358
3359                         if (ret)
3360                                 goto out;
3361
3362                         if (extent_logical + extent_len <
3363                             key.objectid + bytes) {
3364                                 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK) {
3365                                         /*
3366                                          * loop until we find next data stripe
3367                                          * or we have finished all stripes.
3368                                          */
3369 loop:
3370                                         physical += map->stripe_len;
3371                                         ret = get_raid56_logic_offset(physical,
3372                                                         num, map, &logical,
3373                                                         &stripe_logical);
3374                                         logical += base;
3375
3376                                         if (ret && physical < physical_end) {
3377                                                 stripe_logical += base;
3378                                                 stripe_end = stripe_logical +
3379                                                                 increment;
3380                                                 ret = scrub_raid56_parity(sctx,
3381                                                         map, scrub_dev, ppath,
3382                                                         stripe_logical,
3383                                                         stripe_end);
3384                                                 if (ret)
3385                                                         goto out;
3386                                                 goto loop;
3387                                         }
3388                                 } else {
3389                                         physical += map->stripe_len;
3390                                         logical += increment;
3391                                 }
3392                                 if (logical < key.objectid + bytes) {
3393                                         cond_resched();
3394                                         goto again;
3395                                 }
3396
3397                                 if (physical >= physical_end) {
3398                                         stop_loop = 1;
3399                                         break;
3400                                 }
3401                         }
3402 next:
3403                         path->slots[0]++;
3404                 }
3405                 btrfs_release_path(path);
3406 skip:
3407                 logical += increment;
3408                 physical += map->stripe_len;
3409                 spin_lock(&sctx->stat_lock);
3410                 if (stop_loop)
3411                         sctx->stat.last_physical = map->stripes[num].physical +
3412                                                    length;
3413                 else
3414                         sctx->stat.last_physical = physical;
3415                 spin_unlock(&sctx->stat_lock);
3416                 if (stop_loop)
3417                         break;
3418         }
3419 out:
3420         /* push queued extents */
3421         scrub_submit(sctx);
3422         mutex_lock(&sctx->wr_ctx.wr_lock);
3423         scrub_wr_submit(sctx);
3424         mutex_unlock(&sctx->wr_ctx.wr_lock);
3425
3426         blk_finish_plug(&plug);
3427         btrfs_free_path(path);
3428         btrfs_free_path(ppath);
3429         return ret < 0 ? ret : 0;
3430 }
3431
3432 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
3433                                           struct btrfs_device *scrub_dev,
3434                                           u64 chunk_offset, u64 length,
3435                                           u64 dev_offset, int is_dev_replace)
3436 {
3437         struct btrfs_mapping_tree *map_tree =
3438                 &sctx->dev_root->fs_info->mapping_tree;
3439         struct map_lookup *map;
3440         struct extent_map *em;
3441         int i;
3442         int ret = 0;
3443
3444         read_lock(&map_tree->map_tree.lock);
3445         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
3446         read_unlock(&map_tree->map_tree.lock);
3447
3448         if (!em)
3449                 return -EINVAL;
3450
3451         map = (struct map_lookup *)em->bdev;
3452         if (em->start != chunk_offset)
3453                 goto out;
3454
3455         if (em->len < length)
3456                 goto out;
3457
3458         for (i = 0; i < map->num_stripes; ++i) {
3459                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
3460                     map->stripes[i].physical == dev_offset) {
3461                         ret = scrub_stripe(sctx, map, scrub_dev, i,
3462                                            chunk_offset, length,
3463                                            is_dev_replace);
3464                         if (ret)
3465                                 goto out;
3466                 }
3467         }
3468 out:
3469         free_extent_map(em);
3470
3471         return ret;
3472 }
3473
3474 static noinline_for_stack
3475 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
3476                            struct btrfs_device *scrub_dev, u64 start, u64 end,
3477                            int is_dev_replace)
3478 {
3479         struct btrfs_dev_extent *dev_extent = NULL;
3480         struct btrfs_path *path;
3481         struct btrfs_root *root = sctx->dev_root;
3482         struct btrfs_fs_info *fs_info = root->fs_info;
3483         u64 length;
3484         u64 chunk_offset;
3485         int ret = 0;
3486         int slot;
3487         struct extent_buffer *l;
3488         struct btrfs_key key;
3489         struct btrfs_key found_key;
3490         struct btrfs_block_group_cache *cache;
3491         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
3492
3493         path = btrfs_alloc_path();
3494         if (!path)
3495                 return -ENOMEM;
3496
3497         path->reada = 2;
3498         path->search_commit_root = 1;
3499         path->skip_locking = 1;
3500
3501         key.objectid = scrub_dev->devid;
3502         key.offset = 0ull;
3503         key.type = BTRFS_DEV_EXTENT_KEY;
3504
3505         while (1) {
3506                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3507                 if (ret < 0)
3508                         break;
3509                 if (ret > 0) {
3510                         if (path->slots[0] >=
3511                             btrfs_header_nritems(path->nodes[0])) {
3512                                 ret = btrfs_next_leaf(root, path);
3513                                 if (ret < 0)
3514                                         break;
3515                                 if (ret > 0) {
3516                                         ret = 0;
3517                                         break;
3518                                 }
3519                         } else {
3520                                 ret = 0;
3521                         }
3522                 }
3523
3524                 l = path->nodes[0];
3525                 slot = path->slots[0];
3526
3527                 btrfs_item_key_to_cpu(l, &found_key, slot);
3528
3529                 if (found_key.objectid != scrub_dev->devid)
3530                         break;
3531
3532                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
3533                         break;
3534
3535                 if (found_key.offset >= end)
3536                         break;
3537
3538                 if (found_key.offset < key.offset)
3539                         break;
3540
3541                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
3542                 length = btrfs_dev_extent_length(l, dev_extent);
3543
3544                 if (found_key.offset + length <= start)
3545                         goto skip;
3546
3547                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
3548
3549                 /*
3550                  * get a reference on the corresponding block group to prevent
3551                  * the chunk from going away while we scrub it
3552                  */
3553                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
3554
3555                 /* some chunks are removed but not committed to disk yet,
3556                  * continue scrubbing */
3557                 if (!cache)
3558                         goto skip;
3559
3560                 /*
3561                  * we need call btrfs_inc_block_group_ro() with scrubs_paused,
3562                  * to avoid deadlock caused by:
3563                  * btrfs_inc_block_group_ro()
3564                  * -> btrfs_wait_for_commit()
3565                  * -> btrfs_commit_transaction()
3566                  * -> btrfs_scrub_pause()
3567                  */
3568                 scrub_pause_on(fs_info);
3569                 ret = btrfs_inc_block_group_ro(root, cache);
3570                 scrub_pause_off(fs_info);
3571                 if (ret) {
3572                         btrfs_put_block_group(cache);
3573                         break;
3574                 }
3575
3576                 dev_replace->cursor_right = found_key.offset + length;
3577                 dev_replace->cursor_left = found_key.offset;
3578                 dev_replace->item_needs_writeback = 1;
3579                 ret = scrub_chunk(sctx, scrub_dev, chunk_offset, length,
3580                                   found_key.offset, is_dev_replace);
3581
3582                 /*
3583                  * flush, submit all pending read and write bios, afterwards
3584                  * wait for them.
3585                  * Note that in the dev replace case, a read request causes
3586                  * write requests that are submitted in the read completion
3587                  * worker. Therefore in the current situation, it is required
3588                  * that all write requests are flushed, so that all read and
3589                  * write requests are really completed when bios_in_flight
3590                  * changes to 0.
3591                  */
3592                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
3593                 scrub_submit(sctx);
3594                 mutex_lock(&sctx->wr_ctx.wr_lock);
3595                 scrub_wr_submit(sctx);
3596                 mutex_unlock(&sctx->wr_ctx.wr_lock);
3597
3598                 wait_event(sctx->list_wait,
3599                            atomic_read(&sctx->bios_in_flight) == 0);
3600
3601                 scrub_pause_on(fs_info);
3602
3603                 /*
3604                  * must be called before we decrease @scrub_paused.
3605                  * make sure we don't block transaction commit while
3606                  * we are waiting pending workers finished.
3607                  */
3608                 wait_event(sctx->list_wait,
3609                            atomic_read(&sctx->workers_pending) == 0);
3610                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
3611
3612                 scrub_pause_off(fs_info);
3613
3614                 btrfs_dec_block_group_ro(root, cache);
3615
3616                 btrfs_put_block_group(cache);
3617                 if (ret)
3618                         break;
3619                 if (is_dev_replace &&
3620                     atomic64_read(&dev_replace->num_write_errors) > 0) {
3621                         ret = -EIO;
3622                         break;
3623                 }
3624                 if (sctx->stat.malloc_errors > 0) {
3625                         ret = -ENOMEM;
3626                         break;
3627                 }
3628
3629                 dev_replace->cursor_left = dev_replace->cursor_right;
3630                 dev_replace->item_needs_writeback = 1;
3631 skip:
3632                 key.offset = found_key.offset + length;
3633                 btrfs_release_path(path);
3634         }
3635
3636         btrfs_free_path(path);
3637
3638         return ret;
3639 }
3640
3641 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
3642                                            struct btrfs_device *scrub_dev)
3643 {
3644         int     i;
3645         u64     bytenr;
3646         u64     gen;
3647         int     ret;
3648         struct btrfs_root *root = sctx->dev_root;
3649
3650         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
3651                 return -EIO;
3652
3653         /* Seed devices of a new filesystem has their own generation. */
3654         if (scrub_dev->fs_devices != root->fs_info->fs_devices)
3655                 gen = scrub_dev->generation;
3656         else
3657                 gen = root->fs_info->last_trans_committed;
3658
3659         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
3660                 bytenr = btrfs_sb_offset(i);
3661                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
3662                     scrub_dev->commit_total_bytes)
3663                         break;
3664
3665                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
3666                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
3667                                   NULL, 1, bytenr);
3668                 if (ret)
3669                         return ret;
3670         }
3671         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3672
3673         return 0;
3674 }
3675
3676 /*
3677  * get a reference count on fs_info->scrub_workers. start worker if necessary
3678  */
3679 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
3680                                                 int is_dev_replace)
3681 {
3682         unsigned int flags = WQ_FREEZABLE | WQ_UNBOUND;
3683         int max_active = fs_info->thread_pool_size;
3684
3685         if (fs_info->scrub_workers_refcnt == 0) {
3686                 if (is_dev_replace)
3687                         fs_info->scrub_workers =
3688                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
3689                                                       1, 4);
3690                 else
3691                         fs_info->scrub_workers =
3692                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
3693                                                       max_active, 4);
3694                 if (!fs_info->scrub_workers)
3695                         goto fail_scrub_workers;
3696
3697                 fs_info->scrub_wr_completion_workers =
3698                         btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
3699                                               max_active, 2);
3700                 if (!fs_info->scrub_wr_completion_workers)
3701                         goto fail_scrub_wr_completion_workers;
3702
3703                 fs_info->scrub_nocow_workers =
3704                         btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
3705                 if (!fs_info->scrub_nocow_workers)
3706                         goto fail_scrub_nocow_workers;
3707                 fs_info->scrub_parity_workers =
3708                         btrfs_alloc_workqueue("btrfs-scrubparity", flags,
3709                                               max_active, 2);
3710                 if (!fs_info->scrub_parity_workers)
3711                         goto fail_scrub_parity_workers;
3712         }
3713         ++fs_info->scrub_workers_refcnt;
3714         return 0;
3715
3716 fail_scrub_parity_workers:
3717         btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3718 fail_scrub_nocow_workers:
3719         btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3720 fail_scrub_wr_completion_workers:
3721         btrfs_destroy_workqueue(fs_info->scrub_workers);
3722 fail_scrub_workers:
3723         return -ENOMEM;
3724 }
3725
3726 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
3727 {
3728         if (--fs_info->scrub_workers_refcnt == 0) {
3729                 btrfs_destroy_workqueue(fs_info->scrub_workers);
3730                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
3731                 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
3732                 btrfs_destroy_workqueue(fs_info->scrub_parity_workers);
3733         }
3734         WARN_ON(fs_info->scrub_workers_refcnt < 0);
3735 }
3736
3737 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
3738                     u64 end, struct btrfs_scrub_progress *progress,
3739                     int readonly, int is_dev_replace)
3740 {
3741         struct scrub_ctx *sctx;
3742         int ret;
3743         struct btrfs_device *dev;
3744         struct rcu_string *name;
3745
3746         if (btrfs_fs_closing(fs_info))
3747                 return -EINVAL;
3748
3749         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
3750                 /*
3751                  * in this case scrub is unable to calculate the checksum
3752                  * the way scrub is implemented. Do not handle this
3753                  * situation at all because it won't ever happen.
3754                  */
3755                 btrfs_err(fs_info,
3756                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
3757                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
3758                 return -EINVAL;
3759         }
3760
3761         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
3762                 /* not supported for data w/o checksums */
3763                 btrfs_err(fs_info,
3764                            "scrub: size assumption sectorsize != PAGE_SIZE "
3765                            "(%d != %lu) fails",
3766                        fs_info->chunk_root->sectorsize, PAGE_SIZE);
3767                 return -EINVAL;
3768         }
3769
3770         if (fs_info->chunk_root->nodesize >
3771             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
3772             fs_info->chunk_root->sectorsize >
3773             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
3774                 /*
3775                  * would exhaust the array bounds of pagev member in
3776                  * struct scrub_block
3777                  */
3778                 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
3779                            "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
3780                        fs_info->chunk_root->nodesize,
3781                        SCRUB_MAX_PAGES_PER_BLOCK,
3782                        fs_info->chunk_root->sectorsize,
3783                        SCRUB_MAX_PAGES_PER_BLOCK);
3784                 return -EINVAL;
3785         }
3786
3787
3788         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3789         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3790         if (!dev || (dev->missing && !is_dev_replace)) {
3791                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3792                 return -ENODEV;
3793         }
3794
3795         if (!is_dev_replace && !readonly && !dev->writeable) {
3796                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3797                 rcu_read_lock();
3798                 name = rcu_dereference(dev->name);
3799                 btrfs_err(fs_info, "scrub: device %s is not writable",
3800                           name->str);
3801                 rcu_read_unlock();
3802                 return -EROFS;
3803         }
3804
3805         mutex_lock(&fs_info->scrub_lock);
3806         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
3807                 mutex_unlock(&fs_info->scrub_lock);
3808                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3809                 return -EIO;
3810         }
3811
3812         btrfs_dev_replace_lock(&fs_info->dev_replace);
3813         if (dev->scrub_device ||
3814             (!is_dev_replace &&
3815              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
3816                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
3817                 mutex_unlock(&fs_info->scrub_lock);
3818                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3819                 return -EINPROGRESS;
3820         }
3821         btrfs_dev_replace_unlock(&fs_info->dev_replace);
3822
3823         ret = scrub_workers_get(fs_info, is_dev_replace);
3824         if (ret) {
3825                 mutex_unlock(&fs_info->scrub_lock);
3826                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3827                 return ret;
3828         }
3829
3830         sctx = scrub_setup_ctx(dev, is_dev_replace);
3831         if (IS_ERR(sctx)) {
3832                 mutex_unlock(&fs_info->scrub_lock);
3833                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3834                 scrub_workers_put(fs_info);
3835                 return PTR_ERR(sctx);
3836         }
3837         sctx->readonly = readonly;
3838         dev->scrub_device = sctx;
3839         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3840
3841         /*
3842          * checking @scrub_pause_req here, we can avoid
3843          * race between committing transaction and scrubbing.
3844          */
3845         __scrub_blocked_if_needed(fs_info);
3846         atomic_inc(&fs_info->scrubs_running);
3847         mutex_unlock(&fs_info->scrub_lock);
3848
3849         if (!is_dev_replace) {
3850                 /*
3851                  * by holding device list mutex, we can
3852                  * kick off writing super in log tree sync.
3853                  */
3854                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3855                 ret = scrub_supers(sctx, dev);
3856                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3857         }
3858
3859         if (!ret)
3860                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3861                                              is_dev_replace);
3862
3863         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3864         atomic_dec(&fs_info->scrubs_running);
3865         wake_up(&fs_info->scrub_pause_wait);
3866
3867         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3868
3869         if (progress)
3870                 memcpy(progress, &sctx->stat, sizeof(*progress));
3871
3872         mutex_lock(&fs_info->scrub_lock);
3873         dev->scrub_device = NULL;
3874         scrub_workers_put(fs_info);
3875         mutex_unlock(&fs_info->scrub_lock);
3876
3877         scrub_put_ctx(sctx);
3878
3879         return ret;
3880 }
3881
3882 void btrfs_scrub_pause(struct btrfs_root *root)
3883 {
3884         struct btrfs_fs_info *fs_info = root->fs_info;
3885
3886         mutex_lock(&fs_info->scrub_lock);
3887         atomic_inc(&fs_info->scrub_pause_req);
3888         while (atomic_read(&fs_info->scrubs_paused) !=
3889                atomic_read(&fs_info->scrubs_running)) {
3890                 mutex_unlock(&fs_info->scrub_lock);
3891                 wait_event(fs_info->scrub_pause_wait,
3892                            atomic_read(&fs_info->scrubs_paused) ==
3893                            atomic_read(&fs_info->scrubs_running));
3894                 mutex_lock(&fs_info->scrub_lock);
3895         }
3896         mutex_unlock(&fs_info->scrub_lock);
3897 }
3898
3899 void btrfs_scrub_continue(struct btrfs_root *root)
3900 {
3901         struct btrfs_fs_info *fs_info = root->fs_info;
3902
3903         atomic_dec(&fs_info->scrub_pause_req);
3904         wake_up(&fs_info->scrub_pause_wait);
3905 }
3906
3907 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3908 {
3909         mutex_lock(&fs_info->scrub_lock);
3910         if (!atomic_read(&fs_info->scrubs_running)) {
3911                 mutex_unlock(&fs_info->scrub_lock);
3912                 return -ENOTCONN;
3913         }
3914
3915         atomic_inc(&fs_info->scrub_cancel_req);
3916         while (atomic_read(&fs_info->scrubs_running)) {
3917                 mutex_unlock(&fs_info->scrub_lock);
3918                 wait_event(fs_info->scrub_pause_wait,
3919                            atomic_read(&fs_info->scrubs_running) == 0);
3920                 mutex_lock(&fs_info->scrub_lock);
3921         }
3922         atomic_dec(&fs_info->scrub_cancel_req);
3923         mutex_unlock(&fs_info->scrub_lock);
3924
3925         return 0;
3926 }
3927
3928 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3929                            struct btrfs_device *dev)
3930 {
3931         struct scrub_ctx *sctx;
3932
3933         mutex_lock(&fs_info->scrub_lock);
3934         sctx = dev->scrub_device;
3935         if (!sctx) {
3936                 mutex_unlock(&fs_info->scrub_lock);
3937                 return -ENOTCONN;
3938         }
3939         atomic_inc(&sctx->cancel_req);
3940         while (dev->scrub_device) {
3941                 mutex_unlock(&fs_info->scrub_lock);
3942                 wait_event(fs_info->scrub_pause_wait,
3943                            dev->scrub_device == NULL);
3944                 mutex_lock(&fs_info->scrub_lock);
3945         }
3946         mutex_unlock(&fs_info->scrub_lock);
3947
3948         return 0;
3949 }
3950
3951 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3952                          struct btrfs_scrub_progress *progress)
3953 {
3954         struct btrfs_device *dev;
3955         struct scrub_ctx *sctx = NULL;
3956
3957         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3958         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3959         if (dev)
3960                 sctx = dev->scrub_device;
3961         if (sctx)
3962                 memcpy(progress, &sctx->stat, sizeof(*progress));
3963         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3964
3965         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3966 }
3967
3968 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3969                                u64 extent_logical, u64 extent_len,
3970                                u64 *extent_physical,
3971                                struct btrfs_device **extent_dev,
3972                                int *extent_mirror_num)
3973 {
3974         u64 mapped_length;
3975         struct btrfs_bio *bbio = NULL;
3976         int ret;
3977
3978         mapped_length = extent_len;
3979         ret = btrfs_map_block(fs_info, READ, extent_logical,
3980                               &mapped_length, &bbio, 0);
3981         if (ret || !bbio || mapped_length < extent_len ||
3982             !bbio->stripes[0].dev->bdev) {
3983                 btrfs_put_bbio(bbio);
3984                 return;
3985         }
3986
3987         *extent_physical = bbio->stripes[0].physical;
3988         *extent_mirror_num = bbio->mirror_num;
3989         *extent_dev = bbio->stripes[0].dev;
3990         btrfs_put_bbio(bbio);
3991 }
3992
3993 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3994                               struct scrub_wr_ctx *wr_ctx,
3995                               struct btrfs_fs_info *fs_info,
3996                               struct btrfs_device *dev,
3997                               int is_dev_replace)
3998 {
3999         WARN_ON(wr_ctx->wr_curr_bio != NULL);
4000
4001         mutex_init(&wr_ctx->wr_lock);
4002         wr_ctx->wr_curr_bio = NULL;
4003         if (!is_dev_replace)
4004                 return 0;
4005
4006         WARN_ON(!dev->bdev);
4007         wr_ctx->pages_per_wr_bio = SCRUB_PAGES_PER_WR_BIO;
4008         wr_ctx->tgtdev = dev;
4009         atomic_set(&wr_ctx->flush_all_writes, 0);
4010         return 0;
4011 }
4012
4013 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
4014 {
4015         mutex_lock(&wr_ctx->wr_lock);
4016         kfree(wr_ctx->wr_curr_bio);
4017         wr_ctx->wr_curr_bio = NULL;
4018         mutex_unlock(&wr_ctx->wr_lock);
4019 }
4020
4021 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
4022                             int mirror_num, u64 physical_for_dev_replace)
4023 {
4024         struct scrub_copy_nocow_ctx *nocow_ctx;
4025         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
4026
4027         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
4028         if (!nocow_ctx) {
4029                 spin_lock(&sctx->stat_lock);
4030                 sctx->stat.malloc_errors++;
4031                 spin_unlock(&sctx->stat_lock);
4032                 return -ENOMEM;
4033         }
4034
4035         scrub_pending_trans_workers_inc(sctx);
4036
4037         nocow_ctx->sctx = sctx;
4038         nocow_ctx->logical = logical;
4039         nocow_ctx->len = len;
4040         nocow_ctx->mirror_num = mirror_num;
4041         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
4042         btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
4043                         copy_nocow_pages_worker, NULL, NULL);
4044         INIT_LIST_HEAD(&nocow_ctx->inodes);
4045         btrfs_queue_work(fs_info->scrub_nocow_workers,
4046                          &nocow_ctx->work);
4047
4048         return 0;
4049 }
4050
4051 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
4052 {
4053         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
4054         struct scrub_nocow_inode *nocow_inode;
4055
4056         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
4057         if (!nocow_inode)
4058                 return -ENOMEM;
4059         nocow_inode->inum = inum;
4060         nocow_inode->offset = offset;
4061         nocow_inode->root = root;
4062         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
4063         return 0;
4064 }
4065
4066 #define COPY_COMPLETE 1
4067
4068 static void copy_nocow_pages_worker(struct btrfs_work *work)
4069 {
4070         struct scrub_copy_nocow_ctx *nocow_ctx =
4071                 container_of(work, struct scrub_copy_nocow_ctx, work);
4072         struct scrub_ctx *sctx = nocow_ctx->sctx;
4073         u64 logical = nocow_ctx->logical;
4074         u64 len = nocow_ctx->len;
4075         int mirror_num = nocow_ctx->mirror_num;
4076         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4077         int ret;
4078         struct btrfs_trans_handle *trans = NULL;
4079         struct btrfs_fs_info *fs_info;
4080         struct btrfs_path *path;
4081         struct btrfs_root *root;
4082         int not_written = 0;
4083
4084         fs_info = sctx->dev_root->fs_info;
4085         root = fs_info->extent_root;
4086
4087         path = btrfs_alloc_path();
4088         if (!path) {
4089                 spin_lock(&sctx->stat_lock);
4090                 sctx->stat.malloc_errors++;
4091                 spin_unlock(&sctx->stat_lock);
4092                 not_written = 1;
4093                 goto out;
4094         }
4095
4096         trans = btrfs_join_transaction(root);
4097         if (IS_ERR(trans)) {
4098                 not_written = 1;
4099                 goto out;
4100         }
4101
4102         ret = iterate_inodes_from_logical(logical, fs_info, path,
4103                                           record_inode_for_nocow, nocow_ctx);
4104         if (ret != 0 && ret != -ENOENT) {
4105                 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
4106                         "phys %llu, len %llu, mir %u, ret %d",
4107                         logical, physical_for_dev_replace, len, mirror_num,
4108                         ret);
4109                 not_written = 1;
4110                 goto out;
4111         }
4112
4113         btrfs_end_transaction(trans, root);
4114         trans = NULL;
4115         while (!list_empty(&nocow_ctx->inodes)) {
4116                 struct scrub_nocow_inode *entry;
4117                 entry = list_first_entry(&nocow_ctx->inodes,
4118                                          struct scrub_nocow_inode,
4119                                          list);
4120                 list_del_init(&entry->list);
4121                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
4122                                                  entry->root, nocow_ctx);
4123                 kfree(entry);
4124                 if (ret == COPY_COMPLETE) {
4125                         ret = 0;
4126                         break;
4127                 } else if (ret) {
4128                         break;
4129                 }
4130         }
4131 out:
4132         while (!list_empty(&nocow_ctx->inodes)) {
4133                 struct scrub_nocow_inode *entry;
4134                 entry = list_first_entry(&nocow_ctx->inodes,
4135                                          struct scrub_nocow_inode,
4136                                          list);
4137                 list_del_init(&entry->list);
4138                 kfree(entry);
4139         }
4140         if (trans && !IS_ERR(trans))
4141                 btrfs_end_transaction(trans, root);
4142         if (not_written)
4143                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
4144                                             num_uncorrectable_read_errors);
4145
4146         btrfs_free_path(path);
4147         kfree(nocow_ctx);
4148
4149         scrub_pending_trans_workers_dec(sctx);
4150 }
4151
4152 static int check_extent_to_block(struct inode *inode, u64 start, u64 len,
4153                                  u64 logical)
4154 {
4155         struct extent_state *cached_state = NULL;
4156         struct btrfs_ordered_extent *ordered;
4157         struct extent_io_tree *io_tree;
4158         struct extent_map *em;
4159         u64 lockstart = start, lockend = start + len - 1;
4160         int ret = 0;
4161
4162         io_tree = &BTRFS_I(inode)->io_tree;
4163
4164         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
4165         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
4166         if (ordered) {
4167                 btrfs_put_ordered_extent(ordered);
4168                 ret = 1;
4169                 goto out_unlock;
4170         }
4171
4172         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
4173         if (IS_ERR(em)) {
4174                 ret = PTR_ERR(em);
4175                 goto out_unlock;
4176         }
4177
4178         /*
4179          * This extent does not actually cover the logical extent anymore,
4180          * move on to the next inode.
4181          */
4182         if (em->block_start > logical ||
4183             em->block_start + em->block_len < logical + len) {
4184                 free_extent_map(em);
4185                 ret = 1;
4186                 goto out_unlock;
4187         }
4188         free_extent_map(em);
4189
4190 out_unlock:
4191         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
4192                              GFP_NOFS);
4193         return ret;
4194 }
4195
4196 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
4197                                       struct scrub_copy_nocow_ctx *nocow_ctx)
4198 {
4199         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
4200         struct btrfs_key key;
4201         struct inode *inode;
4202         struct page *page;
4203         struct btrfs_root *local_root;
4204         struct extent_io_tree *io_tree;
4205         u64 physical_for_dev_replace;
4206         u64 nocow_ctx_logical;
4207         u64 len = nocow_ctx->len;
4208         unsigned long index;
4209         int srcu_index;
4210         int ret = 0;
4211         int err = 0;
4212
4213         key.objectid = root;
4214         key.type = BTRFS_ROOT_ITEM_KEY;
4215         key.offset = (u64)-1;
4216
4217         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
4218
4219         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
4220         if (IS_ERR(local_root)) {
4221                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4222                 return PTR_ERR(local_root);
4223         }
4224
4225         key.type = BTRFS_INODE_ITEM_KEY;
4226         key.objectid = inum;
4227         key.offset = 0;
4228         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
4229         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
4230         if (IS_ERR(inode))
4231                 return PTR_ERR(inode);
4232
4233         /* Avoid truncate/dio/punch hole.. */
4234         mutex_lock(&inode->i_mutex);
4235         inode_dio_wait(inode);
4236
4237         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
4238         io_tree = &BTRFS_I(inode)->io_tree;
4239         nocow_ctx_logical = nocow_ctx->logical;
4240
4241         ret = check_extent_to_block(inode, offset, len, nocow_ctx_logical);
4242         if (ret) {
4243                 ret = ret > 0 ? 0 : ret;
4244                 goto out;
4245         }
4246
4247         while (len >= PAGE_CACHE_SIZE) {
4248                 index = offset >> PAGE_CACHE_SHIFT;
4249 again:
4250                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
4251                 if (!page) {
4252                         btrfs_err(fs_info, "find_or_create_page() failed");
4253                         ret = -ENOMEM;
4254                         goto out;
4255                 }
4256
4257                 if (PageUptodate(page)) {
4258                         if (PageDirty(page))
4259                                 goto next_page;
4260                 } else {
4261                         ClearPageError(page);
4262                         err = extent_read_full_page(io_tree, page,
4263                                                            btrfs_get_extent,
4264                                                            nocow_ctx->mirror_num);
4265                         if (err) {
4266                                 ret = err;
4267                                 goto next_page;
4268                         }
4269
4270                         lock_page(page);
4271                         /*
4272                          * If the page has been remove from the page cache,
4273                          * the data on it is meaningless, because it may be
4274                          * old one, the new data may be written into the new
4275                          * page in the page cache.
4276                          */
4277                         if (page->mapping != inode->i_mapping) {
4278                                 unlock_page(page);
4279                                 page_cache_release(page);
4280                                 goto again;
4281                         }
4282                         if (!PageUptodate(page)) {
4283                                 ret = -EIO;
4284                                 goto next_page;
4285                         }
4286                 }
4287
4288                 ret = check_extent_to_block(inode, offset, len,
4289                                             nocow_ctx_logical);
4290                 if (ret) {
4291                         ret = ret > 0 ? 0 : ret;
4292                         goto next_page;
4293                 }
4294
4295                 err = write_page_nocow(nocow_ctx->sctx,
4296                                        physical_for_dev_replace, page);
4297                 if (err)
4298                         ret = err;
4299 next_page:
4300                 unlock_page(page);
4301                 page_cache_release(page);
4302
4303                 if (ret)
4304                         break;
4305
4306                 offset += PAGE_CACHE_SIZE;
4307                 physical_for_dev_replace += PAGE_CACHE_SIZE;
4308                 nocow_ctx_logical += PAGE_CACHE_SIZE;
4309                 len -= PAGE_CACHE_SIZE;
4310         }
4311         ret = COPY_COMPLETE;
4312 out:
4313         mutex_unlock(&inode->i_mutex);
4314         iput(inode);
4315         return ret;
4316 }
4317
4318 static int write_page_nocow(struct scrub_ctx *sctx,
4319                             u64 physical_for_dev_replace, struct page *page)
4320 {
4321         struct bio *bio;
4322         struct btrfs_device *dev;
4323         int ret;
4324
4325         dev = sctx->wr_ctx.tgtdev;
4326         if (!dev)
4327                 return -EIO;
4328         if (!dev->bdev) {
4329                 btrfs_warn_rl(dev->dev_root->fs_info,
4330                         "scrub write_page_nocow(bdev == NULL) is unexpected");
4331                 return -EIO;
4332         }
4333         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
4334         if (!bio) {
4335                 spin_lock(&sctx->stat_lock);
4336                 sctx->stat.malloc_errors++;
4337                 spin_unlock(&sctx->stat_lock);
4338                 return -ENOMEM;
4339         }
4340         bio->bi_iter.bi_size = 0;
4341         bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
4342         bio->bi_bdev = dev->bdev;
4343         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
4344         if (ret != PAGE_CACHE_SIZE) {
4345 leave_with_eio:
4346                 bio_put(bio);
4347                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
4348                 return -EIO;
4349         }
4350
4351         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
4352                 goto leave_with_eio;
4353
4354         bio_put(bio);
4355         return 0;
4356 }