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_page {
  67         struct scrub_block      *sblock;
  68         struct page             *page;
  69         struct btrfs_device     *dev;
  70         u64                     flags;  /* extent flags */
  71         u64                     generation;
  72         u64                     logical;
  73         u64                     physical;
  74         u64                     physical_for_dev_replace;
  75         atomic_t                ref_count;
  76         struct {
  77                 unsigned int    mirror_num:8;
  78                 unsigned int    have_csum:1;
  79                 unsigned int    io_error:1;
  80         };
  81         u8                      csum[BTRFS_CSUM_SIZE];
  82 };
  83
  84 struct scrub_bio {
  85         int                     index;
  86         struct scrub_ctx        *sctx;
  87         struct btrfs_device     *dev;
  88         struct bio              *bio;
  89         int                     err;
  90         u64                     logical;
  91         u64                     physical;
  92 #if SCRUB_PAGES_PER_WR_BIO >= SCRUB_PAGES_PER_RD_BIO
  93         struct scrub_page       *pagev[SCRUB_PAGES_PER_WR_BIO];
  94 #else
  95         struct scrub_page       *pagev[SCRUB_PAGES_PER_RD_BIO];
  96 #endif
  97         int                     page_count;
  98         int                     next_free;
  99         struct btrfs_work       work;
 100 };
 101
 102 struct scrub_block {
 103         struct scrub_page       *pagev[SCRUB_MAX_PAGES_PER_BLOCK];
 104         int                     page_count;
 105         atomic_t                outstanding_pages;
 106         atomic_t                ref_count; /* free mem on transition to zero */
 107         struct scrub_ctx        *sctx;
 108         struct {
 109                 unsigned int    header_error:1;
 110                 unsigned int    checksum_error:1;
 111                 unsigned int    no_io_error_seen:1;
 112                 unsigned int    generation_error:1; /* also sets header_error */
 113         };
 114 };
 115
 116 struct scrub_wr_ctx {
 117         struct scrub_bio *wr_curr_bio;
 118         struct btrfs_device *tgtdev;
 119         int pages_per_wr_bio; /* <= SCRUB_PAGES_PER_WR_BIO */
 120         atomic_t flush_all_writes;
 121         struct mutex wr_lock;
 122 };
 123
 124 struct scrub_ctx {
 125         struct scrub_bio        *bios[SCRUB_BIOS_PER_SCTX];
 126         struct btrfs_root       *dev_root;
 127         int                     first_free;
 128         int                     curr;
 129         atomic_t                bios_in_flight;
 130         atomic_t                workers_pending;
 131         spinlock_t              list_lock;
 132         wait_queue_head_t       list_wait;
 133         u16                     csum_size;
 134         struct list_head        csum_list;
 135         atomic_t                cancel_req;
 136         int                     readonly;
 137         int                     pages_per_rd_bio;
 138         u32                     sectorsize;
 139         u32                     nodesize;
 140         u32                     leafsize;
 141
 142         int                     is_dev_replace;
 143         struct scrub_wr_ctx     wr_ctx;
 144
 145         /*
 146          * statistics
 147          */
 148         struct btrfs_scrub_progress stat;
 149         spinlock_t              stat_lock;
 150 };
 151
 152 struct scrub_fixup_nodatasum {
 153         struct scrub_ctx        *sctx;
 154         struct btrfs_device     *dev;
 155         u64                     logical;
 156         struct btrfs_root       *root;
 157         struct btrfs_work       work;
 158         int                     mirror_num;
 159 };
 160
 161 struct scrub_copy_nocow_ctx {
 162         struct scrub_ctx        *sctx;
 163         u64                     logical;
 164         u64                     len;
 165         int                     mirror_num;
 166         u64                     physical_for_dev_replace;
 167         struct btrfs_work       work;
 168 };
 169
 170 struct scrub_warning {
 171         struct btrfs_path       *path;
 172         u64                     extent_item_size;
 173         char                    *scratch_buf;
 174         char                    *msg_buf;
 175         const char              *errstr;
 176         sector_t                sector;
 177         u64                     logical;
 178         struct btrfs_device     *dev;
 179         int                     msg_bufsize;
 180         int                     scratch_bufsize;
 181 };
 182
 183
 184 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 185 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 186 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 187 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 188 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 189 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
 190                                      struct btrfs_fs_info *fs_info,
 191                                      struct scrub_block *original_sblock,
 192                                      u64 length, u64 logical,
 193                                      struct scrub_block *sblocks_for_recheck);
 194 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 195                                 struct scrub_block *sblock, int is_metadata,
 196                                 int have_csum, u8 *csum, u64 generation,
 197                                 u16 csum_size);
 198 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 199                                          struct scrub_block *sblock,
 200                                          int is_metadata, int have_csum,
 201                                          const u8 *csum, u64 generation,
 202                                          u16 csum_size);
 203 static void scrub_complete_bio_end_io(struct bio *bio, int err);
 204 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 205                                              struct scrub_block *sblock_good,
 206                                              int force_write);
 207 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 208                                             struct scrub_block *sblock_good,
 209                                             int page_num, int force_write);
 210 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 211 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 212                                            int page_num);
 213 static int scrub_checksum_data(struct scrub_block *sblock);
 214 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 215 static int scrub_checksum_super(struct scrub_block *sblock);
 216 static void scrub_block_get(struct scrub_block *sblock);
 217 static void scrub_block_put(struct scrub_block *sblock);
 218 static void scrub_page_get(struct scrub_page *spage);
 219 static void scrub_page_put(struct scrub_page *spage);
 220 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 221                                     struct scrub_page *spage);
 222 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 223                        u64 physical, struct btrfs_device *dev, u64 flags,
 224                        u64 gen, int mirror_num, u8 *csum, int force,
 225                        u64 physical_for_dev_replace);
 226 static void scrub_bio_end_io(struct bio *bio, int err);
 227 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 228 static void scrub_block_complete(struct scrub_block *sblock);
 229 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 230                                u64 extent_logical, u64 extent_len,
 231                                u64 *extent_physical,
 232                                struct btrfs_device **extent_dev,
 233                                int *extent_mirror_num);
 234 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
 235                               struct scrub_wr_ctx *wr_ctx,
 236                               struct btrfs_fs_info *fs_info,
 237                               struct btrfs_device *dev,
 238                               int is_dev_replace);
 239 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
 240 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 241                                     struct scrub_page *spage);
 242 static void scrub_wr_submit(struct scrub_ctx *sctx);
 243 static void scrub_wr_bio_end_io(struct bio *bio, int err);
 244 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 245 static int write_page_nocow(struct scrub_ctx *sctx,
 246                             u64 physical_for_dev_replace, struct page *page);
 247 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 248                                       void *ctx);
 249 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 250                             int mirror_num, u64 physical_for_dev_replace);
 251 static void copy_nocow_pages_worker(struct btrfs_work *work);
 252
 253
 254 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 255 {
 256         atomic_inc(&sctx->bios_in_flight);
 257 }
 258
 259 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 260 {
 261         atomic_dec(&sctx->bios_in_flight);
 262         wake_up(&sctx->list_wait);
 263 }
 264
 265 /*
 266  * used for workers that require transaction commits (i.e., for the
 267  * NOCOW case)
 268  */
 269 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 270 {
 271         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 272
 273         /*
 274          * increment scrubs_running to prevent cancel requests from
 275          * completing as long as a worker is running. we must also
 276          * increment scrubs_paused to prevent deadlocking on pause
 277          * requests used for transactions commits (as the worker uses a
 278          * transaction context). it is safe to regard the worker
 279          * as paused for all matters practical. effectively, we only
 280          * avoid cancellation requests from completing.
 281          */
 282         mutex_lock(&fs_info->scrub_lock);
 283         atomic_inc(&fs_info->scrubs_running);
 284         atomic_inc(&fs_info->scrubs_paused);
 285         mutex_unlock(&fs_info->scrub_lock);
 286         atomic_inc(&sctx->workers_pending);
 287 }
 288
 289 /* used for workers that require transaction commits */
 290 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 291 {
 292         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 293
 294         /*
 295          * see scrub_pending_trans_workers_inc() why we're pretending
 296          * to be paused in the scrub counters
 297          */
 298         mutex_lock(&fs_info->scrub_lock);
 299         atomic_dec(&fs_info->scrubs_running);
 300         atomic_dec(&fs_info->scrubs_paused);
 301         mutex_unlock(&fs_info->scrub_lock);
 302         atomic_dec(&sctx->workers_pending);
 303         wake_up(&fs_info->scrub_pause_wait);
 304         wake_up(&sctx->list_wait);
 305 }
 306
 307 static void scrub_free_csums(struct scrub_ctx *sctx)
 308 {
 309         while (!list_empty(&sctx->csum_list)) {
 310                 struct btrfs_ordered_sum *sum;
 311                 sum = list_first_entry(&sctx->csum_list,
 312                                        struct btrfs_ordered_sum, list);
 313                 list_del(&sum->list);
 314                 kfree(sum);
 315         }
 316 }
 317
 318 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 319 {
 320         int i;
 321
 322         if (!sctx)
 323                 return;
 324
 325         scrub_free_wr_ctx(&sctx->wr_ctx);
 326
 327         /* this can happen when scrub is cancelled */
 328         if (sctx->curr != -1) {
 329                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 330
 331                 for (i = 0; i < sbio->page_count; i++) {
 332                         WARN_ON(!sbio->pagev[i]->page);
 333                         scrub_block_put(sbio->pagev[i]->sblock);
 334                 }
 335                 bio_put(sbio->bio);
 336         }
 337
 338         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 339                 struct scrub_bio *sbio = sctx->bios[i];
 340
 341                 if (!sbio)
 342                         break;
 343                 kfree(sbio);
 344         }
 345
 346         scrub_free_csums(sctx);
 347         kfree(sctx);
 348 }
 349
 350 static noinline_for_stack
 351 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 352 {
 353         struct scrub_ctx *sctx;
 354         int             i;
 355         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 356         int pages_per_rd_bio;
 357         int ret;
 358
 359         /*
 360          * the setting of pages_per_rd_bio is correct for scrub but might
 361          * be wrong for the dev_replace code where we might read from
 362          * different devices in the initial huge bios. However, that
 363          * code is able to correctly handle the case when adding a page
 364          * to a bio fails.
 365          */
 366         if (dev->bdev)
 367                 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
 368                                          bio_get_nr_vecs(dev->bdev));
 369         else
 370                 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 371         sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
 372         if (!sctx)
 373                 goto nomem;
 374         sctx->is_dev_replace = is_dev_replace;
 375         sctx->pages_per_rd_bio = pages_per_rd_bio;
 376         sctx->curr = -1;
 377         sctx->dev_root = dev->dev_root;
 378         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 379                 struct scrub_bio *sbio;
 380
 381                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 382                 if (!sbio)
 383                         goto nomem;
 384                 sctx->bios[i] = sbio;
 385
 386                 sbio->index = i;
 387                 sbio->sctx = sctx;
 388                 sbio->page_count = 0;
 389                 sbio->work.func = scrub_bio_end_io_worker;
 390
 391                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 392                         sctx->bios[i]->next_free = i + 1;
 393                 else
 394                         sctx->bios[i]->next_free = -1;
 395         }
 396         sctx->first_free = 0;
 397         sctx->nodesize = dev->dev_root->nodesize;
 398         sctx->leafsize = dev->dev_root->leafsize;
 399         sctx->sectorsize = dev->dev_root->sectorsize;
 400         atomic_set(&sctx->bios_in_flight, 0);
 401         atomic_set(&sctx->workers_pending, 0);
 402         atomic_set(&sctx->cancel_req, 0);
 403         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 404         INIT_LIST_HEAD(&sctx->csum_list);
 405
 406         spin_lock_init(&sctx->list_lock);
 407         spin_lock_init(&sctx->stat_lock);
 408         init_waitqueue_head(&sctx->list_wait);
 409
 410         ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 411                                  fs_info->dev_replace.tgtdev, is_dev_replace);
 412         if (ret) {
 413                 scrub_free_ctx(sctx);
 414                 return ERR_PTR(ret);
 415         }
 416         return sctx;
 417
 418 nomem:
 419         scrub_free_ctx(sctx);
 420         return ERR_PTR(-ENOMEM);
 421 }
 422
 423 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 424                                      void *warn_ctx)
 425 {
 426         u64 isize;
 427         u32 nlink;
 428         int ret;
 429         int i;
 430         struct extent_buffer *eb;
 431         struct btrfs_inode_item *inode_item;
 432         struct scrub_warning *swarn = warn_ctx;
 433         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 434         struct inode_fs_paths *ipath = NULL;
 435         struct btrfs_root *local_root;
 436         struct btrfs_key root_key;
 437
 438         root_key.objectid = root;
 439         root_key.type = BTRFS_ROOT_ITEM_KEY;
 440         root_key.offset = (u64)-1;
 441         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 442         if (IS_ERR(local_root)) {
 443                 ret = PTR_ERR(local_root);
 444                 goto err;
 445         }
 446
 447         ret = inode_item_info(inum, 0, local_root, swarn->path);
 448         if (ret) {
 449                 btrfs_release_path(swarn->path);
 450                 goto err;
 451         }
 452
 453         eb = swarn->path->nodes[0];
 454         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 455                                         struct btrfs_inode_item);
 456         isize = btrfs_inode_size(eb, inode_item);
 457         nlink = btrfs_inode_nlink(eb, inode_item);
 458         btrfs_release_path(swarn->path);
 459
 460         ipath = init_ipath(4096, local_root, swarn->path);
 461         if (IS_ERR(ipath)) {
 462                 ret = PTR_ERR(ipath);
 463                 ipath = NULL;
 464                 goto err;
 465         }
 466         ret = paths_from_inode(inum, ipath);
 467
 468         if (ret < 0)
 469                 goto err;
 470
 471         /*
 472          * we deliberately ignore the bit ipath might have been too small to
 473          * hold all of the paths here
 474          */
 475         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 476                 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 477                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 478                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 479                         swarn->logical, rcu_str_deref(swarn->dev->name),
 480                         (unsigned long long)swarn->sector, root, inum, offset,
 481                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 482                         (char *)(unsigned long)ipath->fspath->val[i]);
 483
 484         free_ipath(ipath);
 485         return 0;
 486
 487 err:
 488         printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 489                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 490                 "resolving failed with ret=%d\n", swarn->errstr,
 491                 swarn->logical, rcu_str_deref(swarn->dev->name),
 492                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 493
 494         free_ipath(ipath);
 495         return 0;
 496 }
 497
 498 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 499 {
 500         struct btrfs_device *dev;
 501         struct btrfs_fs_info *fs_info;
 502         struct btrfs_path *path;
 503         struct btrfs_key found_key;
 504         struct extent_buffer *eb;
 505         struct btrfs_extent_item *ei;
 506         struct scrub_warning swarn;
 507         unsigned long ptr = 0;
 508         u64 extent_item_pos;
 509         u64 flags = 0;
 510         u64 ref_root;
 511         u32 item_size;
 512         u8 ref_level;
 513         const int bufsize = 4096;
 514         int ret;
 515
 516         WARN_ON(sblock->page_count < 1);
 517         dev = sblock->pagev[0]->dev;
 518         fs_info = sblock->sctx->dev_root->fs_info;
 519
 520         path = btrfs_alloc_path();
 521
 522         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 523         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 524         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 525         swarn.logical = sblock->pagev[0]->logical;
 526         swarn.errstr = errstr;
 527         swarn.dev = NULL;
 528         swarn.msg_bufsize = bufsize;
 529         swarn.scratch_bufsize = bufsize;
 530
 531         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 532                 goto out;
 533
 534         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 535                                   &flags);
 536         if (ret < 0)
 537                 goto out;
 538
 539         extent_item_pos = swarn.logical - found_key.objectid;
 540         swarn.extent_item_size = found_key.offset;
 541
 542         eb = path->nodes[0];
 543         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 544         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 545
 546         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 547                 do {
 548                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 549                                                         &ref_root, &ref_level);
 550                         printk_in_rcu(KERN_WARNING
 551                                 "btrfs: %s at logical %llu on dev %s, "
 552                                 "sector %llu: metadata %s (level %d) in tree "
 553                                 "%llu\n", errstr, swarn.logical,
 554                                 rcu_str_deref(dev->name),
 555                                 (unsigned long long)swarn.sector,
 556                                 ref_level ? "node" : "leaf",
 557                                 ret < 0 ? -1 : ref_level,
 558                                 ret < 0 ? -1 : ref_root);
 559                 } while (ret != 1);
 560                 btrfs_release_path(path);
 561         } else {
 562                 btrfs_release_path(path);
 563                 swarn.path = path;
 564                 swarn.dev = dev;
 565                 iterate_extent_inodes(fs_info, found_key.objectid,
 566                                         extent_item_pos, 1,
 567                                         scrub_print_warning_inode, &swarn);
 568         }
 569
 570 out:
 571         btrfs_free_path(path);
 572         kfree(swarn.scratch_buf);
 573         kfree(swarn.msg_buf);
 574 }
 575
 576 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 577 {
 578         struct page *page = NULL;
 579         unsigned long index;
 580         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 581         int ret;
 582         int corrected = 0;
 583         struct btrfs_key key;
 584         struct inode *inode = NULL;
 585         struct btrfs_fs_info *fs_info;
 586         u64 end = offset + PAGE_SIZE - 1;
 587         struct btrfs_root *local_root;
 588         int srcu_index;
 589
 590         key.objectid = root;
 591         key.type = BTRFS_ROOT_ITEM_KEY;
 592         key.offset = (u64)-1;
 593
 594         fs_info = fixup->root->fs_info;
 595         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 596
 597         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 598         if (IS_ERR(local_root)) {
 599                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 600                 return PTR_ERR(local_root);
 601         }
 602
 603         key.type = BTRFS_INODE_ITEM_KEY;
 604         key.objectid = inum;
 605         key.offset = 0;
 606         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 607         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 608         if (IS_ERR(inode))
 609                 return PTR_ERR(inode);
 610
 611         index = offset >> PAGE_CACHE_SHIFT;
 612
 613         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 614         if (!page) {
 615                 ret = -ENOMEM;
 616                 goto out;
 617         }
 618
 619         if (PageUptodate(page)) {
 620                 if (PageDirty(page)) {
 621                         /*
 622                          * we need to write the data to the defect sector. the
 623                          * data that was in that sector is not in memory,
 624                          * because the page was modified. we must not write the
 625                          * modified page to that sector.
 626                          *
 627                          * TODO: what could be done here: wait for the delalloc
 628                          *       runner to write out that page (might involve
 629                          *       COW) and see whether the sector is still
 630                          *       referenced afterwards.
 631                          *
 632                          * For the meantime, we'll treat this error
 633                          * incorrectable, although there is a chance that a
 634                          * later scrub will find the bad sector again and that
 635                          * there's no dirty page in memory, then.
 636                          */
 637                         ret = -EIO;
 638                         goto out;
 639                 }
 640                 fs_info = BTRFS_I(inode)->root->fs_info;
 641                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 642                                         fixup->logical, page,
 643                                         fixup->mirror_num);
 644                 unlock_page(page);
 645                 corrected = !ret;
 646         } else {
 647                 /*
 648                  * we need to get good data first. the general readpage path
 649                  * will call repair_io_failure for us, we just have to make
 650                  * sure we read the bad mirror.
 651                  */
 652                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 653                                         EXTENT_DAMAGED, GFP_NOFS);
 654                 if (ret) {
 655                         /* set_extent_bits should give proper error */
 656                         WARN_ON(ret > 0);
 657                         if (ret > 0)
 658                                 ret = -EFAULT;
 659                         goto out;
 660                 }
 661
 662                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 663                                                 btrfs_get_extent,
 664                                                 fixup->mirror_num);
 665                 wait_on_page_locked(page);
 666
 667                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 668                                                 end, EXTENT_DAMAGED, 0, NULL);
 669                 if (!corrected)
 670                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 671                                                 EXTENT_DAMAGED, GFP_NOFS);
 672         }
 673
 674 out:
 675         if (page)
 676                 put_page(page);
 677         if (inode)
 678                 iput(inode);
 679
 680         if (ret < 0)
 681                 return ret;
 682
 683         if (ret == 0 && corrected) {
 684                 /*
 685                  * we only need to call readpage for one of the inodes belonging
 686                  * to this extent. so make iterate_extent_inodes stop
 687                  */
 688                 return 1;
 689         }
 690
 691         return -EIO;
 692 }
 693
 694 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 695 {
 696         int ret;
 697         struct scrub_fixup_nodatasum *fixup;
 698         struct scrub_ctx *sctx;
 699         struct btrfs_trans_handle *trans = NULL;
 700         struct btrfs_fs_info *fs_info;
 701         struct btrfs_path *path;
 702         int uncorrectable = 0;
 703
 704         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 705         sctx = fixup->sctx;
 706         fs_info = fixup->root->fs_info;
 707
 708         path = btrfs_alloc_path();
 709         if (!path) {
 710                 spin_lock(&sctx->stat_lock);
 711                 ++sctx->stat.malloc_errors;
 712                 spin_unlock(&sctx->stat_lock);
 713                 uncorrectable = 1;
 714                 goto out;
 715         }
 716
 717         trans = btrfs_join_transaction(fixup->root);
 718         if (IS_ERR(trans)) {
 719                 uncorrectable = 1;
 720                 goto out;
 721         }
 722
 723         /*
 724          * the idea is to trigger a regular read through the standard path. we
 725          * read a page from the (failed) logical address by specifying the
 726          * corresponding copynum of the failed sector. thus, that readpage is
 727          * expected to fail.
 728          * that is the point where on-the-fly error correction will kick in
 729          * (once it's finished) and rewrite the failed sector if a good copy
 730          * can be found.
 731          */
 732         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 733                                                 path, scrub_fixup_readpage,
 734                                                 fixup);
 735         if (ret < 0) {
 736                 uncorrectable = 1;
 737                 goto out;
 738         }
 739         WARN_ON(ret != 1);
 740
 741         spin_lock(&sctx->stat_lock);
 742         ++sctx->stat.corrected_errors;
 743         spin_unlock(&sctx->stat_lock);
 744
 745 out:
 746         if (trans && !IS_ERR(trans))
 747                 btrfs_end_transaction(trans, fixup->root);
 748         if (uncorrectable) {
 749                 spin_lock(&sctx->stat_lock);
 750                 ++sctx->stat.uncorrectable_errors;
 751                 spin_unlock(&sctx->stat_lock);
 752                 btrfs_dev_replace_stats_inc(
 753                         &sctx->dev_root->fs_info->dev_replace.
 754                         num_uncorrectable_read_errors);
 755                 printk_ratelimited_in_rcu(KERN_ERR
 756                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 757                         (unsigned long long)fixup->logical,
 758                         rcu_str_deref(fixup->dev->name));
 759         }
 760
 761         btrfs_free_path(path);
 762         kfree(fixup);
 763
 764         scrub_pending_trans_workers_dec(sctx);
 765 }
 766
 767 /*
 768  * scrub_handle_errored_block gets called when either verification of the
 769  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 770  * case, this function handles all pages in the bio, even though only one
 771  * may be bad.
 772  * The goal of this function is to repair the errored block by using the
 773  * contents of one of the mirrors.
 774  */
 775 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 776 {
 777         struct scrub_ctx *sctx = sblock_to_check->sctx;
 778         struct btrfs_device *dev;
 779         struct btrfs_fs_info *fs_info;
 780         u64 length;
 781         u64 logical;
 782         u64 generation;
 783         unsigned int failed_mirror_index;
 784         unsigned int is_metadata;
 785         unsigned int have_csum;
 786         u8 *csum;
 787         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 788         struct scrub_block *sblock_bad;
 789         int ret;
 790         int mirror_index;
 791         int page_num;
 792         int success;
 793         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 794                                       DEFAULT_RATELIMIT_BURST);
 795
 796         BUG_ON(sblock_to_check->page_count < 1);
 797         fs_info = sctx->dev_root->fs_info;
 798         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 799                 /*
 800                  * if we find an error in a super block, we just report it.
 801                  * They will get written with the next transaction commit
 802                  * anyway
 803                  */
 804                 spin_lock(&sctx->stat_lock);
 805                 ++sctx->stat.super_errors;
 806                 spin_unlock(&sctx->stat_lock);
 807                 return 0;
 808         }
 809         length = sblock_to_check->page_count * PAGE_SIZE;
 810         logical = sblock_to_check->pagev[0]->logical;
 811         generation = sblock_to_check->pagev[0]->generation;
 812         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 813         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 814         is_metadata = !(sblock_to_check->pagev[0]->flags &
 815                         BTRFS_EXTENT_FLAG_DATA);
 816         have_csum = sblock_to_check->pagev[0]->have_csum;
 817         csum = sblock_to_check->pagev[0]->csum;
 818         dev = sblock_to_check->pagev[0]->dev;
 819
 820         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 821                 sblocks_for_recheck = NULL;
 822                 goto nodatasum_case;
 823         }
 824
 825         /*
 826          * read all mirrors one after the other. This includes to
 827          * re-read the extent or metadata block that failed (that was
 828          * the cause that this fixup code is called) another time,
 829          * page by page this time in order to know which pages
 830          * caused I/O errors and which ones are good (for all mirrors).
 831          * It is the goal to handle the situation when more than one
 832          * mirror contains I/O errors, but the errors do not
 833          * overlap, i.e. the data can be repaired by selecting the
 834          * pages from those mirrors without I/O error on the
 835          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 836          * would be that mirror #1 has an I/O error on the first page,
 837          * the second page is good, and mirror #2 has an I/O error on
 838          * the second page, but the first page is good.
 839          * Then the first page of the first mirror can be repaired by
 840          * taking the first page of the second mirror, and the
 841          * second page of the second mirror can be repaired by
 842          * copying the contents of the 2nd page of the 1st mirror.
 843          * One more note: if the pages of one mirror contain I/O
 844          * errors, the checksum cannot be verified. In order to get
 845          * the best data for repairing, the first attempt is to find
 846          * a mirror without I/O errors and with a validated checksum.
 847          * Only if this is not possible, the pages are picked from
 848          * mirrors with I/O errors without considering the checksum.
 849          * If the latter is the case, at the end, the checksum of the
 850          * repaired area is verified in order to correctly maintain
 851          * the statistics.
 852          */
 853
 854         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 855                                      sizeof(*sblocks_for_recheck),
 856                                      GFP_NOFS);
 857         if (!sblocks_for_recheck) {
 858                 spin_lock(&sctx->stat_lock);
 859                 sctx->stat.malloc_errors++;
 860                 sctx->stat.read_errors++;
 861                 sctx->stat.uncorrectable_errors++;
 862                 spin_unlock(&sctx->stat_lock);
 863                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 864                 goto out;
 865         }
 866
 867         /* setup the context, map the logical blocks and alloc the pages */
 868         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 869                                         logical, sblocks_for_recheck);
 870         if (ret) {
 871                 spin_lock(&sctx->stat_lock);
 872                 sctx->stat.read_errors++;
 873                 sctx->stat.uncorrectable_errors++;
 874                 spin_unlock(&sctx->stat_lock);
 875                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 876                 goto out;
 877         }
 878         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 879         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 880
 881         /* build and submit the bios for the failed mirror, check checksums */
 882         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 883                             csum, generation, sctx->csum_size);
 884
 885         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 886             sblock_bad->no_io_error_seen) {
 887                 /*
 888                  * the error disappeared after reading page by page, or
 889                  * the area was part of a huge bio and other parts of the
 890                  * bio caused I/O errors, or the block layer merged several
 891                  * read requests into one and the error is caused by a
 892                  * different bio (usually one of the two latter cases is
 893                  * the cause)
 894                  */
 895                 spin_lock(&sctx->stat_lock);
 896                 sctx->stat.unverified_errors++;
 897                 spin_unlock(&sctx->stat_lock);
 898
 899                 if (sctx->is_dev_replace)
 900                         scrub_write_block_to_dev_replace(sblock_bad);
 901                 goto out;
 902         }
 903
 904         if (!sblock_bad->no_io_error_seen) {
 905                 spin_lock(&sctx->stat_lock);
 906                 sctx->stat.read_errors++;
 907                 spin_unlock(&sctx->stat_lock);
 908                 if (__ratelimit(&_rs))
 909                         scrub_print_warning("i/o error", sblock_to_check);
 910                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 911         } else if (sblock_bad->checksum_error) {
 912                 spin_lock(&sctx->stat_lock);
 913                 sctx->stat.csum_errors++;
 914                 spin_unlock(&sctx->stat_lock);
 915                 if (__ratelimit(&_rs))
 916                         scrub_print_warning("checksum error", sblock_to_check);
 917                 btrfs_dev_stat_inc_and_print(dev,
 918                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 919         } else if (sblock_bad->header_error) {
 920                 spin_lock(&sctx->stat_lock);
 921                 sctx->stat.verify_errors++;
 922                 spin_unlock(&sctx->stat_lock);
 923                 if (__ratelimit(&_rs))
 924                         scrub_print_warning("checksum/header error",
 925                                             sblock_to_check);
 926                 if (sblock_bad->generation_error)
 927                         btrfs_dev_stat_inc_and_print(dev,
 928                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 929                 else
 930                         btrfs_dev_stat_inc_and_print(dev,
 931                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 932         }
 933
 934         if (sctx->readonly && !sctx->is_dev_replace)
 935                 goto did_not_correct_error;
 936
 937         if (!is_metadata && !have_csum) {
 938                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 939
 940 nodatasum_case:
 941                 WARN_ON(sctx->is_dev_replace);
 942
 943                 /*
 944                  * !is_metadata and !have_csum, this means that the data
 945                  * might not be COW'ed, that it might be modified
 946                  * concurrently. The general strategy to work on the
 947                  * commit root does not help in the case when COW is not
 948                  * used.
 949                  */
 950                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 951                 if (!fixup_nodatasum)
 952                         goto did_not_correct_error;
 953                 fixup_nodatasum->sctx = sctx;
 954                 fixup_nodatasum->dev = dev;
 955                 fixup_nodatasum->logical = logical;
 956                 fixup_nodatasum->root = fs_info->extent_root;
 957                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 958                 scrub_pending_trans_workers_inc(sctx);
 959                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 960                 btrfs_queue_worker(&fs_info->scrub_workers,
 961                                    &fixup_nodatasum->work);
 962                 goto out;
 963         }
 964
 965         /*
 966          * now build and submit the bios for the other mirrors, check
 967          * checksums.
 968          * First try to pick the mirror which is completely without I/O
 969          * errors and also does not have a checksum error.
 970          * If one is found, and if a checksum is present, the full block
 971          * that is known to contain an error is rewritten. Afterwards
 972          * the block is known to be corrected.
 973          * If a mirror is found which is completely correct, and no
 974          * checksum is present, only those pages are rewritten that had
 975          * an I/O error in the block to be repaired, since it cannot be
 976          * determined, which copy of the other pages is better (and it
 977          * could happen otherwise that a correct page would be
 978          * overwritten by a bad one).
 979          */
 980         for (mirror_index = 0;
 981              mirror_index < BTRFS_MAX_MIRRORS &&
 982              sblocks_for_recheck[mirror_index].page_count > 0;
 983              mirror_index++) {
 984                 struct scrub_block *sblock_other;
 985
 986                 if (mirror_index == failed_mirror_index)
 987                         continue;
 988                 sblock_other = sblocks_for_recheck + mirror_index;
 989
 990                 /* build and submit the bios, check checksums */
 991                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
 992                                     have_csum, csum, generation,
 993                                     sctx->csum_size);
 994
 995                 if (!sblock_other->header_error &&
 996                     !sblock_other->checksum_error &&
 997                     sblock_other->no_io_error_seen) {
 998                         if (sctx->is_dev_replace) {
 999                                 scrub_write_block_to_dev_replace(sblock_other);
1000                         } else {
1001                                 int force_write = is_metadata || have_csum;
1002
1003                                 ret = scrub_repair_block_from_good_copy(
1004                                                 sblock_bad, sblock_other,
1005                                                 force_write);
1006                         }
1007                         if (0 == ret)
1008                                 goto corrected_error;
1009                 }
1010         }
1011
1012         /*
1013          * for dev_replace, pick good pages and write to the target device.
1014          */
1015         if (sctx->is_dev_replace) {
1016                 success = 1;
1017                 for (page_num = 0; page_num < sblock_bad->page_count;
1018                      page_num++) {
1019                         int sub_success;
1020
1021                         sub_success = 0;
1022                         for (mirror_index = 0;
1023                              mirror_index < BTRFS_MAX_MIRRORS &&
1024                              sblocks_for_recheck[mirror_index].page_count > 0;
1025                              mirror_index++) {
1026                                 struct scrub_block *sblock_other =
1027                                         sblocks_for_recheck + mirror_index;
1028                                 struct scrub_page *page_other =
1029                                         sblock_other->pagev[page_num];
1030
1031                                 if (!page_other->io_error) {
1032                                         ret = scrub_write_page_to_dev_replace(
1033                                                         sblock_other, page_num);
1034                                         if (ret == 0) {
1035                                                 /* succeeded for this page */
1036                                                 sub_success = 1;
1037                                                 break;
1038                                         } else {
1039                                                 btrfs_dev_replace_stats_inc(
1040                                                         &sctx->dev_root->
1041                                                         fs_info->dev_replace.
1042                                                         num_write_errors);
1043                                         }
1044                                 }
1045                         }
1046
1047                         if (!sub_success) {
1048                                 /*
1049                                  * did not find a mirror to fetch the page
1050                                  * from. scrub_write_page_to_dev_replace()
1051                                  * handles this case (page->io_error), by
1052                                  * filling the block with zeros before
1053                                  * submitting the write request
1054                                  */
1055                                 success = 0;
1056                                 ret = scrub_write_page_to_dev_replace(
1057                                                 sblock_bad, page_num);
1058                                 if (ret)
1059                                         btrfs_dev_replace_stats_inc(
1060                                                 &sctx->dev_root->fs_info->
1061                                                 dev_replace.num_write_errors);
1062                         }
1063                 }
1064
1065                 goto out;
1066         }
1067
1068         /*
1069          * for regular scrub, repair those pages that are errored.
1070          * In case of I/O errors in the area that is supposed to be
1071          * repaired, continue by picking good copies of those pages.
1072          * Select the good pages from mirrors to rewrite bad pages from
1073          * the area to fix. Afterwards verify the checksum of the block
1074          * that is supposed to be repaired. This verification step is
1075          * only done for the purpose of statistic counting and for the
1076          * final scrub report, whether errors remain.
1077          * A perfect algorithm could make use of the checksum and try
1078          * all possible combinations of pages from the different mirrors
1079          * until the checksum verification succeeds. For example, when
1080          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1081          * of mirror #2 is readable but the final checksum test fails,
1082          * then the 2nd page of mirror #3 could be tried, whether now
1083          * the final checksum succeedes. But this would be a rare
1084          * exception and is therefore not implemented. At least it is
1085          * avoided that the good copy is overwritten.
1086          * A more useful improvement would be to pick the sectors
1087          * without I/O error based on sector sizes (512 bytes on legacy
1088          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1089          * mirror could be repaired by taking 512 byte of a different
1090          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1091          * area are unreadable.
1092          */
1093
1094         /* can only fix I/O errors from here on */
1095         if (sblock_bad->no_io_error_seen)
1096                 goto did_not_correct_error;
1097
1098         success = 1;
1099         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1100                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1101
1102                 if (!page_bad->io_error)
1103                         continue;
1104
1105                 for (mirror_index = 0;
1106                      mirror_index < BTRFS_MAX_MIRRORS &&
1107                      sblocks_for_recheck[mirror_index].page_count > 0;
1108                      mirror_index++) {
1109                         struct scrub_block *sblock_other = sblocks_for_recheck +
1110                                                            mirror_index;
1111                         struct scrub_page *page_other = sblock_other->pagev[
1112                                                         page_num];
1113
1114                         if (!page_other->io_error) {
1115                                 ret = scrub_repair_page_from_good_copy(
1116                                         sblock_bad, sblock_other, page_num, 0);
1117                                 if (0 == ret) {
1118                                         page_bad->io_error = 0;
1119                                         break; /* succeeded for this page */
1120                                 }
1121                         }
1122                 }
1123
1124                 if (page_bad->io_error) {
1125                         /* did not find a mirror to copy the page from */
1126                         success = 0;
1127                 }
1128         }
1129
1130         if (success) {
1131                 if (is_metadata || have_csum) {
1132                         /*
1133                          * need to verify the checksum now that all
1134                          * sectors on disk are repaired (the write
1135                          * request for data to be repaired is on its way).
1136                          * Just be lazy and use scrub_recheck_block()
1137                          * which re-reads the data before the checksum
1138                          * is verified, but most likely the data comes out
1139                          * of the page cache.
1140                          */
1141                         scrub_recheck_block(fs_info, sblock_bad,
1142                                             is_metadata, have_csum, csum,
1143                                             generation, sctx->csum_size);
1144                         if (!sblock_bad->header_error &&
1145                             !sblock_bad->checksum_error &&
1146                             sblock_bad->no_io_error_seen)
1147                                 goto corrected_error;
1148                         else
1149                                 goto did_not_correct_error;
1150                 } else {
1151 corrected_error:
1152                         spin_lock(&sctx->stat_lock);
1153                         sctx->stat.corrected_errors++;
1154                         spin_unlock(&sctx->stat_lock);
1155                         printk_ratelimited_in_rcu(KERN_ERR
1156                                 "btrfs: fixed up error at logical %llu on dev %s\n",
1157                                 (unsigned long long)logical,
1158                                 rcu_str_deref(dev->name));
1159                 }
1160         } else {
1161 did_not_correct_error:
1162                 spin_lock(&sctx->stat_lock);
1163                 sctx->stat.uncorrectable_errors++;
1164                 spin_unlock(&sctx->stat_lock);
1165                 printk_ratelimited_in_rcu(KERN_ERR
1166                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1167                         (unsigned long long)logical,
1168                         rcu_str_deref(dev->name));
1169         }
1170
1171 out:
1172         if (sblocks_for_recheck) {
1173                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1174                      mirror_index++) {
1175                         struct scrub_block *sblock = sblocks_for_recheck +
1176                                                      mirror_index;
1177                         int page_index;
1178
1179                         for (page_index = 0; page_index < sblock->page_count;
1180                              page_index++) {
1181                                 sblock->pagev[page_index]->sblock = NULL;
1182                                 scrub_page_put(sblock->pagev[page_index]);
1183                         }
1184                 }
1185                 kfree(sblocks_for_recheck);
1186         }
1187
1188         return 0;
1189 }
1190
1191 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1192                                      struct btrfs_fs_info *fs_info,
1193                                      struct scrub_block *original_sblock,
1194                                      u64 length, u64 logical,
1195                                      struct scrub_block *sblocks_for_recheck)
1196 {
1197         int page_index;
1198         int mirror_index;
1199         int ret;
1200
1201         /*
1202          * note: the two members ref_count and outstanding_pages
1203          * are not used (and not set) in the blocks that are used for
1204          * the recheck procedure
1205          */
1206
1207         page_index = 0;
1208         while (length > 0) {
1209                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1210                 u64 mapped_length = sublen;
1211                 struct btrfs_bio *bbio = NULL;
1212
1213                 /*
1214                  * with a length of PAGE_SIZE, each returned stripe
1215                  * represents one mirror
1216                  */
1217                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1218                                       &mapped_length, &bbio, 0);
1219                 if (ret || !bbio || mapped_length < sublen) {
1220                         kfree(bbio);
1221                         return -EIO;
1222                 }
1223
1224                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1225                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1226                      mirror_index++) {
1227                         struct scrub_block *sblock;
1228                         struct scrub_page *page;
1229
1230                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1231                                 continue;
1232
1233                         sblock = sblocks_for_recheck + mirror_index;
1234                         sblock->sctx = sctx;
1235                         page = kzalloc(sizeof(*page), GFP_NOFS);
1236                         if (!page) {
1237 leave_nomem:
1238                                 spin_lock(&sctx->stat_lock);
1239                                 sctx->stat.malloc_errors++;
1240                                 spin_unlock(&sctx->stat_lock);
1241                                 kfree(bbio);
1242                                 return -ENOMEM;
1243                         }
1244                         scrub_page_get(page);
1245                         sblock->pagev[page_index] = page;
1246                         page->logical = logical;
1247                         page->physical = bbio->stripes[mirror_index].physical;
1248                         BUG_ON(page_index >= original_sblock->page_count);
1249                         page->physical_for_dev_replace =
1250                                 original_sblock->pagev[page_index]->
1251                                 physical_for_dev_replace;
1252                         /* for missing devices, dev->bdev is NULL */
1253                         page->dev = bbio->stripes[mirror_index].dev;
1254                         page->mirror_num = mirror_index + 1;
1255                         sblock->page_count++;
1256                         page->page = alloc_page(GFP_NOFS);
1257                         if (!page->page)
1258                                 goto leave_nomem;
1259                 }
1260                 kfree(bbio);
1261                 length -= sublen;
1262                 logical += sublen;
1263                 page_index++;
1264         }
1265
1266         return 0;
1267 }
1268
1269 /*
1270  * this function will check the on disk data for checksum errors, header
1271  * errors and read I/O errors. If any I/O errors happen, the exact pages
1272  * which are errored are marked as being bad. The goal is to enable scrub
1273  * to take those pages that are not errored from all the mirrors so that
1274  * the pages that are errored in the just handled mirror can be repaired.
1275  */
1276 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1277                                 struct scrub_block *sblock, int is_metadata,
1278                                 int have_csum, u8 *csum, u64 generation,
1279                                 u16 csum_size)
1280 {
1281         int page_num;
1282
1283         sblock->no_io_error_seen = 1;
1284         sblock->header_error = 0;
1285         sblock->checksum_error = 0;
1286
1287         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1288                 struct bio *bio;
1289                 struct scrub_page *page = sblock->pagev[page_num];
1290                 DECLARE_COMPLETION_ONSTACK(complete);
1291
1292                 if (page->dev->bdev == NULL) {
1293                         page->io_error = 1;
1294                         sblock->no_io_error_seen = 0;
1295                         continue;
1296                 }
1297
1298                 WARN_ON(!page->page);
1299                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1300                 if (!bio) {
1301                         page->io_error = 1;
1302                         sblock->no_io_error_seen = 0;
1303                         continue;
1304                 }
1305                 bio->bi_bdev = page->dev->bdev;
1306                 bio->bi_sector = page->physical >> 9;
1307                 bio->bi_end_io = scrub_complete_bio_end_io;
1308                 bio->bi_private = &complete;
1309
1310                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1311                 btrfsic_submit_bio(READ, bio);
1312
1313                 /* this will also unplug the queue */
1314                 wait_for_completion(&complete);
1315
1316                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1317                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1318                         sblock->no_io_error_seen = 0;
1319                 bio_put(bio);
1320         }
1321
1322         if (sblock->no_io_error_seen)
1323                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1324                                              have_csum, csum, generation,
1325                                              csum_size);
1326
1327         return;
1328 }
1329
1330 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1331                                          struct scrub_block *sblock,
1332                                          int is_metadata, int have_csum,
1333                                          const u8 *csum, u64 generation,
1334                                          u16 csum_size)
1335 {
1336         int page_num;
1337         u8 calculated_csum[BTRFS_CSUM_SIZE];
1338         u32 crc = ~(u32)0;
1339         void *mapped_buffer;
1340
1341         WARN_ON(!sblock->pagev[0]->page);
1342         if (is_metadata) {
1343                 struct btrfs_header *h;
1344
1345                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1346                 h = (struct btrfs_header *)mapped_buffer;
1347
1348                 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1349                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1350                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1351                            BTRFS_UUID_SIZE)) {
1352                         sblock->header_error = 1;
1353                 } else if (generation != le64_to_cpu(h->generation)) {
1354                         sblock->header_error = 1;
1355                         sblock->generation_error = 1;
1356                 }
1357                 csum = h->csum;
1358         } else {
1359                 if (!have_csum)
1360                         return;
1361
1362                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1363         }
1364
1365         for (page_num = 0;;) {
1366                 if (page_num == 0 && is_metadata)
1367                         crc = btrfs_csum_data(
1368                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1369                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1370                 else
1371                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1372
1373                 kunmap_atomic(mapped_buffer);
1374                 page_num++;
1375                 if (page_num >= sblock->page_count)
1376                         break;
1377                 WARN_ON(!sblock->pagev[page_num]->page);
1378
1379                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1380         }
1381
1382         btrfs_csum_final(crc, calculated_csum);
1383         if (memcmp(calculated_csum, csum, csum_size))
1384                 sblock->checksum_error = 1;
1385 }
1386
1387 static void scrub_complete_bio_end_io(struct bio *bio, int err)
1388 {
1389         complete((struct completion *)bio->bi_private);
1390 }
1391
1392 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1393                                              struct scrub_block *sblock_good,
1394                                              int force_write)
1395 {
1396         int page_num;
1397         int ret = 0;
1398
1399         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1400                 int ret_sub;
1401
1402                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1403                                                            sblock_good,
1404                                                            page_num,
1405                                                            force_write);
1406                 if (ret_sub)
1407                         ret = ret_sub;
1408         }
1409
1410         return ret;
1411 }
1412
1413 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1414                                             struct scrub_block *sblock_good,
1415                                             int page_num, int force_write)
1416 {
1417         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1418         struct scrub_page *page_good = sblock_good->pagev[page_num];
1419
1420         BUG_ON(page_bad->page == NULL);
1421         BUG_ON(page_good->page == NULL);
1422         if (force_write || sblock_bad->header_error ||
1423             sblock_bad->checksum_error || page_bad->io_error) {
1424                 struct bio *bio;
1425                 int ret;
1426                 DECLARE_COMPLETION_ONSTACK(complete);
1427
1428                 if (!page_bad->dev->bdev) {
1429                         printk_ratelimited(KERN_WARNING
1430                                 "btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1431                         return -EIO;
1432                 }
1433
1434                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1435                 if (!bio)
1436                         return -EIO;
1437                 bio->bi_bdev = page_bad->dev->bdev;
1438                 bio->bi_sector = page_bad->physical >> 9;
1439                 bio->bi_end_io = scrub_complete_bio_end_io;
1440                 bio->bi_private = &complete;
1441
1442                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1443                 if (PAGE_SIZE != ret) {
1444                         bio_put(bio);
1445                         return -EIO;
1446                 }
1447                 btrfsic_submit_bio(WRITE, bio);
1448
1449                 /* this will also unplug the queue */
1450                 wait_for_completion(&complete);
1451                 if (!bio_flagged(bio, BIO_UPTODATE)) {
1452                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1453                                 BTRFS_DEV_STAT_WRITE_ERRS);
1454                         btrfs_dev_replace_stats_inc(
1455                                 &sblock_bad->sctx->dev_root->fs_info->
1456                                 dev_replace.num_write_errors);
1457                         bio_put(bio);
1458                         return -EIO;
1459                 }
1460                 bio_put(bio);
1461         }
1462
1463         return 0;
1464 }
1465
1466 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1467 {
1468         int page_num;
1469
1470         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1471                 int ret;
1472
1473                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1474                 if (ret)
1475                         btrfs_dev_replace_stats_inc(
1476                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1477                                 num_write_errors);
1478         }
1479 }
1480
1481 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1482                                            int page_num)
1483 {
1484         struct scrub_page *spage = sblock->pagev[page_num];
1485
1486         BUG_ON(spage->page == NULL);
1487         if (spage->io_error) {
1488                 void *mapped_buffer = kmap_atomic(spage->page);
1489
1490                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1491                 flush_dcache_page(spage->page);
1492                 kunmap_atomic(mapped_buffer);
1493         }
1494         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1495 }
1496
1497 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1498                                     struct scrub_page *spage)
1499 {
1500         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1501         struct scrub_bio *sbio;
1502         int ret;
1503
1504         mutex_lock(&wr_ctx->wr_lock);
1505 again:
1506         if (!wr_ctx->wr_curr_bio) {
1507                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1508                                               GFP_NOFS);
1509                 if (!wr_ctx->wr_curr_bio) {
1510                         mutex_unlock(&wr_ctx->wr_lock);
1511                         return -ENOMEM;
1512                 }
1513                 wr_ctx->wr_curr_bio->sctx = sctx;
1514                 wr_ctx->wr_curr_bio->page_count = 0;
1515         }
1516         sbio = wr_ctx->wr_curr_bio;
1517         if (sbio->page_count == 0) {
1518                 struct bio *bio;
1519
1520                 sbio->physical = spage->physical_for_dev_replace;
1521                 sbio->logical = spage->logical;
1522                 sbio->dev = wr_ctx->tgtdev;
1523                 bio = sbio->bio;
1524                 if (!bio) {
1525                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1526                         if (!bio) {
1527                                 mutex_unlock(&wr_ctx->wr_lock);
1528                                 return -ENOMEM;
1529                         }
1530                         sbio->bio = bio;
1531                 }
1532
1533                 bio->bi_private = sbio;
1534                 bio->bi_end_io = scrub_wr_bio_end_io;
1535                 bio->bi_bdev = sbio->dev->bdev;
1536                 bio->bi_sector = sbio->physical >> 9;
1537                 sbio->err = 0;
1538         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1539                    spage->physical_for_dev_replace ||
1540                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1541                    spage->logical) {
1542                 scrub_wr_submit(sctx);
1543                 goto again;
1544         }
1545
1546         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1547         if (ret != PAGE_SIZE) {
1548                 if (sbio->page_count < 1) {
1549                         bio_put(sbio->bio);
1550                         sbio->bio = NULL;
1551                         mutex_unlock(&wr_ctx->wr_lock);
1552                         return -EIO;
1553                 }
1554                 scrub_wr_submit(sctx);
1555                 goto again;
1556         }
1557
1558         sbio->pagev[sbio->page_count] = spage;
1559         scrub_page_get(spage);
1560         sbio->page_count++;
1561         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1562                 scrub_wr_submit(sctx);
1563         mutex_unlock(&wr_ctx->wr_lock);
1564
1565         return 0;
1566 }
1567
1568 static void scrub_wr_submit(struct scrub_ctx *sctx)
1569 {
1570         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1571         struct scrub_bio *sbio;
1572
1573         if (!wr_ctx->wr_curr_bio)
1574                 return;
1575
1576         sbio = wr_ctx->wr_curr_bio;
1577         wr_ctx->wr_curr_bio = NULL;
1578         WARN_ON(!sbio->bio->bi_bdev);
1579         scrub_pending_bio_inc(sctx);
1580         /* process all writes in a single worker thread. Then the block layer
1581          * orders the requests before sending them to the driver which
1582          * doubled the write performance on spinning disks when measured
1583          * with Linux 3.5 */
1584         btrfsic_submit_bio(WRITE, sbio->bio);
1585 }
1586
1587 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1588 {
1589         struct scrub_bio *sbio = bio->bi_private;
1590         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1591
1592         sbio->err = err;
1593         sbio->bio = bio;
1594
1595         sbio->work.func = scrub_wr_bio_end_io_worker;
1596         btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1597 }
1598
1599 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1600 {
1601         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1602         struct scrub_ctx *sctx = sbio->sctx;
1603         int i;
1604
1605         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1606         if (sbio->err) {
1607                 struct btrfs_dev_replace *dev_replace =
1608                         &sbio->sctx->dev_root->fs_info->dev_replace;
1609
1610                 for (i = 0; i < sbio->page_count; i++) {
1611                         struct scrub_page *spage = sbio->pagev[i];
1612
1613                         spage->io_error = 1;
1614                         btrfs_dev_replace_stats_inc(&dev_replace->
1615                                                     num_write_errors);
1616                 }
1617         }
1618
1619         for (i = 0; i < sbio->page_count; i++)
1620                 scrub_page_put(sbio->pagev[i]);
1621
1622         bio_put(sbio->bio);
1623         kfree(sbio);
1624         scrub_pending_bio_dec(sctx);
1625 }
1626
1627 static int scrub_checksum(struct scrub_block *sblock)
1628 {
1629         u64 flags;
1630         int ret;
1631
1632         WARN_ON(sblock->page_count < 1);
1633         flags = sblock->pagev[0]->flags;
1634         ret = 0;
1635         if (flags & BTRFS_EXTENT_FLAG_DATA)
1636                 ret = scrub_checksum_data(sblock);
1637         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1638                 ret = scrub_checksum_tree_block(sblock);
1639         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1640                 (void)scrub_checksum_super(sblock);
1641         else
1642                 WARN_ON(1);
1643         if (ret)
1644                 scrub_handle_errored_block(sblock);
1645
1646         return ret;
1647 }
1648
1649 static int scrub_checksum_data(struct scrub_block *sblock)
1650 {
1651         struct scrub_ctx *sctx = sblock->sctx;
1652         u8 csum[BTRFS_CSUM_SIZE];
1653         u8 *on_disk_csum;
1654         struct page *page;
1655         void *buffer;
1656         u32 crc = ~(u32)0;
1657         int fail = 0;
1658         u64 len;
1659         int index;
1660
1661         BUG_ON(sblock->page_count < 1);
1662         if (!sblock->pagev[0]->have_csum)
1663                 return 0;
1664
1665         on_disk_csum = sblock->pagev[0]->csum;
1666         page = sblock->pagev[0]->page;
1667         buffer = kmap_atomic(page);
1668
1669         len = sctx->sectorsize;
1670         index = 0;
1671         for (;;) {
1672                 u64 l = min_t(u64, len, PAGE_SIZE);
1673
1674                 crc = btrfs_csum_data(buffer, crc, l);
1675                 kunmap_atomic(buffer);
1676                 len -= l;
1677                 if (len == 0)
1678                         break;
1679                 index++;
1680                 BUG_ON(index >= sblock->page_count);
1681                 BUG_ON(!sblock->pagev[index]->page);
1682                 page = sblock->pagev[index]->page;
1683                 buffer = kmap_atomic(page);
1684         }
1685
1686         btrfs_csum_final(crc, csum);
1687         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1688                 fail = 1;
1689
1690         return fail;
1691 }
1692
1693 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1694 {
1695         struct scrub_ctx *sctx = sblock->sctx;
1696         struct btrfs_header *h;
1697         struct btrfs_root *root = sctx->dev_root;
1698         struct btrfs_fs_info *fs_info = root->fs_info;
1699         u8 calculated_csum[BTRFS_CSUM_SIZE];
1700         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1701         struct page *page;
1702         void *mapped_buffer;
1703         u64 mapped_size;
1704         void *p;
1705         u32 crc = ~(u32)0;
1706         int fail = 0;
1707         int crc_fail = 0;
1708         u64 len;
1709         int index;
1710
1711         BUG_ON(sblock->page_count < 1);
1712         page = sblock->pagev[0]->page;
1713         mapped_buffer = kmap_atomic(page);
1714         h = (struct btrfs_header *)mapped_buffer;
1715         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1716
1717         /*
1718          * we don't use the getter functions here, as we
1719          * a) don't have an extent buffer and
1720          * b) the page is already kmapped
1721          */
1722
1723         if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1724                 ++fail;
1725
1726         if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1727                 ++fail;
1728
1729         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1730                 ++fail;
1731
1732         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1733                    BTRFS_UUID_SIZE))
1734                 ++fail;
1735
1736         WARN_ON(sctx->nodesize != sctx->leafsize);
1737         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1738         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1739         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1740         index = 0;
1741         for (;;) {
1742                 u64 l = min_t(u64, len, mapped_size);
1743
1744                 crc = btrfs_csum_data(p, crc, l);
1745                 kunmap_atomic(mapped_buffer);
1746                 len -= l;
1747                 if (len == 0)
1748                         break;
1749                 index++;
1750                 BUG_ON(index >= sblock->page_count);
1751                 BUG_ON(!sblock->pagev[index]->page);
1752                 page = sblock->pagev[index]->page;
1753                 mapped_buffer = kmap_atomic(page);
1754                 mapped_size = PAGE_SIZE;
1755                 p = mapped_buffer;
1756         }
1757
1758         btrfs_csum_final(crc, calculated_csum);
1759         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1760                 ++crc_fail;
1761
1762         return fail || crc_fail;
1763 }
1764
1765 static int scrub_checksum_super(struct scrub_block *sblock)
1766 {
1767         struct btrfs_super_block *s;
1768         struct scrub_ctx *sctx = sblock->sctx;
1769         struct btrfs_root *root = sctx->dev_root;
1770         struct btrfs_fs_info *fs_info = root->fs_info;
1771         u8 calculated_csum[BTRFS_CSUM_SIZE];
1772         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1773         struct page *page;
1774         void *mapped_buffer;
1775         u64 mapped_size;
1776         void *p;
1777         u32 crc = ~(u32)0;
1778         int fail_gen = 0;
1779         int fail_cor = 0;
1780         u64 len;
1781         int index;
1782
1783         BUG_ON(sblock->page_count < 1);
1784         page = sblock->pagev[0]->page;
1785         mapped_buffer = kmap_atomic(page);
1786         s = (struct btrfs_super_block *)mapped_buffer;
1787         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1788
1789         if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1790                 ++fail_cor;
1791
1792         if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1793                 ++fail_gen;
1794
1795         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1796                 ++fail_cor;
1797
1798         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1799         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1800         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1801         index = 0;
1802         for (;;) {
1803                 u64 l = min_t(u64, len, mapped_size);
1804
1805                 crc = btrfs_csum_data(p, crc, l);
1806                 kunmap_atomic(mapped_buffer);
1807                 len -= l;
1808                 if (len == 0)
1809                         break;
1810                 index++;
1811                 BUG_ON(index >= sblock->page_count);
1812                 BUG_ON(!sblock->pagev[index]->page);
1813                 page = sblock->pagev[index]->page;
1814                 mapped_buffer = kmap_atomic(page);
1815                 mapped_size = PAGE_SIZE;
1816                 p = mapped_buffer;
1817         }
1818
1819         btrfs_csum_final(crc, calculated_csum);
1820         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1821                 ++fail_cor;
1822
1823         if (fail_cor + fail_gen) {
1824                 /*
1825                  * if we find an error in a super block, we just report it.
1826                  * They will get written with the next transaction commit
1827                  * anyway
1828                  */
1829                 spin_lock(&sctx->stat_lock);
1830                 ++sctx->stat.super_errors;
1831                 spin_unlock(&sctx->stat_lock);
1832                 if (fail_cor)
1833                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1834                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1835                 else
1836                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1837                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1838         }
1839
1840         return fail_cor + fail_gen;
1841 }
1842
1843 static void scrub_block_get(struct scrub_block *sblock)
1844 {
1845         atomic_inc(&sblock->ref_count);
1846 }
1847
1848 static void scrub_block_put(struct scrub_block *sblock)
1849 {
1850         if (atomic_dec_and_test(&sblock->ref_count)) {
1851                 int i;
1852
1853                 for (i = 0; i < sblock->page_count; i++)
1854                         scrub_page_put(sblock->pagev[i]);
1855                 kfree(sblock);
1856         }
1857 }
1858
1859 static void scrub_page_get(struct scrub_page *spage)
1860 {
1861         atomic_inc(&spage->ref_count);
1862 }
1863
1864 static void scrub_page_put(struct scrub_page *spage)
1865 {
1866         if (atomic_dec_and_test(&spage->ref_count)) {
1867                 if (spage->page)
1868                         __free_page(spage->page);
1869                 kfree(spage);
1870         }
1871 }
1872
1873 static void scrub_submit(struct scrub_ctx *sctx)
1874 {
1875         struct scrub_bio *sbio;
1876
1877         if (sctx->curr == -1)
1878                 return;
1879
1880         sbio = sctx->bios[sctx->curr];
1881         sctx->curr = -1;
1882         scrub_pending_bio_inc(sctx);
1883
1884         if (!sbio->bio->bi_bdev) {
1885                 /*
1886                  * this case should not happen. If btrfs_map_block() is
1887                  * wrong, it could happen for dev-replace operations on
1888                  * missing devices when no mirrors are available, but in
1889                  * this case it should already fail the mount.
1890                  * This case is handled correctly (but _very_ slowly).
1891                  */
1892                 printk_ratelimited(KERN_WARNING
1893                         "btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1894                 bio_endio(sbio->bio, -EIO);
1895         } else {
1896                 btrfsic_submit_bio(READ, sbio->bio);
1897         }
1898 }
1899
1900 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1901                                     struct scrub_page *spage)
1902 {
1903         struct scrub_block *sblock = spage->sblock;
1904         struct scrub_bio *sbio;
1905         int ret;
1906
1907 again:
1908         /*
1909          * grab a fresh bio or wait for one to become available
1910          */
1911         while (sctx->curr == -1) {
1912                 spin_lock(&sctx->list_lock);
1913                 sctx->curr = sctx->first_free;
1914                 if (sctx->curr != -1) {
1915                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1916                         sctx->bios[sctx->curr]->next_free = -1;
1917                         sctx->bios[sctx->curr]->page_count = 0;
1918                         spin_unlock(&sctx->list_lock);
1919                 } else {
1920                         spin_unlock(&sctx->list_lock);
1921                         wait_event(sctx->list_wait, sctx->first_free != -1);
1922                 }
1923         }
1924         sbio = sctx->bios[sctx->curr];
1925         if (sbio->page_count == 0) {
1926                 struct bio *bio;
1927
1928                 sbio->physical = spage->physical;
1929                 sbio->logical = spage->logical;
1930                 sbio->dev = spage->dev;
1931                 bio = sbio->bio;
1932                 if (!bio) {
1933                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1934                         if (!bio)
1935                                 return -ENOMEM;
1936                         sbio->bio = bio;
1937                 }
1938
1939                 bio->bi_private = sbio;
1940                 bio->bi_end_io = scrub_bio_end_io;
1941                 bio->bi_bdev = sbio->dev->bdev;
1942                 bio->bi_sector = sbio->physical >> 9;
1943                 sbio->err = 0;
1944         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1945                    spage->physical ||
1946                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1947                    spage->logical ||
1948                    sbio->dev != spage->dev) {
1949                 scrub_submit(sctx);
1950                 goto again;
1951         }
1952
1953         sbio->pagev[sbio->page_count] = spage;
1954         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1955         if (ret != PAGE_SIZE) {
1956                 if (sbio->page_count < 1) {
1957                         bio_put(sbio->bio);
1958                         sbio->bio = NULL;
1959                         return -EIO;
1960                 }
1961                 scrub_submit(sctx);
1962                 goto again;
1963         }
1964
1965         scrub_block_get(sblock); /* one for the page added to the bio */
1966         atomic_inc(&sblock->outstanding_pages);
1967         sbio->page_count++;
1968         if (sbio->page_count == sctx->pages_per_rd_bio)
1969                 scrub_submit(sctx);
1970
1971         return 0;
1972 }
1973
1974 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1975                        u64 physical, struct btrfs_device *dev, u64 flags,
1976                        u64 gen, int mirror_num, u8 *csum, int force,
1977                        u64 physical_for_dev_replace)
1978 {
1979         struct scrub_block *sblock;
1980         int index;
1981
1982         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1983         if (!sblock) {
1984                 spin_lock(&sctx->stat_lock);
1985                 sctx->stat.malloc_errors++;
1986                 spin_unlock(&sctx->stat_lock);
1987                 return -ENOMEM;
1988         }
1989
1990         /* one ref inside this function, plus one for each page added to
1991          * a bio later on */
1992         atomic_set(&sblock->ref_count, 1);
1993         sblock->sctx = sctx;
1994         sblock->no_io_error_seen = 1;
1995
1996         for (index = 0; len > 0; index++) {
1997                 struct scrub_page *spage;
1998                 u64 l = min_t(u64, len, PAGE_SIZE);
1999
2000                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2001                 if (!spage) {
2002 leave_nomem:
2003                         spin_lock(&sctx->stat_lock);
2004                         sctx->stat.malloc_errors++;
2005                         spin_unlock(&sctx->stat_lock);
2006                         scrub_block_put(sblock);
2007                         return -ENOMEM;
2008                 }
2009                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2010                 scrub_page_get(spage);
2011                 sblock->pagev[index] = spage;
2012                 spage->sblock = sblock;
2013                 spage->dev = dev;
2014                 spage->flags = flags;
2015                 spage->generation = gen;
2016                 spage->logical = logical;
2017                 spage->physical = physical;
2018                 spage->physical_for_dev_replace = physical_for_dev_replace;
2019                 spage->mirror_num = mirror_num;
2020                 if (csum) {
2021                         spage->have_csum = 1;
2022                         memcpy(spage->csum, csum, sctx->csum_size);
2023                 } else {
2024                         spage->have_csum = 0;
2025                 }
2026                 sblock->page_count++;
2027                 spage->page = alloc_page(GFP_NOFS);
2028                 if (!spage->page)
2029                         goto leave_nomem;
2030                 len -= l;
2031                 logical += l;
2032                 physical += l;
2033                 physical_for_dev_replace += l;
2034         }
2035
2036         WARN_ON(sblock->page_count == 0);
2037         for (index = 0; index < sblock->page_count; index++) {
2038                 struct scrub_page *spage = sblock->pagev[index];
2039                 int ret;
2040
2041                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2042                 if (ret) {
2043                         scrub_block_put(sblock);
2044                         return ret;
2045                 }
2046         }
2047
2048         if (force)
2049                 scrub_submit(sctx);
2050
2051         /* last one frees, either here or in bio completion for last page */
2052         scrub_block_put(sblock);
2053         return 0;
2054 }
2055
2056 static void scrub_bio_end_io(struct bio *bio, int err)
2057 {
2058         struct scrub_bio *sbio = bio->bi_private;
2059         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2060
2061         sbio->err = err;
2062         sbio->bio = bio;
2063
2064         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2065 }
2066
2067 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2068 {
2069         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2070         struct scrub_ctx *sctx = sbio->sctx;
2071         int i;
2072
2073         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2074         if (sbio->err) {
2075                 for (i = 0; i < sbio->page_count; i++) {
2076                         struct scrub_page *spage = sbio->pagev[i];
2077
2078                         spage->io_error = 1;
2079                         spage->sblock->no_io_error_seen = 0;
2080                 }
2081         }
2082
2083         /* now complete the scrub_block items that have all pages completed */
2084         for (i = 0; i < sbio->page_count; i++) {
2085                 struct scrub_page *spage = sbio->pagev[i];
2086                 struct scrub_block *sblock = spage->sblock;
2087
2088                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2089                         scrub_block_complete(sblock);
2090                 scrub_block_put(sblock);
2091         }
2092
2093         bio_put(sbio->bio);
2094         sbio->bio = NULL;
2095         spin_lock(&sctx->list_lock);
2096         sbio->next_free = sctx->first_free;
2097         sctx->first_free = sbio->index;
2098         spin_unlock(&sctx->list_lock);
2099
2100         if (sctx->is_dev_replace &&
2101             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2102                 mutex_lock(&sctx->wr_ctx.wr_lock);
2103                 scrub_wr_submit(sctx);
2104                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2105         }
2106
2107         scrub_pending_bio_dec(sctx);
2108 }
2109
2110 static void scrub_block_complete(struct scrub_block *sblock)
2111 {
2112         if (!sblock->no_io_error_seen) {
2113                 scrub_handle_errored_block(sblock);
2114         } else {
2115                 /*
2116                  * if has checksum error, write via repair mechanism in
2117                  * dev replace case, otherwise write here in dev replace
2118                  * case.
2119                  */
2120                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2121                         scrub_write_block_to_dev_replace(sblock);
2122         }
2123 }
2124
2125 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2126                            u8 *csum)
2127 {
2128         struct btrfs_ordered_sum *sum = NULL;
2129         unsigned long index;
2130         unsigned long num_sectors;
2131
2132         while (!list_empty(&sctx->csum_list)) {
2133                 sum = list_first_entry(&sctx->csum_list,
2134                                        struct btrfs_ordered_sum, list);
2135                 if (sum->bytenr > logical)
2136                         return 0;
2137                 if (sum->bytenr + sum->len > logical)
2138                         break;
2139
2140                 ++sctx->stat.csum_discards;
2141                 list_del(&sum->list);
2142                 kfree(sum);
2143                 sum = NULL;
2144         }
2145         if (!sum)
2146                 return 0;
2147
2148         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2149         num_sectors = sum->len / sctx->sectorsize;
2150         memcpy(csum, sum->sums + index, sctx->csum_size);
2151         if (index == num_sectors - 1) {
2152                 list_del(&sum->list);
2153                 kfree(sum);
2154         }
2155         return 1;
2156 }
2157
2158 /* scrub extent tries to collect up to 64 kB for each bio */
2159 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2160                         u64 physical, struct btrfs_device *dev, u64 flags,
2161                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2162 {
2163         int ret;
2164         u8 csum[BTRFS_CSUM_SIZE];
2165         u32 blocksize;
2166
2167         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2168                 blocksize = sctx->sectorsize;
2169                 spin_lock(&sctx->stat_lock);
2170                 sctx->stat.data_extents_scrubbed++;
2171                 sctx->stat.data_bytes_scrubbed += len;
2172                 spin_unlock(&sctx->stat_lock);
2173         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2174                 WARN_ON(sctx->nodesize != sctx->leafsize);
2175                 blocksize = sctx->nodesize;
2176                 spin_lock(&sctx->stat_lock);
2177                 sctx->stat.tree_extents_scrubbed++;
2178                 sctx->stat.tree_bytes_scrubbed += len;
2179                 spin_unlock(&sctx->stat_lock);
2180         } else {
2181                 blocksize = sctx->sectorsize;
2182                 WARN_ON(1);
2183         }
2184
2185         while (len) {
2186                 u64 l = min_t(u64, len, blocksize);
2187                 int have_csum = 0;
2188
2189                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2190                         /* push csums to sbio */
2191                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2192                         if (have_csum == 0)
2193                                 ++sctx->stat.no_csum;
2194                         if (sctx->is_dev_replace && !have_csum) {
2195                                 ret = copy_nocow_pages(sctx, logical, l,
2196                                                        mirror_num,
2197                                                       physical_for_dev_replace);
2198                                 goto behind_scrub_pages;
2199                         }
2200                 }
2201                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2202                                   mirror_num, have_csum ? csum : NULL, 0,
2203                                   physical_for_dev_replace);
2204 behind_scrub_pages:
2205                 if (ret)
2206                         return ret;
2207                 len -= l;
2208                 logical += l;
2209                 physical += l;
2210                 physical_for_dev_replace += l;
2211         }
2212         return 0;
2213 }
2214
2215 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2216                                            struct map_lookup *map,
2217                                            struct btrfs_device *scrub_dev,
2218                                            int num, u64 base, u64 length,
2219                                            int is_dev_replace)
2220 {
2221         struct btrfs_path *path;
2222         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2223         struct btrfs_root *root = fs_info->extent_root;
2224         struct btrfs_root *csum_root = fs_info->csum_root;
2225         struct btrfs_extent_item *extent;
2226         struct blk_plug plug;
2227         u64 flags;
2228         int ret;
2229         int slot;
2230         u64 nstripes;
2231         struct extent_buffer *l;
2232         struct btrfs_key key;
2233         u64 physical;
2234         u64 logical;
2235         u64 logic_end;
2236         u64 generation;
2237         int mirror_num;
2238         struct reada_control *reada1;
2239         struct reada_control *reada2;
2240         struct btrfs_key key_start;
2241         struct btrfs_key key_end;
2242         u64 increment = map->stripe_len;
2243         u64 offset;
2244         u64 extent_logical;
2245         u64 extent_physical;
2246         u64 extent_len;
2247         struct btrfs_device *extent_dev;
2248         int extent_mirror_num;
2249         int stop_loop;
2250
2251         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2252                          BTRFS_BLOCK_GROUP_RAID6)) {
2253                 if (num >= nr_data_stripes(map)) {
2254                         return 0;
2255                 }
2256         }
2257
2258         nstripes = length;
2259         offset = 0;
2260         do_div(nstripes, map->stripe_len);
2261         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2262                 offset = map->stripe_len * num;
2263                 increment = map->stripe_len * map->num_stripes;
2264                 mirror_num = 1;
2265         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2266                 int factor = map->num_stripes / map->sub_stripes;
2267                 offset = map->stripe_len * (num / map->sub_stripes);
2268                 increment = map->stripe_len * factor;
2269                 mirror_num = num % map->sub_stripes + 1;
2270         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2271                 increment = map->stripe_len;
2272                 mirror_num = num % map->num_stripes + 1;
2273         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2274                 increment = map->stripe_len;
2275                 mirror_num = num % map->num_stripes + 1;
2276         } else {
2277                 increment = map->stripe_len;
2278                 mirror_num = 1;
2279         }
2280
2281         path = btrfs_alloc_path();
2282         if (!path)
2283                 return -ENOMEM;
2284
2285         /*
2286          * work on commit root. The related disk blocks are static as
2287          * long as COW is applied. This means, it is save to rewrite
2288          * them to repair disk errors without any race conditions
2289          */
2290         path->search_commit_root = 1;
2291         path->skip_locking = 1;
2292
2293         /*
2294          * trigger the readahead for extent tree csum tree and wait for
2295          * completion. During readahead, the scrub is officially paused
2296          * to not hold off transaction commits
2297          */
2298         logical = base + offset;
2299
2300         wait_event(sctx->list_wait,
2301                    atomic_read(&sctx->bios_in_flight) == 0);
2302         atomic_inc(&fs_info->scrubs_paused);
2303         wake_up(&fs_info->scrub_pause_wait);
2304
2305         /* FIXME it might be better to start readahead at commit root */
2306         key_start.objectid = logical;
2307         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2308         key_start.offset = (u64)0;
2309         key_end.objectid = base + offset + nstripes * increment;
2310         key_end.type = BTRFS_METADATA_ITEM_KEY;
2311         key_end.offset = (u64)-1;
2312         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2313
2314         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2315         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2316         key_start.offset = logical;
2317         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2318         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2319         key_end.offset = base + offset + nstripes * increment;
2320         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2321
2322         if (!IS_ERR(reada1))
2323                 btrfs_reada_wait(reada1);
2324         if (!IS_ERR(reada2))
2325                 btrfs_reada_wait(reada2);
2326
2327         mutex_lock(&fs_info->scrub_lock);
2328         while (atomic_read(&fs_info->scrub_pause_req)) {
2329                 mutex_unlock(&fs_info->scrub_lock);
2330                 wait_event(fs_info->scrub_pause_wait,
2331                    atomic_read(&fs_info->scrub_pause_req) == 0);
2332                 mutex_lock(&fs_info->scrub_lock);
2333         }
2334         atomic_dec(&fs_info->scrubs_paused);
2335         mutex_unlock(&fs_info->scrub_lock);
2336         wake_up(&fs_info->scrub_pause_wait);
2337
2338         /*
2339          * collect all data csums for the stripe to avoid seeking during
2340          * the scrub. This might currently (crc32) end up to be about 1MB
2341          */
2342         blk_start_plug(&plug);
2343
2344         /*
2345          * now find all extents for each stripe and scrub them
2346          */
2347         logical = base + offset;
2348         physical = map->stripes[num].physical;
2349         logic_end = logical + increment * nstripes;
2350         ret = 0;
2351         while (logical < logic_end) {
2352                 /*
2353                  * canceled?
2354                  */
2355                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2356                     atomic_read(&sctx->cancel_req)) {
2357                         ret = -ECANCELED;
2358                         goto out;
2359                 }
2360                 /*
2361                  * check to see if we have to pause
2362                  */
2363                 if (atomic_read(&fs_info->scrub_pause_req)) {
2364                         /* push queued extents */
2365                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2366                         scrub_submit(sctx);
2367                         mutex_lock(&sctx->wr_ctx.wr_lock);
2368                         scrub_wr_submit(sctx);
2369                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2370                         wait_event(sctx->list_wait,
2371                                    atomic_read(&sctx->bios_in_flight) == 0);
2372                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2373                         atomic_inc(&fs_info->scrubs_paused);
2374                         wake_up(&fs_info->scrub_pause_wait);
2375                         mutex_lock(&fs_info->scrub_lock);
2376                         while (atomic_read(&fs_info->scrub_pause_req)) {
2377                                 mutex_unlock(&fs_info->scrub_lock);
2378                                 wait_event(fs_info->scrub_pause_wait,
2379                                    atomic_read(&fs_info->scrub_pause_req) == 0);
2380                                 mutex_lock(&fs_info->scrub_lock);
2381                         }
2382                         atomic_dec(&fs_info->scrubs_paused);
2383                         mutex_unlock(&fs_info->scrub_lock);
2384                         wake_up(&fs_info->scrub_pause_wait);
2385                 }
2386
2387                 key.objectid = logical;
2388                 key.type = BTRFS_EXTENT_ITEM_KEY;
2389                 key.offset = (u64)-1;
2390
2391                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2392                 if (ret < 0)
2393                         goto out;
2394
2395                 if (ret > 0) {
2396                         ret = btrfs_previous_item(root, path, 0,
2397                                                   BTRFS_EXTENT_ITEM_KEY);
2398                         if (ret < 0)
2399                                 goto out;
2400                         if (ret > 0) {
2401                                 /* there's no smaller item, so stick with the
2402                                  * larger one */
2403                                 btrfs_release_path(path);
2404                                 ret = btrfs_search_slot(NULL, root, &key,
2405                                                         path, 0, 0);
2406                                 if (ret < 0)
2407                                         goto out;
2408                         }
2409                 }
2410
2411                 stop_loop = 0;
2412                 while (1) {
2413                         u64 bytes;
2414
2415                         l = path->nodes[0];
2416                         slot = path->slots[0];
2417                         if (slot >= btrfs_header_nritems(l)) {
2418                                 ret = btrfs_next_leaf(root, path);
2419                                 if (ret == 0)
2420                                         continue;
2421                                 if (ret < 0)
2422                                         goto out;
2423
2424                                 stop_loop = 1;
2425                                 break;
2426                         }
2427                         btrfs_item_key_to_cpu(l, &key, slot);
2428
2429                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2430                                 bytes = root->leafsize;
2431                         else
2432                                 bytes = key.offset;
2433
2434                         if (key.objectid + bytes <= logical)
2435                                 goto next;
2436
2437                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2438                             key.type != BTRFS_METADATA_ITEM_KEY)
2439                                 goto next;
2440
2441                         if (key.objectid >= logical + map->stripe_len) {
2442                                 /* out of this device extent */
2443                                 if (key.objectid >= logic_end)
2444                                         stop_loop = 1;
2445                                 break;
2446                         }
2447
2448                         extent = btrfs_item_ptr(l, slot,
2449                                                 struct btrfs_extent_item);
2450                         flags = btrfs_extent_flags(l, extent);
2451                         generation = btrfs_extent_generation(l, extent);
2452
2453                         if (key.objectid < logical &&
2454                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2455                                 printk(KERN_ERR
2456                                        "btrfs scrub: tree block %llu spanning "
2457                                        "stripes, ignored. logical=%llu\n",
2458                                        (unsigned long long)key.objectid,
2459                                        (unsigned long long)logical);
2460                                 goto next;
2461                         }
2462
2463 again:
2464                         extent_logical = key.objectid;
2465                         extent_len = bytes;
2466
2467                         /*
2468                          * trim extent to this stripe
2469                          */
2470                         if (extent_logical < logical) {
2471                                 extent_len -= logical - extent_logical;
2472                                 extent_logical = logical;
2473                         }
2474                         if (extent_logical + extent_len >
2475                             logical + map->stripe_len) {
2476                                 extent_len = logical + map->stripe_len -
2477                                              extent_logical;
2478                         }
2479
2480                         extent_physical = extent_logical - logical + physical;
2481                         extent_dev = scrub_dev;
2482                         extent_mirror_num = mirror_num;
2483                         if (is_dev_replace)
2484                                 scrub_remap_extent(fs_info, extent_logical,
2485                                                    extent_len, &extent_physical,
2486                                                    &extent_dev,
2487                                                    &extent_mirror_num);
2488
2489                         ret = btrfs_lookup_csums_range(csum_root, logical,
2490                                                 logical + map->stripe_len - 1,
2491                                                 &sctx->csum_list, 1);
2492                         if (ret)
2493                                 goto out;
2494
2495                         ret = scrub_extent(sctx, extent_logical, extent_len,
2496                                            extent_physical, extent_dev, flags,
2497                                            generation, extent_mirror_num,
2498                                            extent_logical - logical + physical);
2499                         if (ret)
2500                                 goto out;
2501
2502                         scrub_free_csums(sctx);
2503                         if (extent_logical + extent_len <
2504                             key.objectid + bytes) {
2505                                 logical += increment;
2506                                 physical += map->stripe_len;
2507
2508                                 if (logical < key.objectid + bytes) {
2509                                         cond_resched();
2510                                         goto again;
2511                                 }
2512
2513                                 if (logical >= logic_end) {
2514                                         stop_loop = 1;
2515                                         break;
2516                                 }
2517                         }
2518 next:
2519                         path->slots[0]++;
2520                 }
2521                 btrfs_release_path(path);
2522                 logical += increment;
2523                 physical += map->stripe_len;
2524                 spin_lock(&sctx->stat_lock);
2525                 if (stop_loop)
2526                         sctx->stat.last_physical = map->stripes[num].physical +
2527                                                    length;
2528                 else
2529                         sctx->stat.last_physical = physical;
2530                 spin_unlock(&sctx->stat_lock);
2531                 if (stop_loop)
2532                         break;
2533         }
2534 out:
2535         /* push queued extents */
2536         scrub_submit(sctx);
2537         mutex_lock(&sctx->wr_ctx.wr_lock);
2538         scrub_wr_submit(sctx);
2539         mutex_unlock(&sctx->wr_ctx.wr_lock);
2540
2541         blk_finish_plug(&plug);
2542         btrfs_free_path(path);
2543         return ret < 0 ? ret : 0;
2544 }
2545
2546 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2547                                           struct btrfs_device *scrub_dev,
2548                                           u64 chunk_tree, u64 chunk_objectid,
2549                                           u64 chunk_offset, u64 length,
2550                                           u64 dev_offset, int is_dev_replace)
2551 {
2552         struct btrfs_mapping_tree *map_tree =
2553                 &sctx->dev_root->fs_info->mapping_tree;
2554         struct map_lookup *map;
2555         struct extent_map *em;
2556         int i;
2557         int ret = 0;
2558
2559         read_lock(&map_tree->map_tree.lock);
2560         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2561         read_unlock(&map_tree->map_tree.lock);
2562
2563         if (!em)
2564                 return -EINVAL;
2565
2566         map = (struct map_lookup *)em->bdev;
2567         if (em->start != chunk_offset)
2568                 goto out;
2569
2570         if (em->len < length)
2571                 goto out;
2572
2573         for (i = 0; i < map->num_stripes; ++i) {
2574                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2575                     map->stripes[i].physical == dev_offset) {
2576                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2577                                            chunk_offset, length,
2578                                            is_dev_replace);
2579                         if (ret)
2580                                 goto out;
2581                 }
2582         }
2583 out:
2584         free_extent_map(em);
2585
2586         return ret;
2587 }
2588
2589 static noinline_for_stack
2590 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2591                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2592                            int is_dev_replace)
2593 {
2594         struct btrfs_dev_extent *dev_extent = NULL;
2595         struct btrfs_path *path;
2596         struct btrfs_root *root = sctx->dev_root;
2597         struct btrfs_fs_info *fs_info = root->fs_info;
2598         u64 length;
2599         u64 chunk_tree;
2600         u64 chunk_objectid;
2601         u64 chunk_offset;
2602         int ret;
2603         int slot;
2604         struct extent_buffer *l;
2605         struct btrfs_key key;
2606         struct btrfs_key found_key;
2607         struct btrfs_block_group_cache *cache;
2608         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2609
2610         path = btrfs_alloc_path();
2611         if (!path)
2612                 return -ENOMEM;
2613
2614         path->reada = 2;
2615         path->search_commit_root = 1;
2616         path->skip_locking = 1;
2617
2618         key.objectid = scrub_dev->devid;
2619         key.offset = 0ull;
2620         key.type = BTRFS_DEV_EXTENT_KEY;
2621
2622         while (1) {
2623                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2624                 if (ret < 0)
2625                         break;
2626                 if (ret > 0) {
2627                         if (path->slots[0] >=
2628                             btrfs_header_nritems(path->nodes[0])) {
2629                                 ret = btrfs_next_leaf(root, path);
2630                                 if (ret)
2631                                         break;
2632                         }
2633                 }
2634
2635                 l = path->nodes[0];
2636                 slot = path->slots[0];
2637
2638                 btrfs_item_key_to_cpu(l, &found_key, slot);
2639
2640                 if (found_key.objectid != scrub_dev->devid)
2641                         break;
2642
2643                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2644                         break;
2645
2646                 if (found_key.offset >= end)
2647                         break;
2648
2649                 if (found_key.offset < key.offset)
2650                         break;
2651
2652                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2653                 length = btrfs_dev_extent_length(l, dev_extent);
2654
2655                 if (found_key.offset + length <= start) {
2656                         key.offset = found_key.offset + length;
2657                         btrfs_release_path(path);
2658                         continue;
2659                 }
2660
2661                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2662                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2663                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2664
2665                 /*
2666                  * get a reference on the corresponding block group to prevent
2667                  * the chunk from going away while we scrub it
2668                  */
2669                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2670                 if (!cache) {
2671                         ret = -ENOENT;
2672                         break;
2673                 }
2674                 dev_replace->cursor_right = found_key.offset + length;
2675                 dev_replace->cursor_left = found_key.offset;
2676                 dev_replace->item_needs_writeback = 1;
2677                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2678                                   chunk_offset, length, found_key.offset,
2679                                   is_dev_replace);
2680
2681                 /*
2682                  * flush, submit all pending read and write bios, afterwards
2683                  * wait for them.
2684                  * Note that in the dev replace case, a read request causes
2685                  * write requests that are submitted in the read completion
2686                  * worker. Therefore in the current situation, it is required
2687                  * that all write requests are flushed, so that all read and
2688                  * write requests are really completed when bios_in_flight
2689                  * changes to 0.
2690                  */
2691                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2692                 scrub_submit(sctx);
2693                 mutex_lock(&sctx->wr_ctx.wr_lock);
2694                 scrub_wr_submit(sctx);
2695                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2696
2697                 wait_event(sctx->list_wait,
2698                            atomic_read(&sctx->bios_in_flight) == 0);
2699                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2700                 atomic_inc(&fs_info->scrubs_paused);
2701                 wake_up(&fs_info->scrub_pause_wait);
2702                 wait_event(sctx->list_wait,
2703                            atomic_read(&sctx->workers_pending) == 0);
2704
2705                 mutex_lock(&fs_info->scrub_lock);
2706                 while (atomic_read(&fs_info->scrub_pause_req)) {
2707                         mutex_unlock(&fs_info->scrub_lock);
2708                         wait_event(fs_info->scrub_pause_wait,
2709                            atomic_read(&fs_info->scrub_pause_req) == 0);
2710                         mutex_lock(&fs_info->scrub_lock);
2711                 }
2712                 atomic_dec(&fs_info->scrubs_paused);
2713                 mutex_unlock(&fs_info->scrub_lock);
2714                 wake_up(&fs_info->scrub_pause_wait);
2715
2716                 dev_replace->cursor_left = dev_replace->cursor_right;
2717                 dev_replace->item_needs_writeback = 1;
2718                 btrfs_put_block_group(cache);
2719                 if (ret)
2720                         break;
2721                 if (is_dev_replace &&
2722                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2723                         ret = -EIO;
2724                         break;
2725                 }
2726                 if (sctx->stat.malloc_errors > 0) {
2727                         ret = -ENOMEM;
2728                         break;
2729                 }
2730
2731                 key.offset = found_key.offset + length;
2732                 btrfs_release_path(path);
2733         }
2734
2735         btrfs_free_path(path);
2736
2737         /*
2738          * ret can still be 1 from search_slot or next_leaf,
2739          * that's not an error
2740          */
2741         return ret < 0 ? ret : 0;
2742 }
2743
2744 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2745                                            struct btrfs_device *scrub_dev)
2746 {
2747         int     i;
2748         u64     bytenr;
2749         u64     gen;
2750         int     ret;
2751         struct btrfs_root *root = sctx->dev_root;
2752
2753         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2754                 return -EIO;
2755
2756         gen = root->fs_info->last_trans_committed;
2757
2758         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2759                 bytenr = btrfs_sb_offset(i);
2760                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2761                         break;
2762
2763                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2764                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2765                                   NULL, 1, bytenr);
2766                 if (ret)
2767                         return ret;
2768         }
2769         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2770
2771         return 0;
2772 }
2773
2774 /*
2775  * get a reference count on fs_info->scrub_workers. start worker if necessary
2776  */
2777 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2778                                                 int is_dev_replace)
2779 {
2780         int ret = 0;
2781
2782         mutex_lock(&fs_info->scrub_lock);
2783         if (fs_info->scrub_workers_refcnt == 0) {
2784                 if (is_dev_replace)
2785                         btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2786                                         &fs_info->generic_worker);
2787                 else
2788                         btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2789                                         fs_info->thread_pool_size,
2790                                         &fs_info->generic_worker);
2791                 fs_info->scrub_workers.idle_thresh = 4;
2792                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2793                 if (ret)
2794                         goto out;
2795                 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2796                                    "scrubwrc",
2797                                    fs_info->thread_pool_size,
2798                                    &fs_info->generic_worker);
2799                 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2800                 ret = btrfs_start_workers(
2801                                 &fs_info->scrub_wr_completion_workers);
2802                 if (ret)
2803                         goto out;
2804                 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2805                                    &fs_info->generic_worker);
2806                 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2807                 if (ret)
2808                         goto out;
2809         }
2810         ++fs_info->scrub_workers_refcnt;
2811 out:
2812         mutex_unlock(&fs_info->scrub_lock);
2813
2814         return ret;
2815 }
2816
2817 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2818 {
2819         mutex_lock(&fs_info->scrub_lock);
2820         if (--fs_info->scrub_workers_refcnt == 0) {
2821                 btrfs_stop_workers(&fs_info->scrub_workers);
2822                 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2823                 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2824         }
2825         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2826         mutex_unlock(&fs_info->scrub_lock);
2827 }
2828
2829 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2830                     u64 end, struct btrfs_scrub_progress *progress,
2831                     int readonly, int is_dev_replace)
2832 {
2833         struct scrub_ctx *sctx;
2834         int ret;
2835         struct btrfs_device *dev;
2836
2837         if (btrfs_fs_closing(fs_info))
2838                 return -EINVAL;
2839
2840         /*
2841          * check some assumptions
2842          */
2843         if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2844                 printk(KERN_ERR
2845                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2846                        fs_info->chunk_root->nodesize,
2847                        fs_info->chunk_root->leafsize);
2848                 return -EINVAL;
2849         }
2850
2851         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2852                 /*
2853                  * in this case scrub is unable to calculate the checksum
2854                  * the way scrub is implemented. Do not handle this
2855                  * situation at all because it won't ever happen.
2856                  */
2857                 printk(KERN_ERR
2858                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2859                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2860                 return -EINVAL;
2861         }
2862
2863         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2864                 /* not supported for data w/o checksums */
2865                 printk(KERN_ERR
2866                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2867                        fs_info->chunk_root->sectorsize,
2868                        (unsigned long long)PAGE_SIZE);
2869                 return -EINVAL;
2870         }
2871
2872         if (fs_info->chunk_root->nodesize >
2873             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2874             fs_info->chunk_root->sectorsize >
2875             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2876                 /*
2877                  * would exhaust the array bounds of pagev member in
2878                  * struct scrub_block
2879                  */
2880                 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2881                        fs_info->chunk_root->nodesize,
2882                        SCRUB_MAX_PAGES_PER_BLOCK,
2883                        fs_info->chunk_root->sectorsize,
2884                        SCRUB_MAX_PAGES_PER_BLOCK);
2885                 return -EINVAL;
2886         }
2887
2888         ret = scrub_workers_get(fs_info, is_dev_replace);
2889         if (ret)
2890                 return ret;
2891
2892         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2893         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2894         if (!dev || (dev->missing && !is_dev_replace)) {
2895                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2896                 scrub_workers_put(fs_info);
2897                 return -ENODEV;
2898         }
2899         mutex_lock(&fs_info->scrub_lock);
2900
2901         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2902                 mutex_unlock(&fs_info->scrub_lock);
2903                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2904                 scrub_workers_put(fs_info);
2905                 return -EIO;
2906         }
2907
2908         btrfs_dev_replace_lock(&fs_info->dev_replace);
2909         if (dev->scrub_device ||
2910             (!is_dev_replace &&
2911              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2912                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2913                 mutex_unlock(&fs_info->scrub_lock);
2914                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2915                 scrub_workers_put(fs_info);
2916                 return -EINPROGRESS;
2917         }
2918         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2919         sctx = scrub_setup_ctx(dev, is_dev_replace);
2920         if (IS_ERR(sctx)) {
2921                 mutex_unlock(&fs_info->scrub_lock);
2922                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2923                 scrub_workers_put(fs_info);
2924                 return PTR_ERR(sctx);
2925         }
2926         sctx->readonly = readonly;
2927         dev->scrub_device = sctx;
2928
2929         atomic_inc(&fs_info->scrubs_running);
2930         mutex_unlock(&fs_info->scrub_lock);
2931         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2932
2933         if (!is_dev_replace) {
2934                 down_read(&fs_info->scrub_super_lock);
2935                 ret = scrub_supers(sctx, dev);
2936                 up_read(&fs_info->scrub_super_lock);
2937         }
2938
2939         if (!ret)
2940                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2941                                              is_dev_replace);
2942
2943         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2944         atomic_dec(&fs_info->scrubs_running);
2945         wake_up(&fs_info->scrub_pause_wait);
2946
2947         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2948
2949         if (progress)
2950                 memcpy(progress, &sctx->stat, sizeof(*progress));
2951
2952         mutex_lock(&fs_info->scrub_lock);
2953         dev->scrub_device = NULL;
2954         mutex_unlock(&fs_info->scrub_lock);
2955
2956         scrub_free_ctx(sctx);
2957         scrub_workers_put(fs_info);
2958
2959         return ret;
2960 }
2961
2962 void btrfs_scrub_pause(struct btrfs_root *root)
2963 {
2964         struct btrfs_fs_info *fs_info = root->fs_info;
2965
2966         mutex_lock(&fs_info->scrub_lock);
2967         atomic_inc(&fs_info->scrub_pause_req);
2968         while (atomic_read(&fs_info->scrubs_paused) !=
2969                atomic_read(&fs_info->scrubs_running)) {
2970                 mutex_unlock(&fs_info->scrub_lock);
2971                 wait_event(fs_info->scrub_pause_wait,
2972                            atomic_read(&fs_info->scrubs_paused) ==
2973                            atomic_read(&fs_info->scrubs_running));
2974                 mutex_lock(&fs_info->scrub_lock);
2975         }
2976         mutex_unlock(&fs_info->scrub_lock);
2977 }
2978
2979 void btrfs_scrub_continue(struct btrfs_root *root)
2980 {
2981         struct btrfs_fs_info *fs_info = root->fs_info;
2982
2983         atomic_dec(&fs_info->scrub_pause_req);
2984         wake_up(&fs_info->scrub_pause_wait);
2985 }
2986
2987 void btrfs_scrub_pause_super(struct btrfs_root *root)
2988 {
2989         down_write(&root->fs_info->scrub_super_lock);
2990 }
2991
2992 void btrfs_scrub_continue_super(struct btrfs_root *root)
2993 {
2994         up_write(&root->fs_info->scrub_super_lock);
2995 }
2996
2997 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2998 {
2999         mutex_lock(&fs_info->scrub_lock);
3000         if (!atomic_read(&fs_info->scrubs_running)) {
3001                 mutex_unlock(&fs_info->scrub_lock);
3002                 return -ENOTCONN;
3003         }
3004
3005         atomic_inc(&fs_info->scrub_cancel_req);
3006         while (atomic_read(&fs_info->scrubs_running)) {
3007                 mutex_unlock(&fs_info->scrub_lock);
3008                 wait_event(fs_info->scrub_pause_wait,
3009                            atomic_read(&fs_info->scrubs_running) == 0);
3010                 mutex_lock(&fs_info->scrub_lock);
3011         }
3012         atomic_dec(&fs_info->scrub_cancel_req);
3013         mutex_unlock(&fs_info->scrub_lock);
3014
3015         return 0;
3016 }
3017
3018 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3019                            struct btrfs_device *dev)
3020 {
3021         struct scrub_ctx *sctx;
3022
3023         mutex_lock(&fs_info->scrub_lock);
3024         sctx = dev->scrub_device;
3025         if (!sctx) {
3026                 mutex_unlock(&fs_info->scrub_lock);
3027                 return -ENOTCONN;
3028         }
3029         atomic_inc(&sctx->cancel_req);
3030         while (dev->scrub_device) {
3031                 mutex_unlock(&fs_info->scrub_lock);
3032                 wait_event(fs_info->scrub_pause_wait,
3033                            dev->scrub_device == NULL);
3034                 mutex_lock(&fs_info->scrub_lock);
3035         }
3036         mutex_unlock(&fs_info->scrub_lock);
3037
3038         return 0;
3039 }
3040
3041 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3042                          struct btrfs_scrub_progress *progress)
3043 {
3044         struct btrfs_device *dev;
3045         struct scrub_ctx *sctx = NULL;
3046
3047         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3048         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3049         if (dev)
3050                 sctx = dev->scrub_device;
3051         if (sctx)
3052                 memcpy(progress, &sctx->stat, sizeof(*progress));
3053         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3054
3055         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3056 }
3057
3058 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3059                                u64 extent_logical, u64 extent_len,
3060                                u64 *extent_physical,
3061                                struct btrfs_device **extent_dev,
3062                                int *extent_mirror_num)
3063 {
3064         u64 mapped_length;
3065         struct btrfs_bio *bbio = NULL;
3066         int ret;
3067
3068         mapped_length = extent_len;
3069         ret = btrfs_map_block(fs_info, READ, extent_logical,
3070                               &mapped_length, &bbio, 0);
3071         if (ret || !bbio || mapped_length < extent_len ||
3072             !bbio->stripes[0].dev->bdev) {
3073                 kfree(bbio);
3074                 return;
3075         }
3076
3077         *extent_physical = bbio->stripes[0].physical;
3078         *extent_mirror_num = bbio->mirror_num;
3079         *extent_dev = bbio->stripes[0].dev;
3080         kfree(bbio);
3081 }
3082
3083 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3084                               struct scrub_wr_ctx *wr_ctx,
3085                               struct btrfs_fs_info *fs_info,
3086                               struct btrfs_device *dev,
3087                               int is_dev_replace)
3088 {
3089         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3090
3091         mutex_init(&wr_ctx->wr_lock);
3092         wr_ctx->wr_curr_bio = NULL;
3093         if (!is_dev_replace)
3094                 return 0;
3095
3096         WARN_ON(!dev->bdev);
3097         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3098                                          bio_get_nr_vecs(dev->bdev));
3099         wr_ctx->tgtdev = dev;
3100         atomic_set(&wr_ctx->flush_all_writes, 0);
3101         return 0;
3102 }
3103
3104 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3105 {
3106         mutex_lock(&wr_ctx->wr_lock);
3107         kfree(wr_ctx->wr_curr_bio);
3108         wr_ctx->wr_curr_bio = NULL;
3109         mutex_unlock(&wr_ctx->wr_lock);
3110 }
3111
3112 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3113                             int mirror_num, u64 physical_for_dev_replace)
3114 {
3115         struct scrub_copy_nocow_ctx *nocow_ctx;
3116         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3117
3118         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3119         if (!nocow_ctx) {
3120                 spin_lock(&sctx->stat_lock);
3121                 sctx->stat.malloc_errors++;
3122                 spin_unlock(&sctx->stat_lock);
3123                 return -ENOMEM;
3124         }
3125
3126         scrub_pending_trans_workers_inc(sctx);
3127
3128         nocow_ctx->sctx = sctx;
3129         nocow_ctx->logical = logical;
3130         nocow_ctx->len = len;
3131         nocow_ctx->mirror_num = mirror_num;
3132         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3133         nocow_ctx->work.func = copy_nocow_pages_worker;
3134         btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3135                            &nocow_ctx->work);
3136
3137         return 0;
3138 }
3139
3140 static void copy_nocow_pages_worker(struct btrfs_work *work)
3141 {
3142         struct scrub_copy_nocow_ctx *nocow_ctx =
3143                 container_of(work, struct scrub_copy_nocow_ctx, work);
3144         struct scrub_ctx *sctx = nocow_ctx->sctx;
3145         u64 logical = nocow_ctx->logical;
3146         u64 len = nocow_ctx->len;
3147         int mirror_num = nocow_ctx->mirror_num;
3148         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3149         int ret;
3150         struct btrfs_trans_handle *trans = NULL;
3151         struct btrfs_fs_info *fs_info;
3152         struct btrfs_path *path;
3153         struct btrfs_root *root;
3154         int not_written = 0;
3155
3156         fs_info = sctx->dev_root->fs_info;
3157         root = fs_info->extent_root;
3158
3159         path = btrfs_alloc_path();
3160         if (!path) {
3161                 spin_lock(&sctx->stat_lock);
3162                 sctx->stat.malloc_errors++;
3163                 spin_unlock(&sctx->stat_lock);
3164                 not_written = 1;
3165                 goto out;
3166         }
3167
3168         trans = btrfs_join_transaction(root);
3169         if (IS_ERR(trans)) {
3170                 not_written = 1;
3171                 goto out;
3172         }
3173
3174         ret = iterate_inodes_from_logical(logical, fs_info, path,
3175                                           copy_nocow_pages_for_inode,
3176                                           nocow_ctx);
3177         if (ret != 0 && ret != -ENOENT) {
3178                 pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3179                         (unsigned long long)logical,
3180                         (unsigned long long)physical_for_dev_replace,
3181                         (unsigned long long)len,
3182                         (unsigned long long)mirror_num, ret);
3183                 not_written = 1;
3184                 goto out;
3185         }
3186
3187 out:
3188         if (trans && !IS_ERR(trans))
3189                 btrfs_end_transaction(trans, root);
3190         if (not_written)
3191                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3192                                             num_uncorrectable_read_errors);
3193
3194         btrfs_free_path(path);
3195         kfree(nocow_ctx);
3196
3197         scrub_pending_trans_workers_dec(sctx);
3198 }
3199
3200 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3201 {
3202         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3203         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3204         struct btrfs_key key;
3205         struct inode *inode;
3206         struct page *page;
3207         struct btrfs_root *local_root;
3208         u64 physical_for_dev_replace;
3209         u64 len;
3210         unsigned long index;
3211         int srcu_index;
3212         int ret;
3213         int err;
3214
3215         key.objectid = root;
3216         key.type = BTRFS_ROOT_ITEM_KEY;
3217         key.offset = (u64)-1;
3218
3219         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3220
3221         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3222         if (IS_ERR(local_root)) {
3223                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3224                 return PTR_ERR(local_root);
3225         }
3226
3227         if (btrfs_root_refs(&local_root->root_item) == 0) {
3228                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3229                 return -ENOENT;
3230         }
3231
3232         key.type = BTRFS_INODE_ITEM_KEY;
3233         key.objectid = inum;
3234         key.offset = 0;
3235         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3236         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3237         if (IS_ERR(inode))
3238                 return PTR_ERR(inode);
3239
3240         /* Avoid truncate/dio/punch hole.. */
3241         mutex_lock(&inode->i_mutex);
3242         inode_dio_wait(inode);
3243
3244         ret = 0;
3245         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3246         len = nocow_ctx->len;
3247         while (len >= PAGE_CACHE_SIZE) {
3248                 index = offset >> PAGE_CACHE_SHIFT;
3249 again:
3250                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3251                 if (!page) {
3252                         pr_err("find_or_create_page() failed\n");
3253                         ret = -ENOMEM;
3254                         goto out;
3255                 }
3256
3257                 if (PageUptodate(page)) {
3258                         if (PageDirty(page))
3259                                 goto next_page;
3260                 } else {
3261                         ClearPageError(page);
3262                         err = extent_read_full_page(&BTRFS_I(inode)->
3263                                                          io_tree,
3264                                                         page, btrfs_get_extent,
3265                                                         nocow_ctx->mirror_num);
3266                         if (err) {
3267                                 ret = err;
3268                                 goto next_page;
3269                         }
3270
3271                         lock_page(page);
3272                         /*
3273                          * If the page has been remove from the page cache,
3274                          * the data on it is meaningless, because it may be
3275                          * old one, the new data may be written into the new
3276                          * page in the page cache.
3277                          */
3278                         if (page->mapping != inode->i_mapping) {
3279                                 page_cache_release(page);
3280                                 goto again;
3281                         }
3282                         if (!PageUptodate(page)) {
3283                                 ret = -EIO;
3284                                 goto next_page;
3285                         }
3286                 }
3287                 err = write_page_nocow(nocow_ctx->sctx,
3288                                        physical_for_dev_replace, page);
3289                 if (err)
3290                         ret = err;
3291 next_page:
3292                 unlock_page(page);
3293                 page_cache_release(page);
3294
3295                 if (ret)
3296                         break;
3297
3298                 offset += PAGE_CACHE_SIZE;
3299                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3300                 len -= PAGE_CACHE_SIZE;
3301         }
3302 out:
3303         mutex_unlock(&inode->i_mutex);
3304         iput(inode);
3305         return ret;
3306 }
3307
3308 static int write_page_nocow(struct scrub_ctx *sctx,
3309                             u64 physical_for_dev_replace, struct page *page)
3310 {
3311         struct bio *bio;
3312         struct btrfs_device *dev;
3313         int ret;
3314         DECLARE_COMPLETION_ONSTACK(compl);
3315
3316         dev = sctx->wr_ctx.tgtdev;
3317         if (!dev)
3318                 return -EIO;
3319         if (!dev->bdev) {
3320                 printk_ratelimited(KERN_WARNING
3321                         "btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3322                 return -EIO;
3323         }
3324         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3325         if (!bio) {
3326                 spin_lock(&sctx->stat_lock);
3327                 sctx->stat.malloc_errors++;
3328                 spin_unlock(&sctx->stat_lock);
3329                 return -ENOMEM;
3330         }
3331         bio->bi_private = &compl;
3332         bio->bi_end_io = scrub_complete_bio_end_io;
3333         bio->bi_size = 0;
3334         bio->bi_sector = physical_for_dev_replace >> 9;
3335         bio->bi_bdev = dev->bdev;
3336         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3337         if (ret != PAGE_CACHE_SIZE) {
3338 leave_with_eio:
3339                 bio_put(bio);
3340                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3341                 return -EIO;
3342         }
3343         btrfsic_submit_bio(WRITE_SYNC, bio);
3344         wait_for_completion(&compl);
3345
3346         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3347                 goto leave_with_eio;
3348
3349         bio_put(bio);
3350         return 0;
3351 }