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