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