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         btrfs_release_path(path);
 546
 547         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 548                 do {
 549                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 550                                                         &ref_root, &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         } else {
 562                 swarn.path = path;
 563                 swarn.dev = dev;
 564                 iterate_extent_inodes(fs_info, found_key.objectid,
 565                                         extent_item_pos, 1,
 566                                         scrub_print_warning_inode, &swarn);
 567         }
 568
 569 out:
 570         btrfs_free_path(path);
 571         kfree(swarn.scratch_buf);
 572         kfree(swarn.msg_buf);
 573 }
 574
 575 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 576 {
 577         struct page *page = NULL;
 578         unsigned long index;
 579         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 580         int ret;
 581         int corrected = 0;
 582         struct btrfs_key key;
 583         struct inode *inode = NULL;
 584         struct btrfs_fs_info *fs_info;
 585         u64 end = offset + PAGE_SIZE - 1;
 586         struct btrfs_root *local_root;
 587         int srcu_index;
 588
 589         key.objectid = root;
 590         key.type = BTRFS_ROOT_ITEM_KEY;
 591         key.offset = (u64)-1;
 592
 593         fs_info = fixup->root->fs_info;
 594         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 595
 596         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 597         if (IS_ERR(local_root)) {
 598                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 599                 return PTR_ERR(local_root);
 600         }
 601
 602         key.type = BTRFS_INODE_ITEM_KEY;
 603         key.objectid = inum;
 604         key.offset = 0;
 605         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 606         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 607         if (IS_ERR(inode))
 608                 return PTR_ERR(inode);
 609
 610         index = offset >> PAGE_CACHE_SHIFT;
 611
 612         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 613         if (!page) {
 614                 ret = -ENOMEM;
 615                 goto out;
 616         }
 617
 618         if (PageUptodate(page)) {
 619                 if (PageDirty(page)) {
 620                         /*
 621                          * we need to write the data to the defect sector. the
 622                          * data that was in that sector is not in memory,
 623                          * because the page was modified. we must not write the
 624                          * modified page to that sector.
 625                          *
 626                          * TODO: what could be done here: wait for the delalloc
 627                          *       runner to write out that page (might involve
 628                          *       COW) and see whether the sector is still
 629                          *       referenced afterwards.
 630                          *
 631                          * For the meantime, we'll treat this error
 632                          * incorrectable, although there is a chance that a
 633                          * later scrub will find the bad sector again and that
 634                          * there's no dirty page in memory, then.
 635                          */
 636                         ret = -EIO;
 637                         goto out;
 638                 }
 639                 fs_info = BTRFS_I(inode)->root->fs_info;
 640                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 641                                         fixup->logical, page,
 642                                         fixup->mirror_num);
 643                 unlock_page(page);
 644                 corrected = !ret;
 645         } else {
 646                 /*
 647                  * we need to get good data first. the general readpage path
 648                  * will call repair_io_failure for us, we just have to make
 649                  * sure we read the bad mirror.
 650                  */
 651                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 652                                         EXTENT_DAMAGED, GFP_NOFS);
 653                 if (ret) {
 654                         /* set_extent_bits should give proper error */
 655                         WARN_ON(ret > 0);
 656                         if (ret > 0)
 657                                 ret = -EFAULT;
 658                         goto out;
 659                 }
 660
 661                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 662                                                 btrfs_get_extent,
 663                                                 fixup->mirror_num);
 664                 wait_on_page_locked(page);
 665
 666                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 667                                                 end, EXTENT_DAMAGED, 0, NULL);
 668                 if (!corrected)
 669                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 670                                                 EXTENT_DAMAGED, GFP_NOFS);
 671         }
 672
 673 out:
 674         if (page)
 675                 put_page(page);
 676         if (inode)
 677                 iput(inode);
 678
 679         if (ret < 0)
 680                 return ret;
 681
 682         if (ret == 0 && corrected) {
 683                 /*
 684                  * we only need to call readpage for one of the inodes belonging
 685                  * to this extent. so make iterate_extent_inodes stop
 686                  */
 687                 return 1;
 688         }
 689
 690         return -EIO;
 691 }
 692
 693 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 694 {
 695         int ret;
 696         struct scrub_fixup_nodatasum *fixup;
 697         struct scrub_ctx *sctx;
 698         struct btrfs_trans_handle *trans = NULL;
 699         struct btrfs_fs_info *fs_info;
 700         struct btrfs_path *path;
 701         int uncorrectable = 0;
 702
 703         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 704         sctx = fixup->sctx;
 705         fs_info = fixup->root->fs_info;
 706
 707         path = btrfs_alloc_path();
 708         if (!path) {
 709                 spin_lock(&sctx->stat_lock);
 710                 ++sctx->stat.malloc_errors;
 711                 spin_unlock(&sctx->stat_lock);
 712                 uncorrectable = 1;
 713                 goto out;
 714         }
 715
 716         trans = btrfs_join_transaction(fixup->root);
 717         if (IS_ERR(trans)) {
 718                 uncorrectable = 1;
 719                 goto out;
 720         }
 721
 722         /*
 723          * the idea is to trigger a regular read through the standard path. we
 724          * read a page from the (failed) logical address by specifying the
 725          * corresponding copynum of the failed sector. thus, that readpage is
 726          * expected to fail.
 727          * that is the point where on-the-fly error correction will kick in
 728          * (once it's finished) and rewrite the failed sector if a good copy
 729          * can be found.
 730          */
 731         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 732                                                 path, scrub_fixup_readpage,
 733                                                 fixup);
 734         if (ret < 0) {
 735                 uncorrectable = 1;
 736                 goto out;
 737         }
 738         WARN_ON(ret != 1);
 739
 740         spin_lock(&sctx->stat_lock);
 741         ++sctx->stat.corrected_errors;
 742         spin_unlock(&sctx->stat_lock);
 743
 744 out:
 745         if (trans && !IS_ERR(trans))
 746                 btrfs_end_transaction(trans, fixup->root);
 747         if (uncorrectable) {
 748                 spin_lock(&sctx->stat_lock);
 749                 ++sctx->stat.uncorrectable_errors;
 750                 spin_unlock(&sctx->stat_lock);
 751                 btrfs_dev_replace_stats_inc(
 752                         &sctx->dev_root->fs_info->dev_replace.
 753                         num_uncorrectable_read_errors);
 754                 printk_ratelimited_in_rcu(KERN_ERR
 755                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 756                         (unsigned long long)fixup->logical,
 757                         rcu_str_deref(fixup->dev->name));
 758         }
 759
 760         btrfs_free_path(path);
 761         kfree(fixup);
 762
 763         scrub_pending_trans_workers_dec(sctx);
 764 }
 765
 766 /*
 767  * scrub_handle_errored_block gets called when either verification of the
 768  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 769  * case, this function handles all pages in the bio, even though only one
 770  * may be bad.
 771  * The goal of this function is to repair the errored block by using the
 772  * contents of one of the mirrors.
 773  */
 774 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 775 {
 776         struct scrub_ctx *sctx = sblock_to_check->sctx;
 777         struct btrfs_device *dev;
 778         struct btrfs_fs_info *fs_info;
 779         u64 length;
 780         u64 logical;
 781         u64 generation;
 782         unsigned int failed_mirror_index;
 783         unsigned int is_metadata;
 784         unsigned int have_csum;
 785         u8 *csum;
 786         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 787         struct scrub_block *sblock_bad;
 788         int ret;
 789         int mirror_index;
 790         int page_num;
 791         int success;
 792         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 793                                       DEFAULT_RATELIMIT_BURST);
 794
 795         BUG_ON(sblock_to_check->page_count < 1);
 796         fs_info = sctx->dev_root->fs_info;
 797         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 798                 /*
 799                  * if we find an error in a super block, we just report it.
 800                  * They will get written with the next transaction commit
 801                  * anyway
 802                  */
 803                 spin_lock(&sctx->stat_lock);
 804                 ++sctx->stat.super_errors;
 805                 spin_unlock(&sctx->stat_lock);
 806                 return 0;
 807         }
 808         length = sblock_to_check->page_count * PAGE_SIZE;
 809         logical = sblock_to_check->pagev[0]->logical;
 810         generation = sblock_to_check->pagev[0]->generation;
 811         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 812         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 813         is_metadata = !(sblock_to_check->pagev[0]->flags &
 814                         BTRFS_EXTENT_FLAG_DATA);
 815         have_csum = sblock_to_check->pagev[0]->have_csum;
 816         csum = sblock_to_check->pagev[0]->csum;
 817         dev = sblock_to_check->pagev[0]->dev;
 818
 819         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 820                 sblocks_for_recheck = NULL;
 821                 goto nodatasum_case;
 822         }
 823
 824         /*
 825          * read all mirrors one after the other. This includes to
 826          * re-read the extent or metadata block that failed (that was
 827          * the cause that this fixup code is called) another time,
 828          * page by page this time in order to know which pages
 829          * caused I/O errors and which ones are good (for all mirrors).
 830          * It is the goal to handle the situation when more than one
 831          * mirror contains I/O errors, but the errors do not
 832          * overlap, i.e. the data can be repaired by selecting the
 833          * pages from those mirrors without I/O error on the
 834          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 835          * would be that mirror #1 has an I/O error on the first page,
 836          * the second page is good, and mirror #2 has an I/O error on
 837          * the second page, but the first page is good.
 838          * Then the first page of the first mirror can be repaired by
 839          * taking the first page of the second mirror, and the
 840          * second page of the second mirror can be repaired by
 841          * copying the contents of the 2nd page of the 1st mirror.
 842          * One more note: if the pages of one mirror contain I/O
 843          * errors, the checksum cannot be verified. In order to get
 844          * the best data for repairing, the first attempt is to find
 845          * a mirror without I/O errors and with a validated checksum.
 846          * Only if this is not possible, the pages are picked from
 847          * mirrors with I/O errors without considering the checksum.
 848          * If the latter is the case, at the end, the checksum of the
 849          * repaired area is verified in order to correctly maintain
 850          * the statistics.
 851          */
 852
 853         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 854                                      sizeof(*sblocks_for_recheck),
 855                                      GFP_NOFS);
 856         if (!sblocks_for_recheck) {
 857                 spin_lock(&sctx->stat_lock);
 858                 sctx->stat.malloc_errors++;
 859                 sctx->stat.read_errors++;
 860                 sctx->stat.uncorrectable_errors++;
 861                 spin_unlock(&sctx->stat_lock);
 862                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 863                 goto out;
 864         }
 865
 866         /* setup the context, map the logical blocks and alloc the pages */
 867         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 868                                         logical, sblocks_for_recheck);
 869         if (ret) {
 870                 spin_lock(&sctx->stat_lock);
 871                 sctx->stat.read_errors++;
 872                 sctx->stat.uncorrectable_errors++;
 873                 spin_unlock(&sctx->stat_lock);
 874                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 875                 goto out;
 876         }
 877         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 878         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 879
 880         /* build and submit the bios for the failed mirror, check checksums */
 881         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 882                             csum, generation, sctx->csum_size);
 883
 884         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 885             sblock_bad->no_io_error_seen) {
 886                 /*
 887                  * the error disappeared after reading page by page, or
 888                  * the area was part of a huge bio and other parts of the
 889                  * bio caused I/O errors, or the block layer merged several
 890                  * read requests into one and the error is caused by a
 891                  * different bio (usually one of the two latter cases is
 892                  * the cause)
 893                  */
 894                 spin_lock(&sctx->stat_lock);
 895                 sctx->stat.unverified_errors++;
 896                 spin_unlock(&sctx->stat_lock);
 897
 898                 if (sctx->is_dev_replace)
 899                         scrub_write_block_to_dev_replace(sblock_bad);
 900                 goto out;
 901         }
 902
 903         if (!sblock_bad->no_io_error_seen) {
 904                 spin_lock(&sctx->stat_lock);
 905                 sctx->stat.read_errors++;
 906                 spin_unlock(&sctx->stat_lock);
 907                 if (__ratelimit(&_rs))
 908                         scrub_print_warning("i/o error", sblock_to_check);
 909                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 910         } else if (sblock_bad->checksum_error) {
 911                 spin_lock(&sctx->stat_lock);
 912                 sctx->stat.csum_errors++;
 913                 spin_unlock(&sctx->stat_lock);
 914                 if (__ratelimit(&_rs))
 915                         scrub_print_warning("checksum error", sblock_to_check);
 916                 btrfs_dev_stat_inc_and_print(dev,
 917                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 918         } else if (sblock_bad->header_error) {
 919                 spin_lock(&sctx->stat_lock);
 920                 sctx->stat.verify_errors++;
 921                 spin_unlock(&sctx->stat_lock);
 922                 if (__ratelimit(&_rs))
 923                         scrub_print_warning("checksum/header error",
 924                                             sblock_to_check);
 925                 if (sblock_bad->generation_error)
 926                         btrfs_dev_stat_inc_and_print(dev,
 927                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 928                 else
 929                         btrfs_dev_stat_inc_and_print(dev,
 930                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 931         }
 932
 933         if (sctx->readonly && !sctx->is_dev_replace)
 934                 goto did_not_correct_error;
 935
 936         if (!is_metadata && !have_csum) {
 937                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 938
 939 nodatasum_case:
 940                 WARN_ON(sctx->is_dev_replace);
 941
 942                 /*
 943                  * !is_metadata and !have_csum, this means that the data
 944                  * might not be COW'ed, that it might be modified
 945                  * concurrently. The general strategy to work on the
 946                  * commit root does not help in the case when COW is not
 947                  * used.
 948                  */
 949                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 950                 if (!fixup_nodatasum)
 951                         goto did_not_correct_error;
 952                 fixup_nodatasum->sctx = sctx;
 953                 fixup_nodatasum->dev = dev;
 954                 fixup_nodatasum->logical = logical;
 955                 fixup_nodatasum->root = fs_info->extent_root;
 956                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 957                 scrub_pending_trans_workers_inc(sctx);
 958                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 959                 btrfs_queue_worker(&fs_info->scrub_workers,
 960                                    &fixup_nodatasum->work);
 961                 goto out;
 962         }
 963
 964         /*
 965          * now build and submit the bios for the other mirrors, check
 966          * checksums.
 967          * First try to pick the mirror which is completely without I/O
 968          * errors and also does not have a checksum error.
 969          * If one is found, and if a checksum is present, the full block
 970          * that is known to contain an error is rewritten. Afterwards
 971          * the block is known to be corrected.
 972          * If a mirror is found which is completely correct, and no
 973          * checksum is present, only those pages are rewritten that had
 974          * an I/O error in the block to be repaired, since it cannot be
 975          * determined, which copy of the other pages is better (and it
 976          * could happen otherwise that a correct page would be
 977          * overwritten by a bad one).
 978          */
 979         for (mirror_index = 0;
 980              mirror_index < BTRFS_MAX_MIRRORS &&
 981              sblocks_for_recheck[mirror_index].page_count > 0;
 982              mirror_index++) {
 983                 struct scrub_block *sblock_other;
 984
 985                 if (mirror_index == failed_mirror_index)
 986                         continue;
 987                 sblock_other = sblocks_for_recheck + mirror_index;
 988
 989                 /* build and submit the bios, check checksums */
 990                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
 991                                     have_csum, csum, generation,
 992                                     sctx->csum_size);
 993
 994                 if (!sblock_other->header_error &&
 995                     !sblock_other->checksum_error &&
 996                     sblock_other->no_io_error_seen) {
 997                         if (sctx->is_dev_replace) {
 998                                 scrub_write_block_to_dev_replace(sblock_other);
 999                         } else {
1000                                 int force_write = is_metadata || have_csum;
1001
1002                                 ret = scrub_repair_block_from_good_copy(
1003                                                 sblock_bad, sblock_other,
1004                                                 force_write);
1005                         }
1006                         if (0 == ret)
1007                                 goto corrected_error;
1008                 }
1009         }
1010
1011         /*
1012          * for dev_replace, pick good pages and write to the target device.
1013          */
1014         if (sctx->is_dev_replace) {
1015                 success = 1;
1016                 for (page_num = 0; page_num < sblock_bad->page_count;
1017                      page_num++) {
1018                         int sub_success;
1019
1020                         sub_success = 0;
1021                         for (mirror_index = 0;
1022                              mirror_index < BTRFS_MAX_MIRRORS &&
1023                              sblocks_for_recheck[mirror_index].page_count > 0;
1024                              mirror_index++) {
1025                                 struct scrub_block *sblock_other =
1026                                         sblocks_for_recheck + mirror_index;
1027                                 struct scrub_page *page_other =
1028                                         sblock_other->pagev[page_num];
1029
1030                                 if (!page_other->io_error) {
1031                                         ret = scrub_write_page_to_dev_replace(
1032                                                         sblock_other, page_num);
1033                                         if (ret == 0) {
1034                                                 /* succeeded for this page */
1035                                                 sub_success = 1;
1036                                                 break;
1037                                         } else {
1038                                                 btrfs_dev_replace_stats_inc(
1039                                                         &sctx->dev_root->
1040                                                         fs_info->dev_replace.
1041                                                         num_write_errors);
1042                                         }
1043                                 }
1044                         }
1045
1046                         if (!sub_success) {
1047                                 /*
1048                                  * did not find a mirror to fetch the page
1049                                  * from. scrub_write_page_to_dev_replace()
1050                                  * handles this case (page->io_error), by
1051                                  * filling the block with zeros before
1052                                  * submitting the write request
1053                                  */
1054                                 success = 0;
1055                                 ret = scrub_write_page_to_dev_replace(
1056                                                 sblock_bad, page_num);
1057                                 if (ret)
1058                                         btrfs_dev_replace_stats_inc(
1059                                                 &sctx->dev_root->fs_info->
1060                                                 dev_replace.num_write_errors);
1061                         }
1062                 }
1063
1064                 goto out;
1065         }
1066
1067         /*
1068          * for regular scrub, repair those pages that are errored.
1069          * In case of I/O errors in the area that is supposed to be
1070          * repaired, continue by picking good copies of those pages.
1071          * Select the good pages from mirrors to rewrite bad pages from
1072          * the area to fix. Afterwards verify the checksum of the block
1073          * that is supposed to be repaired. This verification step is
1074          * only done for the purpose of statistic counting and for the
1075          * final scrub report, whether errors remain.
1076          * A perfect algorithm could make use of the checksum and try
1077          * all possible combinations of pages from the different mirrors
1078          * until the checksum verification succeeds. For example, when
1079          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1080          * of mirror #2 is readable but the final checksum test fails,
1081          * then the 2nd page of mirror #3 could be tried, whether now
1082          * the final checksum succeedes. But this would be a rare
1083          * exception and is therefore not implemented. At least it is
1084          * avoided that the good copy is overwritten.
1085          * A more useful improvement would be to pick the sectors
1086          * without I/O error based on sector sizes (512 bytes on legacy
1087          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1088          * mirror could be repaired by taking 512 byte of a different
1089          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1090          * area are unreadable.
1091          */
1092
1093         /* can only fix I/O errors from here on */
1094         if (sblock_bad->no_io_error_seen)
1095                 goto did_not_correct_error;
1096
1097         success = 1;
1098         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1099                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1100
1101                 if (!page_bad->io_error)
1102                         continue;
1103
1104                 for (mirror_index = 0;
1105                      mirror_index < BTRFS_MAX_MIRRORS &&
1106                      sblocks_for_recheck[mirror_index].page_count > 0;
1107                      mirror_index++) {
1108                         struct scrub_block *sblock_other = sblocks_for_recheck +
1109                                                            mirror_index;
1110                         struct scrub_page *page_other = sblock_other->pagev[
1111                                                         page_num];
1112
1113                         if (!page_other->io_error) {
1114                                 ret = scrub_repair_page_from_good_copy(
1115                                         sblock_bad, sblock_other, page_num, 0);
1116                                 if (0 == ret) {
1117                                         page_bad->io_error = 0;
1118                                         break; /* succeeded for this page */
1119                                 }
1120                         }
1121                 }
1122
1123                 if (page_bad->io_error) {
1124                         /* did not find a mirror to copy the page from */
1125                         success = 0;
1126                 }
1127         }
1128
1129         if (success) {
1130                 if (is_metadata || have_csum) {
1131                         /*
1132                          * need to verify the checksum now that all
1133                          * sectors on disk are repaired (the write
1134                          * request for data to be repaired is on its way).
1135                          * Just be lazy and use scrub_recheck_block()
1136                          * which re-reads the data before the checksum
1137                          * is verified, but most likely the data comes out
1138                          * of the page cache.
1139                          */
1140                         scrub_recheck_block(fs_info, sblock_bad,
1141                                             is_metadata, have_csum, csum,
1142                                             generation, sctx->csum_size);
1143                         if (!sblock_bad->header_error &&
1144                             !sblock_bad->checksum_error &&
1145                             sblock_bad->no_io_error_seen)
1146                                 goto corrected_error;
1147                         else
1148                                 goto did_not_correct_error;
1149                 } else {
1150 corrected_error:
1151                         spin_lock(&sctx->stat_lock);
1152                         sctx->stat.corrected_errors++;
1153                         spin_unlock(&sctx->stat_lock);
1154                         printk_ratelimited_in_rcu(KERN_ERR
1155                                 "btrfs: fixed up error at logical %llu on dev %s\n",
1156                                 (unsigned long long)logical,
1157                                 rcu_str_deref(dev->name));
1158                 }
1159         } else {
1160 did_not_correct_error:
1161                 spin_lock(&sctx->stat_lock);
1162                 sctx->stat.uncorrectable_errors++;
1163                 spin_unlock(&sctx->stat_lock);
1164                 printk_ratelimited_in_rcu(KERN_ERR
1165                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1166                         (unsigned long long)logical,
1167                         rcu_str_deref(dev->name));
1168         }
1169
1170 out:
1171         if (sblocks_for_recheck) {
1172                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1173                      mirror_index++) {
1174                         struct scrub_block *sblock = sblocks_for_recheck +
1175                                                      mirror_index;
1176                         int page_index;
1177
1178                         for (page_index = 0; page_index < sblock->page_count;
1179                              page_index++) {
1180                                 sblock->pagev[page_index]->sblock = NULL;
1181                                 scrub_page_put(sblock->pagev[page_index]);
1182                         }
1183                 }
1184                 kfree(sblocks_for_recheck);
1185         }
1186
1187         return 0;
1188 }
1189
1190 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1191                                      struct btrfs_fs_info *fs_info,
1192                                      struct scrub_block *original_sblock,
1193                                      u64 length, u64 logical,
1194                                      struct scrub_block *sblocks_for_recheck)
1195 {
1196         int page_index;
1197         int mirror_index;
1198         int ret;
1199
1200         /*
1201          * note: the two members ref_count and outstanding_pages
1202          * are not used (and not set) in the blocks that are used for
1203          * the recheck procedure
1204          */
1205
1206         page_index = 0;
1207         while (length > 0) {
1208                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1209                 u64 mapped_length = sublen;
1210                 struct btrfs_bio *bbio = NULL;
1211
1212                 /*
1213                  * with a length of PAGE_SIZE, each returned stripe
1214                  * represents one mirror
1215                  */
1216                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1217                                       &mapped_length, &bbio, 0);
1218                 if (ret || !bbio || mapped_length < sublen) {
1219                         kfree(bbio);
1220                         return -EIO;
1221                 }
1222
1223                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1224                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1225                      mirror_index++) {
1226                         struct scrub_block *sblock;
1227                         struct scrub_page *page;
1228
1229                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1230                                 continue;
1231
1232                         sblock = sblocks_for_recheck + mirror_index;
1233                         sblock->sctx = sctx;
1234                         page = kzalloc(sizeof(*page), GFP_NOFS);
1235                         if (!page) {
1236 leave_nomem:
1237                                 spin_lock(&sctx->stat_lock);
1238                                 sctx->stat.malloc_errors++;
1239                                 spin_unlock(&sctx->stat_lock);
1240                                 kfree(bbio);
1241                                 return -ENOMEM;
1242                         }
1243                         scrub_page_get(page);
1244                         sblock->pagev[page_index] = page;
1245                         page->logical = logical;
1246                         page->physical = bbio->stripes[mirror_index].physical;
1247                         BUG_ON(page_index >= original_sblock->page_count);
1248                         page->physical_for_dev_replace =
1249                                 original_sblock->pagev[page_index]->
1250                                 physical_for_dev_replace;
1251                         /* for missing devices, dev->bdev is NULL */
1252                         page->dev = bbio->stripes[mirror_index].dev;
1253                         page->mirror_num = mirror_index + 1;
1254                         sblock->page_count++;
1255                         page->page = alloc_page(GFP_NOFS);
1256                         if (!page->page)
1257                                 goto leave_nomem;
1258                 }
1259                 kfree(bbio);
1260                 length -= sublen;
1261                 logical += sublen;
1262                 page_index++;
1263         }
1264
1265         return 0;
1266 }
1267
1268 /*
1269  * this function will check the on disk data for checksum errors, header
1270  * errors and read I/O errors. If any I/O errors happen, the exact pages
1271  * which are errored are marked as being bad. The goal is to enable scrub
1272  * to take those pages that are not errored from all the mirrors so that
1273  * the pages that are errored in the just handled mirror can be repaired.
1274  */
1275 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1276                                 struct scrub_block *sblock, int is_metadata,
1277                                 int have_csum, u8 *csum, u64 generation,
1278                                 u16 csum_size)
1279 {
1280         int page_num;
1281
1282         sblock->no_io_error_seen = 1;
1283         sblock->header_error = 0;
1284         sblock->checksum_error = 0;
1285
1286         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1287                 struct bio *bio;
1288                 struct scrub_page *page = sblock->pagev[page_num];
1289                 DECLARE_COMPLETION_ONSTACK(complete);
1290
1291                 if (page->dev->bdev == NULL) {
1292                         page->io_error = 1;
1293                         sblock->no_io_error_seen = 0;
1294                         continue;
1295                 }
1296
1297                 WARN_ON(!page->page);
1298                 bio = bio_alloc(GFP_NOFS, 1);
1299                 if (!bio) {
1300                         page->io_error = 1;
1301                         sblock->no_io_error_seen = 0;
1302                         continue;
1303                 }
1304                 bio->bi_bdev = page->dev->bdev;
1305                 bio->bi_sector = page->physical >> 9;
1306                 bio->bi_end_io = scrub_complete_bio_end_io;
1307                 bio->bi_private = &complete;
1308
1309                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1310                 btrfsic_submit_bio(READ, bio);
1311
1312                 /* this will also unplug the queue */
1313                 wait_for_completion(&complete);
1314
1315                 page->io_error = !test_bit(BIO_UPTODATE, &bio->bi_flags);
1316                 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
1317                         sblock->no_io_error_seen = 0;
1318                 bio_put(bio);
1319         }
1320
1321         if (sblock->no_io_error_seen)
1322                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1323                                              have_csum, csum, generation,
1324                                              csum_size);
1325
1326         return;
1327 }
1328
1329 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1330                                          struct scrub_block *sblock,
1331                                          int is_metadata, int have_csum,
1332                                          const u8 *csum, u64 generation,
1333                                          u16 csum_size)
1334 {
1335         int page_num;
1336         u8 calculated_csum[BTRFS_CSUM_SIZE];
1337         u32 crc = ~(u32)0;
1338         struct btrfs_root *root = fs_info->extent_root;
1339         void *mapped_buffer;
1340
1341         WARN_ON(!sblock->pagev[0]->page);
1342         if (is_metadata) {
1343                 struct btrfs_header *h;
1344
1345                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1346                 h = (struct btrfs_header *)mapped_buffer;
1347
1348                 if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr) ||
1349                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1350                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1351                            BTRFS_UUID_SIZE)) {
1352                         sblock->header_error = 1;
1353                 } else if (generation != le64_to_cpu(h->generation)) {
1354                         sblock->header_error = 1;
1355                         sblock->generation_error = 1;
1356                 }
1357                 csum = h->csum;
1358         } else {
1359                 if (!have_csum)
1360                         return;
1361
1362                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1363         }
1364
1365         for (page_num = 0;;) {
1366                 if (page_num == 0 && is_metadata)
1367                         crc = btrfs_csum_data(root,
1368                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1369                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1370                 else
1371                         crc = btrfs_csum_data(root, mapped_buffer, crc,
1372                                               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 = 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 = 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         struct btrfs_root *root = sctx->dev_root;
1660         u64 len;
1661         int index;
1662
1663         BUG_ON(sblock->page_count < 1);
1664         if (!sblock->pagev[0]->have_csum)
1665                 return 0;
1666
1667         on_disk_csum = sblock->pagev[0]->csum;
1668         page = sblock->pagev[0]->page;
1669         buffer = kmap_atomic(page);
1670
1671         len = sctx->sectorsize;
1672         index = 0;
1673         for (;;) {
1674                 u64 l = min_t(u64, len, PAGE_SIZE);
1675
1676                 crc = btrfs_csum_data(root, buffer, crc, l);
1677                 kunmap_atomic(buffer);
1678                 len -= l;
1679                 if (len == 0)
1680                         break;
1681                 index++;
1682                 BUG_ON(index >= sblock->page_count);
1683                 BUG_ON(!sblock->pagev[index]->page);
1684                 page = sblock->pagev[index]->page;
1685                 buffer = kmap_atomic(page);
1686         }
1687
1688         btrfs_csum_final(crc, csum);
1689         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1690                 fail = 1;
1691
1692         return fail;
1693 }
1694
1695 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1696 {
1697         struct scrub_ctx *sctx = sblock->sctx;
1698         struct btrfs_header *h;
1699         struct btrfs_root *root = sctx->dev_root;
1700         struct btrfs_fs_info *fs_info = root->fs_info;
1701         u8 calculated_csum[BTRFS_CSUM_SIZE];
1702         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1703         struct page *page;
1704         void *mapped_buffer;
1705         u64 mapped_size;
1706         void *p;
1707         u32 crc = ~(u32)0;
1708         int fail = 0;
1709         int crc_fail = 0;
1710         u64 len;
1711         int index;
1712
1713         BUG_ON(sblock->page_count < 1);
1714         page = sblock->pagev[0]->page;
1715         mapped_buffer = kmap_atomic(page);
1716         h = (struct btrfs_header *)mapped_buffer;
1717         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1718
1719         /*
1720          * we don't use the getter functions here, as we
1721          * a) don't have an extent buffer and
1722          * b) the page is already kmapped
1723          */
1724
1725         if (sblock->pagev[0]->logical != le64_to_cpu(h->bytenr))
1726                 ++fail;
1727
1728         if (sblock->pagev[0]->generation != le64_to_cpu(h->generation))
1729                 ++fail;
1730
1731         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1732                 ++fail;
1733
1734         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1735                    BTRFS_UUID_SIZE))
1736                 ++fail;
1737
1738         WARN_ON(sctx->nodesize != sctx->leafsize);
1739         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1740         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1741         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1742         index = 0;
1743         for (;;) {
1744                 u64 l = min_t(u64, len, mapped_size);
1745
1746                 crc = btrfs_csum_data(root, p, crc, l);
1747                 kunmap_atomic(mapped_buffer);
1748                 len -= l;
1749                 if (len == 0)
1750                         break;
1751                 index++;
1752                 BUG_ON(index >= sblock->page_count);
1753                 BUG_ON(!sblock->pagev[index]->page);
1754                 page = sblock->pagev[index]->page;
1755                 mapped_buffer = kmap_atomic(page);
1756                 mapped_size = PAGE_SIZE;
1757                 p = mapped_buffer;
1758         }
1759
1760         btrfs_csum_final(crc, calculated_csum);
1761         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1762                 ++crc_fail;
1763
1764         return fail || crc_fail;
1765 }
1766
1767 static int scrub_checksum_super(struct scrub_block *sblock)
1768 {
1769         struct btrfs_super_block *s;
1770         struct scrub_ctx *sctx = sblock->sctx;
1771         struct btrfs_root *root = sctx->dev_root;
1772         struct btrfs_fs_info *fs_info = root->fs_info;
1773         u8 calculated_csum[BTRFS_CSUM_SIZE];
1774         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1775         struct page *page;
1776         void *mapped_buffer;
1777         u64 mapped_size;
1778         void *p;
1779         u32 crc = ~(u32)0;
1780         int fail_gen = 0;
1781         int fail_cor = 0;
1782         u64 len;
1783         int index;
1784
1785         BUG_ON(sblock->page_count < 1);
1786         page = sblock->pagev[0]->page;
1787         mapped_buffer = kmap_atomic(page);
1788         s = (struct btrfs_super_block *)mapped_buffer;
1789         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1790
1791         if (sblock->pagev[0]->logical != le64_to_cpu(s->bytenr))
1792                 ++fail_cor;
1793
1794         if (sblock->pagev[0]->generation != le64_to_cpu(s->generation))
1795                 ++fail_gen;
1796
1797         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1798                 ++fail_cor;
1799
1800         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1801         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1802         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1803         index = 0;
1804         for (;;) {
1805                 u64 l = min_t(u64, len, mapped_size);
1806
1807                 crc = btrfs_csum_data(root, p, crc, l);
1808                 kunmap_atomic(mapped_buffer);
1809                 len -= l;
1810                 if (len == 0)
1811                         break;
1812                 index++;
1813                 BUG_ON(index >= sblock->page_count);
1814                 BUG_ON(!sblock->pagev[index]->page);
1815                 page = sblock->pagev[index]->page;
1816                 mapped_buffer = kmap_atomic(page);
1817                 mapped_size = PAGE_SIZE;
1818                 p = mapped_buffer;
1819         }
1820
1821         btrfs_csum_final(crc, calculated_csum);
1822         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1823                 ++fail_cor;
1824
1825         if (fail_cor + fail_gen) {
1826                 /*
1827                  * if we find an error in a super block, we just report it.
1828                  * They will get written with the next transaction commit
1829                  * anyway
1830                  */
1831                 spin_lock(&sctx->stat_lock);
1832                 ++sctx->stat.super_errors;
1833                 spin_unlock(&sctx->stat_lock);
1834                 if (fail_cor)
1835                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1836                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1837                 else
1838                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1839                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1840         }
1841
1842         return fail_cor + fail_gen;
1843 }
1844
1845 static void scrub_block_get(struct scrub_block *sblock)
1846 {
1847         atomic_inc(&sblock->ref_count);
1848 }
1849
1850 static void scrub_block_put(struct scrub_block *sblock)
1851 {
1852         if (atomic_dec_and_test(&sblock->ref_count)) {
1853                 int i;
1854
1855                 for (i = 0; i < sblock->page_count; i++)
1856                         scrub_page_put(sblock->pagev[i]);
1857                 kfree(sblock);
1858         }
1859 }
1860
1861 static void scrub_page_get(struct scrub_page *spage)
1862 {
1863         atomic_inc(&spage->ref_count);
1864 }
1865
1866 static void scrub_page_put(struct scrub_page *spage)
1867 {
1868         if (atomic_dec_and_test(&spage->ref_count)) {
1869                 if (spage->page)
1870                         __free_page(spage->page);
1871                 kfree(spage);
1872         }
1873 }
1874
1875 static void scrub_submit(struct scrub_ctx *sctx)
1876 {
1877         struct scrub_bio *sbio;
1878
1879         if (sctx->curr == -1)
1880                 return;
1881
1882         sbio = sctx->bios[sctx->curr];
1883         sctx->curr = -1;
1884         scrub_pending_bio_inc(sctx);
1885
1886         if (!sbio->bio->bi_bdev) {
1887                 /*
1888                  * this case should not happen. If btrfs_map_block() is
1889                  * wrong, it could happen for dev-replace operations on
1890                  * missing devices when no mirrors are available, but in
1891                  * this case it should already fail the mount.
1892                  * This case is handled correctly (but _very_ slowly).
1893                  */
1894                 printk_ratelimited(KERN_WARNING
1895                         "btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1896                 bio_endio(sbio->bio, -EIO);
1897         } else {
1898                 btrfsic_submit_bio(READ, sbio->bio);
1899         }
1900 }
1901
1902 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1903                                     struct scrub_page *spage)
1904 {
1905         struct scrub_block *sblock = spage->sblock;
1906         struct scrub_bio *sbio;
1907         int ret;
1908
1909 again:
1910         /*
1911          * grab a fresh bio or wait for one to become available
1912          */
1913         while (sctx->curr == -1) {
1914                 spin_lock(&sctx->list_lock);
1915                 sctx->curr = sctx->first_free;
1916                 if (sctx->curr != -1) {
1917                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1918                         sctx->bios[sctx->curr]->next_free = -1;
1919                         sctx->bios[sctx->curr]->page_count = 0;
1920                         spin_unlock(&sctx->list_lock);
1921                 } else {
1922                         spin_unlock(&sctx->list_lock);
1923                         wait_event(sctx->list_wait, sctx->first_free != -1);
1924                 }
1925         }
1926         sbio = sctx->bios[sctx->curr];
1927         if (sbio->page_count == 0) {
1928                 struct bio *bio;
1929
1930                 sbio->physical = spage->physical;
1931                 sbio->logical = spage->logical;
1932                 sbio->dev = spage->dev;
1933                 bio = sbio->bio;
1934                 if (!bio) {
1935                         bio = bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1936                         if (!bio)
1937                                 return -ENOMEM;
1938                         sbio->bio = bio;
1939                 }
1940
1941                 bio->bi_private = sbio;
1942                 bio->bi_end_io = scrub_bio_end_io;
1943                 bio->bi_bdev = sbio->dev->bdev;
1944                 bio->bi_sector = sbio->physical >> 9;
1945                 sbio->err = 0;
1946         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1947                    spage->physical ||
1948                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1949                    spage->logical ||
1950                    sbio->dev != spage->dev) {
1951                 scrub_submit(sctx);
1952                 goto again;
1953         }
1954
1955         sbio->pagev[sbio->page_count] = spage;
1956         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1957         if (ret != PAGE_SIZE) {
1958                 if (sbio->page_count < 1) {
1959                         bio_put(sbio->bio);
1960                         sbio->bio = NULL;
1961                         return -EIO;
1962                 }
1963                 scrub_submit(sctx);
1964                 goto again;
1965         }
1966
1967         scrub_block_get(sblock); /* one for the page added to the bio */
1968         atomic_inc(&sblock->outstanding_pages);
1969         sbio->page_count++;
1970         if (sbio->page_count == sctx->pages_per_rd_bio)
1971                 scrub_submit(sctx);
1972
1973         return 0;
1974 }
1975
1976 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1977                        u64 physical, struct btrfs_device *dev, u64 flags,
1978                        u64 gen, int mirror_num, u8 *csum, int force,
1979                        u64 physical_for_dev_replace)
1980 {
1981         struct scrub_block *sblock;
1982         int index;
1983
1984         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1985         if (!sblock) {
1986                 spin_lock(&sctx->stat_lock);
1987                 sctx->stat.malloc_errors++;
1988                 spin_unlock(&sctx->stat_lock);
1989                 return -ENOMEM;
1990         }
1991
1992         /* one ref inside this function, plus one for each page added to
1993          * a bio later on */
1994         atomic_set(&sblock->ref_count, 1);
1995         sblock->sctx = sctx;
1996         sblock->no_io_error_seen = 1;
1997
1998         for (index = 0; len > 0; index++) {
1999                 struct scrub_page *spage;
2000                 u64 l = min_t(u64, len, PAGE_SIZE);
2001
2002                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2003                 if (!spage) {
2004 leave_nomem:
2005                         spin_lock(&sctx->stat_lock);
2006                         sctx->stat.malloc_errors++;
2007                         spin_unlock(&sctx->stat_lock);
2008                         scrub_block_put(sblock);
2009                         return -ENOMEM;
2010                 }
2011                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2012                 scrub_page_get(spage);
2013                 sblock->pagev[index] = spage;
2014                 spage->sblock = sblock;
2015                 spage->dev = dev;
2016                 spage->flags = flags;
2017                 spage->generation = gen;
2018                 spage->logical = logical;
2019                 spage->physical = physical;
2020                 spage->physical_for_dev_replace = physical_for_dev_replace;
2021                 spage->mirror_num = mirror_num;
2022                 if (csum) {
2023                         spage->have_csum = 1;
2024                         memcpy(spage->csum, csum, sctx->csum_size);
2025                 } else {
2026                         spage->have_csum = 0;
2027                 }
2028                 sblock->page_count++;
2029                 spage->page = alloc_page(GFP_NOFS);
2030                 if (!spage->page)
2031                         goto leave_nomem;
2032                 len -= l;
2033                 logical += l;
2034                 physical += l;
2035                 physical_for_dev_replace += l;
2036         }
2037
2038         WARN_ON(sblock->page_count == 0);
2039         for (index = 0; index < sblock->page_count; index++) {
2040                 struct scrub_page *spage = sblock->pagev[index];
2041                 int ret;
2042
2043                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2044                 if (ret) {
2045                         scrub_block_put(sblock);
2046                         return ret;
2047                 }
2048         }
2049
2050         if (force)
2051                 scrub_submit(sctx);
2052
2053         /* last one frees, either here or in bio completion for last page */
2054         scrub_block_put(sblock);
2055         return 0;
2056 }
2057
2058 static void scrub_bio_end_io(struct bio *bio, int err)
2059 {
2060         struct scrub_bio *sbio = bio->bi_private;
2061         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2062
2063         sbio->err = err;
2064         sbio->bio = bio;
2065
2066         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2067 }
2068
2069 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2070 {
2071         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2072         struct scrub_ctx *sctx = sbio->sctx;
2073         int i;
2074
2075         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2076         if (sbio->err) {
2077                 for (i = 0; i < sbio->page_count; i++) {
2078                         struct scrub_page *spage = sbio->pagev[i];
2079
2080                         spage->io_error = 1;
2081                         spage->sblock->no_io_error_seen = 0;
2082                 }
2083         }
2084
2085         /* now complete the scrub_block items that have all pages completed */
2086         for (i = 0; i < sbio->page_count; i++) {
2087                 struct scrub_page *spage = sbio->pagev[i];
2088                 struct scrub_block *sblock = spage->sblock;
2089
2090                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2091                         scrub_block_complete(sblock);
2092                 scrub_block_put(sblock);
2093         }
2094
2095         bio_put(sbio->bio);
2096         sbio->bio = NULL;
2097         spin_lock(&sctx->list_lock);
2098         sbio->next_free = sctx->first_free;
2099         sctx->first_free = sbio->index;
2100         spin_unlock(&sctx->list_lock);
2101
2102         if (sctx->is_dev_replace &&
2103             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2104                 mutex_lock(&sctx->wr_ctx.wr_lock);
2105                 scrub_wr_submit(sctx);
2106                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2107         }
2108
2109         scrub_pending_bio_dec(sctx);
2110 }
2111
2112 static void scrub_block_complete(struct scrub_block *sblock)
2113 {
2114         if (!sblock->no_io_error_seen) {
2115                 scrub_handle_errored_block(sblock);
2116         } else {
2117                 /*
2118                  * if has checksum error, write via repair mechanism in
2119                  * dev replace case, otherwise write here in dev replace
2120                  * case.
2121                  */
2122                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2123                         scrub_write_block_to_dev_replace(sblock);
2124         }
2125 }
2126
2127 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2128                            u8 *csum)
2129 {
2130         struct btrfs_ordered_sum *sum = NULL;
2131         int ret = 0;
2132         unsigned long i;
2133         unsigned long num_sectors;
2134
2135         while (!list_empty(&sctx->csum_list)) {
2136                 sum = list_first_entry(&sctx->csum_list,
2137                                        struct btrfs_ordered_sum, list);
2138                 if (sum->bytenr > logical)
2139                         return 0;
2140                 if (sum->bytenr + sum->len > logical)
2141                         break;
2142
2143                 ++sctx->stat.csum_discards;
2144                 list_del(&sum->list);
2145                 kfree(sum);
2146                 sum = NULL;
2147         }
2148         if (!sum)
2149                 return 0;
2150
2151         num_sectors = sum->len / sctx->sectorsize;
2152         for (i = 0; i < num_sectors; ++i) {
2153                 if (sum->sums[i].bytenr == logical) {
2154                         memcpy(csum, &sum->sums[i].sum, sctx->csum_size);
2155                         ret = 1;
2156                         break;
2157                 }
2158         }
2159         if (ret && i == num_sectors - 1) {
2160                 list_del(&sum->list);
2161                 kfree(sum);
2162         }
2163         return ret;
2164 }
2165
2166 /* scrub extent tries to collect up to 64 kB for each bio */
2167 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2168                         u64 physical, struct btrfs_device *dev, u64 flags,
2169                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2170 {
2171         int ret;
2172         u8 csum[BTRFS_CSUM_SIZE];
2173         u32 blocksize;
2174
2175         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2176                 blocksize = sctx->sectorsize;
2177                 spin_lock(&sctx->stat_lock);
2178                 sctx->stat.data_extents_scrubbed++;
2179                 sctx->stat.data_bytes_scrubbed += len;
2180                 spin_unlock(&sctx->stat_lock);
2181         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2182                 WARN_ON(sctx->nodesize != sctx->leafsize);
2183                 blocksize = sctx->nodesize;
2184                 spin_lock(&sctx->stat_lock);
2185                 sctx->stat.tree_extents_scrubbed++;
2186                 sctx->stat.tree_bytes_scrubbed += len;
2187                 spin_unlock(&sctx->stat_lock);
2188         } else {
2189                 blocksize = sctx->sectorsize;
2190                 WARN_ON(1);
2191         }
2192
2193         while (len) {
2194                 u64 l = min_t(u64, len, blocksize);
2195                 int have_csum = 0;
2196
2197                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2198                         /* push csums to sbio */
2199                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2200                         if (have_csum == 0)
2201                                 ++sctx->stat.no_csum;
2202                         if (sctx->is_dev_replace && !have_csum) {
2203                                 ret = copy_nocow_pages(sctx, logical, l,
2204                                                        mirror_num,
2205                                                       physical_for_dev_replace);
2206                                 goto behind_scrub_pages;
2207                         }
2208                 }
2209                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2210                                   mirror_num, have_csum ? csum : NULL, 0,
2211                                   physical_for_dev_replace);
2212 behind_scrub_pages:
2213                 if (ret)
2214                         return ret;
2215                 len -= l;
2216                 logical += l;
2217                 physical += l;
2218                 physical_for_dev_replace += l;
2219         }
2220         return 0;
2221 }
2222
2223 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2224                                            struct map_lookup *map,
2225                                            struct btrfs_device *scrub_dev,
2226                                            int num, u64 base, u64 length,
2227                                            int is_dev_replace)
2228 {
2229         struct btrfs_path *path;
2230         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2231         struct btrfs_root *root = fs_info->extent_root;
2232         struct btrfs_root *csum_root = fs_info->csum_root;
2233         struct btrfs_extent_item *extent;
2234         struct blk_plug plug;
2235         u64 flags;
2236         int ret;
2237         int slot;
2238         int i;
2239         u64 nstripes;
2240         struct extent_buffer *l;
2241         struct btrfs_key key;
2242         u64 physical;
2243         u64 logical;
2244         u64 generation;
2245         int mirror_num;
2246         struct reada_control *reada1;
2247         struct reada_control *reada2;
2248         struct btrfs_key key_start;
2249         struct btrfs_key key_end;
2250         u64 increment = map->stripe_len;
2251         u64 offset;
2252         u64 extent_logical;
2253         u64 extent_physical;
2254         u64 extent_len;
2255         struct btrfs_device *extent_dev;
2256         int extent_mirror_num;
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_EXTENT_ITEM_KEY;
2318         key_end.offset = (u64)0;
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         ret = 0;
2357         for (i = 0; i < nstripes; ++i) {
2358                 /*
2359                  * canceled?
2360                  */
2361                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2362                     atomic_read(&sctx->cancel_req)) {
2363                         ret = -ECANCELED;
2364                         goto out;
2365                 }
2366                 /*
2367                  * check to see if we have to pause
2368                  */
2369                 if (atomic_read(&fs_info->scrub_pause_req)) {
2370                         /* push queued extents */
2371                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2372                         scrub_submit(sctx);
2373                         mutex_lock(&sctx->wr_ctx.wr_lock);
2374                         scrub_wr_submit(sctx);
2375                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2376                         wait_event(sctx->list_wait,
2377                                    atomic_read(&sctx->bios_in_flight) == 0);
2378                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2379                         atomic_inc(&fs_info->scrubs_paused);
2380                         wake_up(&fs_info->scrub_pause_wait);
2381                         mutex_lock(&fs_info->scrub_lock);
2382                         while (atomic_read(&fs_info->scrub_pause_req)) {
2383                                 mutex_unlock(&fs_info->scrub_lock);
2384                                 wait_event(fs_info->scrub_pause_wait,
2385                                    atomic_read(&fs_info->scrub_pause_req) == 0);
2386                                 mutex_lock(&fs_info->scrub_lock);
2387                         }
2388                         atomic_dec(&fs_info->scrubs_paused);
2389                         mutex_unlock(&fs_info->scrub_lock);
2390                         wake_up(&fs_info->scrub_pause_wait);
2391                 }
2392
2393                 ret = btrfs_lookup_csums_range(csum_root, logical,
2394                                                logical + map->stripe_len - 1,
2395                                                &sctx->csum_list, 1);
2396                 if (ret)
2397                         goto out;
2398
2399                 key.objectid = logical;
2400                 key.type = BTRFS_EXTENT_ITEM_KEY;
2401                 key.offset = (u64)0;
2402
2403                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2404                 if (ret < 0)
2405                         goto out;
2406                 if (ret > 0) {
2407                         ret = btrfs_previous_item(root, path, 0,
2408                                                   BTRFS_EXTENT_ITEM_KEY);
2409                         if (ret < 0)
2410                                 goto out;
2411                         if (ret > 0) {
2412                                 /* there's no smaller item, so stick with the
2413                                  * larger one */
2414                                 btrfs_release_path(path);
2415                                 ret = btrfs_search_slot(NULL, root, &key,
2416                                                         path, 0, 0);
2417                                 if (ret < 0)
2418                                         goto out;
2419                         }
2420                 }
2421
2422                 while (1) {
2423                         l = path->nodes[0];
2424                         slot = path->slots[0];
2425                         if (slot >= btrfs_header_nritems(l)) {
2426                                 ret = btrfs_next_leaf(root, path);
2427                                 if (ret == 0)
2428                                         continue;
2429                                 if (ret < 0)
2430                                         goto out;
2431
2432                                 break;
2433                         }
2434                         btrfs_item_key_to_cpu(l, &key, slot);
2435
2436                         if (key.objectid + key.offset <= logical)
2437                                 goto next;
2438
2439                         if (key.objectid >= logical + map->stripe_len)
2440                                 break;
2441
2442                         if (btrfs_key_type(&key) != BTRFS_EXTENT_ITEM_KEY)
2443                                 goto next;
2444
2445                         extent = btrfs_item_ptr(l, slot,
2446                                                 struct btrfs_extent_item);
2447                         flags = btrfs_extent_flags(l, extent);
2448                         generation = btrfs_extent_generation(l, extent);
2449
2450                         if (key.objectid < logical &&
2451                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2452                                 printk(KERN_ERR
2453                                        "btrfs scrub: tree block %llu spanning "
2454                                        "stripes, ignored. logical=%llu\n",
2455                                        (unsigned long long)key.objectid,
2456                                        (unsigned long long)logical);
2457                                 goto next;
2458                         }
2459
2460                         /*
2461                          * trim extent to this stripe
2462                          */
2463                         if (key.objectid < logical) {
2464                                 key.offset -= logical - key.objectid;
2465                                 key.objectid = logical;
2466                         }
2467                         if (key.objectid + key.offset >
2468                             logical + map->stripe_len) {
2469                                 key.offset = logical + map->stripe_len -
2470                                              key.objectid;
2471                         }
2472
2473                         extent_logical = key.objectid;
2474                         extent_physical = key.objectid - logical + physical;
2475                         extent_len = key.offset;
2476                         extent_dev = scrub_dev;
2477                         extent_mirror_num = mirror_num;
2478                         if (is_dev_replace)
2479                                 scrub_remap_extent(fs_info, extent_logical,
2480                                                    extent_len, &extent_physical,
2481                                                    &extent_dev,
2482                                                    &extent_mirror_num);
2483                         ret = scrub_extent(sctx, extent_logical, extent_len,
2484                                            extent_physical, extent_dev, flags,
2485                                            generation, extent_mirror_num,
2486                                            key.objectid - logical + physical);
2487                         if (ret)
2488                                 goto out;
2489
2490 next:
2491                         path->slots[0]++;
2492                 }
2493                 btrfs_release_path(path);
2494                 logical += increment;
2495                 physical += map->stripe_len;
2496                 spin_lock(&sctx->stat_lock);
2497                 sctx->stat.last_physical = physical;
2498                 spin_unlock(&sctx->stat_lock);
2499         }
2500 out:
2501         /* push queued extents */
2502         scrub_submit(sctx);
2503         mutex_lock(&sctx->wr_ctx.wr_lock);
2504         scrub_wr_submit(sctx);
2505         mutex_unlock(&sctx->wr_ctx.wr_lock);
2506
2507         blk_finish_plug(&plug);
2508         btrfs_free_path(path);
2509         return ret < 0 ? ret : 0;
2510 }
2511
2512 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2513                                           struct btrfs_device *scrub_dev,
2514                                           u64 chunk_tree, u64 chunk_objectid,
2515                                           u64 chunk_offset, u64 length,
2516                                           u64 dev_offset, int is_dev_replace)
2517 {
2518         struct btrfs_mapping_tree *map_tree =
2519                 &sctx->dev_root->fs_info->mapping_tree;
2520         struct map_lookup *map;
2521         struct extent_map *em;
2522         int i;
2523         int ret = 0;
2524
2525         read_lock(&map_tree->map_tree.lock);
2526         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2527         read_unlock(&map_tree->map_tree.lock);
2528
2529         if (!em)
2530                 return -EINVAL;
2531
2532         map = (struct map_lookup *)em->bdev;
2533         if (em->start != chunk_offset)
2534                 goto out;
2535
2536         if (em->len < length)
2537                 goto out;
2538
2539         for (i = 0; i < map->num_stripes; ++i) {
2540                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2541                     map->stripes[i].physical == dev_offset) {
2542                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2543                                            chunk_offset, length,
2544                                            is_dev_replace);
2545                         if (ret)
2546                                 goto out;
2547                 }
2548         }
2549 out:
2550         free_extent_map(em);
2551
2552         return ret;
2553 }
2554
2555 static noinline_for_stack
2556 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2557                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2558                            int is_dev_replace)
2559 {
2560         struct btrfs_dev_extent *dev_extent = NULL;
2561         struct btrfs_path *path;
2562         struct btrfs_root *root = sctx->dev_root;
2563         struct btrfs_fs_info *fs_info = root->fs_info;
2564         u64 length;
2565         u64 chunk_tree;
2566         u64 chunk_objectid;
2567         u64 chunk_offset;
2568         int ret;
2569         int slot;
2570         struct extent_buffer *l;
2571         struct btrfs_key key;
2572         struct btrfs_key found_key;
2573         struct btrfs_block_group_cache *cache;
2574         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2575
2576         path = btrfs_alloc_path();
2577         if (!path)
2578                 return -ENOMEM;
2579
2580         path->reada = 2;
2581         path->search_commit_root = 1;
2582         path->skip_locking = 1;
2583
2584         key.objectid = scrub_dev->devid;
2585         key.offset = 0ull;
2586         key.type = BTRFS_DEV_EXTENT_KEY;
2587
2588         while (1) {
2589                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2590                 if (ret < 0)
2591                         break;
2592                 if (ret > 0) {
2593                         if (path->slots[0] >=
2594                             btrfs_header_nritems(path->nodes[0])) {
2595                                 ret = btrfs_next_leaf(root, path);
2596                                 if (ret)
2597                                         break;
2598                         }
2599                 }
2600
2601                 l = path->nodes[0];
2602                 slot = path->slots[0];
2603
2604                 btrfs_item_key_to_cpu(l, &found_key, slot);
2605
2606                 if (found_key.objectid != scrub_dev->devid)
2607                         break;
2608
2609                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2610                         break;
2611
2612                 if (found_key.offset >= end)
2613                         break;
2614
2615                 if (found_key.offset < key.offset)
2616                         break;
2617
2618                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2619                 length = btrfs_dev_extent_length(l, dev_extent);
2620
2621                 if (found_key.offset + length <= start) {
2622                         key.offset = found_key.offset + length;
2623                         btrfs_release_path(path);
2624                         continue;
2625                 }
2626
2627                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2628                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2629                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2630
2631                 /*
2632                  * get a reference on the corresponding block group to prevent
2633                  * the chunk from going away while we scrub it
2634                  */
2635                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2636                 if (!cache) {
2637                         ret = -ENOENT;
2638                         break;
2639                 }
2640                 dev_replace->cursor_right = found_key.offset + length;
2641                 dev_replace->cursor_left = found_key.offset;
2642                 dev_replace->item_needs_writeback = 1;
2643                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2644                                   chunk_offset, length, found_key.offset,
2645                                   is_dev_replace);
2646
2647                 /*
2648                  * flush, submit all pending read and write bios, afterwards
2649                  * wait for them.
2650                  * Note that in the dev replace case, a read request causes
2651                  * write requests that are submitted in the read completion
2652                  * worker. Therefore in the current situation, it is required
2653                  * that all write requests are flushed, so that all read and
2654                  * write requests are really completed when bios_in_flight
2655                  * changes to 0.
2656                  */
2657                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2658                 scrub_submit(sctx);
2659                 mutex_lock(&sctx->wr_ctx.wr_lock);
2660                 scrub_wr_submit(sctx);
2661                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2662
2663                 wait_event(sctx->list_wait,
2664                            atomic_read(&sctx->bios_in_flight) == 0);
2665                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2666                 atomic_inc(&fs_info->scrubs_paused);
2667                 wake_up(&fs_info->scrub_pause_wait);
2668                 wait_event(sctx->list_wait,
2669                            atomic_read(&sctx->workers_pending) == 0);
2670
2671                 mutex_lock(&fs_info->scrub_lock);
2672                 while (atomic_read(&fs_info->scrub_pause_req)) {
2673                         mutex_unlock(&fs_info->scrub_lock);
2674                         wait_event(fs_info->scrub_pause_wait,
2675                            atomic_read(&fs_info->scrub_pause_req) == 0);
2676                         mutex_lock(&fs_info->scrub_lock);
2677                 }
2678                 atomic_dec(&fs_info->scrubs_paused);
2679                 mutex_unlock(&fs_info->scrub_lock);
2680                 wake_up(&fs_info->scrub_pause_wait);
2681
2682                 dev_replace->cursor_left = dev_replace->cursor_right;
2683                 dev_replace->item_needs_writeback = 1;
2684                 btrfs_put_block_group(cache);
2685                 if (ret)
2686                         break;
2687                 if (is_dev_replace &&
2688                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2689                         ret = -EIO;
2690                         break;
2691                 }
2692                 if (sctx->stat.malloc_errors > 0) {
2693                         ret = -ENOMEM;
2694                         break;
2695                 }
2696
2697                 key.offset = found_key.offset + length;
2698                 btrfs_release_path(path);
2699         }
2700
2701         btrfs_free_path(path);
2702
2703         /*
2704          * ret can still be 1 from search_slot or next_leaf,
2705          * that's not an error
2706          */
2707         return ret < 0 ? ret : 0;
2708 }
2709
2710 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2711                                            struct btrfs_device *scrub_dev)
2712 {
2713         int     i;
2714         u64     bytenr;
2715         u64     gen;
2716         int     ret;
2717         struct btrfs_root *root = sctx->dev_root;
2718
2719         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2720                 return -EIO;
2721
2722         gen = root->fs_info->last_trans_committed;
2723
2724         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2725                 bytenr = btrfs_sb_offset(i);
2726                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2727                         break;
2728
2729                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2730                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2731                                   NULL, 1, bytenr);
2732                 if (ret)
2733                         return ret;
2734         }
2735         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2736
2737         return 0;
2738 }
2739
2740 /*
2741  * get a reference count on fs_info->scrub_workers. start worker if necessary
2742  */
2743 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2744                                                 int is_dev_replace)
2745 {
2746         int ret = 0;
2747
2748         mutex_lock(&fs_info->scrub_lock);
2749         if (fs_info->scrub_workers_refcnt == 0) {
2750                 if (is_dev_replace)
2751                         btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2752                                         &fs_info->generic_worker);
2753                 else
2754                         btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2755                                         fs_info->thread_pool_size,
2756                                         &fs_info->generic_worker);
2757                 fs_info->scrub_workers.idle_thresh = 4;
2758                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2759                 if (ret)
2760                         goto out;
2761                 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2762                                    "scrubwrc",
2763                                    fs_info->thread_pool_size,
2764                                    &fs_info->generic_worker);
2765                 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2766                 ret = btrfs_start_workers(
2767                                 &fs_info->scrub_wr_completion_workers);
2768                 if (ret)
2769                         goto out;
2770                 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2771                                    &fs_info->generic_worker);
2772                 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2773                 if (ret)
2774                         goto out;
2775         }
2776         ++fs_info->scrub_workers_refcnt;
2777 out:
2778         mutex_unlock(&fs_info->scrub_lock);
2779
2780         return ret;
2781 }
2782
2783 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2784 {
2785         mutex_lock(&fs_info->scrub_lock);
2786         if (--fs_info->scrub_workers_refcnt == 0) {
2787                 btrfs_stop_workers(&fs_info->scrub_workers);
2788                 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2789                 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2790         }
2791         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2792         mutex_unlock(&fs_info->scrub_lock);
2793 }
2794
2795 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2796                     u64 end, struct btrfs_scrub_progress *progress,
2797                     int readonly, int is_dev_replace)
2798 {
2799         struct scrub_ctx *sctx;
2800         int ret;
2801         struct btrfs_device *dev;
2802
2803         if (btrfs_fs_closing(fs_info))
2804                 return -EINVAL;
2805
2806         /*
2807          * check some assumptions
2808          */
2809         if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2810                 printk(KERN_ERR
2811                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2812                        fs_info->chunk_root->nodesize,
2813                        fs_info->chunk_root->leafsize);
2814                 return -EINVAL;
2815         }
2816
2817         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2818                 /*
2819                  * in this case scrub is unable to calculate the checksum
2820                  * the way scrub is implemented. Do not handle this
2821                  * situation at all because it won't ever happen.
2822                  */
2823                 printk(KERN_ERR
2824                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2825                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2826                 return -EINVAL;
2827         }
2828
2829         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2830                 /* not supported for data w/o checksums */
2831                 printk(KERN_ERR
2832                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lld) fails\n",
2833                        fs_info->chunk_root->sectorsize,
2834                        (unsigned long long)PAGE_SIZE);
2835                 return -EINVAL;
2836         }
2837
2838         if (fs_info->chunk_root->nodesize >
2839             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2840             fs_info->chunk_root->sectorsize >
2841             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2842                 /*
2843                  * would exhaust the array bounds of pagev member in
2844                  * struct scrub_block
2845                  */
2846                 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2847                        fs_info->chunk_root->nodesize,
2848                        SCRUB_MAX_PAGES_PER_BLOCK,
2849                        fs_info->chunk_root->sectorsize,
2850                        SCRUB_MAX_PAGES_PER_BLOCK);
2851                 return -EINVAL;
2852         }
2853
2854         ret = scrub_workers_get(fs_info, is_dev_replace);
2855         if (ret)
2856                 return ret;
2857
2858         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2859         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2860         if (!dev || (dev->missing && !is_dev_replace)) {
2861                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2862                 scrub_workers_put(fs_info);
2863                 return -ENODEV;
2864         }
2865         mutex_lock(&fs_info->scrub_lock);
2866
2867         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2868                 mutex_unlock(&fs_info->scrub_lock);
2869                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2870                 scrub_workers_put(fs_info);
2871                 return -EIO;
2872         }
2873
2874         btrfs_dev_replace_lock(&fs_info->dev_replace);
2875         if (dev->scrub_device ||
2876             (!is_dev_replace &&
2877              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2878                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2879                 mutex_unlock(&fs_info->scrub_lock);
2880                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2881                 scrub_workers_put(fs_info);
2882                 return -EINPROGRESS;
2883         }
2884         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2885         sctx = scrub_setup_ctx(dev, is_dev_replace);
2886         if (IS_ERR(sctx)) {
2887                 mutex_unlock(&fs_info->scrub_lock);
2888                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2889                 scrub_workers_put(fs_info);
2890                 return PTR_ERR(sctx);
2891         }
2892         sctx->readonly = readonly;
2893         dev->scrub_device = sctx;
2894
2895         atomic_inc(&fs_info->scrubs_running);
2896         mutex_unlock(&fs_info->scrub_lock);
2897         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2898
2899         if (!is_dev_replace) {
2900                 down_read(&fs_info->scrub_super_lock);
2901                 ret = scrub_supers(sctx, dev);
2902                 up_read(&fs_info->scrub_super_lock);
2903         }
2904
2905         if (!ret)
2906                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2907                                              is_dev_replace);
2908
2909         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2910         atomic_dec(&fs_info->scrubs_running);
2911         wake_up(&fs_info->scrub_pause_wait);
2912
2913         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2914
2915         if (progress)
2916                 memcpy(progress, &sctx->stat, sizeof(*progress));
2917
2918         mutex_lock(&fs_info->scrub_lock);
2919         dev->scrub_device = NULL;
2920         mutex_unlock(&fs_info->scrub_lock);
2921
2922         scrub_free_ctx(sctx);
2923         scrub_workers_put(fs_info);
2924
2925         return ret;
2926 }
2927
2928 void btrfs_scrub_pause(struct btrfs_root *root)
2929 {
2930         struct btrfs_fs_info *fs_info = root->fs_info;
2931
2932         mutex_lock(&fs_info->scrub_lock);
2933         atomic_inc(&fs_info->scrub_pause_req);
2934         while (atomic_read(&fs_info->scrubs_paused) !=
2935                atomic_read(&fs_info->scrubs_running)) {
2936                 mutex_unlock(&fs_info->scrub_lock);
2937                 wait_event(fs_info->scrub_pause_wait,
2938                            atomic_read(&fs_info->scrubs_paused) ==
2939                            atomic_read(&fs_info->scrubs_running));
2940                 mutex_lock(&fs_info->scrub_lock);
2941         }
2942         mutex_unlock(&fs_info->scrub_lock);
2943 }
2944
2945 void btrfs_scrub_continue(struct btrfs_root *root)
2946 {
2947         struct btrfs_fs_info *fs_info = root->fs_info;
2948
2949         atomic_dec(&fs_info->scrub_pause_req);
2950         wake_up(&fs_info->scrub_pause_wait);
2951 }
2952
2953 void btrfs_scrub_pause_super(struct btrfs_root *root)
2954 {
2955         down_write(&root->fs_info->scrub_super_lock);
2956 }
2957
2958 void btrfs_scrub_continue_super(struct btrfs_root *root)
2959 {
2960         up_write(&root->fs_info->scrub_super_lock);
2961 }
2962
2963 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2964 {
2965         mutex_lock(&fs_info->scrub_lock);
2966         if (!atomic_read(&fs_info->scrubs_running)) {
2967                 mutex_unlock(&fs_info->scrub_lock);
2968                 return -ENOTCONN;
2969         }
2970
2971         atomic_inc(&fs_info->scrub_cancel_req);
2972         while (atomic_read(&fs_info->scrubs_running)) {
2973                 mutex_unlock(&fs_info->scrub_lock);
2974                 wait_event(fs_info->scrub_pause_wait,
2975                            atomic_read(&fs_info->scrubs_running) == 0);
2976                 mutex_lock(&fs_info->scrub_lock);
2977         }
2978         atomic_dec(&fs_info->scrub_cancel_req);
2979         mutex_unlock(&fs_info->scrub_lock);
2980
2981         return 0;
2982 }
2983
2984 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2985                            struct btrfs_device *dev)
2986 {
2987         struct scrub_ctx *sctx;
2988
2989         mutex_lock(&fs_info->scrub_lock);
2990         sctx = dev->scrub_device;
2991         if (!sctx) {
2992                 mutex_unlock(&fs_info->scrub_lock);
2993                 return -ENOTCONN;
2994         }
2995         atomic_inc(&sctx->cancel_req);
2996         while (dev->scrub_device) {
2997                 mutex_unlock(&fs_info->scrub_lock);
2998                 wait_event(fs_info->scrub_pause_wait,
2999                            dev->scrub_device == NULL);
3000                 mutex_lock(&fs_info->scrub_lock);
3001         }
3002         mutex_unlock(&fs_info->scrub_lock);
3003
3004         return 0;
3005 }
3006
3007 int btrfs_scrub_cancel_devid(struct btrfs_root *root, u64 devid)
3008 {
3009         struct btrfs_fs_info *fs_info = root->fs_info;
3010         struct btrfs_device *dev;
3011         int ret;
3012
3013         /*
3014          * we have to hold the device_list_mutex here so the device
3015          * does not go away in cancel_dev. FIXME: find a better solution
3016          */
3017         mutex_lock(&fs_info->fs_devices->device_list_mutex);
3018         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
3019         if (!dev) {
3020                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3021                 return -ENODEV;
3022         }
3023         ret = btrfs_scrub_cancel_dev(fs_info, dev);
3024         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3025
3026         return ret;
3027 }
3028
3029 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3030                          struct btrfs_scrub_progress *progress)
3031 {
3032         struct btrfs_device *dev;
3033         struct scrub_ctx *sctx = NULL;
3034
3035         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3036         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3037         if (dev)
3038                 sctx = dev->scrub_device;
3039         if (sctx)
3040                 memcpy(progress, &sctx->stat, sizeof(*progress));
3041         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3042
3043         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3044 }
3045
3046 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3047                                u64 extent_logical, u64 extent_len,
3048                                u64 *extent_physical,
3049                                struct btrfs_device **extent_dev,
3050                                int *extent_mirror_num)
3051 {
3052         u64 mapped_length;
3053         struct btrfs_bio *bbio = NULL;
3054         int ret;
3055
3056         mapped_length = extent_len;
3057         ret = btrfs_map_block(fs_info, READ, extent_logical,
3058                               &mapped_length, &bbio, 0);
3059         if (ret || !bbio || mapped_length < extent_len ||
3060             !bbio->stripes[0].dev->bdev) {
3061                 kfree(bbio);
3062                 return;
3063         }
3064
3065         *extent_physical = bbio->stripes[0].physical;
3066         *extent_mirror_num = bbio->mirror_num;
3067         *extent_dev = bbio->stripes[0].dev;
3068         kfree(bbio);
3069 }
3070
3071 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3072                               struct scrub_wr_ctx *wr_ctx,
3073                               struct btrfs_fs_info *fs_info,
3074                               struct btrfs_device *dev,
3075                               int is_dev_replace)
3076 {
3077         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3078
3079         mutex_init(&wr_ctx->wr_lock);
3080         wr_ctx->wr_curr_bio = NULL;
3081         if (!is_dev_replace)
3082                 return 0;
3083
3084         WARN_ON(!dev->bdev);
3085         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3086                                          bio_get_nr_vecs(dev->bdev));
3087         wr_ctx->tgtdev = dev;
3088         atomic_set(&wr_ctx->flush_all_writes, 0);
3089         return 0;
3090 }
3091
3092 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3093 {
3094         mutex_lock(&wr_ctx->wr_lock);
3095         kfree(wr_ctx->wr_curr_bio);
3096         wr_ctx->wr_curr_bio = NULL;
3097         mutex_unlock(&wr_ctx->wr_lock);
3098 }
3099
3100 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3101                             int mirror_num, u64 physical_for_dev_replace)
3102 {
3103         struct scrub_copy_nocow_ctx *nocow_ctx;
3104         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3105
3106         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3107         if (!nocow_ctx) {
3108                 spin_lock(&sctx->stat_lock);
3109                 sctx->stat.malloc_errors++;
3110                 spin_unlock(&sctx->stat_lock);
3111                 return -ENOMEM;
3112         }
3113
3114         scrub_pending_trans_workers_inc(sctx);
3115
3116         nocow_ctx->sctx = sctx;
3117         nocow_ctx->logical = logical;
3118         nocow_ctx->len = len;
3119         nocow_ctx->mirror_num = mirror_num;
3120         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3121         nocow_ctx->work.func = copy_nocow_pages_worker;
3122         btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3123                            &nocow_ctx->work);
3124
3125         return 0;
3126 }
3127
3128 static void copy_nocow_pages_worker(struct btrfs_work *work)
3129 {
3130         struct scrub_copy_nocow_ctx *nocow_ctx =
3131                 container_of(work, struct scrub_copy_nocow_ctx, work);
3132         struct scrub_ctx *sctx = nocow_ctx->sctx;
3133         u64 logical = nocow_ctx->logical;
3134         u64 len = nocow_ctx->len;
3135         int mirror_num = nocow_ctx->mirror_num;
3136         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3137         int ret;
3138         struct btrfs_trans_handle *trans = NULL;
3139         struct btrfs_fs_info *fs_info;
3140         struct btrfs_path *path;
3141         struct btrfs_root *root;
3142         int not_written = 0;
3143
3144         fs_info = sctx->dev_root->fs_info;
3145         root = fs_info->extent_root;
3146
3147         path = btrfs_alloc_path();
3148         if (!path) {
3149                 spin_lock(&sctx->stat_lock);
3150                 sctx->stat.malloc_errors++;
3151                 spin_unlock(&sctx->stat_lock);
3152                 not_written = 1;
3153                 goto out;
3154         }
3155
3156         trans = btrfs_join_transaction(root);
3157         if (IS_ERR(trans)) {
3158                 not_written = 1;
3159                 goto out;
3160         }
3161
3162         ret = iterate_inodes_from_logical(logical, fs_info, path,
3163                                           copy_nocow_pages_for_inode,
3164                                           nocow_ctx);
3165         if (ret != 0 && ret != -ENOENT) {
3166                 pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %llu, ret %d\n",
3167                         (unsigned long long)logical,
3168                         (unsigned long long)physical_for_dev_replace,
3169                         (unsigned long long)len,
3170                         (unsigned long long)mirror_num, ret);
3171                 not_written = 1;
3172                 goto out;
3173         }
3174
3175 out:
3176         if (trans && !IS_ERR(trans))
3177                 btrfs_end_transaction(trans, root);
3178         if (not_written)
3179                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3180                                             num_uncorrectable_read_errors);
3181
3182         btrfs_free_path(path);
3183         kfree(nocow_ctx);
3184
3185         scrub_pending_trans_workers_dec(sctx);
3186 }
3187
3188 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root, void *ctx)
3189 {
3190         unsigned long index;
3191         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3192         int ret = 0;
3193         struct btrfs_key key;
3194         struct inode *inode = NULL;
3195         struct btrfs_root *local_root;
3196         u64 physical_for_dev_replace;
3197         u64 len;
3198         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3199         int srcu_index;
3200
3201         key.objectid = root;
3202         key.type = BTRFS_ROOT_ITEM_KEY;
3203         key.offset = (u64)-1;
3204
3205         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3206
3207         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3208         if (IS_ERR(local_root)) {
3209                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3210                 return PTR_ERR(local_root);
3211         }
3212
3213         key.type = BTRFS_INODE_ITEM_KEY;
3214         key.objectid = inum;
3215         key.offset = 0;
3216         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3217         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3218         if (IS_ERR(inode))
3219                 return PTR_ERR(inode);
3220
3221         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3222         len = nocow_ctx->len;
3223         while (len >= PAGE_CACHE_SIZE) {
3224                 struct page *page = NULL;
3225                 int ret_sub;
3226
3227                 index = offset >> PAGE_CACHE_SHIFT;
3228
3229                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3230                 if (!page) {
3231                         pr_err("find_or_create_page() failed\n");
3232                         ret = -ENOMEM;
3233                         goto next_page;
3234                 }
3235
3236                 if (PageUptodate(page)) {
3237                         if (PageDirty(page))
3238                                 goto next_page;
3239                 } else {
3240                         ClearPageError(page);
3241                         ret_sub = extent_read_full_page(&BTRFS_I(inode)->
3242                                                          io_tree,
3243                                                         page, btrfs_get_extent,
3244                                                         nocow_ctx->mirror_num);
3245                         if (ret_sub) {
3246                                 ret = ret_sub;
3247                                 goto next_page;
3248                         }
3249                         wait_on_page_locked(page);
3250                         if (!PageUptodate(page)) {
3251                                 ret = -EIO;
3252                                 goto next_page;
3253                         }
3254                 }
3255                 ret_sub = write_page_nocow(nocow_ctx->sctx,
3256                                            physical_for_dev_replace, page);
3257                 if (ret_sub) {
3258                         ret = ret_sub;
3259                         goto next_page;
3260                 }
3261
3262 next_page:
3263                 if (page) {
3264                         unlock_page(page);
3265                         put_page(page);
3266                 }
3267                 offset += PAGE_CACHE_SIZE;
3268                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3269                 len -= PAGE_CACHE_SIZE;
3270         }
3271
3272         if (inode)
3273                 iput(inode);
3274         return ret;
3275 }
3276
3277 static int write_page_nocow(struct scrub_ctx *sctx,
3278                             u64 physical_for_dev_replace, struct page *page)
3279 {
3280         struct bio *bio;
3281         struct btrfs_device *dev;
3282         int ret;
3283         DECLARE_COMPLETION_ONSTACK(compl);
3284
3285         dev = sctx->wr_ctx.tgtdev;
3286         if (!dev)
3287                 return -EIO;
3288         if (!dev->bdev) {
3289                 printk_ratelimited(KERN_WARNING
3290                         "btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3291                 return -EIO;
3292         }
3293         bio = bio_alloc(GFP_NOFS, 1);
3294         if (!bio) {
3295                 spin_lock(&sctx->stat_lock);
3296                 sctx->stat.malloc_errors++;
3297                 spin_unlock(&sctx->stat_lock);
3298                 return -ENOMEM;
3299         }
3300         bio->bi_private = &compl;
3301         bio->bi_end_io = scrub_complete_bio_end_io;
3302         bio->bi_size = 0;
3303         bio->bi_sector = physical_for_dev_replace >> 9;
3304         bio->bi_bdev = dev->bdev;
3305         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3306         if (ret != PAGE_CACHE_SIZE) {
3307 leave_with_eio:
3308                 bio_put(bio);
3309                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3310                 return -EIO;
3311         }
3312         btrfsic_submit_bio(WRITE_SYNC, bio);
3313         wait_for_completion(&compl);
3314
3315         if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
3316                 goto leave_with_eio;
3317
3318         bio_put(bio);
3319         return 0;
3320 }