Pileus Git - ~andy/linux/blob - 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_nocow_inode {
 162         u64                     inum;
 163         u64                     offset;
 164         u64                     root;
 165         struct list_head        list;
 166 };
 167
 168 struct scrub_copy_nocow_ctx {
 169         struct scrub_ctx        *sctx;
 170         u64                     logical;
 171         u64                     len;
 172         int                     mirror_num;
 173         u64                     physical_for_dev_replace;
 174         struct list_head        inodes;
 175         struct btrfs_work       work;
 176 };
 177
 178 struct scrub_warning {
 179         struct btrfs_path       *path;
 180         u64                     extent_item_size;
 181         char                    *scratch_buf;
 182         char                    *msg_buf;
 183         const char              *errstr;
 184         sector_t                sector;
 185         u64                     logical;
 186         struct btrfs_device     *dev;
 187         int                     msg_bufsize;
 188         int                     scratch_bufsize;
 189 };
 190
 191
 192 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 193 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 194 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 195 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 196 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 197 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
 198                                      struct btrfs_fs_info *fs_info,
 199                                      struct scrub_block *original_sblock,
 200                                      u64 length, u64 logical,
 201                                      struct scrub_block *sblocks_for_recheck);
 202 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 203                                 struct scrub_block *sblock, int is_metadata,
 204                                 int have_csum, u8 *csum, u64 generation,
 205                                 u16 csum_size);
 206 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 207                                          struct scrub_block *sblock,
 208                                          int is_metadata, int have_csum,
 209                                          const u8 *csum, u64 generation,
 210                                          u16 csum_size);
 211 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 212                                              struct scrub_block *sblock_good,
 213                                              int force_write);
 214 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 215                                             struct scrub_block *sblock_good,
 216                                             int page_num, int force_write);
 217 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 218 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 219                                            int page_num);
 220 static int scrub_checksum_data(struct scrub_block *sblock);
 221 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 222 static int scrub_checksum_super(struct scrub_block *sblock);
 223 static void scrub_block_get(struct scrub_block *sblock);
 224 static void scrub_block_put(struct scrub_block *sblock);
 225 static void scrub_page_get(struct scrub_page *spage);
 226 static void scrub_page_put(struct scrub_page *spage);
 227 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 228                                     struct scrub_page *spage);
 229 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 230                        u64 physical, struct btrfs_device *dev, u64 flags,
 231                        u64 gen, int mirror_num, u8 *csum, int force,
 232                        u64 physical_for_dev_replace);
 233 static void scrub_bio_end_io(struct bio *bio, int err);
 234 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 235 static void scrub_block_complete(struct scrub_block *sblock);
 236 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 237                                u64 extent_logical, u64 extent_len,
 238                                u64 *extent_physical,
 239                                struct btrfs_device **extent_dev,
 240                                int *extent_mirror_num);
 241 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
 242                               struct scrub_wr_ctx *wr_ctx,
 243                               struct btrfs_fs_info *fs_info,
 244                               struct btrfs_device *dev,
 245                               int is_dev_replace);
 246 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
 247 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 248                                     struct scrub_page *spage);
 249 static void scrub_wr_submit(struct scrub_ctx *sctx);
 250 static void scrub_wr_bio_end_io(struct bio *bio, int err);
 251 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 252 static int write_page_nocow(struct scrub_ctx *sctx,
 253                             u64 physical_for_dev_replace, struct page *page);
 254 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 255                                       struct scrub_copy_nocow_ctx *ctx);
 256 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 257                             int mirror_num, u64 physical_for_dev_replace);
 258 static void copy_nocow_pages_worker(struct btrfs_work *work);
 259
 260
 261 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 262 {
 263         atomic_inc(&sctx->bios_in_flight);
 264 }
 265
 266 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 267 {
 268         atomic_dec(&sctx->bios_in_flight);
 269         wake_up(&sctx->list_wait);
 270 }
 271
 272 /*
 273  * used for workers that require transaction commits (i.e., for the
 274  * NOCOW case)
 275  */
 276 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 277 {
 278         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 279
 280         /*
 281          * increment scrubs_running to prevent cancel requests from
 282          * completing as long as a worker is running. we must also
 283          * increment scrubs_paused to prevent deadlocking on pause
 284          * requests used for transactions commits (as the worker uses a
 285          * transaction context). it is safe to regard the worker
 286          * as paused for all matters practical. effectively, we only
 287          * avoid cancellation requests from completing.
 288          */
 289         mutex_lock(&fs_info->scrub_lock);
 290         atomic_inc(&fs_info->scrubs_running);
 291         atomic_inc(&fs_info->scrubs_paused);
 292         mutex_unlock(&fs_info->scrub_lock);
 293         atomic_inc(&sctx->workers_pending);
 294 }
 295
 296 /* used for workers that require transaction commits */
 297 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 298 {
 299         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 300
 301         /*
 302          * see scrub_pending_trans_workers_inc() why we're pretending
 303          * to be paused in the scrub counters
 304          */
 305         mutex_lock(&fs_info->scrub_lock);
 306         atomic_dec(&fs_info->scrubs_running);
 307         atomic_dec(&fs_info->scrubs_paused);
 308         mutex_unlock(&fs_info->scrub_lock);
 309         atomic_dec(&sctx->workers_pending);
 310         wake_up(&fs_info->scrub_pause_wait);
 311         wake_up(&sctx->list_wait);
 312 }
 313
 314 static void scrub_free_csums(struct scrub_ctx *sctx)
 315 {
 316         while (!list_empty(&sctx->csum_list)) {
 317                 struct btrfs_ordered_sum *sum;
 318                 sum = list_first_entry(&sctx->csum_list,
 319                                        struct btrfs_ordered_sum, list);
 320                 list_del(&sum->list);
 321                 kfree(sum);
 322         }
 323 }
 324
 325 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 326 {
 327         int i;
 328
 329         if (!sctx)
 330                 return;
 331
 332         scrub_free_wr_ctx(&sctx->wr_ctx);
 333
 334         /* this can happen when scrub is cancelled */
 335         if (sctx->curr != -1) {
 336                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 337
 338                 for (i = 0; i < sbio->page_count; i++) {
 339                         WARN_ON(!sbio->pagev[i]->page);
 340                         scrub_block_put(sbio->pagev[i]->sblock);
 341                 }
 342                 bio_put(sbio->bio);
 343         }
 344
 345         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 346                 struct scrub_bio *sbio = sctx->bios[i];
 347
 348                 if (!sbio)
 349                         break;
 350                 kfree(sbio);
 351         }
 352
 353         scrub_free_csums(sctx);
 354         kfree(sctx);
 355 }
 356
 357 static noinline_for_stack
 358 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 359 {
 360         struct scrub_ctx *sctx;
 361         int             i;
 362         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 363         int pages_per_rd_bio;
 364         int ret;
 365
 366         /*
 367          * the setting of pages_per_rd_bio is correct for scrub but might
 368          * be wrong for the dev_replace code where we might read from
 369          * different devices in the initial huge bios. However, that
 370          * code is able to correctly handle the case when adding a page
 371          * to a bio fails.
 372          */
 373         if (dev->bdev)
 374                 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
 375                                          bio_get_nr_vecs(dev->bdev));
 376         else
 377                 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 378         sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
 379         if (!sctx)
 380                 goto nomem;
 381         sctx->is_dev_replace = is_dev_replace;
 382         sctx->pages_per_rd_bio = pages_per_rd_bio;
 383         sctx->curr = -1;
 384         sctx->dev_root = dev->dev_root;
 385         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 386                 struct scrub_bio *sbio;
 387
 388                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 389                 if (!sbio)
 390                         goto nomem;
 391                 sctx->bios[i] = sbio;
 392
 393                 sbio->index = i;
 394                 sbio->sctx = sctx;
 395                 sbio->page_count = 0;
 396                 sbio->work.func = scrub_bio_end_io_worker;
 397
 398                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 399                         sctx->bios[i]->next_free = i + 1;
 400                 else
 401                         sctx->bios[i]->next_free = -1;
 402         }
 403         sctx->first_free = 0;
 404         sctx->nodesize = dev->dev_root->nodesize;
 405         sctx->leafsize = dev->dev_root->leafsize;
 406         sctx->sectorsize = dev->dev_root->sectorsize;
 407         atomic_set(&sctx->bios_in_flight, 0);
 408         atomic_set(&sctx->workers_pending, 0);
 409         atomic_set(&sctx->cancel_req, 0);
 410         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 411         INIT_LIST_HEAD(&sctx->csum_list);
 412
 413         spin_lock_init(&sctx->list_lock);
 414         spin_lock_init(&sctx->stat_lock);
 415         init_waitqueue_head(&sctx->list_wait);
 416
 417         ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 418                                  fs_info->dev_replace.tgtdev, is_dev_replace);
 419         if (ret) {
 420                 scrub_free_ctx(sctx);
 421                 return ERR_PTR(ret);
 422         }
 423         return sctx;
 424
 425 nomem:
 426         scrub_free_ctx(sctx);
 427         return ERR_PTR(-ENOMEM);
 428 }
 429
 430 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 431                                      void *warn_ctx)
 432 {
 433         u64 isize;
 434         u32 nlink;
 435         int ret;
 436         int i;
 437         struct extent_buffer *eb;
 438         struct btrfs_inode_item *inode_item;
 439         struct scrub_warning *swarn = warn_ctx;
 440         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 441         struct inode_fs_paths *ipath = NULL;
 442         struct btrfs_root *local_root;
 443         struct btrfs_key root_key;
 444
 445         root_key.objectid = root;
 446         root_key.type = BTRFS_ROOT_ITEM_KEY;
 447         root_key.offset = (u64)-1;
 448         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 449         if (IS_ERR(local_root)) {
 450                 ret = PTR_ERR(local_root);
 451                 goto err;
 452         }
 453
 454         ret = inode_item_info(inum, 0, local_root, swarn->path);
 455         if (ret) {
 456                 btrfs_release_path(swarn->path);
 457                 goto err;
 458         }
 459
 460         eb = swarn->path->nodes[0];
 461         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 462                                         struct btrfs_inode_item);
 463         isize = btrfs_inode_size(eb, inode_item);
 464         nlink = btrfs_inode_nlink(eb, inode_item);
 465         btrfs_release_path(swarn->path);
 466
 467         ipath = init_ipath(4096, local_root, swarn->path);
 468         if (IS_ERR(ipath)) {
 469                 ret = PTR_ERR(ipath);
 470                 ipath = NULL;
 471                 goto err;
 472         }
 473         ret = paths_from_inode(inum, ipath);
 474
 475         if (ret < 0)
 476                 goto err;
 477
 478         /*
 479          * we deliberately ignore the bit ipath might have been too small to
 480          * hold all of the paths here
 481          */
 482         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 483                 printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 484                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 485                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 486                         swarn->logical, rcu_str_deref(swarn->dev->name),
 487                         (unsigned long long)swarn->sector, root, inum, offset,
 488                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 489                         (char *)(unsigned long)ipath->fspath->val[i]);
 490
 491         free_ipath(ipath);
 492         return 0;
 493
 494 err:
 495         printk_in_rcu(KERN_WARNING "btrfs: %s at logical %llu on dev "
 496                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 497                 "resolving failed with ret=%d\n", swarn->errstr,
 498                 swarn->logical, rcu_str_deref(swarn->dev->name),
 499                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 500
 501         free_ipath(ipath);
 502         return 0;
 503 }
 504
 505 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 506 {
 507         struct btrfs_device *dev;
 508         struct btrfs_fs_info *fs_info;
 509         struct btrfs_path *path;
 510         struct btrfs_key found_key;
 511         struct extent_buffer *eb;
 512         struct btrfs_extent_item *ei;
 513         struct scrub_warning swarn;
 514         unsigned long ptr = 0;
 515         u64 extent_item_pos;
 516         u64 flags = 0;
 517         u64 ref_root;
 518         u32 item_size;
 519         u8 ref_level;
 520         const int bufsize = 4096;
 521         int ret;
 522
 523         WARN_ON(sblock->page_count < 1);
 524         dev = sblock->pagev[0]->dev;
 525         fs_info = sblock->sctx->dev_root->fs_info;
 526
 527         path = btrfs_alloc_path();
 528
 529         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 530         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 531         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 532         swarn.logical = sblock->pagev[0]->logical;
 533         swarn.errstr = errstr;
 534         swarn.dev = NULL;
 535         swarn.msg_bufsize = bufsize;
 536         swarn.scratch_bufsize = bufsize;
 537
 538         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 539                 goto out;
 540
 541         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 542                                   &flags);
 543         if (ret < 0)
 544                 goto out;
 545
 546         extent_item_pos = swarn.logical - found_key.objectid;
 547         swarn.extent_item_size = found_key.offset;
 548
 549         eb = path->nodes[0];
 550         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 551         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 552
 553         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 554                 do {
 555                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 556                                                         &ref_root, &ref_level);
 557                         printk_in_rcu(KERN_WARNING
 558                                 "btrfs: %s at logical %llu on dev %s, "
 559                                 "sector %llu: metadata %s (level %d) in tree "
 560                                 "%llu\n", errstr, swarn.logical,
 561                                 rcu_str_deref(dev->name),
 562                                 (unsigned long long)swarn.sector,
 563                                 ref_level ? "node" : "leaf",
 564                                 ret < 0 ? -1 : ref_level,
 565                                 ret < 0 ? -1 : ref_root);
 566                 } while (ret != 1);
 567                 btrfs_release_path(path);
 568         } else {
 569                 btrfs_release_path(path);
 570                 swarn.path = path;
 571                 swarn.dev = dev;
 572                 iterate_extent_inodes(fs_info, found_key.objectid,
 573                                         extent_item_pos, 1,
 574                                         scrub_print_warning_inode, &swarn);
 575         }
 576
 577 out:
 578         btrfs_free_path(path);
 579         kfree(swarn.scratch_buf);
 580         kfree(swarn.msg_buf);
 581 }
 582
 583 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 584 {
 585         struct page *page = NULL;
 586         unsigned long index;
 587         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 588         int ret;
 589         int corrected = 0;
 590         struct btrfs_key key;
 591         struct inode *inode = NULL;
 592         struct btrfs_fs_info *fs_info;
 593         u64 end = offset + PAGE_SIZE - 1;
 594         struct btrfs_root *local_root;
 595         int srcu_index;
 596
 597         key.objectid = root;
 598         key.type = BTRFS_ROOT_ITEM_KEY;
 599         key.offset = (u64)-1;
 600
 601         fs_info = fixup->root->fs_info;
 602         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 603
 604         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 605         if (IS_ERR(local_root)) {
 606                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 607                 return PTR_ERR(local_root);
 608         }
 609
 610         key.type = BTRFS_INODE_ITEM_KEY;
 611         key.objectid = inum;
 612         key.offset = 0;
 613         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 614         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 615         if (IS_ERR(inode))
 616                 return PTR_ERR(inode);
 617
 618         index = offset >> PAGE_CACHE_SHIFT;
 619
 620         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 621         if (!page) {
 622                 ret = -ENOMEM;
 623                 goto out;
 624         }
 625
 626         if (PageUptodate(page)) {
 627                 if (PageDirty(page)) {
 628                         /*
 629                          * we need to write the data to the defect sector. the
 630                          * data that was in that sector is not in memory,
 631                          * because the page was modified. we must not write the
 632                          * modified page to that sector.
 633                          *
 634                          * TODO: what could be done here: wait for the delalloc
 635                          *       runner to write out that page (might involve
 636                          *       COW) and see whether the sector is still
 637                          *       referenced afterwards.
 638                          *
 639                          * For the meantime, we'll treat this error
 640                          * incorrectable, although there is a chance that a
 641                          * later scrub will find the bad sector again and that
 642                          * there's no dirty page in memory, then.
 643                          */
 644                         ret = -EIO;
 645                         goto out;
 646                 }
 647                 fs_info = BTRFS_I(inode)->root->fs_info;
 648                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 649                                         fixup->logical, page,
 650                                         fixup->mirror_num);
 651                 unlock_page(page);
 652                 corrected = !ret;
 653         } else {
 654                 /*
 655                  * we need to get good data first. the general readpage path
 656                  * will call repair_io_failure for us, we just have to make
 657                  * sure we read the bad mirror.
 658                  */
 659                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 660                                         EXTENT_DAMAGED, GFP_NOFS);
 661                 if (ret) {
 662                         /* set_extent_bits should give proper error */
 663                         WARN_ON(ret > 0);
 664                         if (ret > 0)
 665                                 ret = -EFAULT;
 666                         goto out;
 667                 }
 668
 669                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 670                                                 btrfs_get_extent,
 671                                                 fixup->mirror_num);
 672                 wait_on_page_locked(page);
 673
 674                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 675                                                 end, EXTENT_DAMAGED, 0, NULL);
 676                 if (!corrected)
 677                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 678                                                 EXTENT_DAMAGED, GFP_NOFS);
 679         }
 680
 681 out:
 682         if (page)
 683                 put_page(page);
 684         if (inode)
 685                 iput(inode);
 686
 687         if (ret < 0)
 688                 return ret;
 689
 690         if (ret == 0 && corrected) {
 691                 /*
 692                  * we only need to call readpage for one of the inodes belonging
 693                  * to this extent. so make iterate_extent_inodes stop
 694                  */
 695                 return 1;
 696         }
 697
 698         return -EIO;
 699 }
 700
 701 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 702 {
 703         int ret;
 704         struct scrub_fixup_nodatasum *fixup;
 705         struct scrub_ctx *sctx;
 706         struct btrfs_trans_handle *trans = NULL;
 707         struct btrfs_path *path;
 708         int uncorrectable = 0;
 709
 710         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 711         sctx = fixup->sctx;
 712
 713         path = btrfs_alloc_path();
 714         if (!path) {
 715                 spin_lock(&sctx->stat_lock);
 716                 ++sctx->stat.malloc_errors;
 717                 spin_unlock(&sctx->stat_lock);
 718                 uncorrectable = 1;
 719                 goto out;
 720         }
 721
 722         trans = btrfs_join_transaction(fixup->root);
 723         if (IS_ERR(trans)) {
 724                 uncorrectable = 1;
 725                 goto out;
 726         }
 727
 728         /*
 729          * the idea is to trigger a regular read through the standard path. we
 730          * read a page from the (failed) logical address by specifying the
 731          * corresponding copynum of the failed sector. thus, that readpage is
 732          * expected to fail.
 733          * that is the point where on-the-fly error correction will kick in
 734          * (once it's finished) and rewrite the failed sector if a good copy
 735          * can be found.
 736          */
 737         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 738                                                 path, scrub_fixup_readpage,
 739                                                 fixup);
 740         if (ret < 0) {
 741                 uncorrectable = 1;
 742                 goto out;
 743         }
 744         WARN_ON(ret != 1);
 745
 746         spin_lock(&sctx->stat_lock);
 747         ++sctx->stat.corrected_errors;
 748         spin_unlock(&sctx->stat_lock);
 749
 750 out:
 751         if (trans && !IS_ERR(trans))
 752                 btrfs_end_transaction(trans, fixup->root);
 753         if (uncorrectable) {
 754                 spin_lock(&sctx->stat_lock);
 755                 ++sctx->stat.uncorrectable_errors;
 756                 spin_unlock(&sctx->stat_lock);
 757                 btrfs_dev_replace_stats_inc(
 758                         &sctx->dev_root->fs_info->dev_replace.
 759                         num_uncorrectable_read_errors);
 760                 printk_ratelimited_in_rcu(KERN_ERR
 761                         "btrfs: unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 762                         fixup->logical, rcu_str_deref(fixup->dev->name));
 763         }
 764
 765         btrfs_free_path(path);
 766         kfree(fixup);
 767
 768         scrub_pending_trans_workers_dec(sctx);
 769 }
 770
 771 /*
 772  * scrub_handle_errored_block gets called when either verification of the
 773  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 774  * case, this function handles all pages in the bio, even though only one
 775  * may be bad.
 776  * The goal of this function is to repair the errored block by using the
 777  * contents of one of the mirrors.
 778  */
 779 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 780 {
 781         struct scrub_ctx *sctx = sblock_to_check->sctx;
 782         struct btrfs_device *dev;
 783         struct btrfs_fs_info *fs_info;
 784         u64 length;
 785         u64 logical;
 786         u64 generation;
 787         unsigned int failed_mirror_index;
 788         unsigned int is_metadata;
 789         unsigned int have_csum;
 790         u8 *csum;
 791         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 792         struct scrub_block *sblock_bad;
 793         int ret;
 794         int mirror_index;
 795         int page_num;
 796         int success;
 797         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 798                                       DEFAULT_RATELIMIT_BURST);
 799
 800         BUG_ON(sblock_to_check->page_count < 1);
 801         fs_info = sctx->dev_root->fs_info;
 802         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 803                 /*
 804                  * if we find an error in a super block, we just report it.
 805                  * They will get written with the next transaction commit
 806                  * anyway
 807                  */
 808                 spin_lock(&sctx->stat_lock);
 809                 ++sctx->stat.super_errors;
 810                 spin_unlock(&sctx->stat_lock);
 811                 return 0;
 812         }
 813         length = sblock_to_check->page_count * PAGE_SIZE;
 814         logical = sblock_to_check->pagev[0]->logical;
 815         generation = sblock_to_check->pagev[0]->generation;
 816         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 817         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 818         is_metadata = !(sblock_to_check->pagev[0]->flags &
 819                         BTRFS_EXTENT_FLAG_DATA);
 820         have_csum = sblock_to_check->pagev[0]->have_csum;
 821         csum = sblock_to_check->pagev[0]->csum;
 822         dev = sblock_to_check->pagev[0]->dev;
 823
 824         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 825                 sblocks_for_recheck = NULL;
 826                 goto nodatasum_case;
 827         }
 828
 829         /*
 830          * read all mirrors one after the other. This includes to
 831          * re-read the extent or metadata block that failed (that was
 832          * the cause that this fixup code is called) another time,
 833          * page by page this time in order to know which pages
 834          * caused I/O errors and which ones are good (for all mirrors).
 835          * It is the goal to handle the situation when more than one
 836          * mirror contains I/O errors, but the errors do not
 837          * overlap, i.e. the data can be repaired by selecting the
 838          * pages from those mirrors without I/O error on the
 839          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 840          * would be that mirror #1 has an I/O error on the first page,
 841          * the second page is good, and mirror #2 has an I/O error on
 842          * the second page, but the first page is good.
 843          * Then the first page of the first mirror can be repaired by
 844          * taking the first page of the second mirror, and the
 845          * second page of the second mirror can be repaired by
 846          * copying the contents of the 2nd page of the 1st mirror.
 847          * One more note: if the pages of one mirror contain I/O
 848          * errors, the checksum cannot be verified. In order to get
 849          * the best data for repairing, the first attempt is to find
 850          * a mirror without I/O errors and with a validated checksum.
 851          * Only if this is not possible, the pages are picked from
 852          * mirrors with I/O errors without considering the checksum.
 853          * If the latter is the case, at the end, the checksum of the
 854          * repaired area is verified in order to correctly maintain
 855          * the statistics.
 856          */
 857
 858         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 859                                      sizeof(*sblocks_for_recheck),
 860                                      GFP_NOFS);
 861         if (!sblocks_for_recheck) {
 862                 spin_lock(&sctx->stat_lock);
 863                 sctx->stat.malloc_errors++;
 864                 sctx->stat.read_errors++;
 865                 sctx->stat.uncorrectable_errors++;
 866                 spin_unlock(&sctx->stat_lock);
 867                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 868                 goto out;
 869         }
 870
 871         /* setup the context, map the logical blocks and alloc the pages */
 872         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 873                                         logical, sblocks_for_recheck);
 874         if (ret) {
 875                 spin_lock(&sctx->stat_lock);
 876                 sctx->stat.read_errors++;
 877                 sctx->stat.uncorrectable_errors++;
 878                 spin_unlock(&sctx->stat_lock);
 879                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 880                 goto out;
 881         }
 882         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 883         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 884
 885         /* build and submit the bios for the failed mirror, check checksums */
 886         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 887                             csum, generation, sctx->csum_size);
 888
 889         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 890             sblock_bad->no_io_error_seen) {
 891                 /*
 892                  * the error disappeared after reading page by page, or
 893                  * the area was part of a huge bio and other parts of the
 894                  * bio caused I/O errors, or the block layer merged several
 895                  * read requests into one and the error is caused by a
 896                  * different bio (usually one of the two latter cases is
 897                  * the cause)
 898                  */
 899                 spin_lock(&sctx->stat_lock);
 900                 sctx->stat.unverified_errors++;
 901                 spin_unlock(&sctx->stat_lock);
 902
 903                 if (sctx->is_dev_replace)
 904                         scrub_write_block_to_dev_replace(sblock_bad);
 905                 goto out;
 906         }
 907
 908         if (!sblock_bad->no_io_error_seen) {
 909                 spin_lock(&sctx->stat_lock);
 910                 sctx->stat.read_errors++;
 911                 spin_unlock(&sctx->stat_lock);
 912                 if (__ratelimit(&_rs))
 913                         scrub_print_warning("i/o error", sblock_to_check);
 914                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 915         } else if (sblock_bad->checksum_error) {
 916                 spin_lock(&sctx->stat_lock);
 917                 sctx->stat.csum_errors++;
 918                 spin_unlock(&sctx->stat_lock);
 919                 if (__ratelimit(&_rs))
 920                         scrub_print_warning("checksum error", sblock_to_check);
 921                 btrfs_dev_stat_inc_and_print(dev,
 922                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 923         } else if (sblock_bad->header_error) {
 924                 spin_lock(&sctx->stat_lock);
 925                 sctx->stat.verify_errors++;
 926                 spin_unlock(&sctx->stat_lock);
 927                 if (__ratelimit(&_rs))
 928                         scrub_print_warning("checksum/header error",
 929                                             sblock_to_check);
 930                 if (sblock_bad->generation_error)
 931                         btrfs_dev_stat_inc_and_print(dev,
 932                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 933                 else
 934                         btrfs_dev_stat_inc_and_print(dev,
 935                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 936         }
 937
 938         if (sctx->readonly) {
 939                 ASSERT(!sctx->is_dev_replace);
 940                 goto out;
 941         }
 942
 943         if (!is_metadata && !have_csum) {
 944                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 945
 946 nodatasum_case:
 947                 WARN_ON(sctx->is_dev_replace);
 948
 949                 /*
 950                  * !is_metadata and !have_csum, this means that the data
 951                  * might not be COW'ed, that it might be modified
 952                  * concurrently. The general strategy to work on the
 953                  * commit root does not help in the case when COW is not
 954                  * used.
 955                  */
 956                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 957                 if (!fixup_nodatasum)
 958                         goto did_not_correct_error;
 959                 fixup_nodatasum->sctx = sctx;
 960                 fixup_nodatasum->dev = dev;
 961                 fixup_nodatasum->logical = logical;
 962                 fixup_nodatasum->root = fs_info->extent_root;
 963                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 964                 scrub_pending_trans_workers_inc(sctx);
 965                 fixup_nodatasum->work.func = scrub_fixup_nodatasum;
 966                 btrfs_queue_worker(&fs_info->scrub_workers,
 967                                    &fixup_nodatasum->work);
 968                 goto out;
 969         }
 970
 971         /*
 972          * now build and submit the bios for the other mirrors, check
 973          * checksums.
 974          * First try to pick the mirror which is completely without I/O
 975          * errors and also does not have a checksum error.
 976          * If one is found, and if a checksum is present, the full block
 977          * that is known to contain an error is rewritten. Afterwards
 978          * the block is known to be corrected.
 979          * If a mirror is found which is completely correct, and no
 980          * checksum is present, only those pages are rewritten that had
 981          * an I/O error in the block to be repaired, since it cannot be
 982          * determined, which copy of the other pages is better (and it
 983          * could happen otherwise that a correct page would be
 984          * overwritten by a bad one).
 985          */
 986         for (mirror_index = 0;
 987              mirror_index < BTRFS_MAX_MIRRORS &&
 988              sblocks_for_recheck[mirror_index].page_count > 0;
 989              mirror_index++) {
 990                 struct scrub_block *sblock_other;
 991
 992                 if (mirror_index == failed_mirror_index)
 993                         continue;
 994                 sblock_other = sblocks_for_recheck + mirror_index;
 995
 996                 /* build and submit the bios, check checksums */
 997                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
 998                                     have_csum, csum, generation,
 999                                     sctx->csum_size);
1000
1001                 if (!sblock_other->header_error &&
1002                     !sblock_other->checksum_error &&
1003                     sblock_other->no_io_error_seen) {
1004                         if (sctx->is_dev_replace) {
1005                                 scrub_write_block_to_dev_replace(sblock_other);
1006                         } else {
1007                                 int force_write = is_metadata || have_csum;
1008
1009                                 ret = scrub_repair_block_from_good_copy(
1010                                                 sblock_bad, sblock_other,
1011                                                 force_write);
1012                         }
1013                         if (0 == ret)
1014                                 goto corrected_error;
1015                 }
1016         }
1017
1018         /*
1019          * for dev_replace, pick good pages and write to the target device.
1020          */
1021         if (sctx->is_dev_replace) {
1022                 success = 1;
1023                 for (page_num = 0; page_num < sblock_bad->page_count;
1024                      page_num++) {
1025                         int sub_success;
1026
1027                         sub_success = 0;
1028                         for (mirror_index = 0;
1029                              mirror_index < BTRFS_MAX_MIRRORS &&
1030                              sblocks_for_recheck[mirror_index].page_count > 0;
1031                              mirror_index++) {
1032                                 struct scrub_block *sblock_other =
1033                                         sblocks_for_recheck + mirror_index;
1034                                 struct scrub_page *page_other =
1035                                         sblock_other->pagev[page_num];
1036
1037                                 if (!page_other->io_error) {
1038                                         ret = scrub_write_page_to_dev_replace(
1039                                                         sblock_other, page_num);
1040                                         if (ret == 0) {
1041                                                 /* succeeded for this page */
1042                                                 sub_success = 1;
1043                                                 break;
1044                                         } else {
1045                                                 btrfs_dev_replace_stats_inc(
1046                                                         &sctx->dev_root->
1047                                                         fs_info->dev_replace.
1048                                                         num_write_errors);
1049                                         }
1050                                 }
1051                         }
1052
1053                         if (!sub_success) {
1054                                 /*
1055                                  * did not find a mirror to fetch the page
1056                                  * from. scrub_write_page_to_dev_replace()
1057                                  * handles this case (page->io_error), by
1058                                  * filling the block with zeros before
1059                                  * submitting the write request
1060                                  */
1061                                 success = 0;
1062                                 ret = scrub_write_page_to_dev_replace(
1063                                                 sblock_bad, page_num);
1064                                 if (ret)
1065                                         btrfs_dev_replace_stats_inc(
1066                                                 &sctx->dev_root->fs_info->
1067                                                 dev_replace.num_write_errors);
1068                         }
1069                 }
1070
1071                 goto out;
1072         }
1073
1074         /*
1075          * for regular scrub, repair those pages that are errored.
1076          * In case of I/O errors in the area that is supposed to be
1077          * repaired, continue by picking good copies of those pages.
1078          * Select the good pages from mirrors to rewrite bad pages from
1079          * the area to fix. Afterwards verify the checksum of the block
1080          * that is supposed to be repaired. This verification step is
1081          * only done for the purpose of statistic counting and for the
1082          * final scrub report, whether errors remain.
1083          * A perfect algorithm could make use of the checksum and try
1084          * all possible combinations of pages from the different mirrors
1085          * until the checksum verification succeeds. For example, when
1086          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1087          * of mirror #2 is readable but the final checksum test fails,
1088          * then the 2nd page of mirror #3 could be tried, whether now
1089          * the final checksum succeedes. But this would be a rare
1090          * exception and is therefore not implemented. At least it is
1091          * avoided that the good copy is overwritten.
1092          * A more useful improvement would be to pick the sectors
1093          * without I/O error based on sector sizes (512 bytes on legacy
1094          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1095          * mirror could be repaired by taking 512 byte of a different
1096          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1097          * area are unreadable.
1098          */
1099
1100         /* can only fix I/O errors from here on */
1101         if (sblock_bad->no_io_error_seen)
1102                 goto did_not_correct_error;
1103
1104         success = 1;
1105         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1106                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1107
1108                 if (!page_bad->io_error)
1109                         continue;
1110
1111                 for (mirror_index = 0;
1112                      mirror_index < BTRFS_MAX_MIRRORS &&
1113                      sblocks_for_recheck[mirror_index].page_count > 0;
1114                      mirror_index++) {
1115                         struct scrub_block *sblock_other = sblocks_for_recheck +
1116                                                            mirror_index;
1117                         struct scrub_page *page_other = sblock_other->pagev[
1118                                                         page_num];
1119
1120                         if (!page_other->io_error) {
1121                                 ret = scrub_repair_page_from_good_copy(
1122                                         sblock_bad, sblock_other, page_num, 0);
1123                                 if (0 == ret) {
1124                                         page_bad->io_error = 0;
1125                                         break; /* succeeded for this page */
1126                                 }
1127                         }
1128                 }
1129
1130                 if (page_bad->io_error) {
1131                         /* did not find a mirror to copy the page from */
1132                         success = 0;
1133                 }
1134         }
1135
1136         if (success) {
1137                 if (is_metadata || have_csum) {
1138                         /*
1139                          * need to verify the checksum now that all
1140                          * sectors on disk are repaired (the write
1141                          * request for data to be repaired is on its way).
1142                          * Just be lazy and use scrub_recheck_block()
1143                          * which re-reads the data before the checksum
1144                          * is verified, but most likely the data comes out
1145                          * of the page cache.
1146                          */
1147                         scrub_recheck_block(fs_info, sblock_bad,
1148                                             is_metadata, have_csum, csum,
1149                                             generation, sctx->csum_size);
1150                         if (!sblock_bad->header_error &&
1151                             !sblock_bad->checksum_error &&
1152                             sblock_bad->no_io_error_seen)
1153                                 goto corrected_error;
1154                         else
1155                                 goto did_not_correct_error;
1156                 } else {
1157 corrected_error:
1158                         spin_lock(&sctx->stat_lock);
1159                         sctx->stat.corrected_errors++;
1160                         spin_unlock(&sctx->stat_lock);
1161                         printk_ratelimited_in_rcu(KERN_ERR
1162                                 "btrfs: fixed up error at logical %llu on dev %s\n",
1163                                 logical, rcu_str_deref(dev->name));
1164                 }
1165         } else {
1166 did_not_correct_error:
1167                 spin_lock(&sctx->stat_lock);
1168                 sctx->stat.uncorrectable_errors++;
1169                 spin_unlock(&sctx->stat_lock);
1170                 printk_ratelimited_in_rcu(KERN_ERR
1171                         "btrfs: unable to fixup (regular) error at logical %llu on dev %s\n",
1172                         logical, rcu_str_deref(dev->name));
1173         }
1174
1175 out:
1176         if (sblocks_for_recheck) {
1177                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1178                      mirror_index++) {
1179                         struct scrub_block *sblock = sblocks_for_recheck +
1180                                                      mirror_index;
1181                         int page_index;
1182
1183                         for (page_index = 0; page_index < sblock->page_count;
1184                              page_index++) {
1185                                 sblock->pagev[page_index]->sblock = NULL;
1186                                 scrub_page_put(sblock->pagev[page_index]);
1187                         }
1188                 }
1189                 kfree(sblocks_for_recheck);
1190         }
1191
1192         return 0;
1193 }
1194
1195 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1196                                      struct btrfs_fs_info *fs_info,
1197                                      struct scrub_block *original_sblock,
1198                                      u64 length, u64 logical,
1199                                      struct scrub_block *sblocks_for_recheck)
1200 {
1201         int page_index;
1202         int mirror_index;
1203         int ret;
1204
1205         /*
1206          * note: the two members ref_count and outstanding_pages
1207          * are not used (and not set) in the blocks that are used for
1208          * the recheck procedure
1209          */
1210
1211         page_index = 0;
1212         while (length > 0) {
1213                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1214                 u64 mapped_length = sublen;
1215                 struct btrfs_bio *bbio = NULL;
1216
1217                 /*
1218                  * with a length of PAGE_SIZE, each returned stripe
1219                  * represents one mirror
1220                  */
1221                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1222                                       &mapped_length, &bbio, 0);
1223                 if (ret || !bbio || mapped_length < sublen) {
1224                         kfree(bbio);
1225                         return -EIO;
1226                 }
1227
1228                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1229                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1230                      mirror_index++) {
1231                         struct scrub_block *sblock;
1232                         struct scrub_page *page;
1233
1234                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1235                                 continue;
1236
1237                         sblock = sblocks_for_recheck + mirror_index;
1238                         sblock->sctx = sctx;
1239                         page = kzalloc(sizeof(*page), GFP_NOFS);
1240                         if (!page) {
1241 leave_nomem:
1242                                 spin_lock(&sctx->stat_lock);
1243                                 sctx->stat.malloc_errors++;
1244                                 spin_unlock(&sctx->stat_lock);
1245                                 kfree(bbio);
1246                                 return -ENOMEM;
1247                         }
1248                         scrub_page_get(page);
1249                         sblock->pagev[page_index] = page;
1250                         page->logical = logical;
1251                         page->physical = bbio->stripes[mirror_index].physical;
1252                         BUG_ON(page_index >= original_sblock->page_count);
1253                         page->physical_for_dev_replace =
1254                                 original_sblock->pagev[page_index]->
1255                                 physical_for_dev_replace;
1256                         /* for missing devices, dev->bdev is NULL */
1257                         page->dev = bbio->stripes[mirror_index].dev;
1258                         page->mirror_num = mirror_index + 1;
1259                         sblock->page_count++;
1260                         page->page = alloc_page(GFP_NOFS);
1261                         if (!page->page)
1262                                 goto leave_nomem;
1263                 }
1264                 kfree(bbio);
1265                 length -= sublen;
1266                 logical += sublen;
1267                 page_index++;
1268         }
1269
1270         return 0;
1271 }
1272
1273 /*
1274  * this function will check the on disk data for checksum errors, header
1275  * errors and read I/O errors. If any I/O errors happen, the exact pages
1276  * which are errored are marked as being bad. The goal is to enable scrub
1277  * to take those pages that are not errored from all the mirrors so that
1278  * the pages that are errored in the just handled mirror can be repaired.
1279  */
1280 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1281                                 struct scrub_block *sblock, int is_metadata,
1282                                 int have_csum, u8 *csum, u64 generation,
1283                                 u16 csum_size)
1284 {
1285         int page_num;
1286
1287         sblock->no_io_error_seen = 1;
1288         sblock->header_error = 0;
1289         sblock->checksum_error = 0;
1290
1291         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1292                 struct bio *bio;
1293                 struct scrub_page *page = sblock->pagev[page_num];
1294
1295                 if (page->dev->bdev == NULL) {
1296                         page->io_error = 1;
1297                         sblock->no_io_error_seen = 0;
1298                         continue;
1299                 }
1300
1301                 WARN_ON(!page->page);
1302                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1303                 if (!bio) {
1304                         page->io_error = 1;
1305                         sblock->no_io_error_seen = 0;
1306                         continue;
1307                 }
1308                 bio->bi_bdev = page->dev->bdev;
1309                 bio->bi_sector = page->physical >> 9;
1310
1311                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1312                 if (btrfsic_submit_bio_wait(READ, bio))
1313                         sblock->no_io_error_seen = 0;
1314
1315                 bio_put(bio);
1316         }
1317
1318         if (sblock->no_io_error_seen)
1319                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1320                                              have_csum, csum, generation,
1321                                              csum_size);
1322
1323         return;
1324 }
1325
1326 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1327                                          struct scrub_block *sblock,
1328                                          int is_metadata, int have_csum,
1329                                          const u8 *csum, u64 generation,
1330                                          u16 csum_size)
1331 {
1332         int page_num;
1333         u8 calculated_csum[BTRFS_CSUM_SIZE];
1334         u32 crc = ~(u32)0;
1335         void *mapped_buffer;
1336
1337         WARN_ON(!sblock->pagev[0]->page);
1338         if (is_metadata) {
1339                 struct btrfs_header *h;
1340
1341                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1342                 h = (struct btrfs_header *)mapped_buffer;
1343
1344                 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1345                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1346                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1347                            BTRFS_UUID_SIZE)) {
1348                         sblock->header_error = 1;
1349                 } else if (generation != btrfs_stack_header_generation(h)) {
1350                         sblock->header_error = 1;
1351                         sblock->generation_error = 1;
1352                 }
1353                 csum = h->csum;
1354         } else {
1355                 if (!have_csum)
1356                         return;
1357
1358                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1359         }
1360
1361         for (page_num = 0;;) {
1362                 if (page_num == 0 && is_metadata)
1363                         crc = btrfs_csum_data(
1364                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1365                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1366                 else
1367                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1368
1369                 kunmap_atomic(mapped_buffer);
1370                 page_num++;
1371                 if (page_num >= sblock->page_count)
1372                         break;
1373                 WARN_ON(!sblock->pagev[page_num]->page);
1374
1375                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1376         }
1377
1378         btrfs_csum_final(crc, calculated_csum);
1379         if (memcmp(calculated_csum, csum, csum_size))
1380                 sblock->checksum_error = 1;
1381 }
1382
1383 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1384                                              struct scrub_block *sblock_good,
1385                                              int force_write)
1386 {
1387         int page_num;
1388         int ret = 0;
1389
1390         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1391                 int ret_sub;
1392
1393                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1394                                                            sblock_good,
1395                                                            page_num,
1396                                                            force_write);
1397                 if (ret_sub)
1398                         ret = ret_sub;
1399         }
1400
1401         return ret;
1402 }
1403
1404 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1405                                             struct scrub_block *sblock_good,
1406                                             int page_num, int force_write)
1407 {
1408         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1409         struct scrub_page *page_good = sblock_good->pagev[page_num];
1410
1411         BUG_ON(page_bad->page == NULL);
1412         BUG_ON(page_good->page == NULL);
1413         if (force_write || sblock_bad->header_error ||
1414             sblock_bad->checksum_error || page_bad->io_error) {
1415                 struct bio *bio;
1416                 int ret;
1417
1418                 if (!page_bad->dev->bdev) {
1419                         printk_ratelimited(KERN_WARNING
1420                                 "btrfs: scrub_repair_page_from_good_copy(bdev == NULL) is unexpected!\n");
1421                         return -EIO;
1422                 }
1423
1424                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1425                 if (!bio)
1426                         return -EIO;
1427                 bio->bi_bdev = page_bad->dev->bdev;
1428                 bio->bi_sector = page_bad->physical >> 9;
1429
1430                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1431                 if (PAGE_SIZE != ret) {
1432                         bio_put(bio);
1433                         return -EIO;
1434                 }
1435
1436                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1437                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1438                                 BTRFS_DEV_STAT_WRITE_ERRS);
1439                         btrfs_dev_replace_stats_inc(
1440                                 &sblock_bad->sctx->dev_root->fs_info->
1441                                 dev_replace.num_write_errors);
1442                         bio_put(bio);
1443                         return -EIO;
1444                 }
1445                 bio_put(bio);
1446         }
1447
1448         return 0;
1449 }
1450
1451 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1452 {
1453         int page_num;
1454
1455         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1456                 int ret;
1457
1458                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1459                 if (ret)
1460                         btrfs_dev_replace_stats_inc(
1461                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1462                                 num_write_errors);
1463         }
1464 }
1465
1466 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1467                                            int page_num)
1468 {
1469         struct scrub_page *spage = sblock->pagev[page_num];
1470
1471         BUG_ON(spage->page == NULL);
1472         if (spage->io_error) {
1473                 void *mapped_buffer = kmap_atomic(spage->page);
1474
1475                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1476                 flush_dcache_page(spage->page);
1477                 kunmap_atomic(mapped_buffer);
1478         }
1479         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1480 }
1481
1482 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1483                                     struct scrub_page *spage)
1484 {
1485         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1486         struct scrub_bio *sbio;
1487         int ret;
1488
1489         mutex_lock(&wr_ctx->wr_lock);
1490 again:
1491         if (!wr_ctx->wr_curr_bio) {
1492                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1493                                               GFP_NOFS);
1494                 if (!wr_ctx->wr_curr_bio) {
1495                         mutex_unlock(&wr_ctx->wr_lock);
1496                         return -ENOMEM;
1497                 }
1498                 wr_ctx->wr_curr_bio->sctx = sctx;
1499                 wr_ctx->wr_curr_bio->page_count = 0;
1500         }
1501         sbio = wr_ctx->wr_curr_bio;
1502         if (sbio->page_count == 0) {
1503                 struct bio *bio;
1504
1505                 sbio->physical = spage->physical_for_dev_replace;
1506                 sbio->logical = spage->logical;
1507                 sbio->dev = wr_ctx->tgtdev;
1508                 bio = sbio->bio;
1509                 if (!bio) {
1510                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1511                         if (!bio) {
1512                                 mutex_unlock(&wr_ctx->wr_lock);
1513                                 return -ENOMEM;
1514                         }
1515                         sbio->bio = bio;
1516                 }
1517
1518                 bio->bi_private = sbio;
1519                 bio->bi_end_io = scrub_wr_bio_end_io;
1520                 bio->bi_bdev = sbio->dev->bdev;
1521                 bio->bi_sector = sbio->physical >> 9;
1522                 sbio->err = 0;
1523         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1524                    spage->physical_for_dev_replace ||
1525                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1526                    spage->logical) {
1527                 scrub_wr_submit(sctx);
1528                 goto again;
1529         }
1530
1531         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1532         if (ret != PAGE_SIZE) {
1533                 if (sbio->page_count < 1) {
1534                         bio_put(sbio->bio);
1535                         sbio->bio = NULL;
1536                         mutex_unlock(&wr_ctx->wr_lock);
1537                         return -EIO;
1538                 }
1539                 scrub_wr_submit(sctx);
1540                 goto again;
1541         }
1542
1543         sbio->pagev[sbio->page_count] = spage;
1544         scrub_page_get(spage);
1545         sbio->page_count++;
1546         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1547                 scrub_wr_submit(sctx);
1548         mutex_unlock(&wr_ctx->wr_lock);
1549
1550         return 0;
1551 }
1552
1553 static void scrub_wr_submit(struct scrub_ctx *sctx)
1554 {
1555         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1556         struct scrub_bio *sbio;
1557
1558         if (!wr_ctx->wr_curr_bio)
1559                 return;
1560
1561         sbio = wr_ctx->wr_curr_bio;
1562         wr_ctx->wr_curr_bio = NULL;
1563         WARN_ON(!sbio->bio->bi_bdev);
1564         scrub_pending_bio_inc(sctx);
1565         /* process all writes in a single worker thread. Then the block layer
1566          * orders the requests before sending them to the driver which
1567          * doubled the write performance on spinning disks when measured
1568          * with Linux 3.5 */
1569         btrfsic_submit_bio(WRITE, sbio->bio);
1570 }
1571
1572 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1573 {
1574         struct scrub_bio *sbio = bio->bi_private;
1575         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1576
1577         sbio->err = err;
1578         sbio->bio = bio;
1579
1580         sbio->work.func = scrub_wr_bio_end_io_worker;
1581         btrfs_queue_worker(&fs_info->scrub_wr_completion_workers, &sbio->work);
1582 }
1583
1584 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1585 {
1586         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1587         struct scrub_ctx *sctx = sbio->sctx;
1588         int i;
1589
1590         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1591         if (sbio->err) {
1592                 struct btrfs_dev_replace *dev_replace =
1593                         &sbio->sctx->dev_root->fs_info->dev_replace;
1594
1595                 for (i = 0; i < sbio->page_count; i++) {
1596                         struct scrub_page *spage = sbio->pagev[i];
1597
1598                         spage->io_error = 1;
1599                         btrfs_dev_replace_stats_inc(&dev_replace->
1600                                                     num_write_errors);
1601                 }
1602         }
1603
1604         for (i = 0; i < sbio->page_count; i++)
1605                 scrub_page_put(sbio->pagev[i]);
1606
1607         bio_put(sbio->bio);
1608         kfree(sbio);
1609         scrub_pending_bio_dec(sctx);
1610 }
1611
1612 static int scrub_checksum(struct scrub_block *sblock)
1613 {
1614         u64 flags;
1615         int ret;
1616
1617         WARN_ON(sblock->page_count < 1);
1618         flags = sblock->pagev[0]->flags;
1619         ret = 0;
1620         if (flags & BTRFS_EXTENT_FLAG_DATA)
1621                 ret = scrub_checksum_data(sblock);
1622         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1623                 ret = scrub_checksum_tree_block(sblock);
1624         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1625                 (void)scrub_checksum_super(sblock);
1626         else
1627                 WARN_ON(1);
1628         if (ret)
1629                 scrub_handle_errored_block(sblock);
1630
1631         return ret;
1632 }
1633
1634 static int scrub_checksum_data(struct scrub_block *sblock)
1635 {
1636         struct scrub_ctx *sctx = sblock->sctx;
1637         u8 csum[BTRFS_CSUM_SIZE];
1638         u8 *on_disk_csum;
1639         struct page *page;
1640         void *buffer;
1641         u32 crc = ~(u32)0;
1642         int fail = 0;
1643         u64 len;
1644         int index;
1645
1646         BUG_ON(sblock->page_count < 1);
1647         if (!sblock->pagev[0]->have_csum)
1648                 return 0;
1649
1650         on_disk_csum = sblock->pagev[0]->csum;
1651         page = sblock->pagev[0]->page;
1652         buffer = kmap_atomic(page);
1653
1654         len = sctx->sectorsize;
1655         index = 0;
1656         for (;;) {
1657                 u64 l = min_t(u64, len, PAGE_SIZE);
1658
1659                 crc = btrfs_csum_data(buffer, crc, l);
1660                 kunmap_atomic(buffer);
1661                 len -= l;
1662                 if (len == 0)
1663                         break;
1664                 index++;
1665                 BUG_ON(index >= sblock->page_count);
1666                 BUG_ON(!sblock->pagev[index]->page);
1667                 page = sblock->pagev[index]->page;
1668                 buffer = kmap_atomic(page);
1669         }
1670
1671         btrfs_csum_final(crc, csum);
1672         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1673                 fail = 1;
1674
1675         return fail;
1676 }
1677
1678 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1679 {
1680         struct scrub_ctx *sctx = sblock->sctx;
1681         struct btrfs_header *h;
1682         struct btrfs_root *root = sctx->dev_root;
1683         struct btrfs_fs_info *fs_info = root->fs_info;
1684         u8 calculated_csum[BTRFS_CSUM_SIZE];
1685         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1686         struct page *page;
1687         void *mapped_buffer;
1688         u64 mapped_size;
1689         void *p;
1690         u32 crc = ~(u32)0;
1691         int fail = 0;
1692         int crc_fail = 0;
1693         u64 len;
1694         int index;
1695
1696         BUG_ON(sblock->page_count < 1);
1697         page = sblock->pagev[0]->page;
1698         mapped_buffer = kmap_atomic(page);
1699         h = (struct btrfs_header *)mapped_buffer;
1700         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1701
1702         /*
1703          * we don't use the getter functions here, as we
1704          * a) don't have an extent buffer and
1705          * b) the page is already kmapped
1706          */
1707
1708         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1709                 ++fail;
1710
1711         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1712                 ++fail;
1713
1714         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1715                 ++fail;
1716
1717         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1718                    BTRFS_UUID_SIZE))
1719                 ++fail;
1720
1721         WARN_ON(sctx->nodesize != sctx->leafsize);
1722         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1723         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1724         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1725         index = 0;
1726         for (;;) {
1727                 u64 l = min_t(u64, len, mapped_size);
1728
1729                 crc = btrfs_csum_data(p, crc, l);
1730                 kunmap_atomic(mapped_buffer);
1731                 len -= l;
1732                 if (len == 0)
1733                         break;
1734                 index++;
1735                 BUG_ON(index >= sblock->page_count);
1736                 BUG_ON(!sblock->pagev[index]->page);
1737                 page = sblock->pagev[index]->page;
1738                 mapped_buffer = kmap_atomic(page);
1739                 mapped_size = PAGE_SIZE;
1740                 p = mapped_buffer;
1741         }
1742
1743         btrfs_csum_final(crc, calculated_csum);
1744         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1745                 ++crc_fail;
1746
1747         return fail || crc_fail;
1748 }
1749
1750 static int scrub_checksum_super(struct scrub_block *sblock)
1751 {
1752         struct btrfs_super_block *s;
1753         struct scrub_ctx *sctx = sblock->sctx;
1754         struct btrfs_root *root = sctx->dev_root;
1755         struct btrfs_fs_info *fs_info = root->fs_info;
1756         u8 calculated_csum[BTRFS_CSUM_SIZE];
1757         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1758         struct page *page;
1759         void *mapped_buffer;
1760         u64 mapped_size;
1761         void *p;
1762         u32 crc = ~(u32)0;
1763         int fail_gen = 0;
1764         int fail_cor = 0;
1765         u64 len;
1766         int index;
1767
1768         BUG_ON(sblock->page_count < 1);
1769         page = sblock->pagev[0]->page;
1770         mapped_buffer = kmap_atomic(page);
1771         s = (struct btrfs_super_block *)mapped_buffer;
1772         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1773
1774         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1775                 ++fail_cor;
1776
1777         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1778                 ++fail_gen;
1779
1780         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1781                 ++fail_cor;
1782
1783         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1784         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1785         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1786         index = 0;
1787         for (;;) {
1788                 u64 l = min_t(u64, len, mapped_size);
1789
1790                 crc = btrfs_csum_data(p, crc, l);
1791                 kunmap_atomic(mapped_buffer);
1792                 len -= l;
1793                 if (len == 0)
1794                         break;
1795                 index++;
1796                 BUG_ON(index >= sblock->page_count);
1797                 BUG_ON(!sblock->pagev[index]->page);
1798                 page = sblock->pagev[index]->page;
1799                 mapped_buffer = kmap_atomic(page);
1800                 mapped_size = PAGE_SIZE;
1801                 p = mapped_buffer;
1802         }
1803
1804         btrfs_csum_final(crc, calculated_csum);
1805         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1806                 ++fail_cor;
1807
1808         if (fail_cor + fail_gen) {
1809                 /*
1810                  * if we find an error in a super block, we just report it.
1811                  * They will get written with the next transaction commit
1812                  * anyway
1813                  */
1814                 spin_lock(&sctx->stat_lock);
1815                 ++sctx->stat.super_errors;
1816                 spin_unlock(&sctx->stat_lock);
1817                 if (fail_cor)
1818                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1819                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1820                 else
1821                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1822                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1823         }
1824
1825         return fail_cor + fail_gen;
1826 }
1827
1828 static void scrub_block_get(struct scrub_block *sblock)
1829 {
1830         atomic_inc(&sblock->ref_count);
1831 }
1832
1833 static void scrub_block_put(struct scrub_block *sblock)
1834 {
1835         if (atomic_dec_and_test(&sblock->ref_count)) {
1836                 int i;
1837
1838                 for (i = 0; i < sblock->page_count; i++)
1839                         scrub_page_put(sblock->pagev[i]);
1840                 kfree(sblock);
1841         }
1842 }
1843
1844 static void scrub_page_get(struct scrub_page *spage)
1845 {
1846         atomic_inc(&spage->ref_count);
1847 }
1848
1849 static void scrub_page_put(struct scrub_page *spage)
1850 {
1851         if (atomic_dec_and_test(&spage->ref_count)) {
1852                 if (spage->page)
1853                         __free_page(spage->page);
1854                 kfree(spage);
1855         }
1856 }
1857
1858 static void scrub_submit(struct scrub_ctx *sctx)
1859 {
1860         struct scrub_bio *sbio;
1861
1862         if (sctx->curr == -1)
1863                 return;
1864
1865         sbio = sctx->bios[sctx->curr];
1866         sctx->curr = -1;
1867         scrub_pending_bio_inc(sctx);
1868
1869         if (!sbio->bio->bi_bdev) {
1870                 /*
1871                  * this case should not happen. If btrfs_map_block() is
1872                  * wrong, it could happen for dev-replace operations on
1873                  * missing devices when no mirrors are available, but in
1874                  * this case it should already fail the mount.
1875                  * This case is handled correctly (but _very_ slowly).
1876                  */
1877                 printk_ratelimited(KERN_WARNING
1878                         "btrfs: scrub_submit(bio bdev == NULL) is unexpected!\n");
1879                 bio_endio(sbio->bio, -EIO);
1880         } else {
1881                 btrfsic_submit_bio(READ, sbio->bio);
1882         }
1883 }
1884
1885 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1886                                     struct scrub_page *spage)
1887 {
1888         struct scrub_block *sblock = spage->sblock;
1889         struct scrub_bio *sbio;
1890         int ret;
1891
1892 again:
1893         /*
1894          * grab a fresh bio or wait for one to become available
1895          */
1896         while (sctx->curr == -1) {
1897                 spin_lock(&sctx->list_lock);
1898                 sctx->curr = sctx->first_free;
1899                 if (sctx->curr != -1) {
1900                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1901                         sctx->bios[sctx->curr]->next_free = -1;
1902                         sctx->bios[sctx->curr]->page_count = 0;
1903                         spin_unlock(&sctx->list_lock);
1904                 } else {
1905                         spin_unlock(&sctx->list_lock);
1906                         wait_event(sctx->list_wait, sctx->first_free != -1);
1907                 }
1908         }
1909         sbio = sctx->bios[sctx->curr];
1910         if (sbio->page_count == 0) {
1911                 struct bio *bio;
1912
1913                 sbio->physical = spage->physical;
1914                 sbio->logical = spage->logical;
1915                 sbio->dev = spage->dev;
1916                 bio = sbio->bio;
1917                 if (!bio) {
1918                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1919                         if (!bio)
1920                                 return -ENOMEM;
1921                         sbio->bio = bio;
1922                 }
1923
1924                 bio->bi_private = sbio;
1925                 bio->bi_end_io = scrub_bio_end_io;
1926                 bio->bi_bdev = sbio->dev->bdev;
1927                 bio->bi_sector = sbio->physical >> 9;
1928                 sbio->err = 0;
1929         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1930                    spage->physical ||
1931                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1932                    spage->logical ||
1933                    sbio->dev != spage->dev) {
1934                 scrub_submit(sctx);
1935                 goto again;
1936         }
1937
1938         sbio->pagev[sbio->page_count] = spage;
1939         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1940         if (ret != PAGE_SIZE) {
1941                 if (sbio->page_count < 1) {
1942                         bio_put(sbio->bio);
1943                         sbio->bio = NULL;
1944                         return -EIO;
1945                 }
1946                 scrub_submit(sctx);
1947                 goto again;
1948         }
1949
1950         scrub_block_get(sblock); /* one for the page added to the bio */
1951         atomic_inc(&sblock->outstanding_pages);
1952         sbio->page_count++;
1953         if (sbio->page_count == sctx->pages_per_rd_bio)
1954                 scrub_submit(sctx);
1955
1956         return 0;
1957 }
1958
1959 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1960                        u64 physical, struct btrfs_device *dev, u64 flags,
1961                        u64 gen, int mirror_num, u8 *csum, int force,
1962                        u64 physical_for_dev_replace)
1963 {
1964         struct scrub_block *sblock;
1965         int index;
1966
1967         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
1968         if (!sblock) {
1969                 spin_lock(&sctx->stat_lock);
1970                 sctx->stat.malloc_errors++;
1971                 spin_unlock(&sctx->stat_lock);
1972                 return -ENOMEM;
1973         }
1974
1975         /* one ref inside this function, plus one for each page added to
1976          * a bio later on */
1977         atomic_set(&sblock->ref_count, 1);
1978         sblock->sctx = sctx;
1979         sblock->no_io_error_seen = 1;
1980
1981         for (index = 0; len > 0; index++) {
1982                 struct scrub_page *spage;
1983                 u64 l = min_t(u64, len, PAGE_SIZE);
1984
1985                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
1986                 if (!spage) {
1987 leave_nomem:
1988                         spin_lock(&sctx->stat_lock);
1989                         sctx->stat.malloc_errors++;
1990                         spin_unlock(&sctx->stat_lock);
1991                         scrub_block_put(sblock);
1992                         return -ENOMEM;
1993                 }
1994                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
1995                 scrub_page_get(spage);
1996                 sblock->pagev[index] = spage;
1997                 spage->sblock = sblock;
1998                 spage->dev = dev;
1999                 spage->flags = flags;
2000                 spage->generation = gen;
2001                 spage->logical = logical;
2002                 spage->physical = physical;
2003                 spage->physical_for_dev_replace = physical_for_dev_replace;
2004                 spage->mirror_num = mirror_num;
2005                 if (csum) {
2006                         spage->have_csum = 1;
2007                         memcpy(spage->csum, csum, sctx->csum_size);
2008                 } else {
2009                         spage->have_csum = 0;
2010                 }
2011                 sblock->page_count++;
2012                 spage->page = alloc_page(GFP_NOFS);
2013                 if (!spage->page)
2014                         goto leave_nomem;
2015                 len -= l;
2016                 logical += l;
2017                 physical += l;
2018                 physical_for_dev_replace += l;
2019         }
2020
2021         WARN_ON(sblock->page_count == 0);
2022         for (index = 0; index < sblock->page_count; index++) {
2023                 struct scrub_page *spage = sblock->pagev[index];
2024                 int ret;
2025
2026                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2027                 if (ret) {
2028                         scrub_block_put(sblock);
2029                         return ret;
2030                 }
2031         }
2032
2033         if (force)
2034                 scrub_submit(sctx);
2035
2036         /* last one frees, either here or in bio completion for last page */
2037         scrub_block_put(sblock);
2038         return 0;
2039 }
2040
2041 static void scrub_bio_end_io(struct bio *bio, int err)
2042 {
2043         struct scrub_bio *sbio = bio->bi_private;
2044         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2045
2046         sbio->err = err;
2047         sbio->bio = bio;
2048
2049         btrfs_queue_worker(&fs_info->scrub_workers, &sbio->work);
2050 }
2051
2052 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2053 {
2054         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2055         struct scrub_ctx *sctx = sbio->sctx;
2056         int i;
2057
2058         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2059         if (sbio->err) {
2060                 for (i = 0; i < sbio->page_count; i++) {
2061                         struct scrub_page *spage = sbio->pagev[i];
2062
2063                         spage->io_error = 1;
2064                         spage->sblock->no_io_error_seen = 0;
2065                 }
2066         }
2067
2068         /* now complete the scrub_block items that have all pages completed */
2069         for (i = 0; i < sbio->page_count; i++) {
2070                 struct scrub_page *spage = sbio->pagev[i];
2071                 struct scrub_block *sblock = spage->sblock;
2072
2073                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2074                         scrub_block_complete(sblock);
2075                 scrub_block_put(sblock);
2076         }
2077
2078         bio_put(sbio->bio);
2079         sbio->bio = NULL;
2080         spin_lock(&sctx->list_lock);
2081         sbio->next_free = sctx->first_free;
2082         sctx->first_free = sbio->index;
2083         spin_unlock(&sctx->list_lock);
2084
2085         if (sctx->is_dev_replace &&
2086             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2087                 mutex_lock(&sctx->wr_ctx.wr_lock);
2088                 scrub_wr_submit(sctx);
2089                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2090         }
2091
2092         scrub_pending_bio_dec(sctx);
2093 }
2094
2095 static void scrub_block_complete(struct scrub_block *sblock)
2096 {
2097         if (!sblock->no_io_error_seen) {
2098                 scrub_handle_errored_block(sblock);
2099         } else {
2100                 /*
2101                  * if has checksum error, write via repair mechanism in
2102                  * dev replace case, otherwise write here in dev replace
2103                  * case.
2104                  */
2105                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2106                         scrub_write_block_to_dev_replace(sblock);
2107         }
2108 }
2109
2110 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2111                            u8 *csum)
2112 {
2113         struct btrfs_ordered_sum *sum = NULL;
2114         unsigned long index;
2115         unsigned long num_sectors;
2116
2117         while (!list_empty(&sctx->csum_list)) {
2118                 sum = list_first_entry(&sctx->csum_list,
2119                                        struct btrfs_ordered_sum, list);
2120                 if (sum->bytenr > logical)
2121                         return 0;
2122                 if (sum->bytenr + sum->len > logical)
2123                         break;
2124
2125                 ++sctx->stat.csum_discards;
2126                 list_del(&sum->list);
2127                 kfree(sum);
2128                 sum = NULL;
2129         }
2130         if (!sum)
2131                 return 0;
2132
2133         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2134         num_sectors = sum->len / sctx->sectorsize;
2135         memcpy(csum, sum->sums + index, sctx->csum_size);
2136         if (index == num_sectors - 1) {
2137                 list_del(&sum->list);
2138                 kfree(sum);
2139         }
2140         return 1;
2141 }
2142
2143 /* scrub extent tries to collect up to 64 kB for each bio */
2144 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2145                         u64 physical, struct btrfs_device *dev, u64 flags,
2146                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2147 {
2148         int ret;
2149         u8 csum[BTRFS_CSUM_SIZE];
2150         u32 blocksize;
2151
2152         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2153                 blocksize = sctx->sectorsize;
2154                 spin_lock(&sctx->stat_lock);
2155                 sctx->stat.data_extents_scrubbed++;
2156                 sctx->stat.data_bytes_scrubbed += len;
2157                 spin_unlock(&sctx->stat_lock);
2158         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2159                 WARN_ON(sctx->nodesize != sctx->leafsize);
2160                 blocksize = sctx->nodesize;
2161                 spin_lock(&sctx->stat_lock);
2162                 sctx->stat.tree_extents_scrubbed++;
2163                 sctx->stat.tree_bytes_scrubbed += len;
2164                 spin_unlock(&sctx->stat_lock);
2165         } else {
2166                 blocksize = sctx->sectorsize;
2167                 WARN_ON(1);
2168         }
2169
2170         while (len) {
2171                 u64 l = min_t(u64, len, blocksize);
2172                 int have_csum = 0;
2173
2174                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2175                         /* push csums to sbio */
2176                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2177                         if (have_csum == 0)
2178                                 ++sctx->stat.no_csum;
2179                         if (sctx->is_dev_replace && !have_csum) {
2180                                 ret = copy_nocow_pages(sctx, logical, l,
2181                                                        mirror_num,
2182                                                       physical_for_dev_replace);
2183                                 goto behind_scrub_pages;
2184                         }
2185                 }
2186                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2187                                   mirror_num, have_csum ? csum : NULL, 0,
2188                                   physical_for_dev_replace);
2189 behind_scrub_pages:
2190                 if (ret)
2191                         return ret;
2192                 len -= l;
2193                 logical += l;
2194                 physical += l;
2195                 physical_for_dev_replace += l;
2196         }
2197         return 0;
2198 }
2199
2200 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2201                                            struct map_lookup *map,
2202                                            struct btrfs_device *scrub_dev,
2203                                            int num, u64 base, u64 length,
2204                                            int is_dev_replace)
2205 {
2206         struct btrfs_path *path;
2207         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2208         struct btrfs_root *root = fs_info->extent_root;
2209         struct btrfs_root *csum_root = fs_info->csum_root;
2210         struct btrfs_extent_item *extent;
2211         struct blk_plug plug;
2212         u64 flags;
2213         int ret;
2214         int slot;
2215         u64 nstripes;
2216         struct extent_buffer *l;
2217         struct btrfs_key key;
2218         u64 physical;
2219         u64 logical;
2220         u64 logic_end;
2221         u64 generation;
2222         int mirror_num;
2223         struct reada_control *reada1;
2224         struct reada_control *reada2;
2225         struct btrfs_key key_start;
2226         struct btrfs_key key_end;
2227         u64 increment = map->stripe_len;
2228         u64 offset;
2229         u64 extent_logical;
2230         u64 extent_physical;
2231         u64 extent_len;
2232         struct btrfs_device *extent_dev;
2233         int extent_mirror_num;
2234         int stop_loop;
2235
2236         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2237                          BTRFS_BLOCK_GROUP_RAID6)) {
2238                 if (num >= nr_data_stripes(map)) {
2239                         return 0;
2240                 }
2241         }
2242
2243         nstripes = length;
2244         offset = 0;
2245         do_div(nstripes, map->stripe_len);
2246         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2247                 offset = map->stripe_len * num;
2248                 increment = map->stripe_len * map->num_stripes;
2249                 mirror_num = 1;
2250         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2251                 int factor = map->num_stripes / map->sub_stripes;
2252                 offset = map->stripe_len * (num / map->sub_stripes);
2253                 increment = map->stripe_len * factor;
2254                 mirror_num = num % map->sub_stripes + 1;
2255         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2256                 increment = map->stripe_len;
2257                 mirror_num = num % map->num_stripes + 1;
2258         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2259                 increment = map->stripe_len;
2260                 mirror_num = num % map->num_stripes + 1;
2261         } else {
2262                 increment = map->stripe_len;
2263                 mirror_num = 1;
2264         }
2265
2266         path = btrfs_alloc_path();
2267         if (!path)
2268                 return -ENOMEM;
2269
2270         /*
2271          * work on commit root. The related disk blocks are static as
2272          * long as COW is applied. This means, it is save to rewrite
2273          * them to repair disk errors without any race conditions
2274          */
2275         path->search_commit_root = 1;
2276         path->skip_locking = 1;
2277
2278         /*
2279          * trigger the readahead for extent tree csum tree and wait for
2280          * completion. During readahead, the scrub is officially paused
2281          * to not hold off transaction commits
2282          */
2283         logical = base + offset;
2284
2285         wait_event(sctx->list_wait,
2286                    atomic_read(&sctx->bios_in_flight) == 0);
2287         atomic_inc(&fs_info->scrubs_paused);
2288         wake_up(&fs_info->scrub_pause_wait);
2289
2290         /* FIXME it might be better to start readahead at commit root */
2291         key_start.objectid = logical;
2292         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2293         key_start.offset = (u64)0;
2294         key_end.objectid = base + offset + nstripes * increment;
2295         key_end.type = BTRFS_METADATA_ITEM_KEY;
2296         key_end.offset = (u64)-1;
2297         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2298
2299         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2300         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2301         key_start.offset = logical;
2302         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2303         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2304         key_end.offset = base + offset + nstripes * increment;
2305         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2306
2307         if (!IS_ERR(reada1))
2308                 btrfs_reada_wait(reada1);
2309         if (!IS_ERR(reada2))
2310                 btrfs_reada_wait(reada2);
2311
2312         mutex_lock(&fs_info->scrub_lock);
2313         while (atomic_read(&fs_info->scrub_pause_req)) {
2314                 mutex_unlock(&fs_info->scrub_lock);
2315                 wait_event(fs_info->scrub_pause_wait,
2316                    atomic_read(&fs_info->scrub_pause_req) == 0);
2317                 mutex_lock(&fs_info->scrub_lock);
2318         }
2319         atomic_dec(&fs_info->scrubs_paused);
2320         mutex_unlock(&fs_info->scrub_lock);
2321         wake_up(&fs_info->scrub_pause_wait);
2322
2323         /*
2324          * collect all data csums for the stripe to avoid seeking during
2325          * the scrub. This might currently (crc32) end up to be about 1MB
2326          */
2327         blk_start_plug(&plug);
2328
2329         /*
2330          * now find all extents for each stripe and scrub them
2331          */
2332         logical = base + offset;
2333         physical = map->stripes[num].physical;
2334         logic_end = logical + increment * nstripes;
2335         ret = 0;
2336         while (logical < logic_end) {
2337                 /*
2338                  * canceled?
2339                  */
2340                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2341                     atomic_read(&sctx->cancel_req)) {
2342                         ret = -ECANCELED;
2343                         goto out;
2344                 }
2345                 /*
2346                  * check to see if we have to pause
2347                  */
2348                 if (atomic_read(&fs_info->scrub_pause_req)) {
2349                         /* push queued extents */
2350                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2351                         scrub_submit(sctx);
2352                         mutex_lock(&sctx->wr_ctx.wr_lock);
2353                         scrub_wr_submit(sctx);
2354                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2355                         wait_event(sctx->list_wait,
2356                                    atomic_read(&sctx->bios_in_flight) == 0);
2357                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2358                         atomic_inc(&fs_info->scrubs_paused);
2359                         wake_up(&fs_info->scrub_pause_wait);
2360                         mutex_lock(&fs_info->scrub_lock);
2361                         while (atomic_read(&fs_info->scrub_pause_req)) {
2362                                 mutex_unlock(&fs_info->scrub_lock);
2363                                 wait_event(fs_info->scrub_pause_wait,
2364                                    atomic_read(&fs_info->scrub_pause_req) == 0);
2365                                 mutex_lock(&fs_info->scrub_lock);
2366                         }
2367                         atomic_dec(&fs_info->scrubs_paused);
2368                         mutex_unlock(&fs_info->scrub_lock);
2369                         wake_up(&fs_info->scrub_pause_wait);
2370                 }
2371
2372                 key.objectid = logical;
2373                 key.type = BTRFS_EXTENT_ITEM_KEY;
2374                 key.offset = (u64)-1;
2375
2376                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2377                 if (ret < 0)
2378                         goto out;
2379
2380                 if (ret > 0) {
2381                         ret = btrfs_previous_item(root, path, 0,
2382                                                   BTRFS_EXTENT_ITEM_KEY);
2383                         if (ret < 0)
2384                                 goto out;
2385                         if (ret > 0) {
2386                                 /* there's no smaller item, so stick with the
2387                                  * larger one */
2388                                 btrfs_release_path(path);
2389                                 ret = btrfs_search_slot(NULL, root, &key,
2390                                                         path, 0, 0);
2391                                 if (ret < 0)
2392                                         goto out;
2393                         }
2394                 }
2395
2396                 stop_loop = 0;
2397                 while (1) {
2398                         u64 bytes;
2399
2400                         l = path->nodes[0];
2401                         slot = path->slots[0];
2402                         if (slot >= btrfs_header_nritems(l)) {
2403                                 ret = btrfs_next_leaf(root, path);
2404                                 if (ret == 0)
2405                                         continue;
2406                                 if (ret < 0)
2407                                         goto out;
2408
2409                                 stop_loop = 1;
2410                                 break;
2411                         }
2412                         btrfs_item_key_to_cpu(l, &key, slot);
2413
2414                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2415                                 bytes = root->leafsize;
2416                         else
2417                                 bytes = key.offset;
2418
2419                         if (key.objectid + bytes <= logical)
2420                                 goto next;
2421
2422                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2423                             key.type != BTRFS_METADATA_ITEM_KEY)
2424                                 goto next;
2425
2426                         if (key.objectid >= logical + map->stripe_len) {
2427                                 /* out of this device extent */
2428                                 if (key.objectid >= logic_end)
2429                                         stop_loop = 1;
2430                                 break;
2431                         }
2432
2433                         extent = btrfs_item_ptr(l, slot,
2434                                                 struct btrfs_extent_item);
2435                         flags = btrfs_extent_flags(l, extent);
2436                         generation = btrfs_extent_generation(l, extent);
2437
2438                         if (key.objectid < logical &&
2439                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2440                                 printk(KERN_ERR
2441                                        "btrfs scrub: tree block %llu spanning "
2442                                        "stripes, ignored. logical=%llu\n",
2443                                        key.objectid, logical);
2444                                 goto next;
2445                         }
2446
2447 again:
2448                         extent_logical = key.objectid;
2449                         extent_len = bytes;
2450
2451                         /*
2452                          * trim extent to this stripe
2453                          */
2454                         if (extent_logical < logical) {
2455                                 extent_len -= logical - extent_logical;
2456                                 extent_logical = logical;
2457                         }
2458                         if (extent_logical + extent_len >
2459                             logical + map->stripe_len) {
2460                                 extent_len = logical + map->stripe_len -
2461                                              extent_logical;
2462                         }
2463
2464                         extent_physical = extent_logical - logical + physical;
2465                         extent_dev = scrub_dev;
2466                         extent_mirror_num = mirror_num;
2467                         if (is_dev_replace)
2468                                 scrub_remap_extent(fs_info, extent_logical,
2469                                                    extent_len, &extent_physical,
2470                                                    &extent_dev,
2471                                                    &extent_mirror_num);
2472
2473                         ret = btrfs_lookup_csums_range(csum_root, logical,
2474                                                 logical + map->stripe_len - 1,
2475                                                 &sctx->csum_list, 1);
2476                         if (ret)
2477                                 goto out;
2478
2479                         ret = scrub_extent(sctx, extent_logical, extent_len,
2480                                            extent_physical, extent_dev, flags,
2481                                            generation, extent_mirror_num,
2482                                            extent_logical - logical + physical);
2483                         if (ret)
2484                                 goto out;
2485
2486                         scrub_free_csums(sctx);
2487                         if (extent_logical + extent_len <
2488                             key.objectid + bytes) {
2489                                 logical += increment;
2490                                 physical += map->stripe_len;
2491
2492                                 if (logical < key.objectid + bytes) {
2493                                         cond_resched();
2494                                         goto again;
2495                                 }
2496
2497                                 if (logical >= logic_end) {
2498                                         stop_loop = 1;
2499                                         break;
2500                                 }
2501                         }
2502 next:
2503                         path->slots[0]++;
2504                 }
2505                 btrfs_release_path(path);
2506                 logical += increment;
2507                 physical += map->stripe_len;
2508                 spin_lock(&sctx->stat_lock);
2509                 if (stop_loop)
2510                         sctx->stat.last_physical = map->stripes[num].physical +
2511                                                    length;
2512                 else
2513                         sctx->stat.last_physical = physical;
2514                 spin_unlock(&sctx->stat_lock);
2515                 if (stop_loop)
2516                         break;
2517         }
2518 out:
2519         /* push queued extents */
2520         scrub_submit(sctx);
2521         mutex_lock(&sctx->wr_ctx.wr_lock);
2522         scrub_wr_submit(sctx);
2523         mutex_unlock(&sctx->wr_ctx.wr_lock);
2524
2525         blk_finish_plug(&plug);
2526         btrfs_free_path(path);
2527         return ret < 0 ? ret : 0;
2528 }
2529
2530 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2531                                           struct btrfs_device *scrub_dev,
2532                                           u64 chunk_tree, u64 chunk_objectid,
2533                                           u64 chunk_offset, u64 length,
2534                                           u64 dev_offset, int is_dev_replace)
2535 {
2536         struct btrfs_mapping_tree *map_tree =
2537                 &sctx->dev_root->fs_info->mapping_tree;
2538         struct map_lookup *map;
2539         struct extent_map *em;
2540         int i;
2541         int ret = 0;
2542
2543         read_lock(&map_tree->map_tree.lock);
2544         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2545         read_unlock(&map_tree->map_tree.lock);
2546
2547         if (!em)
2548                 return -EINVAL;
2549
2550         map = (struct map_lookup *)em->bdev;
2551         if (em->start != chunk_offset)
2552                 goto out;
2553
2554         if (em->len < length)
2555                 goto out;
2556
2557         for (i = 0; i < map->num_stripes; ++i) {
2558                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2559                     map->stripes[i].physical == dev_offset) {
2560                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2561                                            chunk_offset, length,
2562                                            is_dev_replace);
2563                         if (ret)
2564                                 goto out;
2565                 }
2566         }
2567 out:
2568         free_extent_map(em);
2569
2570         return ret;
2571 }
2572
2573 static noinline_for_stack
2574 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2575                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2576                            int is_dev_replace)
2577 {
2578         struct btrfs_dev_extent *dev_extent = NULL;
2579         struct btrfs_path *path;
2580         struct btrfs_root *root = sctx->dev_root;
2581         struct btrfs_fs_info *fs_info = root->fs_info;
2582         u64 length;
2583         u64 chunk_tree;
2584         u64 chunk_objectid;
2585         u64 chunk_offset;
2586         int ret;
2587         int slot;
2588         struct extent_buffer *l;
2589         struct btrfs_key key;
2590         struct btrfs_key found_key;
2591         struct btrfs_block_group_cache *cache;
2592         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2593
2594         path = btrfs_alloc_path();
2595         if (!path)
2596                 return -ENOMEM;
2597
2598         path->reada = 2;
2599         path->search_commit_root = 1;
2600         path->skip_locking = 1;
2601
2602         key.objectid = scrub_dev->devid;
2603         key.offset = 0ull;
2604         key.type = BTRFS_DEV_EXTENT_KEY;
2605
2606         while (1) {
2607                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2608                 if (ret < 0)
2609                         break;
2610                 if (ret > 0) {
2611                         if (path->slots[0] >=
2612                             btrfs_header_nritems(path->nodes[0])) {
2613                                 ret = btrfs_next_leaf(root, path);
2614                                 if (ret)
2615                                         break;
2616                         }
2617                 }
2618
2619                 l = path->nodes[0];
2620                 slot = path->slots[0];
2621
2622                 btrfs_item_key_to_cpu(l, &found_key, slot);
2623
2624                 if (found_key.objectid != scrub_dev->devid)
2625                         break;
2626
2627                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2628                         break;
2629
2630                 if (found_key.offset >= end)
2631                         break;
2632
2633                 if (found_key.offset < key.offset)
2634                         break;
2635
2636                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2637                 length = btrfs_dev_extent_length(l, dev_extent);
2638
2639                 if (found_key.offset + length <= start) {
2640                         key.offset = found_key.offset + length;
2641                         btrfs_release_path(path);
2642                         continue;
2643                 }
2644
2645                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2646                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2647                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2648
2649                 /*
2650                  * get a reference on the corresponding block group to prevent
2651                  * the chunk from going away while we scrub it
2652                  */
2653                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2654                 if (!cache) {
2655                         ret = -ENOENT;
2656                         break;
2657                 }
2658                 dev_replace->cursor_right = found_key.offset + length;
2659                 dev_replace->cursor_left = found_key.offset;
2660                 dev_replace->item_needs_writeback = 1;
2661                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2662                                   chunk_offset, length, found_key.offset,
2663                                   is_dev_replace);
2664
2665                 /*
2666                  * flush, submit all pending read and write bios, afterwards
2667                  * wait for them.
2668                  * Note that in the dev replace case, a read request causes
2669                  * write requests that are submitted in the read completion
2670                  * worker. Therefore in the current situation, it is required
2671                  * that all write requests are flushed, so that all read and
2672                  * write requests are really completed when bios_in_flight
2673                  * changes to 0.
2674                  */
2675                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2676                 scrub_submit(sctx);
2677                 mutex_lock(&sctx->wr_ctx.wr_lock);
2678                 scrub_wr_submit(sctx);
2679                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2680
2681                 wait_event(sctx->list_wait,
2682                            atomic_read(&sctx->bios_in_flight) == 0);
2683                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2684                 atomic_inc(&fs_info->scrubs_paused);
2685                 wake_up(&fs_info->scrub_pause_wait);
2686                 wait_event(sctx->list_wait,
2687                            atomic_read(&sctx->workers_pending) == 0);
2688
2689                 mutex_lock(&fs_info->scrub_lock);
2690                 while (atomic_read(&fs_info->scrub_pause_req)) {
2691                         mutex_unlock(&fs_info->scrub_lock);
2692                         wait_event(fs_info->scrub_pause_wait,
2693                            atomic_read(&fs_info->scrub_pause_req) == 0);
2694                         mutex_lock(&fs_info->scrub_lock);
2695                 }
2696                 atomic_dec(&fs_info->scrubs_paused);
2697                 mutex_unlock(&fs_info->scrub_lock);
2698                 wake_up(&fs_info->scrub_pause_wait);
2699
2700                 btrfs_put_block_group(cache);
2701                 if (ret)
2702                         break;
2703                 if (is_dev_replace &&
2704                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2705                         ret = -EIO;
2706                         break;
2707                 }
2708                 if (sctx->stat.malloc_errors > 0) {
2709                         ret = -ENOMEM;
2710                         break;
2711                 }
2712
2713                 dev_replace->cursor_left = dev_replace->cursor_right;
2714                 dev_replace->item_needs_writeback = 1;
2715
2716                 key.offset = found_key.offset + length;
2717                 btrfs_release_path(path);
2718         }
2719
2720         btrfs_free_path(path);
2721
2722         /*
2723          * ret can still be 1 from search_slot or next_leaf,
2724          * that's not an error
2725          */
2726         return ret < 0 ? ret : 0;
2727 }
2728
2729 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2730                                            struct btrfs_device *scrub_dev)
2731 {
2732         int     i;
2733         u64     bytenr;
2734         u64     gen;
2735         int     ret;
2736         struct btrfs_root *root = sctx->dev_root;
2737
2738         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2739                 return -EIO;
2740
2741         gen = root->fs_info->last_trans_committed;
2742
2743         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2744                 bytenr = btrfs_sb_offset(i);
2745                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2746                         break;
2747
2748                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2749                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2750                                   NULL, 1, bytenr);
2751                 if (ret)
2752                         return ret;
2753         }
2754         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2755
2756         return 0;
2757 }
2758
2759 /*
2760  * get a reference count on fs_info->scrub_workers. start worker if necessary
2761  */
2762 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2763                                                 int is_dev_replace)
2764 {
2765         int ret = 0;
2766
2767         if (fs_info->scrub_workers_refcnt == 0) {
2768                 if (is_dev_replace)
2769                         btrfs_init_workers(&fs_info->scrub_workers, "scrub", 1,
2770                                         &fs_info->generic_worker);
2771                 else
2772                         btrfs_init_workers(&fs_info->scrub_workers, "scrub",
2773                                         fs_info->thread_pool_size,
2774                                         &fs_info->generic_worker);
2775                 fs_info->scrub_workers.idle_thresh = 4;
2776                 ret = btrfs_start_workers(&fs_info->scrub_workers);
2777                 if (ret)
2778                         goto out;
2779                 btrfs_init_workers(&fs_info->scrub_wr_completion_workers,
2780                                    "scrubwrc",
2781                                    fs_info->thread_pool_size,
2782                                    &fs_info->generic_worker);
2783                 fs_info->scrub_wr_completion_workers.idle_thresh = 2;
2784                 ret = btrfs_start_workers(
2785                                 &fs_info->scrub_wr_completion_workers);
2786                 if (ret)
2787                         goto out;
2788                 btrfs_init_workers(&fs_info->scrub_nocow_workers, "scrubnc", 1,
2789                                    &fs_info->generic_worker);
2790                 ret = btrfs_start_workers(&fs_info->scrub_nocow_workers);
2791                 if (ret)
2792                         goto out;
2793         }
2794         ++fs_info->scrub_workers_refcnt;
2795 out:
2796         return ret;
2797 }
2798
2799 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2800 {
2801         if (--fs_info->scrub_workers_refcnt == 0) {
2802                 btrfs_stop_workers(&fs_info->scrub_workers);
2803                 btrfs_stop_workers(&fs_info->scrub_wr_completion_workers);
2804                 btrfs_stop_workers(&fs_info->scrub_nocow_workers);
2805         }
2806         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2807 }
2808
2809 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2810                     u64 end, struct btrfs_scrub_progress *progress,
2811                     int readonly, int is_dev_replace)
2812 {
2813         struct scrub_ctx *sctx;
2814         int ret;
2815         struct btrfs_device *dev;
2816
2817         if (btrfs_fs_closing(fs_info))
2818                 return -EINVAL;
2819
2820         /*
2821          * check some assumptions
2822          */
2823         if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2824                 printk(KERN_ERR
2825                        "btrfs_scrub: size assumption nodesize == leafsize (%d == %d) fails\n",
2826                        fs_info->chunk_root->nodesize,
2827                        fs_info->chunk_root->leafsize);
2828                 return -EINVAL;
2829         }
2830
2831         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2832                 /*
2833                  * in this case scrub is unable to calculate the checksum
2834                  * the way scrub is implemented. Do not handle this
2835                  * situation at all because it won't ever happen.
2836                  */
2837                 printk(KERN_ERR
2838                        "btrfs_scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails\n",
2839                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2840                 return -EINVAL;
2841         }
2842
2843         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2844                 /* not supported for data w/o checksums */
2845                 printk(KERN_ERR
2846                        "btrfs_scrub: size assumption sectorsize != PAGE_SIZE (%d != %lu) fails\n",
2847                        fs_info->chunk_root->sectorsize, PAGE_SIZE);
2848                 return -EINVAL;
2849         }
2850
2851         if (fs_info->chunk_root->nodesize >
2852             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2853             fs_info->chunk_root->sectorsize >
2854             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2855                 /*
2856                  * would exhaust the array bounds of pagev member in
2857                  * struct scrub_block
2858                  */
2859                 pr_err("btrfs_scrub: size assumption nodesize and sectorsize <= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails\n",
2860                        fs_info->chunk_root->nodesize,
2861                        SCRUB_MAX_PAGES_PER_BLOCK,
2862                        fs_info->chunk_root->sectorsize,
2863                        SCRUB_MAX_PAGES_PER_BLOCK);
2864                 return -EINVAL;
2865         }
2866
2867
2868         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2869         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2870         if (!dev || (dev->missing && !is_dev_replace)) {
2871                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2872                 return -ENODEV;
2873         }
2874
2875         mutex_lock(&fs_info->scrub_lock);
2876         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2877                 mutex_unlock(&fs_info->scrub_lock);
2878                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2879                 return -EIO;
2880         }
2881
2882         btrfs_dev_replace_lock(&fs_info->dev_replace);
2883         if (dev->scrub_device ||
2884             (!is_dev_replace &&
2885              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2886                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2887                 mutex_unlock(&fs_info->scrub_lock);
2888                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2889                 return -EINPROGRESS;
2890         }
2891         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2892
2893         ret = scrub_workers_get(fs_info, is_dev_replace);
2894         if (ret) {
2895                 mutex_unlock(&fs_info->scrub_lock);
2896                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2897                 return ret;
2898         }
2899
2900         sctx = scrub_setup_ctx(dev, is_dev_replace);
2901         if (IS_ERR(sctx)) {
2902                 mutex_unlock(&fs_info->scrub_lock);
2903                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2904                 scrub_workers_put(fs_info);
2905                 return PTR_ERR(sctx);
2906         }
2907         sctx->readonly = readonly;
2908         dev->scrub_device = sctx;
2909
2910         atomic_inc(&fs_info->scrubs_running);
2911         mutex_unlock(&fs_info->scrub_lock);
2912
2913         if (!is_dev_replace) {
2914                 /*
2915                  * by holding device list mutex, we can
2916                  * kick off writing super in log tree sync.
2917                  */
2918                 ret = scrub_supers(sctx, dev);
2919         }
2920         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2921
2922         if (!ret)
2923                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
2924                                              is_dev_replace);
2925
2926         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2927         atomic_dec(&fs_info->scrubs_running);
2928         wake_up(&fs_info->scrub_pause_wait);
2929
2930         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
2931
2932         if (progress)
2933                 memcpy(progress, &sctx->stat, sizeof(*progress));
2934
2935         mutex_lock(&fs_info->scrub_lock);
2936         dev->scrub_device = NULL;
2937         scrub_workers_put(fs_info);
2938         mutex_unlock(&fs_info->scrub_lock);
2939
2940         scrub_free_ctx(sctx);
2941
2942         return ret;
2943 }
2944
2945 void btrfs_scrub_pause(struct btrfs_root *root)
2946 {
2947         struct btrfs_fs_info *fs_info = root->fs_info;
2948
2949         mutex_lock(&fs_info->scrub_lock);
2950         atomic_inc(&fs_info->scrub_pause_req);
2951         while (atomic_read(&fs_info->scrubs_paused) !=
2952                atomic_read(&fs_info->scrubs_running)) {
2953                 mutex_unlock(&fs_info->scrub_lock);
2954                 wait_event(fs_info->scrub_pause_wait,
2955                            atomic_read(&fs_info->scrubs_paused) ==
2956                            atomic_read(&fs_info->scrubs_running));
2957                 mutex_lock(&fs_info->scrub_lock);
2958         }
2959         mutex_unlock(&fs_info->scrub_lock);
2960 }
2961
2962 void btrfs_scrub_continue(struct btrfs_root *root)
2963 {
2964         struct btrfs_fs_info *fs_info = root->fs_info;
2965
2966         atomic_dec(&fs_info->scrub_pause_req);
2967         wake_up(&fs_info->scrub_pause_wait);
2968 }
2969
2970 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
2971 {
2972         mutex_lock(&fs_info->scrub_lock);
2973         if (!atomic_read(&fs_info->scrubs_running)) {
2974                 mutex_unlock(&fs_info->scrub_lock);
2975                 return -ENOTCONN;
2976         }
2977
2978         atomic_inc(&fs_info->scrub_cancel_req);
2979         while (atomic_read(&fs_info->scrubs_running)) {
2980                 mutex_unlock(&fs_info->scrub_lock);
2981                 wait_event(fs_info->scrub_pause_wait,
2982                            atomic_read(&fs_info->scrubs_running) == 0);
2983                 mutex_lock(&fs_info->scrub_lock);
2984         }
2985         atomic_dec(&fs_info->scrub_cancel_req);
2986         mutex_unlock(&fs_info->scrub_lock);
2987
2988         return 0;
2989 }
2990
2991 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
2992                            struct btrfs_device *dev)
2993 {
2994         struct scrub_ctx *sctx;
2995
2996         mutex_lock(&fs_info->scrub_lock);
2997         sctx = dev->scrub_device;
2998         if (!sctx) {
2999                 mutex_unlock(&fs_info->scrub_lock);
3000                 return -ENOTCONN;
3001         }
3002         atomic_inc(&sctx->cancel_req);
3003         while (dev->scrub_device) {
3004                 mutex_unlock(&fs_info->scrub_lock);
3005                 wait_event(fs_info->scrub_pause_wait,
3006                            dev->scrub_device == NULL);
3007                 mutex_lock(&fs_info->scrub_lock);
3008         }
3009         mutex_unlock(&fs_info->scrub_lock);
3010
3011         return 0;
3012 }
3013
3014 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3015                          struct btrfs_scrub_progress *progress)
3016 {
3017         struct btrfs_device *dev;
3018         struct scrub_ctx *sctx = NULL;
3019
3020         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3021         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3022         if (dev)
3023                 sctx = dev->scrub_device;
3024         if (sctx)
3025                 memcpy(progress, &sctx->stat, sizeof(*progress));
3026         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3027
3028         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3029 }
3030
3031 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3032                                u64 extent_logical, u64 extent_len,
3033                                u64 *extent_physical,
3034                                struct btrfs_device **extent_dev,
3035                                int *extent_mirror_num)
3036 {
3037         u64 mapped_length;
3038         struct btrfs_bio *bbio = NULL;
3039         int ret;
3040
3041         mapped_length = extent_len;
3042         ret = btrfs_map_block(fs_info, READ, extent_logical,
3043                               &mapped_length, &bbio, 0);
3044         if (ret || !bbio || mapped_length < extent_len ||
3045             !bbio->stripes[0].dev->bdev) {
3046                 kfree(bbio);
3047                 return;
3048         }
3049
3050         *extent_physical = bbio->stripes[0].physical;
3051         *extent_mirror_num = bbio->mirror_num;
3052         *extent_dev = bbio->stripes[0].dev;
3053         kfree(bbio);
3054 }
3055
3056 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3057                               struct scrub_wr_ctx *wr_ctx,
3058                               struct btrfs_fs_info *fs_info,
3059                               struct btrfs_device *dev,
3060                               int is_dev_replace)
3061 {
3062         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3063
3064         mutex_init(&wr_ctx->wr_lock);
3065         wr_ctx->wr_curr_bio = NULL;
3066         if (!is_dev_replace)
3067                 return 0;
3068
3069         WARN_ON(!dev->bdev);
3070         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3071                                          bio_get_nr_vecs(dev->bdev));
3072         wr_ctx->tgtdev = dev;
3073         atomic_set(&wr_ctx->flush_all_writes, 0);
3074         return 0;
3075 }
3076
3077 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3078 {
3079         mutex_lock(&wr_ctx->wr_lock);
3080         kfree(wr_ctx->wr_curr_bio);
3081         wr_ctx->wr_curr_bio = NULL;
3082         mutex_unlock(&wr_ctx->wr_lock);
3083 }
3084
3085 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3086                             int mirror_num, u64 physical_for_dev_replace)
3087 {
3088         struct scrub_copy_nocow_ctx *nocow_ctx;
3089         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3090
3091         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3092         if (!nocow_ctx) {
3093                 spin_lock(&sctx->stat_lock);
3094                 sctx->stat.malloc_errors++;
3095                 spin_unlock(&sctx->stat_lock);
3096                 return -ENOMEM;
3097         }
3098
3099         scrub_pending_trans_workers_inc(sctx);
3100
3101         nocow_ctx->sctx = sctx;
3102         nocow_ctx->logical = logical;
3103         nocow_ctx->len = len;
3104         nocow_ctx->mirror_num = mirror_num;
3105         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3106         nocow_ctx->work.func = copy_nocow_pages_worker;
3107         INIT_LIST_HEAD(&nocow_ctx->inodes);
3108         btrfs_queue_worker(&fs_info->scrub_nocow_workers,
3109                            &nocow_ctx->work);
3110
3111         return 0;
3112 }
3113
3114 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3115 {
3116         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3117         struct scrub_nocow_inode *nocow_inode;
3118
3119         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3120         if (!nocow_inode)
3121                 return -ENOMEM;
3122         nocow_inode->inum = inum;
3123         nocow_inode->offset = offset;
3124         nocow_inode->root = root;
3125         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3126         return 0;
3127 }
3128
3129 #define COPY_COMPLETE 1
3130
3131 static void copy_nocow_pages_worker(struct btrfs_work *work)
3132 {
3133         struct scrub_copy_nocow_ctx *nocow_ctx =
3134                 container_of(work, struct scrub_copy_nocow_ctx, work);
3135         struct scrub_ctx *sctx = nocow_ctx->sctx;
3136         u64 logical = nocow_ctx->logical;
3137         u64 len = nocow_ctx->len;
3138         int mirror_num = nocow_ctx->mirror_num;
3139         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3140         int ret;
3141         struct btrfs_trans_handle *trans = NULL;
3142         struct btrfs_fs_info *fs_info;
3143         struct btrfs_path *path;
3144         struct btrfs_root *root;
3145         int not_written = 0;
3146
3147         fs_info = sctx->dev_root->fs_info;
3148         root = fs_info->extent_root;
3149
3150         path = btrfs_alloc_path();
3151         if (!path) {
3152                 spin_lock(&sctx->stat_lock);
3153                 sctx->stat.malloc_errors++;
3154                 spin_unlock(&sctx->stat_lock);
3155                 not_written = 1;
3156                 goto out;
3157         }
3158
3159         trans = btrfs_join_transaction(root);
3160         if (IS_ERR(trans)) {
3161                 not_written = 1;
3162                 goto out;
3163         }
3164
3165         ret = iterate_inodes_from_logical(logical, fs_info, path,
3166                                           record_inode_for_nocow, nocow_ctx);
3167         if (ret != 0 && ret != -ENOENT) {
3168                 pr_warn("iterate_inodes_from_logical() failed: log %llu, phys %llu, len %llu, mir %u, ret %d\n",
3169                         logical, physical_for_dev_replace, len, mirror_num,
3170                         ret);
3171                 not_written = 1;
3172                 goto out;
3173         }
3174
3175         btrfs_end_transaction(trans, root);
3176         trans = NULL;
3177         while (!list_empty(&nocow_ctx->inodes)) {
3178                 struct scrub_nocow_inode *entry;
3179                 entry = list_first_entry(&nocow_ctx->inodes,
3180                                          struct scrub_nocow_inode,
3181                                          list);
3182                 list_del_init(&entry->list);
3183                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3184                                                  entry->root, nocow_ctx);
3185                 kfree(entry);
3186                 if (ret == COPY_COMPLETE) {
3187                         ret = 0;
3188                         break;
3189                 } else if (ret) {
3190                         break;
3191                 }
3192         }
3193 out:
3194         while (!list_empty(&nocow_ctx->inodes)) {
3195                 struct scrub_nocow_inode *entry;
3196                 entry = list_first_entry(&nocow_ctx->inodes,
3197                                          struct scrub_nocow_inode,
3198                                          list);
3199                 list_del_init(&entry->list);
3200                 kfree(entry);
3201         }
3202         if (trans && !IS_ERR(trans))
3203                 btrfs_end_transaction(trans, root);
3204         if (not_written)
3205                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3206                                             num_uncorrectable_read_errors);
3207
3208         btrfs_free_path(path);
3209         kfree(nocow_ctx);
3210
3211         scrub_pending_trans_workers_dec(sctx);
3212 }
3213
3214 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3215                                       struct scrub_copy_nocow_ctx *nocow_ctx)
3216 {
3217         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3218         struct btrfs_key key;
3219         struct inode *inode;
3220         struct page *page;
3221         struct btrfs_root *local_root;
3222         struct btrfs_ordered_extent *ordered;
3223         struct extent_map *em;
3224         struct extent_state *cached_state = NULL;
3225         struct extent_io_tree *io_tree;
3226         u64 physical_for_dev_replace;
3227         u64 len = nocow_ctx->len;
3228         u64 lockstart = offset, lockend = offset + len - 1;
3229         unsigned long index;
3230         int srcu_index;
3231         int ret = 0;
3232         int err = 0;
3233
3234         key.objectid = root;
3235         key.type = BTRFS_ROOT_ITEM_KEY;
3236         key.offset = (u64)-1;
3237
3238         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3239
3240         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3241         if (IS_ERR(local_root)) {
3242                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3243                 return PTR_ERR(local_root);
3244         }
3245
3246         key.type = BTRFS_INODE_ITEM_KEY;
3247         key.objectid = inum;
3248         key.offset = 0;
3249         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3250         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3251         if (IS_ERR(inode))
3252                 return PTR_ERR(inode);
3253
3254         /* Avoid truncate/dio/punch hole.. */
3255         mutex_lock(&inode->i_mutex);
3256         inode_dio_wait(inode);
3257
3258         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3259         io_tree = &BTRFS_I(inode)->io_tree;
3260
3261         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3262         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3263         if (ordered) {
3264                 btrfs_put_ordered_extent(ordered);
3265                 goto out_unlock;
3266         }
3267
3268         em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3269         if (IS_ERR(em)) {
3270                 ret = PTR_ERR(em);
3271                 goto out_unlock;
3272         }
3273
3274         /*
3275          * This extent does not actually cover the logical extent anymore,
3276          * move on to the next inode.
3277          */
3278         if (em->block_start > nocow_ctx->logical ||
3279             em->block_start + em->block_len < nocow_ctx->logical + len) {
3280                 free_extent_map(em);
3281                 goto out_unlock;
3282         }
3283         free_extent_map(em);
3284
3285         while (len >= PAGE_CACHE_SIZE) {
3286                 index = offset >> PAGE_CACHE_SHIFT;
3287 again:
3288                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3289                 if (!page) {
3290                         pr_err("find_or_create_page() failed\n");
3291                         ret = -ENOMEM;
3292                         goto out;
3293                 }
3294
3295                 if (PageUptodate(page)) {
3296                         if (PageDirty(page))
3297                                 goto next_page;
3298                 } else {
3299                         ClearPageError(page);
3300                         err = extent_read_full_page_nolock(io_tree, page,
3301                                                            btrfs_get_extent,
3302                                                            nocow_ctx->mirror_num);
3303                         if (err) {
3304                                 ret = err;
3305                                 goto next_page;
3306                         }
3307
3308                         lock_page(page);
3309                         /*
3310                          * If the page has been remove from the page cache,
3311                          * the data on it is meaningless, because it may be
3312                          * old one, the new data may be written into the new
3313                          * page in the page cache.
3314                          */
3315                         if (page->mapping != inode->i_mapping) {
3316                                 unlock_page(page);
3317                                 page_cache_release(page);
3318                                 goto again;
3319                         }
3320                         if (!PageUptodate(page)) {
3321                                 ret = -EIO;
3322                                 goto next_page;
3323                         }
3324                 }
3325                 err = write_page_nocow(nocow_ctx->sctx,
3326                                        physical_for_dev_replace, page);
3327                 if (err)
3328                         ret = err;
3329 next_page:
3330                 unlock_page(page);
3331                 page_cache_release(page);
3332
3333                 if (ret)
3334                         break;
3335
3336                 offset += PAGE_CACHE_SIZE;
3337                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3338                 len -= PAGE_CACHE_SIZE;
3339         }
3340         ret = COPY_COMPLETE;
3341 out_unlock:
3342         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3343                              GFP_NOFS);
3344 out:
3345         mutex_unlock(&inode->i_mutex);
3346         iput(inode);
3347         return ret;
3348 }
3349
3350 static int write_page_nocow(struct scrub_ctx *sctx,
3351                             u64 physical_for_dev_replace, struct page *page)
3352 {
3353         struct bio *bio;
3354         struct btrfs_device *dev;
3355         int ret;
3356
3357         dev = sctx->wr_ctx.tgtdev;
3358         if (!dev)
3359                 return -EIO;
3360         if (!dev->bdev) {
3361                 printk_ratelimited(KERN_WARNING
3362                         "btrfs: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3363                 return -EIO;
3364         }
3365         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3366         if (!bio) {
3367                 spin_lock(&sctx->stat_lock);
3368                 sctx->stat.malloc_errors++;
3369                 spin_unlock(&sctx->stat_lock);
3370                 return -ENOMEM;
3371         }
3372         bio->bi_size = 0;
3373         bio->bi_sector = physical_for_dev_replace >> 9;
3374         bio->bi_bdev = dev->bdev;
3375         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3376         if (ret != PAGE_CACHE_SIZE) {
3377 leave_with_eio:
3378                 bio_put(bio);
3379                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3380                 return -EIO;
3381         }
3382
3383         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3384                 goto leave_with_eio;
3385
3386         bio_put(bio);
3387         return 0;
3388 }