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 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 260 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 261
 262
 263 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 264 {
 265         atomic_inc(&sctx->bios_in_flight);
 266 }
 267
 268 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 269 {
 270         atomic_dec(&sctx->bios_in_flight);
 271         wake_up(&sctx->list_wait);
 272 }
 273
 274 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 275 {
 276         while (atomic_read(&fs_info->scrub_pause_req)) {
 277                 mutex_unlock(&fs_info->scrub_lock);
 278                 wait_event(fs_info->scrub_pause_wait,
 279                    atomic_read(&fs_info->scrub_pause_req) == 0);
 280                 mutex_lock(&fs_info->scrub_lock);
 281         }
 282 }
 283
 284 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 285 {
 286         atomic_inc(&fs_info->scrubs_paused);
 287         wake_up(&fs_info->scrub_pause_wait);
 288
 289         mutex_lock(&fs_info->scrub_lock);
 290         __scrub_blocked_if_needed(fs_info);
 291         atomic_dec(&fs_info->scrubs_paused);
 292         mutex_unlock(&fs_info->scrub_lock);
 293
 294         wake_up(&fs_info->scrub_pause_wait);
 295 }
 296
 297 /*
 298  * used for workers that require transaction commits (i.e., for the
 299  * NOCOW case)
 300  */
 301 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 302 {
 303         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 304
 305         /*
 306          * increment scrubs_running to prevent cancel requests from
 307          * completing as long as a worker is running. we must also
 308          * increment scrubs_paused to prevent deadlocking on pause
 309          * requests used for transactions commits (as the worker uses a
 310          * transaction context). it is safe to regard the worker
 311          * as paused for all matters practical. effectively, we only
 312          * avoid cancellation requests from completing.
 313          */
 314         mutex_lock(&fs_info->scrub_lock);
 315         atomic_inc(&fs_info->scrubs_running);
 316         atomic_inc(&fs_info->scrubs_paused);
 317         mutex_unlock(&fs_info->scrub_lock);
 318
 319         /*
 320          * check if @scrubs_running=@scrubs_paused condition
 321          * inside wait_event() is not an atomic operation.
 322          * which means we may inc/dec @scrub_running/paused
 323          * at any time. Let's wake up @scrub_pause_wait as
 324          * much as we can to let commit transaction blocked less.
 325          */
 326         wake_up(&fs_info->scrub_pause_wait);
 327
 328         atomic_inc(&sctx->workers_pending);
 329 }
 330
 331 /* used for workers that require transaction commits */
 332 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 333 {
 334         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 335
 336         /*
 337          * see scrub_pending_trans_workers_inc() why we're pretending
 338          * to be paused in the scrub counters
 339          */
 340         mutex_lock(&fs_info->scrub_lock);
 341         atomic_dec(&fs_info->scrubs_running);
 342         atomic_dec(&fs_info->scrubs_paused);
 343         mutex_unlock(&fs_info->scrub_lock);
 344         atomic_dec(&sctx->workers_pending);
 345         wake_up(&fs_info->scrub_pause_wait);
 346         wake_up(&sctx->list_wait);
 347 }
 348
 349 static void scrub_free_csums(struct scrub_ctx *sctx)
 350 {
 351         while (!list_empty(&sctx->csum_list)) {
 352                 struct btrfs_ordered_sum *sum;
 353                 sum = list_first_entry(&sctx->csum_list,
 354                                        struct btrfs_ordered_sum, list);
 355                 list_del(&sum->list);
 356                 kfree(sum);
 357         }
 358 }
 359
 360 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 361 {
 362         int i;
 363
 364         if (!sctx)
 365                 return;
 366
 367         scrub_free_wr_ctx(&sctx->wr_ctx);
 368
 369         /* this can happen when scrub is cancelled */
 370         if (sctx->curr != -1) {
 371                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 372
 373                 for (i = 0; i < sbio->page_count; i++) {
 374                         WARN_ON(!sbio->pagev[i]->page);
 375                         scrub_block_put(sbio->pagev[i]->sblock);
 376                 }
 377                 bio_put(sbio->bio);
 378         }
 379
 380         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 381                 struct scrub_bio *sbio = sctx->bios[i];
 382
 383                 if (!sbio)
 384                         break;
 385                 kfree(sbio);
 386         }
 387
 388         scrub_free_csums(sctx);
 389         kfree(sctx);
 390 }
 391
 392 static noinline_for_stack
 393 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 394 {
 395         struct scrub_ctx *sctx;
 396         int             i;
 397         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 398         int pages_per_rd_bio;
 399         int ret;
 400
 401         /*
 402          * the setting of pages_per_rd_bio is correct for scrub but might
 403          * be wrong for the dev_replace code where we might read from
 404          * different devices in the initial huge bios. However, that
 405          * code is able to correctly handle the case when adding a page
 406          * to a bio fails.
 407          */
 408         if (dev->bdev)
 409                 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
 410                                          bio_get_nr_vecs(dev->bdev));
 411         else
 412                 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 413         sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
 414         if (!sctx)
 415                 goto nomem;
 416         sctx->is_dev_replace = is_dev_replace;
 417         sctx->pages_per_rd_bio = pages_per_rd_bio;
 418         sctx->curr = -1;
 419         sctx->dev_root = dev->dev_root;
 420         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 421                 struct scrub_bio *sbio;
 422
 423                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 424                 if (!sbio)
 425                         goto nomem;
 426                 sctx->bios[i] = sbio;
 427
 428                 sbio->index = i;
 429                 sbio->sctx = sctx;
 430                 sbio->page_count = 0;
 431                 btrfs_init_work(&sbio->work, scrub_bio_end_io_worker,
 432                                 NULL, NULL);
 433
 434                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 435                         sctx->bios[i]->next_free = i + 1;
 436                 else
 437                         sctx->bios[i]->next_free = -1;
 438         }
 439         sctx->first_free = 0;
 440         sctx->nodesize = dev->dev_root->nodesize;
 441         sctx->leafsize = dev->dev_root->leafsize;
 442         sctx->sectorsize = dev->dev_root->sectorsize;
 443         atomic_set(&sctx->bios_in_flight, 0);
 444         atomic_set(&sctx->workers_pending, 0);
 445         atomic_set(&sctx->cancel_req, 0);
 446         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 447         INIT_LIST_HEAD(&sctx->csum_list);
 448
 449         spin_lock_init(&sctx->list_lock);
 450         spin_lock_init(&sctx->stat_lock);
 451         init_waitqueue_head(&sctx->list_wait);
 452
 453         ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 454                                  fs_info->dev_replace.tgtdev, is_dev_replace);
 455         if (ret) {
 456                 scrub_free_ctx(sctx);
 457                 return ERR_PTR(ret);
 458         }
 459         return sctx;
 460
 461 nomem:
 462         scrub_free_ctx(sctx);
 463         return ERR_PTR(-ENOMEM);
 464 }
 465
 466 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 467                                      void *warn_ctx)
 468 {
 469         u64 isize;
 470         u32 nlink;
 471         int ret;
 472         int i;
 473         struct extent_buffer *eb;
 474         struct btrfs_inode_item *inode_item;
 475         struct scrub_warning *swarn = warn_ctx;
 476         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 477         struct inode_fs_paths *ipath = NULL;
 478         struct btrfs_root *local_root;
 479         struct btrfs_key root_key;
 480
 481         root_key.objectid = root;
 482         root_key.type = BTRFS_ROOT_ITEM_KEY;
 483         root_key.offset = (u64)-1;
 484         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 485         if (IS_ERR(local_root)) {
 486                 ret = PTR_ERR(local_root);
 487                 goto err;
 488         }
 489
 490         ret = inode_item_info(inum, 0, local_root, swarn->path);
 491         if (ret) {
 492                 btrfs_release_path(swarn->path);
 493                 goto err;
 494         }
 495
 496         eb = swarn->path->nodes[0];
 497         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 498                                         struct btrfs_inode_item);
 499         isize = btrfs_inode_size(eb, inode_item);
 500         nlink = btrfs_inode_nlink(eb, inode_item);
 501         btrfs_release_path(swarn->path);
 502
 503         ipath = init_ipath(4096, local_root, swarn->path);
 504         if (IS_ERR(ipath)) {
 505                 ret = PTR_ERR(ipath);
 506                 ipath = NULL;
 507                 goto err;
 508         }
 509         ret = paths_from_inode(inum, ipath);
 510
 511         if (ret < 0)
 512                 goto err;
 513
 514         /*
 515          * we deliberately ignore the bit ipath might have been too small to
 516          * hold all of the paths here
 517          */
 518         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 519                 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
 520                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 521                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 522                         swarn->logical, rcu_str_deref(swarn->dev->name),
 523                         (unsigned long long)swarn->sector, root, inum, offset,
 524                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 525                         (char *)(unsigned long)ipath->fspath->val[i]);
 526
 527         free_ipath(ipath);
 528         return 0;
 529
 530 err:
 531         printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
 532                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 533                 "resolving failed with ret=%d\n", swarn->errstr,
 534                 swarn->logical, rcu_str_deref(swarn->dev->name),
 535                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 536
 537         free_ipath(ipath);
 538         return 0;
 539 }
 540
 541 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 542 {
 543         struct btrfs_device *dev;
 544         struct btrfs_fs_info *fs_info;
 545         struct btrfs_path *path;
 546         struct btrfs_key found_key;
 547         struct extent_buffer *eb;
 548         struct btrfs_extent_item *ei;
 549         struct scrub_warning swarn;
 550         unsigned long ptr = 0;
 551         u64 extent_item_pos;
 552         u64 flags = 0;
 553         u64 ref_root;
 554         u32 item_size;
 555         u8 ref_level;
 556         const int bufsize = 4096;
 557         int ret;
 558
 559         WARN_ON(sblock->page_count < 1);
 560         dev = sblock->pagev[0]->dev;
 561         fs_info = sblock->sctx->dev_root->fs_info;
 562
 563         path = btrfs_alloc_path();
 564
 565         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 566         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 567         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 568         swarn.logical = sblock->pagev[0]->logical;
 569         swarn.errstr = errstr;
 570         swarn.dev = NULL;
 571         swarn.msg_bufsize = bufsize;
 572         swarn.scratch_bufsize = bufsize;
 573
 574         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 575                 goto out;
 576
 577         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 578                                   &flags);
 579         if (ret < 0)
 580                 goto out;
 581
 582         extent_item_pos = swarn.logical - found_key.objectid;
 583         swarn.extent_item_size = found_key.offset;
 584
 585         eb = path->nodes[0];
 586         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 587         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 588
 589         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 590                 do {
 591                         ret = tree_backref_for_extent(&ptr, eb, ei, item_size,
 592                                                         &ref_root, &ref_level);
 593                         printk_in_rcu(KERN_WARNING
 594                                 "BTRFS: %s at logical %llu on dev %s, "
 595                                 "sector %llu: metadata %s (level %d) in tree "
 596                                 "%llu\n", errstr, swarn.logical,
 597                                 rcu_str_deref(dev->name),
 598                                 (unsigned long long)swarn.sector,
 599                                 ref_level ? "node" : "leaf",
 600                                 ret < 0 ? -1 : ref_level,
 601                                 ret < 0 ? -1 : ref_root);
 602                 } while (ret != 1);
 603                 btrfs_release_path(path);
 604         } else {
 605                 btrfs_release_path(path);
 606                 swarn.path = path;
 607                 swarn.dev = dev;
 608                 iterate_extent_inodes(fs_info, found_key.objectid,
 609                                         extent_item_pos, 1,
 610                                         scrub_print_warning_inode, &swarn);
 611         }
 612
 613 out:
 614         btrfs_free_path(path);
 615         kfree(swarn.scratch_buf);
 616         kfree(swarn.msg_buf);
 617 }
 618
 619 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 620 {
 621         struct page *page = NULL;
 622         unsigned long index;
 623         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 624         int ret;
 625         int corrected = 0;
 626         struct btrfs_key key;
 627         struct inode *inode = NULL;
 628         struct btrfs_fs_info *fs_info;
 629         u64 end = offset + PAGE_SIZE - 1;
 630         struct btrfs_root *local_root;
 631         int srcu_index;
 632
 633         key.objectid = root;
 634         key.type = BTRFS_ROOT_ITEM_KEY;
 635         key.offset = (u64)-1;
 636
 637         fs_info = fixup->root->fs_info;
 638         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 639
 640         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 641         if (IS_ERR(local_root)) {
 642                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 643                 return PTR_ERR(local_root);
 644         }
 645
 646         key.type = BTRFS_INODE_ITEM_KEY;
 647         key.objectid = inum;
 648         key.offset = 0;
 649         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 650         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 651         if (IS_ERR(inode))
 652                 return PTR_ERR(inode);
 653
 654         index = offset >> PAGE_CACHE_SHIFT;
 655
 656         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 657         if (!page) {
 658                 ret = -ENOMEM;
 659                 goto out;
 660         }
 661
 662         if (PageUptodate(page)) {
 663                 if (PageDirty(page)) {
 664                         /*
 665                          * we need to write the data to the defect sector. the
 666                          * data that was in that sector is not in memory,
 667                          * because the page was modified. we must not write the
 668                          * modified page to that sector.
 669                          *
 670                          * TODO: what could be done here: wait for the delalloc
 671                          *       runner to write out that page (might involve
 672                          *       COW) and see whether the sector is still
 673                          *       referenced afterwards.
 674                          *
 675                          * For the meantime, we'll treat this error
 676                          * incorrectable, although there is a chance that a
 677                          * later scrub will find the bad sector again and that
 678                          * there's no dirty page in memory, then.
 679                          */
 680                         ret = -EIO;
 681                         goto out;
 682                 }
 683                 fs_info = BTRFS_I(inode)->root->fs_info;
 684                 ret = repair_io_failure(fs_info, offset, PAGE_SIZE,
 685                                         fixup->logical, page,
 686                                         fixup->mirror_num);
 687                 unlock_page(page);
 688                 corrected = !ret;
 689         } else {
 690                 /*
 691                  * we need to get good data first. the general readpage path
 692                  * will call repair_io_failure for us, we just have to make
 693                  * sure we read the bad mirror.
 694                  */
 695                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 696                                         EXTENT_DAMAGED, GFP_NOFS);
 697                 if (ret) {
 698                         /* set_extent_bits should give proper error */
 699                         WARN_ON(ret > 0);
 700                         if (ret > 0)
 701                                 ret = -EFAULT;
 702                         goto out;
 703                 }
 704
 705                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 706                                                 btrfs_get_extent,
 707                                                 fixup->mirror_num);
 708                 wait_on_page_locked(page);
 709
 710                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 711                                                 end, EXTENT_DAMAGED, 0, NULL);
 712                 if (!corrected)
 713                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 714                                                 EXTENT_DAMAGED, GFP_NOFS);
 715         }
 716
 717 out:
 718         if (page)
 719                 put_page(page);
 720         if (inode)
 721                 iput(inode);
 722
 723         if (ret < 0)
 724                 return ret;
 725
 726         if (ret == 0 && corrected) {
 727                 /*
 728                  * we only need to call readpage for one of the inodes belonging
 729                  * to this extent. so make iterate_extent_inodes stop
 730                  */
 731                 return 1;
 732         }
 733
 734         return -EIO;
 735 }
 736
 737 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 738 {
 739         int ret;
 740         struct scrub_fixup_nodatasum *fixup;
 741         struct scrub_ctx *sctx;
 742         struct btrfs_trans_handle *trans = NULL;
 743         struct btrfs_path *path;
 744         int uncorrectable = 0;
 745
 746         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 747         sctx = fixup->sctx;
 748
 749         path = btrfs_alloc_path();
 750         if (!path) {
 751                 spin_lock(&sctx->stat_lock);
 752                 ++sctx->stat.malloc_errors;
 753                 spin_unlock(&sctx->stat_lock);
 754                 uncorrectable = 1;
 755                 goto out;
 756         }
 757
 758         trans = btrfs_join_transaction(fixup->root);
 759         if (IS_ERR(trans)) {
 760                 uncorrectable = 1;
 761                 goto out;
 762         }
 763
 764         /*
 765          * the idea is to trigger a regular read through the standard path. we
 766          * read a page from the (failed) logical address by specifying the
 767          * corresponding copynum of the failed sector. thus, that readpage is
 768          * expected to fail.
 769          * that is the point where on-the-fly error correction will kick in
 770          * (once it's finished) and rewrite the failed sector if a good copy
 771          * can be found.
 772          */
 773         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 774                                                 path, scrub_fixup_readpage,
 775                                                 fixup);
 776         if (ret < 0) {
 777                 uncorrectable = 1;
 778                 goto out;
 779         }
 780         WARN_ON(ret != 1);
 781
 782         spin_lock(&sctx->stat_lock);
 783         ++sctx->stat.corrected_errors;
 784         spin_unlock(&sctx->stat_lock);
 785
 786 out:
 787         if (trans && !IS_ERR(trans))
 788                 btrfs_end_transaction(trans, fixup->root);
 789         if (uncorrectable) {
 790                 spin_lock(&sctx->stat_lock);
 791                 ++sctx->stat.uncorrectable_errors;
 792                 spin_unlock(&sctx->stat_lock);
 793                 btrfs_dev_replace_stats_inc(
 794                         &sctx->dev_root->fs_info->dev_replace.
 795                         num_uncorrectable_read_errors);
 796                 printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
 797                     "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 798                         fixup->logical, rcu_str_deref(fixup->dev->name));
 799         }
 800
 801         btrfs_free_path(path);
 802         kfree(fixup);
 803
 804         scrub_pending_trans_workers_dec(sctx);
 805 }
 806
 807 /*
 808  * scrub_handle_errored_block gets called when either verification of the
 809  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 810  * case, this function handles all pages in the bio, even though only one
 811  * may be bad.
 812  * The goal of this function is to repair the errored block by using the
 813  * contents of one of the mirrors.
 814  */
 815 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 816 {
 817         struct scrub_ctx *sctx = sblock_to_check->sctx;
 818         struct btrfs_device *dev;
 819         struct btrfs_fs_info *fs_info;
 820         u64 length;
 821         u64 logical;
 822         u64 generation;
 823         unsigned int failed_mirror_index;
 824         unsigned int is_metadata;
 825         unsigned int have_csum;
 826         u8 *csum;
 827         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 828         struct scrub_block *sblock_bad;
 829         int ret;
 830         int mirror_index;
 831         int page_num;
 832         int success;
 833         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 834                                       DEFAULT_RATELIMIT_BURST);
 835
 836         BUG_ON(sblock_to_check->page_count < 1);
 837         fs_info = sctx->dev_root->fs_info;
 838         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 839                 /*
 840                  * if we find an error in a super block, we just report it.
 841                  * They will get written with the next transaction commit
 842                  * anyway
 843                  */
 844                 spin_lock(&sctx->stat_lock);
 845                 ++sctx->stat.super_errors;
 846                 spin_unlock(&sctx->stat_lock);
 847                 return 0;
 848         }
 849         length = sblock_to_check->page_count * PAGE_SIZE;
 850         logical = sblock_to_check->pagev[0]->logical;
 851         generation = sblock_to_check->pagev[0]->generation;
 852         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 853         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 854         is_metadata = !(sblock_to_check->pagev[0]->flags &
 855                         BTRFS_EXTENT_FLAG_DATA);
 856         have_csum = sblock_to_check->pagev[0]->have_csum;
 857         csum = sblock_to_check->pagev[0]->csum;
 858         dev = sblock_to_check->pagev[0]->dev;
 859
 860         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 861                 sblocks_for_recheck = NULL;
 862                 goto nodatasum_case;
 863         }
 864
 865         /*
 866          * read all mirrors one after the other. This includes to
 867          * re-read the extent or metadata block that failed (that was
 868          * the cause that this fixup code is called) another time,
 869          * page by page this time in order to know which pages
 870          * caused I/O errors and which ones are good (for all mirrors).
 871          * It is the goal to handle the situation when more than one
 872          * mirror contains I/O errors, but the errors do not
 873          * overlap, i.e. the data can be repaired by selecting the
 874          * pages from those mirrors without I/O error on the
 875          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 876          * would be that mirror #1 has an I/O error on the first page,
 877          * the second page is good, and mirror #2 has an I/O error on
 878          * the second page, but the first page is good.
 879          * Then the first page of the first mirror can be repaired by
 880          * taking the first page of the second mirror, and the
 881          * second page of the second mirror can be repaired by
 882          * copying the contents of the 2nd page of the 1st mirror.
 883          * One more note: if the pages of one mirror contain I/O
 884          * errors, the checksum cannot be verified. In order to get
 885          * the best data for repairing, the first attempt is to find
 886          * a mirror without I/O errors and with a validated checksum.
 887          * Only if this is not possible, the pages are picked from
 888          * mirrors with I/O errors without considering the checksum.
 889          * If the latter is the case, at the end, the checksum of the
 890          * repaired area is verified in order to correctly maintain
 891          * the statistics.
 892          */
 893
 894         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 895                                      sizeof(*sblocks_for_recheck),
 896                                      GFP_NOFS);
 897         if (!sblocks_for_recheck) {
 898                 spin_lock(&sctx->stat_lock);
 899                 sctx->stat.malloc_errors++;
 900                 sctx->stat.read_errors++;
 901                 sctx->stat.uncorrectable_errors++;
 902                 spin_unlock(&sctx->stat_lock);
 903                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 904                 goto out;
 905         }
 906
 907         /* setup the context, map the logical blocks and alloc the pages */
 908         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 909                                         logical, sblocks_for_recheck);
 910         if (ret) {
 911                 spin_lock(&sctx->stat_lock);
 912                 sctx->stat.read_errors++;
 913                 sctx->stat.uncorrectable_errors++;
 914                 spin_unlock(&sctx->stat_lock);
 915                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 916                 goto out;
 917         }
 918         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 919         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 920
 921         /* build and submit the bios for the failed mirror, check checksums */
 922         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 923                             csum, generation, sctx->csum_size);
 924
 925         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 926             sblock_bad->no_io_error_seen) {
 927                 /*
 928                  * the error disappeared after reading page by page, or
 929                  * the area was part of a huge bio and other parts of the
 930                  * bio caused I/O errors, or the block layer merged several
 931                  * read requests into one and the error is caused by a
 932                  * different bio (usually one of the two latter cases is
 933                  * the cause)
 934                  */
 935                 spin_lock(&sctx->stat_lock);
 936                 sctx->stat.unverified_errors++;
 937                 spin_unlock(&sctx->stat_lock);
 938
 939                 if (sctx->is_dev_replace)
 940                         scrub_write_block_to_dev_replace(sblock_bad);
 941                 goto out;
 942         }
 943
 944         if (!sblock_bad->no_io_error_seen) {
 945                 spin_lock(&sctx->stat_lock);
 946                 sctx->stat.read_errors++;
 947                 spin_unlock(&sctx->stat_lock);
 948                 if (__ratelimit(&_rs))
 949                         scrub_print_warning("i/o error", sblock_to_check);
 950                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 951         } else if (sblock_bad->checksum_error) {
 952                 spin_lock(&sctx->stat_lock);
 953                 sctx->stat.csum_errors++;
 954                 spin_unlock(&sctx->stat_lock);
 955                 if (__ratelimit(&_rs))
 956                         scrub_print_warning("checksum error", sblock_to_check);
 957                 btrfs_dev_stat_inc_and_print(dev,
 958                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 959         } else if (sblock_bad->header_error) {
 960                 spin_lock(&sctx->stat_lock);
 961                 sctx->stat.verify_errors++;
 962                 spin_unlock(&sctx->stat_lock);
 963                 if (__ratelimit(&_rs))
 964                         scrub_print_warning("checksum/header error",
 965                                             sblock_to_check);
 966                 if (sblock_bad->generation_error)
 967                         btrfs_dev_stat_inc_and_print(dev,
 968                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 969                 else
 970                         btrfs_dev_stat_inc_and_print(dev,
 971                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 972         }
 973
 974         if (sctx->readonly) {
 975                 ASSERT(!sctx->is_dev_replace);
 976                 goto out;
 977         }
 978
 979         if (!is_metadata && !have_csum) {
 980                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 981
 982 nodatasum_case:
 983                 WARN_ON(sctx->is_dev_replace);
 984
 985                 /*
 986                  * !is_metadata and !have_csum, this means that the data
 987                  * might not be COW'ed, that it might be modified
 988                  * concurrently. The general strategy to work on the
 989                  * commit root does not help in the case when COW is not
 990                  * used.
 991                  */
 992                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 993                 if (!fixup_nodatasum)
 994                         goto did_not_correct_error;
 995                 fixup_nodatasum->sctx = sctx;
 996                 fixup_nodatasum->dev = dev;
 997                 fixup_nodatasum->logical = logical;
 998                 fixup_nodatasum->root = fs_info->extent_root;
 999                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
1000                 scrub_pending_trans_workers_inc(sctx);
1001                 btrfs_init_work(&fixup_nodatasum->work, scrub_fixup_nodatasum,
1002                                 NULL, NULL);
1003                 btrfs_queue_work(fs_info->scrub_workers,
1004                                  &fixup_nodatasum->work);
1005                 goto out;
1006         }
1007
1008         /*
1009          * now build and submit the bios for the other mirrors, check
1010          * checksums.
1011          * First try to pick the mirror which is completely without I/O
1012          * errors and also does not have a checksum error.
1013          * If one is found, and if a checksum is present, the full block
1014          * that is known to contain an error is rewritten. Afterwards
1015          * the block is known to be corrected.
1016          * If a mirror is found which is completely correct, and no
1017          * checksum is present, only those pages are rewritten that had
1018          * an I/O error in the block to be repaired, since it cannot be
1019          * determined, which copy of the other pages is better (and it
1020          * could happen otherwise that a correct page would be
1021          * overwritten by a bad one).
1022          */
1023         for (mirror_index = 0;
1024              mirror_index < BTRFS_MAX_MIRRORS &&
1025              sblocks_for_recheck[mirror_index].page_count > 0;
1026              mirror_index++) {
1027                 struct scrub_block *sblock_other;
1028
1029                 if (mirror_index == failed_mirror_index)
1030                         continue;
1031                 sblock_other = sblocks_for_recheck + mirror_index;
1032
1033                 /* build and submit the bios, check checksums */
1034                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1035                                     have_csum, csum, generation,
1036                                     sctx->csum_size);
1037
1038                 if (!sblock_other->header_error &&
1039                     !sblock_other->checksum_error &&
1040                     sblock_other->no_io_error_seen) {
1041                         if (sctx->is_dev_replace) {
1042                                 scrub_write_block_to_dev_replace(sblock_other);
1043                         } else {
1044                                 int force_write = is_metadata || have_csum;
1045
1046                                 ret = scrub_repair_block_from_good_copy(
1047                                                 sblock_bad, sblock_other,
1048                                                 force_write);
1049                         }
1050                         if (0 == ret)
1051                                 goto corrected_error;
1052                 }
1053         }
1054
1055         /*
1056          * for dev_replace, pick good pages and write to the target device.
1057          */
1058         if (sctx->is_dev_replace) {
1059                 success = 1;
1060                 for (page_num = 0; page_num < sblock_bad->page_count;
1061                      page_num++) {
1062                         int sub_success;
1063
1064                         sub_success = 0;
1065                         for (mirror_index = 0;
1066                              mirror_index < BTRFS_MAX_MIRRORS &&
1067                              sblocks_for_recheck[mirror_index].page_count > 0;
1068                              mirror_index++) {
1069                                 struct scrub_block *sblock_other =
1070                                         sblocks_for_recheck + mirror_index;
1071                                 struct scrub_page *page_other =
1072                                         sblock_other->pagev[page_num];
1073
1074                                 if (!page_other->io_error) {
1075                                         ret = scrub_write_page_to_dev_replace(
1076                                                         sblock_other, page_num);
1077                                         if (ret == 0) {
1078                                                 /* succeeded for this page */
1079                                                 sub_success = 1;
1080                                                 break;
1081                                         } else {
1082                                                 btrfs_dev_replace_stats_inc(
1083                                                         &sctx->dev_root->
1084                                                         fs_info->dev_replace.
1085                                                         num_write_errors);
1086                                         }
1087                                 }
1088                         }
1089
1090                         if (!sub_success) {
1091                                 /*
1092                                  * did not find a mirror to fetch the page
1093                                  * from. scrub_write_page_to_dev_replace()
1094                                  * handles this case (page->io_error), by
1095                                  * filling the block with zeros before
1096                                  * submitting the write request
1097                                  */
1098                                 success = 0;
1099                                 ret = scrub_write_page_to_dev_replace(
1100                                                 sblock_bad, page_num);
1101                                 if (ret)
1102                                         btrfs_dev_replace_stats_inc(
1103                                                 &sctx->dev_root->fs_info->
1104                                                 dev_replace.num_write_errors);
1105                         }
1106                 }
1107
1108                 goto out;
1109         }
1110
1111         /*
1112          * for regular scrub, repair those pages that are errored.
1113          * In case of I/O errors in the area that is supposed to be
1114          * repaired, continue by picking good copies of those pages.
1115          * Select the good pages from mirrors to rewrite bad pages from
1116          * the area to fix. Afterwards verify the checksum of the block
1117          * that is supposed to be repaired. This verification step is
1118          * only done for the purpose of statistic counting and for the
1119          * final scrub report, whether errors remain.
1120          * A perfect algorithm could make use of the checksum and try
1121          * all possible combinations of pages from the different mirrors
1122          * until the checksum verification succeeds. For example, when
1123          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1124          * of mirror #2 is readable but the final checksum test fails,
1125          * then the 2nd page of mirror #3 could be tried, whether now
1126          * the final checksum succeedes. But this would be a rare
1127          * exception and is therefore not implemented. At least it is
1128          * avoided that the good copy is overwritten.
1129          * A more useful improvement would be to pick the sectors
1130          * without I/O error based on sector sizes (512 bytes on legacy
1131          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1132          * mirror could be repaired by taking 512 byte of a different
1133          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1134          * area are unreadable.
1135          */
1136
1137         /* can only fix I/O errors from here on */
1138         if (sblock_bad->no_io_error_seen)
1139                 goto did_not_correct_error;
1140
1141         success = 1;
1142         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1143                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1144
1145                 if (!page_bad->io_error)
1146                         continue;
1147
1148                 for (mirror_index = 0;
1149                      mirror_index < BTRFS_MAX_MIRRORS &&
1150                      sblocks_for_recheck[mirror_index].page_count > 0;
1151                      mirror_index++) {
1152                         struct scrub_block *sblock_other = sblocks_for_recheck +
1153                                                            mirror_index;
1154                         struct scrub_page *page_other = sblock_other->pagev[
1155                                                         page_num];
1156
1157                         if (!page_other->io_error) {
1158                                 ret = scrub_repair_page_from_good_copy(
1159                                         sblock_bad, sblock_other, page_num, 0);
1160                                 if (0 == ret) {
1161                                         page_bad->io_error = 0;
1162                                         break; /* succeeded for this page */
1163                                 }
1164                         }
1165                 }
1166
1167                 if (page_bad->io_error) {
1168                         /* did not find a mirror to copy the page from */
1169                         success = 0;
1170                 }
1171         }
1172
1173         if (success) {
1174                 if (is_metadata || have_csum) {
1175                         /*
1176                          * need to verify the checksum now that all
1177                          * sectors on disk are repaired (the write
1178                          * request for data to be repaired is on its way).
1179                          * Just be lazy and use scrub_recheck_block()
1180                          * which re-reads the data before the checksum
1181                          * is verified, but most likely the data comes out
1182                          * of the page cache.
1183                          */
1184                         scrub_recheck_block(fs_info, sblock_bad,
1185                                             is_metadata, have_csum, csum,
1186                                             generation, sctx->csum_size);
1187                         if (!sblock_bad->header_error &&
1188                             !sblock_bad->checksum_error &&
1189                             sblock_bad->no_io_error_seen)
1190                                 goto corrected_error;
1191                         else
1192                                 goto did_not_correct_error;
1193                 } else {
1194 corrected_error:
1195                         spin_lock(&sctx->stat_lock);
1196                         sctx->stat.corrected_errors++;
1197                         spin_unlock(&sctx->stat_lock);
1198                         printk_ratelimited_in_rcu(KERN_ERR
1199                                 "BTRFS: fixed up error at logical %llu on dev %s\n",
1200                                 logical, rcu_str_deref(dev->name));
1201                 }
1202         } else {
1203 did_not_correct_error:
1204                 spin_lock(&sctx->stat_lock);
1205                 sctx->stat.uncorrectable_errors++;
1206                 spin_unlock(&sctx->stat_lock);
1207                 printk_ratelimited_in_rcu(KERN_ERR
1208                         "BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1209                         logical, rcu_str_deref(dev->name));
1210         }
1211
1212 out:
1213         if (sblocks_for_recheck) {
1214                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1215                      mirror_index++) {
1216                         struct scrub_block *sblock = sblocks_for_recheck +
1217                                                      mirror_index;
1218                         int page_index;
1219
1220                         for (page_index = 0; page_index < sblock->page_count;
1221                              page_index++) {
1222                                 sblock->pagev[page_index]->sblock = NULL;
1223                                 scrub_page_put(sblock->pagev[page_index]);
1224                         }
1225                 }
1226                 kfree(sblocks_for_recheck);
1227         }
1228
1229         return 0;
1230 }
1231
1232 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1233                                      struct btrfs_fs_info *fs_info,
1234                                      struct scrub_block *original_sblock,
1235                                      u64 length, u64 logical,
1236                                      struct scrub_block *sblocks_for_recheck)
1237 {
1238         int page_index;
1239         int mirror_index;
1240         int ret;
1241
1242         /*
1243          * note: the two members ref_count and outstanding_pages
1244          * are not used (and not set) in the blocks that are used for
1245          * the recheck procedure
1246          */
1247
1248         page_index = 0;
1249         while (length > 0) {
1250                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1251                 u64 mapped_length = sublen;
1252                 struct btrfs_bio *bbio = NULL;
1253
1254                 /*
1255                  * with a length of PAGE_SIZE, each returned stripe
1256                  * represents one mirror
1257                  */
1258                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1259                                       &mapped_length, &bbio, 0);
1260                 if (ret || !bbio || mapped_length < sublen) {
1261                         kfree(bbio);
1262                         return -EIO;
1263                 }
1264
1265                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1266                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1267                      mirror_index++) {
1268                         struct scrub_block *sblock;
1269                         struct scrub_page *page;
1270
1271                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1272                                 continue;
1273
1274                         sblock = sblocks_for_recheck + mirror_index;
1275                         sblock->sctx = sctx;
1276                         page = kzalloc(sizeof(*page), GFP_NOFS);
1277                         if (!page) {
1278 leave_nomem:
1279                                 spin_lock(&sctx->stat_lock);
1280                                 sctx->stat.malloc_errors++;
1281                                 spin_unlock(&sctx->stat_lock);
1282                                 kfree(bbio);
1283                                 return -ENOMEM;
1284                         }
1285                         scrub_page_get(page);
1286                         sblock->pagev[page_index] = page;
1287                         page->logical = logical;
1288                         page->physical = bbio->stripes[mirror_index].physical;
1289                         BUG_ON(page_index >= original_sblock->page_count);
1290                         page->physical_for_dev_replace =
1291                                 original_sblock->pagev[page_index]->
1292                                 physical_for_dev_replace;
1293                         /* for missing devices, dev->bdev is NULL */
1294                         page->dev = bbio->stripes[mirror_index].dev;
1295                         page->mirror_num = mirror_index + 1;
1296                         sblock->page_count++;
1297                         page->page = alloc_page(GFP_NOFS);
1298                         if (!page->page)
1299                                 goto leave_nomem;
1300                 }
1301                 kfree(bbio);
1302                 length -= sublen;
1303                 logical += sublen;
1304                 page_index++;
1305         }
1306
1307         return 0;
1308 }
1309
1310 /*
1311  * this function will check the on disk data for checksum errors, header
1312  * errors and read I/O errors. If any I/O errors happen, the exact pages
1313  * which are errored are marked as being bad. The goal is to enable scrub
1314  * to take those pages that are not errored from all the mirrors so that
1315  * the pages that are errored in the just handled mirror can be repaired.
1316  */
1317 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1318                                 struct scrub_block *sblock, int is_metadata,
1319                                 int have_csum, u8 *csum, u64 generation,
1320                                 u16 csum_size)
1321 {
1322         int page_num;
1323
1324         sblock->no_io_error_seen = 1;
1325         sblock->header_error = 0;
1326         sblock->checksum_error = 0;
1327
1328         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1329                 struct bio *bio;
1330                 struct scrub_page *page = sblock->pagev[page_num];
1331
1332                 if (page->dev->bdev == NULL) {
1333                         page->io_error = 1;
1334                         sblock->no_io_error_seen = 0;
1335                         continue;
1336                 }
1337
1338                 WARN_ON(!page->page);
1339                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1340                 if (!bio) {
1341                         page->io_error = 1;
1342                         sblock->no_io_error_seen = 0;
1343                         continue;
1344                 }
1345                 bio->bi_bdev = page->dev->bdev;
1346                 bio->bi_iter.bi_sector = page->physical >> 9;
1347
1348                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1349                 if (btrfsic_submit_bio_wait(READ, bio))
1350                         sblock->no_io_error_seen = 0;
1351
1352                 bio_put(bio);
1353         }
1354
1355         if (sblock->no_io_error_seen)
1356                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1357                                              have_csum, csum, generation,
1358                                              csum_size);
1359
1360         return;
1361 }
1362
1363 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1364                                          struct scrub_block *sblock,
1365                                          int is_metadata, int have_csum,
1366                                          const u8 *csum, u64 generation,
1367                                          u16 csum_size)
1368 {
1369         int page_num;
1370         u8 calculated_csum[BTRFS_CSUM_SIZE];
1371         u32 crc = ~(u32)0;
1372         void *mapped_buffer;
1373
1374         WARN_ON(!sblock->pagev[0]->page);
1375         if (is_metadata) {
1376                 struct btrfs_header *h;
1377
1378                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1379                 h = (struct btrfs_header *)mapped_buffer;
1380
1381                 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1382                     memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE) ||
1383                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1384                            BTRFS_UUID_SIZE)) {
1385                         sblock->header_error = 1;
1386                 } else if (generation != btrfs_stack_header_generation(h)) {
1387                         sblock->header_error = 1;
1388                         sblock->generation_error = 1;
1389                 }
1390                 csum = h->csum;
1391         } else {
1392                 if (!have_csum)
1393                         return;
1394
1395                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1396         }
1397
1398         for (page_num = 0;;) {
1399                 if (page_num == 0 && is_metadata)
1400                         crc = btrfs_csum_data(
1401                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1402                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1403                 else
1404                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1405
1406                 kunmap_atomic(mapped_buffer);
1407                 page_num++;
1408                 if (page_num >= sblock->page_count)
1409                         break;
1410                 WARN_ON(!sblock->pagev[page_num]->page);
1411
1412                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1413         }
1414
1415         btrfs_csum_final(crc, calculated_csum);
1416         if (memcmp(calculated_csum, csum, csum_size))
1417                 sblock->checksum_error = 1;
1418 }
1419
1420 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1421                                              struct scrub_block *sblock_good,
1422                                              int force_write)
1423 {
1424         int page_num;
1425         int ret = 0;
1426
1427         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1428                 int ret_sub;
1429
1430                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1431                                                            sblock_good,
1432                                                            page_num,
1433                                                            force_write);
1434                 if (ret_sub)
1435                         ret = ret_sub;
1436         }
1437
1438         return ret;
1439 }
1440
1441 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1442                                             struct scrub_block *sblock_good,
1443                                             int page_num, int force_write)
1444 {
1445         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1446         struct scrub_page *page_good = sblock_good->pagev[page_num];
1447
1448         BUG_ON(page_bad->page == NULL);
1449         BUG_ON(page_good->page == NULL);
1450         if (force_write || sblock_bad->header_error ||
1451             sblock_bad->checksum_error || page_bad->io_error) {
1452                 struct bio *bio;
1453                 int ret;
1454
1455                 if (!page_bad->dev->bdev) {
1456                         printk_ratelimited(KERN_WARNING "BTRFS: "
1457                                 "scrub_repair_page_from_good_copy(bdev == NULL) "
1458                                 "is unexpected!\n");
1459                         return -EIO;
1460                 }
1461
1462                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1463                 if (!bio)
1464                         return -EIO;
1465                 bio->bi_bdev = page_bad->dev->bdev;
1466                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1467
1468                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1469                 if (PAGE_SIZE != ret) {
1470                         bio_put(bio);
1471                         return -EIO;
1472                 }
1473
1474                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1475                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1476                                 BTRFS_DEV_STAT_WRITE_ERRS);
1477                         btrfs_dev_replace_stats_inc(
1478                                 &sblock_bad->sctx->dev_root->fs_info->
1479                                 dev_replace.num_write_errors);
1480                         bio_put(bio);
1481                         return -EIO;
1482                 }
1483                 bio_put(bio);
1484         }
1485
1486         return 0;
1487 }
1488
1489 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1490 {
1491         int page_num;
1492
1493         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1494                 int ret;
1495
1496                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1497                 if (ret)
1498                         btrfs_dev_replace_stats_inc(
1499                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1500                                 num_write_errors);
1501         }
1502 }
1503
1504 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1505                                            int page_num)
1506 {
1507         struct scrub_page *spage = sblock->pagev[page_num];
1508
1509         BUG_ON(spage->page == NULL);
1510         if (spage->io_error) {
1511                 void *mapped_buffer = kmap_atomic(spage->page);
1512
1513                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1514                 flush_dcache_page(spage->page);
1515                 kunmap_atomic(mapped_buffer);
1516         }
1517         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1518 }
1519
1520 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1521                                     struct scrub_page *spage)
1522 {
1523         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1524         struct scrub_bio *sbio;
1525         int ret;
1526
1527         mutex_lock(&wr_ctx->wr_lock);
1528 again:
1529         if (!wr_ctx->wr_curr_bio) {
1530                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1531                                               GFP_NOFS);
1532                 if (!wr_ctx->wr_curr_bio) {
1533                         mutex_unlock(&wr_ctx->wr_lock);
1534                         return -ENOMEM;
1535                 }
1536                 wr_ctx->wr_curr_bio->sctx = sctx;
1537                 wr_ctx->wr_curr_bio->page_count = 0;
1538         }
1539         sbio = wr_ctx->wr_curr_bio;
1540         if (sbio->page_count == 0) {
1541                 struct bio *bio;
1542
1543                 sbio->physical = spage->physical_for_dev_replace;
1544                 sbio->logical = spage->logical;
1545                 sbio->dev = wr_ctx->tgtdev;
1546                 bio = sbio->bio;
1547                 if (!bio) {
1548                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1549                         if (!bio) {
1550                                 mutex_unlock(&wr_ctx->wr_lock);
1551                                 return -ENOMEM;
1552                         }
1553                         sbio->bio = bio;
1554                 }
1555
1556                 bio->bi_private = sbio;
1557                 bio->bi_end_io = scrub_wr_bio_end_io;
1558                 bio->bi_bdev = sbio->dev->bdev;
1559                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1560                 sbio->err = 0;
1561         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1562                    spage->physical_for_dev_replace ||
1563                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1564                    spage->logical) {
1565                 scrub_wr_submit(sctx);
1566                 goto again;
1567         }
1568
1569         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1570         if (ret != PAGE_SIZE) {
1571                 if (sbio->page_count < 1) {
1572                         bio_put(sbio->bio);
1573                         sbio->bio = NULL;
1574                         mutex_unlock(&wr_ctx->wr_lock);
1575                         return -EIO;
1576                 }
1577                 scrub_wr_submit(sctx);
1578                 goto again;
1579         }
1580
1581         sbio->pagev[sbio->page_count] = spage;
1582         scrub_page_get(spage);
1583         sbio->page_count++;
1584         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1585                 scrub_wr_submit(sctx);
1586         mutex_unlock(&wr_ctx->wr_lock);
1587
1588         return 0;
1589 }
1590
1591 static void scrub_wr_submit(struct scrub_ctx *sctx)
1592 {
1593         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1594         struct scrub_bio *sbio;
1595
1596         if (!wr_ctx->wr_curr_bio)
1597                 return;
1598
1599         sbio = wr_ctx->wr_curr_bio;
1600         wr_ctx->wr_curr_bio = NULL;
1601         WARN_ON(!sbio->bio->bi_bdev);
1602         scrub_pending_bio_inc(sctx);
1603         /* process all writes in a single worker thread. Then the block layer
1604          * orders the requests before sending them to the driver which
1605          * doubled the write performance on spinning disks when measured
1606          * with Linux 3.5 */
1607         btrfsic_submit_bio(WRITE, sbio->bio);
1608 }
1609
1610 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1611 {
1612         struct scrub_bio *sbio = bio->bi_private;
1613         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1614
1615         sbio->err = err;
1616         sbio->bio = bio;
1617
1618         btrfs_init_work(&sbio->work, scrub_wr_bio_end_io_worker, NULL, NULL);
1619         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1620 }
1621
1622 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1623 {
1624         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1625         struct scrub_ctx *sctx = sbio->sctx;
1626         int i;
1627
1628         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1629         if (sbio->err) {
1630                 struct btrfs_dev_replace *dev_replace =
1631                         &sbio->sctx->dev_root->fs_info->dev_replace;
1632
1633                 for (i = 0; i < sbio->page_count; i++) {
1634                         struct scrub_page *spage = sbio->pagev[i];
1635
1636                         spage->io_error = 1;
1637                         btrfs_dev_replace_stats_inc(&dev_replace->
1638                                                     num_write_errors);
1639                 }
1640         }
1641
1642         for (i = 0; i < sbio->page_count; i++)
1643                 scrub_page_put(sbio->pagev[i]);
1644
1645         bio_put(sbio->bio);
1646         kfree(sbio);
1647         scrub_pending_bio_dec(sctx);
1648 }
1649
1650 static int scrub_checksum(struct scrub_block *sblock)
1651 {
1652         u64 flags;
1653         int ret;
1654
1655         WARN_ON(sblock->page_count < 1);
1656         flags = sblock->pagev[0]->flags;
1657         ret = 0;
1658         if (flags & BTRFS_EXTENT_FLAG_DATA)
1659                 ret = scrub_checksum_data(sblock);
1660         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1661                 ret = scrub_checksum_tree_block(sblock);
1662         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1663                 (void)scrub_checksum_super(sblock);
1664         else
1665                 WARN_ON(1);
1666         if (ret)
1667                 scrub_handle_errored_block(sblock);
1668
1669         return ret;
1670 }
1671
1672 static int scrub_checksum_data(struct scrub_block *sblock)
1673 {
1674         struct scrub_ctx *sctx = sblock->sctx;
1675         u8 csum[BTRFS_CSUM_SIZE];
1676         u8 *on_disk_csum;
1677         struct page *page;
1678         void *buffer;
1679         u32 crc = ~(u32)0;
1680         int fail = 0;
1681         u64 len;
1682         int index;
1683
1684         BUG_ON(sblock->page_count < 1);
1685         if (!sblock->pagev[0]->have_csum)
1686                 return 0;
1687
1688         on_disk_csum = sblock->pagev[0]->csum;
1689         page = sblock->pagev[0]->page;
1690         buffer = kmap_atomic(page);
1691
1692         len = sctx->sectorsize;
1693         index = 0;
1694         for (;;) {
1695                 u64 l = min_t(u64, len, PAGE_SIZE);
1696
1697                 crc = btrfs_csum_data(buffer, crc, l);
1698                 kunmap_atomic(buffer);
1699                 len -= l;
1700                 if (len == 0)
1701                         break;
1702                 index++;
1703                 BUG_ON(index >= sblock->page_count);
1704                 BUG_ON(!sblock->pagev[index]->page);
1705                 page = sblock->pagev[index]->page;
1706                 buffer = kmap_atomic(page);
1707         }
1708
1709         btrfs_csum_final(crc, csum);
1710         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1711                 fail = 1;
1712
1713         return fail;
1714 }
1715
1716 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1717 {
1718         struct scrub_ctx *sctx = sblock->sctx;
1719         struct btrfs_header *h;
1720         struct btrfs_root *root = sctx->dev_root;
1721         struct btrfs_fs_info *fs_info = root->fs_info;
1722         u8 calculated_csum[BTRFS_CSUM_SIZE];
1723         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1724         struct page *page;
1725         void *mapped_buffer;
1726         u64 mapped_size;
1727         void *p;
1728         u32 crc = ~(u32)0;
1729         int fail = 0;
1730         int crc_fail = 0;
1731         u64 len;
1732         int index;
1733
1734         BUG_ON(sblock->page_count < 1);
1735         page = sblock->pagev[0]->page;
1736         mapped_buffer = kmap_atomic(page);
1737         h = (struct btrfs_header *)mapped_buffer;
1738         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1739
1740         /*
1741          * we don't use the getter functions here, as we
1742          * a) don't have an extent buffer and
1743          * b) the page is already kmapped
1744          */
1745
1746         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1747                 ++fail;
1748
1749         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1750                 ++fail;
1751
1752         if (memcmp(h->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1753                 ++fail;
1754
1755         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1756                    BTRFS_UUID_SIZE))
1757                 ++fail;
1758
1759         WARN_ON(sctx->nodesize != sctx->leafsize);
1760         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1761         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1762         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1763         index = 0;
1764         for (;;) {
1765                 u64 l = min_t(u64, len, mapped_size);
1766
1767                 crc = btrfs_csum_data(p, crc, l);
1768                 kunmap_atomic(mapped_buffer);
1769                 len -= l;
1770                 if (len == 0)
1771                         break;
1772                 index++;
1773                 BUG_ON(index >= sblock->page_count);
1774                 BUG_ON(!sblock->pagev[index]->page);
1775                 page = sblock->pagev[index]->page;
1776                 mapped_buffer = kmap_atomic(page);
1777                 mapped_size = PAGE_SIZE;
1778                 p = mapped_buffer;
1779         }
1780
1781         btrfs_csum_final(crc, calculated_csum);
1782         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1783                 ++crc_fail;
1784
1785         return fail || crc_fail;
1786 }
1787
1788 static int scrub_checksum_super(struct scrub_block *sblock)
1789 {
1790         struct btrfs_super_block *s;
1791         struct scrub_ctx *sctx = sblock->sctx;
1792         struct btrfs_root *root = sctx->dev_root;
1793         struct btrfs_fs_info *fs_info = root->fs_info;
1794         u8 calculated_csum[BTRFS_CSUM_SIZE];
1795         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1796         struct page *page;
1797         void *mapped_buffer;
1798         u64 mapped_size;
1799         void *p;
1800         u32 crc = ~(u32)0;
1801         int fail_gen = 0;
1802         int fail_cor = 0;
1803         u64 len;
1804         int index;
1805
1806         BUG_ON(sblock->page_count < 1);
1807         page = sblock->pagev[0]->page;
1808         mapped_buffer = kmap_atomic(page);
1809         s = (struct btrfs_super_block *)mapped_buffer;
1810         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1811
1812         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1813                 ++fail_cor;
1814
1815         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1816                 ++fail_gen;
1817
1818         if (memcmp(s->fsid, fs_info->fsid, BTRFS_UUID_SIZE))
1819                 ++fail_cor;
1820
1821         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1822         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1823         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1824         index = 0;
1825         for (;;) {
1826                 u64 l = min_t(u64, len, mapped_size);
1827
1828                 crc = btrfs_csum_data(p, crc, l);
1829                 kunmap_atomic(mapped_buffer);
1830                 len -= l;
1831                 if (len == 0)
1832                         break;
1833                 index++;
1834                 BUG_ON(index >= sblock->page_count);
1835                 BUG_ON(!sblock->pagev[index]->page);
1836                 page = sblock->pagev[index]->page;
1837                 mapped_buffer = kmap_atomic(page);
1838                 mapped_size = PAGE_SIZE;
1839                 p = mapped_buffer;
1840         }
1841
1842         btrfs_csum_final(crc, calculated_csum);
1843         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1844                 ++fail_cor;
1845
1846         if (fail_cor + fail_gen) {
1847                 /*
1848                  * if we find an error in a super block, we just report it.
1849                  * They will get written with the next transaction commit
1850                  * anyway
1851                  */
1852                 spin_lock(&sctx->stat_lock);
1853                 ++sctx->stat.super_errors;
1854                 spin_unlock(&sctx->stat_lock);
1855                 if (fail_cor)
1856                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1857                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1858                 else
1859                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1860                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1861         }
1862
1863         return fail_cor + fail_gen;
1864 }
1865
1866 static void scrub_block_get(struct scrub_block *sblock)
1867 {
1868         atomic_inc(&sblock->ref_count);
1869 }
1870
1871 static void scrub_block_put(struct scrub_block *sblock)
1872 {
1873         if (atomic_dec_and_test(&sblock->ref_count)) {
1874                 int i;
1875
1876                 for (i = 0; i < sblock->page_count; i++)
1877                         scrub_page_put(sblock->pagev[i]);
1878                 kfree(sblock);
1879         }
1880 }
1881
1882 static void scrub_page_get(struct scrub_page *spage)
1883 {
1884         atomic_inc(&spage->ref_count);
1885 }
1886
1887 static void scrub_page_put(struct scrub_page *spage)
1888 {
1889         if (atomic_dec_and_test(&spage->ref_count)) {
1890                 if (spage->page)
1891                         __free_page(spage->page);
1892                 kfree(spage);
1893         }
1894 }
1895
1896 static void scrub_submit(struct scrub_ctx *sctx)
1897 {
1898         struct scrub_bio *sbio;
1899
1900         if (sctx->curr == -1)
1901                 return;
1902
1903         sbio = sctx->bios[sctx->curr];
1904         sctx->curr = -1;
1905         scrub_pending_bio_inc(sctx);
1906
1907         if (!sbio->bio->bi_bdev) {
1908                 /*
1909                  * this case should not happen. If btrfs_map_block() is
1910                  * wrong, it could happen for dev-replace operations on
1911                  * missing devices when no mirrors are available, but in
1912                  * this case it should already fail the mount.
1913                  * This case is handled correctly (but _very_ slowly).
1914                  */
1915                 printk_ratelimited(KERN_WARNING
1916                         "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1917                 bio_endio(sbio->bio, -EIO);
1918         } else {
1919                 btrfsic_submit_bio(READ, sbio->bio);
1920         }
1921 }
1922
1923 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1924                                     struct scrub_page *spage)
1925 {
1926         struct scrub_block *sblock = spage->sblock;
1927         struct scrub_bio *sbio;
1928         int ret;
1929
1930 again:
1931         /*
1932          * grab a fresh bio or wait for one to become available
1933          */
1934         while (sctx->curr == -1) {
1935                 spin_lock(&sctx->list_lock);
1936                 sctx->curr = sctx->first_free;
1937                 if (sctx->curr != -1) {
1938                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1939                         sctx->bios[sctx->curr]->next_free = -1;
1940                         sctx->bios[sctx->curr]->page_count = 0;
1941                         spin_unlock(&sctx->list_lock);
1942                 } else {
1943                         spin_unlock(&sctx->list_lock);
1944                         wait_event(sctx->list_wait, sctx->first_free != -1);
1945                 }
1946         }
1947         sbio = sctx->bios[sctx->curr];
1948         if (sbio->page_count == 0) {
1949                 struct bio *bio;
1950
1951                 sbio->physical = spage->physical;
1952                 sbio->logical = spage->logical;
1953                 sbio->dev = spage->dev;
1954                 bio = sbio->bio;
1955                 if (!bio) {
1956                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1957                         if (!bio)
1958                                 return -ENOMEM;
1959                         sbio->bio = bio;
1960                 }
1961
1962                 bio->bi_private = sbio;
1963                 bio->bi_end_io = scrub_bio_end_io;
1964                 bio->bi_bdev = sbio->dev->bdev;
1965                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1966                 sbio->err = 0;
1967         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1968                    spage->physical ||
1969                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1970                    spage->logical ||
1971                    sbio->dev != spage->dev) {
1972                 scrub_submit(sctx);
1973                 goto again;
1974         }
1975
1976         sbio->pagev[sbio->page_count] = spage;
1977         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1978         if (ret != PAGE_SIZE) {
1979                 if (sbio->page_count < 1) {
1980                         bio_put(sbio->bio);
1981                         sbio->bio = NULL;
1982                         return -EIO;
1983                 }
1984                 scrub_submit(sctx);
1985                 goto again;
1986         }
1987
1988         scrub_block_get(sblock); /* one for the page added to the bio */
1989         atomic_inc(&sblock->outstanding_pages);
1990         sbio->page_count++;
1991         if (sbio->page_count == sctx->pages_per_rd_bio)
1992                 scrub_submit(sctx);
1993
1994         return 0;
1995 }
1996
1997 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
1998                        u64 physical, struct btrfs_device *dev, u64 flags,
1999                        u64 gen, int mirror_num, u8 *csum, int force,
2000                        u64 physical_for_dev_replace)
2001 {
2002         struct scrub_block *sblock;
2003         int index;
2004
2005         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2006         if (!sblock) {
2007                 spin_lock(&sctx->stat_lock);
2008                 sctx->stat.malloc_errors++;
2009                 spin_unlock(&sctx->stat_lock);
2010                 return -ENOMEM;
2011         }
2012
2013         /* one ref inside this function, plus one for each page added to
2014          * a bio later on */
2015         atomic_set(&sblock->ref_count, 1);
2016         sblock->sctx = sctx;
2017         sblock->no_io_error_seen = 1;
2018
2019         for (index = 0; len > 0; index++) {
2020                 struct scrub_page *spage;
2021                 u64 l = min_t(u64, len, PAGE_SIZE);
2022
2023                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2024                 if (!spage) {
2025 leave_nomem:
2026                         spin_lock(&sctx->stat_lock);
2027                         sctx->stat.malloc_errors++;
2028                         spin_unlock(&sctx->stat_lock);
2029                         scrub_block_put(sblock);
2030                         return -ENOMEM;
2031                 }
2032                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2033                 scrub_page_get(spage);
2034                 sblock->pagev[index] = spage;
2035                 spage->sblock = sblock;
2036                 spage->dev = dev;
2037                 spage->flags = flags;
2038                 spage->generation = gen;
2039                 spage->logical = logical;
2040                 spage->physical = physical;
2041                 spage->physical_for_dev_replace = physical_for_dev_replace;
2042                 spage->mirror_num = mirror_num;
2043                 if (csum) {
2044                         spage->have_csum = 1;
2045                         memcpy(spage->csum, csum, sctx->csum_size);
2046                 } else {
2047                         spage->have_csum = 0;
2048                 }
2049                 sblock->page_count++;
2050                 spage->page = alloc_page(GFP_NOFS);
2051                 if (!spage->page)
2052                         goto leave_nomem;
2053                 len -= l;
2054                 logical += l;
2055                 physical += l;
2056                 physical_for_dev_replace += l;
2057         }
2058
2059         WARN_ON(sblock->page_count == 0);
2060         for (index = 0; index < sblock->page_count; index++) {
2061                 struct scrub_page *spage = sblock->pagev[index];
2062                 int ret;
2063
2064                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2065                 if (ret) {
2066                         scrub_block_put(sblock);
2067                         return ret;
2068                 }
2069         }
2070
2071         if (force)
2072                 scrub_submit(sctx);
2073
2074         /* last one frees, either here or in bio completion for last page */
2075         scrub_block_put(sblock);
2076         return 0;
2077 }
2078
2079 static void scrub_bio_end_io(struct bio *bio, int err)
2080 {
2081         struct scrub_bio *sbio = bio->bi_private;
2082         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2083
2084         sbio->err = err;
2085         sbio->bio = bio;
2086
2087         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2088 }
2089
2090 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2091 {
2092         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2093         struct scrub_ctx *sctx = sbio->sctx;
2094         int i;
2095
2096         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2097         if (sbio->err) {
2098                 for (i = 0; i < sbio->page_count; i++) {
2099                         struct scrub_page *spage = sbio->pagev[i];
2100
2101                         spage->io_error = 1;
2102                         spage->sblock->no_io_error_seen = 0;
2103                 }
2104         }
2105
2106         /* now complete the scrub_block items that have all pages completed */
2107         for (i = 0; i < sbio->page_count; i++) {
2108                 struct scrub_page *spage = sbio->pagev[i];
2109                 struct scrub_block *sblock = spage->sblock;
2110
2111                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2112                         scrub_block_complete(sblock);
2113                 scrub_block_put(sblock);
2114         }
2115
2116         bio_put(sbio->bio);
2117         sbio->bio = NULL;
2118         spin_lock(&sctx->list_lock);
2119         sbio->next_free = sctx->first_free;
2120         sctx->first_free = sbio->index;
2121         spin_unlock(&sctx->list_lock);
2122
2123         if (sctx->is_dev_replace &&
2124             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2125                 mutex_lock(&sctx->wr_ctx.wr_lock);
2126                 scrub_wr_submit(sctx);
2127                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2128         }
2129
2130         scrub_pending_bio_dec(sctx);
2131 }
2132
2133 static void scrub_block_complete(struct scrub_block *sblock)
2134 {
2135         if (!sblock->no_io_error_seen) {
2136                 scrub_handle_errored_block(sblock);
2137         } else {
2138                 /*
2139                  * if has checksum error, write via repair mechanism in
2140                  * dev replace case, otherwise write here in dev replace
2141                  * case.
2142                  */
2143                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2144                         scrub_write_block_to_dev_replace(sblock);
2145         }
2146 }
2147
2148 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2149                            u8 *csum)
2150 {
2151         struct btrfs_ordered_sum *sum = NULL;
2152         unsigned long index;
2153         unsigned long num_sectors;
2154
2155         while (!list_empty(&sctx->csum_list)) {
2156                 sum = list_first_entry(&sctx->csum_list,
2157                                        struct btrfs_ordered_sum, list);
2158                 if (sum->bytenr > logical)
2159                         return 0;
2160                 if (sum->bytenr + sum->len > logical)
2161                         break;
2162
2163                 ++sctx->stat.csum_discards;
2164                 list_del(&sum->list);
2165                 kfree(sum);
2166                 sum = NULL;
2167         }
2168         if (!sum)
2169                 return 0;
2170
2171         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2172         num_sectors = sum->len / sctx->sectorsize;
2173         memcpy(csum, sum->sums + index, sctx->csum_size);
2174         if (index == num_sectors - 1) {
2175                 list_del(&sum->list);
2176                 kfree(sum);
2177         }
2178         return 1;
2179 }
2180
2181 /* scrub extent tries to collect up to 64 kB for each bio */
2182 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2183                         u64 physical, struct btrfs_device *dev, u64 flags,
2184                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2185 {
2186         int ret;
2187         u8 csum[BTRFS_CSUM_SIZE];
2188         u32 blocksize;
2189
2190         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2191                 blocksize = sctx->sectorsize;
2192                 spin_lock(&sctx->stat_lock);
2193                 sctx->stat.data_extents_scrubbed++;
2194                 sctx->stat.data_bytes_scrubbed += len;
2195                 spin_unlock(&sctx->stat_lock);
2196         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2197                 WARN_ON(sctx->nodesize != sctx->leafsize);
2198                 blocksize = sctx->nodesize;
2199                 spin_lock(&sctx->stat_lock);
2200                 sctx->stat.tree_extents_scrubbed++;
2201                 sctx->stat.tree_bytes_scrubbed += len;
2202                 spin_unlock(&sctx->stat_lock);
2203         } else {
2204                 blocksize = sctx->sectorsize;
2205                 WARN_ON(1);
2206         }
2207
2208         while (len) {
2209                 u64 l = min_t(u64, len, blocksize);
2210                 int have_csum = 0;
2211
2212                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2213                         /* push csums to sbio */
2214                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2215                         if (have_csum == 0)
2216                                 ++sctx->stat.no_csum;
2217                         if (sctx->is_dev_replace && !have_csum) {
2218                                 ret = copy_nocow_pages(sctx, logical, l,
2219                                                        mirror_num,
2220                                                       physical_for_dev_replace);
2221                                 goto behind_scrub_pages;
2222                         }
2223                 }
2224                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2225                                   mirror_num, have_csum ? csum : NULL, 0,
2226                                   physical_for_dev_replace);
2227 behind_scrub_pages:
2228                 if (ret)
2229                         return ret;
2230                 len -= l;
2231                 logical += l;
2232                 physical += l;
2233                 physical_for_dev_replace += l;
2234         }
2235         return 0;
2236 }
2237
2238 /*
2239  * Given a physical address, this will calculate it's
2240  * logical offset. if this is a parity stripe, it will return
2241  * the most left data stripe's logical offset.
2242  *
2243  * return 0 if it is a data stripe, 1 means parity stripe.
2244  */
2245 static int get_raid56_logic_offset(u64 physical, int num,
2246                                    struct map_lookup *map, u64 *offset)
2247 {
2248         int i;
2249         int j = 0;
2250         u64 stripe_nr;
2251         u64 last_offset;
2252         int stripe_index;
2253         int rot;
2254
2255         last_offset = (physical - map->stripes[num].physical) *
2256                       nr_data_stripes(map);
2257         *offset = last_offset;
2258         for (i = 0; i < nr_data_stripes(map); i++) {
2259                 *offset = last_offset + i * map->stripe_len;
2260
2261                 stripe_nr = *offset;
2262                 do_div(stripe_nr, map->stripe_len);
2263                 do_div(stripe_nr, nr_data_stripes(map));
2264
2265                 /* Work out the disk rotation on this stripe-set */
2266                 rot = do_div(stripe_nr, map->num_stripes);
2267                 /* calculate which stripe this data locates */
2268                 rot += i;
2269                 stripe_index = rot % map->num_stripes;
2270                 if (stripe_index == num)
2271                         return 0;
2272                 if (stripe_index < num)
2273                         j++;
2274         }
2275         *offset = last_offset + j * map->stripe_len;
2276         return 1;
2277 }
2278
2279 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2280                                            struct map_lookup *map,
2281                                            struct btrfs_device *scrub_dev,
2282                                            int num, u64 base, u64 length,
2283                                            int is_dev_replace)
2284 {
2285         struct btrfs_path *path;
2286         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2287         struct btrfs_root *root = fs_info->extent_root;
2288         struct btrfs_root *csum_root = fs_info->csum_root;
2289         struct btrfs_extent_item *extent;
2290         struct blk_plug plug;
2291         u64 flags;
2292         int ret;
2293         int slot;
2294         u64 nstripes;
2295         struct extent_buffer *l;
2296         struct btrfs_key key;
2297         u64 physical;
2298         u64 logical;
2299         u64 logic_end;
2300         u64 physical_end;
2301         u64 generation;
2302         int mirror_num;
2303         struct reada_control *reada1;
2304         struct reada_control *reada2;
2305         struct btrfs_key key_start;
2306         struct btrfs_key key_end;
2307         u64 increment = map->stripe_len;
2308         u64 offset;
2309         u64 extent_logical;
2310         u64 extent_physical;
2311         u64 extent_len;
2312         struct btrfs_device *extent_dev;
2313         int extent_mirror_num;
2314         int stop_loop = 0;
2315
2316         nstripes = length;
2317         physical = map->stripes[num].physical;
2318         offset = 0;
2319         do_div(nstripes, map->stripe_len);
2320         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2321                 offset = map->stripe_len * num;
2322                 increment = map->stripe_len * map->num_stripes;
2323                 mirror_num = 1;
2324         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2325                 int factor = map->num_stripes / map->sub_stripes;
2326                 offset = map->stripe_len * (num / map->sub_stripes);
2327                 increment = map->stripe_len * factor;
2328                 mirror_num = num % map->sub_stripes + 1;
2329         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2330                 increment = map->stripe_len;
2331                 mirror_num = num % map->num_stripes + 1;
2332         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2333                 increment = map->stripe_len;
2334                 mirror_num = num % map->num_stripes + 1;
2335         } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2336                                 BTRFS_BLOCK_GROUP_RAID6)) {
2337                 get_raid56_logic_offset(physical, num, map, &offset);
2338                 increment = map->stripe_len * nr_data_stripes(map);
2339                 mirror_num = 1;
2340         } else {
2341                 increment = map->stripe_len;
2342                 mirror_num = 1;
2343         }
2344
2345         path = btrfs_alloc_path();
2346         if (!path)
2347                 return -ENOMEM;
2348
2349         /*
2350          * work on commit root. The related disk blocks are static as
2351          * long as COW is applied. This means, it is save to rewrite
2352          * them to repair disk errors without any race conditions
2353          */
2354         path->search_commit_root = 1;
2355         path->skip_locking = 1;
2356
2357         /*
2358          * trigger the readahead for extent tree csum tree and wait for
2359          * completion. During readahead, the scrub is officially paused
2360          * to not hold off transaction commits
2361          */
2362         logical = base + offset;
2363         physical_end = physical + nstripes * map->stripe_len;
2364         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2365                          BTRFS_BLOCK_GROUP_RAID6)) {
2366                 get_raid56_logic_offset(physical_end, num,
2367                                         map, &logic_end);
2368                 logic_end += base;
2369         } else {
2370                 logic_end = logical + increment * nstripes;
2371         }
2372         wait_event(sctx->list_wait,
2373                    atomic_read(&sctx->bios_in_flight) == 0);
2374         scrub_blocked_if_needed(fs_info);
2375
2376         /* FIXME it might be better to start readahead at commit root */
2377         key_start.objectid = logical;
2378         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2379         key_start.offset = (u64)0;
2380         key_end.objectid = logic_end;
2381         key_end.type = BTRFS_METADATA_ITEM_KEY;
2382         key_end.offset = (u64)-1;
2383         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2384
2385         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2386         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2387         key_start.offset = logical;
2388         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2389         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2390         key_end.offset = logic_end;
2391         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2392
2393         if (!IS_ERR(reada1))
2394                 btrfs_reada_wait(reada1);
2395         if (!IS_ERR(reada2))
2396                 btrfs_reada_wait(reada2);
2397
2398
2399         /*
2400          * collect all data csums for the stripe to avoid seeking during
2401          * the scrub. This might currently (crc32) end up to be about 1MB
2402          */
2403         blk_start_plug(&plug);
2404
2405         /*
2406          * now find all extents for each stripe and scrub them
2407          */
2408         ret = 0;
2409         while (physical < physical_end) {
2410                 /* for raid56, we skip parity stripe */
2411                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2412                                 BTRFS_BLOCK_GROUP_RAID6)) {
2413                         ret = get_raid56_logic_offset(physical, num,
2414                                         map, &logical);
2415                         logical += base;
2416                         if (ret)
2417                                 goto skip;
2418                 }
2419                 /*
2420                  * canceled?
2421                  */
2422                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2423                     atomic_read(&sctx->cancel_req)) {
2424                         ret = -ECANCELED;
2425                         goto out;
2426                 }
2427                 /*
2428                  * check to see if we have to pause
2429                  */
2430                 if (atomic_read(&fs_info->scrub_pause_req)) {
2431                         /* push queued extents */
2432                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2433                         scrub_submit(sctx);
2434                         mutex_lock(&sctx->wr_ctx.wr_lock);
2435                         scrub_wr_submit(sctx);
2436                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2437                         wait_event(sctx->list_wait,
2438                                    atomic_read(&sctx->bios_in_flight) == 0);
2439                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2440                         scrub_blocked_if_needed(fs_info);
2441                 }
2442
2443                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2444                         key.type = BTRFS_METADATA_ITEM_KEY;
2445                 else
2446                         key.type = BTRFS_EXTENT_ITEM_KEY;
2447                 key.objectid = logical;
2448                 key.offset = (u64)-1;
2449
2450                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2451                 if (ret < 0)
2452                         goto out;
2453
2454                 if (ret > 0) {
2455                         ret = btrfs_previous_extent_item(root, path, 0);
2456                         if (ret < 0)
2457                                 goto out;
2458                         if (ret > 0) {
2459                                 /* there's no smaller item, so stick with the
2460                                  * larger one */
2461                                 btrfs_release_path(path);
2462                                 ret = btrfs_search_slot(NULL, root, &key,
2463                                                         path, 0, 0);
2464                                 if (ret < 0)
2465                                         goto out;
2466                         }
2467                 }
2468
2469                 stop_loop = 0;
2470                 while (1) {
2471                         u64 bytes;
2472
2473                         l = path->nodes[0];
2474                         slot = path->slots[0];
2475                         if (slot >= btrfs_header_nritems(l)) {
2476                                 ret = btrfs_next_leaf(root, path);
2477                                 if (ret == 0)
2478                                         continue;
2479                                 if (ret < 0)
2480                                         goto out;
2481
2482                                 stop_loop = 1;
2483                                 break;
2484                         }
2485                         btrfs_item_key_to_cpu(l, &key, slot);
2486
2487                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2488                                 bytes = root->leafsize;
2489                         else
2490                                 bytes = key.offset;
2491
2492                         if (key.objectid + bytes <= logical)
2493                                 goto next;
2494
2495                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2496                             key.type != BTRFS_METADATA_ITEM_KEY)
2497                                 goto next;
2498
2499                         if (key.objectid >= logical + map->stripe_len) {
2500                                 /* out of this device extent */
2501                                 if (key.objectid >= logic_end)
2502                                         stop_loop = 1;
2503                                 break;
2504                         }
2505
2506                         extent = btrfs_item_ptr(l, slot,
2507                                                 struct btrfs_extent_item);
2508                         flags = btrfs_extent_flags(l, extent);
2509                         generation = btrfs_extent_generation(l, extent);
2510
2511                         if (key.objectid < logical &&
2512                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2513                                 btrfs_err(fs_info,
2514                                            "scrub: tree block %llu spanning "
2515                                            "stripes, ignored. logical=%llu",
2516                                        key.objectid, logical);
2517                                 goto next;
2518                         }
2519
2520 again:
2521                         extent_logical = key.objectid;
2522                         extent_len = bytes;
2523
2524                         /*
2525                          * trim extent to this stripe
2526                          */
2527                         if (extent_logical < logical) {
2528                                 extent_len -= logical - extent_logical;
2529                                 extent_logical = logical;
2530                         }
2531                         if (extent_logical + extent_len >
2532                             logical + map->stripe_len) {
2533                                 extent_len = logical + map->stripe_len -
2534                                              extent_logical;
2535                         }
2536
2537                         extent_physical = extent_logical - logical + physical;
2538                         extent_dev = scrub_dev;
2539                         extent_mirror_num = mirror_num;
2540                         if (is_dev_replace)
2541                                 scrub_remap_extent(fs_info, extent_logical,
2542                                                    extent_len, &extent_physical,
2543                                                    &extent_dev,
2544                                                    &extent_mirror_num);
2545
2546                         ret = btrfs_lookup_csums_range(csum_root, logical,
2547                                                 logical + map->stripe_len - 1,
2548                                                 &sctx->csum_list, 1);
2549                         if (ret)
2550                                 goto out;
2551
2552                         ret = scrub_extent(sctx, extent_logical, extent_len,
2553                                            extent_physical, extent_dev, flags,
2554                                            generation, extent_mirror_num,
2555                                            extent_logical - logical + physical);
2556                         if (ret)
2557                                 goto out;
2558
2559                         scrub_free_csums(sctx);
2560                         if (extent_logical + extent_len <
2561                             key.objectid + bytes) {
2562                                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2563                                         BTRFS_BLOCK_GROUP_RAID6)) {
2564                                         /*
2565                                          * loop until we find next data stripe
2566                                          * or we have finished all stripes.
2567                                          */
2568                                         do {
2569                                                 physical += map->stripe_len;
2570                                                 ret = get_raid56_logic_offset(
2571                                                                 physical, num,
2572                                                                 map, &logical);
2573                                                 logical += base;
2574                                         } while (physical < physical_end && ret);
2575                                 } else {
2576                                         physical += map->stripe_len;
2577                                         logical += increment;
2578                                 }
2579                                 if (logical < key.objectid + bytes) {
2580                                         cond_resched();
2581                                         goto again;
2582                                 }
2583
2584                                 if (physical >= physical_end) {
2585                                         stop_loop = 1;
2586                                         break;
2587                                 }
2588                         }
2589 next:
2590                         path->slots[0]++;
2591                 }
2592                 btrfs_release_path(path);
2593 skip:
2594                 logical += increment;
2595                 physical += map->stripe_len;
2596                 spin_lock(&sctx->stat_lock);
2597                 if (stop_loop)
2598                         sctx->stat.last_physical = map->stripes[num].physical +
2599                                                    length;
2600                 else
2601                         sctx->stat.last_physical = physical;
2602                 spin_unlock(&sctx->stat_lock);
2603                 if (stop_loop)
2604                         break;
2605         }
2606 out:
2607         /* push queued extents */
2608         scrub_submit(sctx);
2609         mutex_lock(&sctx->wr_ctx.wr_lock);
2610         scrub_wr_submit(sctx);
2611         mutex_unlock(&sctx->wr_ctx.wr_lock);
2612
2613         blk_finish_plug(&plug);
2614         btrfs_free_path(path);
2615         return ret < 0 ? ret : 0;
2616 }
2617
2618 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2619                                           struct btrfs_device *scrub_dev,
2620                                           u64 chunk_tree, u64 chunk_objectid,
2621                                           u64 chunk_offset, u64 length,
2622                                           u64 dev_offset, int is_dev_replace)
2623 {
2624         struct btrfs_mapping_tree *map_tree =
2625                 &sctx->dev_root->fs_info->mapping_tree;
2626         struct map_lookup *map;
2627         struct extent_map *em;
2628         int i;
2629         int ret = 0;
2630
2631         read_lock(&map_tree->map_tree.lock);
2632         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2633         read_unlock(&map_tree->map_tree.lock);
2634
2635         if (!em)
2636                 return -EINVAL;
2637
2638         map = (struct map_lookup *)em->bdev;
2639         if (em->start != chunk_offset)
2640                 goto out;
2641
2642         if (em->len < length)
2643                 goto out;
2644
2645         for (i = 0; i < map->num_stripes; ++i) {
2646                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2647                     map->stripes[i].physical == dev_offset) {
2648                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2649                                            chunk_offset, length,
2650                                            is_dev_replace);
2651                         if (ret)
2652                                 goto out;
2653                 }
2654         }
2655 out:
2656         free_extent_map(em);
2657
2658         return ret;
2659 }
2660
2661 static noinline_for_stack
2662 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2663                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2664                            int is_dev_replace)
2665 {
2666         struct btrfs_dev_extent *dev_extent = NULL;
2667         struct btrfs_path *path;
2668         struct btrfs_root *root = sctx->dev_root;
2669         struct btrfs_fs_info *fs_info = root->fs_info;
2670         u64 length;
2671         u64 chunk_tree;
2672         u64 chunk_objectid;
2673         u64 chunk_offset;
2674         int ret;
2675         int slot;
2676         struct extent_buffer *l;
2677         struct btrfs_key key;
2678         struct btrfs_key found_key;
2679         struct btrfs_block_group_cache *cache;
2680         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2681
2682         path = btrfs_alloc_path();
2683         if (!path)
2684                 return -ENOMEM;
2685
2686         path->reada = 2;
2687         path->search_commit_root = 1;
2688         path->skip_locking = 1;
2689
2690         key.objectid = scrub_dev->devid;
2691         key.offset = 0ull;
2692         key.type = BTRFS_DEV_EXTENT_KEY;
2693
2694         while (1) {
2695                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2696                 if (ret < 0)
2697                         break;
2698                 if (ret > 0) {
2699                         if (path->slots[0] >=
2700                             btrfs_header_nritems(path->nodes[0])) {
2701                                 ret = btrfs_next_leaf(root, path);
2702                                 if (ret)
2703                                         break;
2704                         }
2705                 }
2706
2707                 l = path->nodes[0];
2708                 slot = path->slots[0];
2709
2710                 btrfs_item_key_to_cpu(l, &found_key, slot);
2711
2712                 if (found_key.objectid != scrub_dev->devid)
2713                         break;
2714
2715                 if (btrfs_key_type(&found_key) != BTRFS_DEV_EXTENT_KEY)
2716                         break;
2717
2718                 if (found_key.offset >= end)
2719                         break;
2720
2721                 if (found_key.offset < key.offset)
2722                         break;
2723
2724                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2725                 length = btrfs_dev_extent_length(l, dev_extent);
2726
2727                 if (found_key.offset + length <= start) {
2728                         key.offset = found_key.offset + length;
2729                         btrfs_release_path(path);
2730                         continue;
2731                 }
2732
2733                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2734                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2735                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2736
2737                 /*
2738                  * get a reference on the corresponding block group to prevent
2739                  * the chunk from going away while we scrub it
2740                  */
2741                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2742                 if (!cache) {
2743                         ret = -ENOENT;
2744                         break;
2745                 }
2746                 dev_replace->cursor_right = found_key.offset + length;
2747                 dev_replace->cursor_left = found_key.offset;
2748                 dev_replace->item_needs_writeback = 1;
2749                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2750                                   chunk_offset, length, found_key.offset,
2751                                   is_dev_replace);
2752
2753                 /*
2754                  * flush, submit all pending read and write bios, afterwards
2755                  * wait for them.
2756                  * Note that in the dev replace case, a read request causes
2757                  * write requests that are submitted in the read completion
2758                  * worker. Therefore in the current situation, it is required
2759                  * that all write requests are flushed, so that all read and
2760                  * write requests are really completed when bios_in_flight
2761                  * changes to 0.
2762                  */
2763                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2764                 scrub_submit(sctx);
2765                 mutex_lock(&sctx->wr_ctx.wr_lock);
2766                 scrub_wr_submit(sctx);
2767                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2768
2769                 wait_event(sctx->list_wait,
2770                            atomic_read(&sctx->bios_in_flight) == 0);
2771                 atomic_inc(&fs_info->scrubs_paused);
2772                 wake_up(&fs_info->scrub_pause_wait);
2773
2774                 /*
2775                  * must be called before we decrease @scrub_paused.
2776                  * make sure we don't block transaction commit while
2777                  * we are waiting pending workers finished.
2778                  */
2779                 wait_event(sctx->list_wait,
2780                            atomic_read(&sctx->workers_pending) == 0);
2781                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2782
2783                 mutex_lock(&fs_info->scrub_lock);
2784                 __scrub_blocked_if_needed(fs_info);
2785                 atomic_dec(&fs_info->scrubs_paused);
2786                 mutex_unlock(&fs_info->scrub_lock);
2787                 wake_up(&fs_info->scrub_pause_wait);
2788
2789                 btrfs_put_block_group(cache);
2790                 if (ret)
2791                         break;
2792                 if (is_dev_replace &&
2793                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2794                         ret = -EIO;
2795                         break;
2796                 }
2797                 if (sctx->stat.malloc_errors > 0) {
2798                         ret = -ENOMEM;
2799                         break;
2800                 }
2801
2802                 dev_replace->cursor_left = dev_replace->cursor_right;
2803                 dev_replace->item_needs_writeback = 1;
2804
2805                 key.offset = found_key.offset + length;
2806                 btrfs_release_path(path);
2807         }
2808
2809         btrfs_free_path(path);
2810
2811         /*
2812          * ret can still be 1 from search_slot or next_leaf,
2813          * that's not an error
2814          */
2815         return ret < 0 ? ret : 0;
2816 }
2817
2818 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2819                                            struct btrfs_device *scrub_dev)
2820 {
2821         int     i;
2822         u64     bytenr;
2823         u64     gen;
2824         int     ret;
2825         struct btrfs_root *root = sctx->dev_root;
2826
2827         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2828                 return -EIO;
2829
2830         gen = root->fs_info->last_trans_committed;
2831
2832         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2833                 bytenr = btrfs_sb_offset(i);
2834                 if (bytenr + BTRFS_SUPER_INFO_SIZE > scrub_dev->total_bytes)
2835                         break;
2836
2837                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2838                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2839                                   NULL, 1, bytenr);
2840                 if (ret)
2841                         return ret;
2842         }
2843         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2844
2845         return 0;
2846 }
2847
2848 /*
2849  * get a reference count on fs_info->scrub_workers. start worker if necessary
2850  */
2851 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2852                                                 int is_dev_replace)
2853 {
2854         int ret = 0;
2855         int flags = WQ_FREEZABLE | WQ_UNBOUND;
2856         int max_active = fs_info->thread_pool_size;
2857
2858         if (fs_info->scrub_workers_refcnt == 0) {
2859                 if (is_dev_replace)
2860                         fs_info->scrub_workers =
2861                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2862                                                       1, 4);
2863                 else
2864                         fs_info->scrub_workers =
2865                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2866                                                       max_active, 4);
2867                 if (!fs_info->scrub_workers) {
2868                         ret = -ENOMEM;
2869                         goto out;
2870                 }
2871                 fs_info->scrub_wr_completion_workers =
2872                         btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
2873                                               max_active, 2);
2874                 if (!fs_info->scrub_wr_completion_workers) {
2875                         ret = -ENOMEM;
2876                         goto out;
2877                 }
2878                 fs_info->scrub_nocow_workers =
2879                         btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
2880                 if (!fs_info->scrub_nocow_workers) {
2881                         ret = -ENOMEM;
2882                         goto out;
2883                 }
2884         }
2885         ++fs_info->scrub_workers_refcnt;
2886 out:
2887         return ret;
2888 }
2889
2890 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2891 {
2892         if (--fs_info->scrub_workers_refcnt == 0) {
2893                 btrfs_destroy_workqueue(fs_info->scrub_workers);
2894                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
2895                 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
2896         }
2897         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2898 }
2899
2900 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2901                     u64 end, struct btrfs_scrub_progress *progress,
2902                     int readonly, int is_dev_replace)
2903 {
2904         struct scrub_ctx *sctx;
2905         int ret;
2906         struct btrfs_device *dev;
2907
2908         if (btrfs_fs_closing(fs_info))
2909                 return -EINVAL;
2910
2911         /*
2912          * check some assumptions
2913          */
2914         if (fs_info->chunk_root->nodesize != fs_info->chunk_root->leafsize) {
2915                 btrfs_err(fs_info,
2916                            "scrub: size assumption nodesize == leafsize (%d == %d) fails",
2917                        fs_info->chunk_root->nodesize,
2918                        fs_info->chunk_root->leafsize);
2919                 return -EINVAL;
2920         }
2921
2922         if (fs_info->chunk_root->nodesize > BTRFS_STRIPE_LEN) {
2923                 /*
2924                  * in this case scrub is unable to calculate the checksum
2925                  * the way scrub is implemented. Do not handle this
2926                  * situation at all because it won't ever happen.
2927                  */
2928                 btrfs_err(fs_info,
2929                            "scrub: size assumption nodesize <= BTRFS_STRIPE_LEN (%d <= %d) fails",
2930                        fs_info->chunk_root->nodesize, BTRFS_STRIPE_LEN);
2931                 return -EINVAL;
2932         }
2933
2934         if (fs_info->chunk_root->sectorsize != PAGE_SIZE) {
2935                 /* not supported for data w/o checksums */
2936                 btrfs_err(fs_info,
2937                            "scrub: size assumption sectorsize != PAGE_SIZE "
2938                            "(%d != %lu) fails",
2939                        fs_info->chunk_root->sectorsize, PAGE_SIZE);
2940                 return -EINVAL;
2941         }
2942
2943         if (fs_info->chunk_root->nodesize >
2944             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK ||
2945             fs_info->chunk_root->sectorsize >
2946             PAGE_SIZE * SCRUB_MAX_PAGES_PER_BLOCK) {
2947                 /*
2948                  * would exhaust the array bounds of pagev member in
2949                  * struct scrub_block
2950                  */
2951                 btrfs_err(fs_info, "scrub: size assumption nodesize and sectorsize "
2952                            "<= SCRUB_MAX_PAGES_PER_BLOCK (%d <= %d && %d <= %d) fails",
2953                        fs_info->chunk_root->nodesize,
2954                        SCRUB_MAX_PAGES_PER_BLOCK,
2955                        fs_info->chunk_root->sectorsize,
2956                        SCRUB_MAX_PAGES_PER_BLOCK);
2957                 return -EINVAL;
2958         }
2959
2960
2961         mutex_lock(&fs_info->fs_devices->device_list_mutex);
2962         dev = btrfs_find_device(fs_info, devid, NULL, NULL);
2963         if (!dev || (dev->missing && !is_dev_replace)) {
2964                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2965                 return -ENODEV;
2966         }
2967
2968         mutex_lock(&fs_info->scrub_lock);
2969         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2970                 mutex_unlock(&fs_info->scrub_lock);
2971                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2972                 return -EIO;
2973         }
2974
2975         btrfs_dev_replace_lock(&fs_info->dev_replace);
2976         if (dev->scrub_device ||
2977             (!is_dev_replace &&
2978              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2979                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2980                 mutex_unlock(&fs_info->scrub_lock);
2981                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2982                 return -EINPROGRESS;
2983         }
2984         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2985
2986         ret = scrub_workers_get(fs_info, is_dev_replace);
2987         if (ret) {
2988                 mutex_unlock(&fs_info->scrub_lock);
2989                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2990                 return ret;
2991         }
2992
2993         sctx = scrub_setup_ctx(dev, is_dev_replace);
2994         if (IS_ERR(sctx)) {
2995                 mutex_unlock(&fs_info->scrub_lock);
2996                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2997                 scrub_workers_put(fs_info);
2998                 return PTR_ERR(sctx);
2999         }
3000         sctx->readonly = readonly;
3001         dev->scrub_device = sctx;
3002         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3003
3004         /*
3005          * checking @scrub_pause_req here, we can avoid
3006          * race between committing transaction and scrubbing.
3007          */
3008         __scrub_blocked_if_needed(fs_info);
3009         atomic_inc(&fs_info->scrubs_running);
3010         mutex_unlock(&fs_info->scrub_lock);
3011
3012         if (!is_dev_replace) {
3013                 /*
3014                  * by holding device list mutex, we can
3015                  * kick off writing super in log tree sync.
3016                  */
3017                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3018                 ret = scrub_supers(sctx, dev);
3019                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3020         }
3021
3022         if (!ret)
3023                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3024                                              is_dev_replace);
3025
3026         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3027         atomic_dec(&fs_info->scrubs_running);
3028         wake_up(&fs_info->scrub_pause_wait);
3029
3030         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3031
3032         if (progress)
3033                 memcpy(progress, &sctx->stat, sizeof(*progress));
3034
3035         mutex_lock(&fs_info->scrub_lock);
3036         dev->scrub_device = NULL;
3037         scrub_workers_put(fs_info);
3038         mutex_unlock(&fs_info->scrub_lock);
3039
3040         scrub_free_ctx(sctx);
3041
3042         return ret;
3043 }
3044
3045 void btrfs_scrub_pause(struct btrfs_root *root)
3046 {
3047         struct btrfs_fs_info *fs_info = root->fs_info;
3048
3049         mutex_lock(&fs_info->scrub_lock);
3050         atomic_inc(&fs_info->scrub_pause_req);
3051         while (atomic_read(&fs_info->scrubs_paused) !=
3052                atomic_read(&fs_info->scrubs_running)) {
3053                 mutex_unlock(&fs_info->scrub_lock);
3054                 wait_event(fs_info->scrub_pause_wait,
3055                            atomic_read(&fs_info->scrubs_paused) ==
3056                            atomic_read(&fs_info->scrubs_running));
3057                 mutex_lock(&fs_info->scrub_lock);
3058         }
3059         mutex_unlock(&fs_info->scrub_lock);
3060 }
3061
3062 void btrfs_scrub_continue(struct btrfs_root *root)
3063 {
3064         struct btrfs_fs_info *fs_info = root->fs_info;
3065
3066         atomic_dec(&fs_info->scrub_pause_req);
3067         wake_up(&fs_info->scrub_pause_wait);
3068 }
3069
3070 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3071 {
3072         mutex_lock(&fs_info->scrub_lock);
3073         if (!atomic_read(&fs_info->scrubs_running)) {
3074                 mutex_unlock(&fs_info->scrub_lock);
3075                 return -ENOTCONN;
3076         }
3077
3078         atomic_inc(&fs_info->scrub_cancel_req);
3079         while (atomic_read(&fs_info->scrubs_running)) {
3080                 mutex_unlock(&fs_info->scrub_lock);
3081                 wait_event(fs_info->scrub_pause_wait,
3082                            atomic_read(&fs_info->scrubs_running) == 0);
3083                 mutex_lock(&fs_info->scrub_lock);
3084         }
3085         atomic_dec(&fs_info->scrub_cancel_req);
3086         mutex_unlock(&fs_info->scrub_lock);
3087
3088         return 0;
3089 }
3090
3091 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3092                            struct btrfs_device *dev)
3093 {
3094         struct scrub_ctx *sctx;
3095
3096         mutex_lock(&fs_info->scrub_lock);
3097         sctx = dev->scrub_device;
3098         if (!sctx) {
3099                 mutex_unlock(&fs_info->scrub_lock);
3100                 return -ENOTCONN;
3101         }
3102         atomic_inc(&sctx->cancel_req);
3103         while (dev->scrub_device) {
3104                 mutex_unlock(&fs_info->scrub_lock);
3105                 wait_event(fs_info->scrub_pause_wait,
3106                            dev->scrub_device == NULL);
3107                 mutex_lock(&fs_info->scrub_lock);
3108         }
3109         mutex_unlock(&fs_info->scrub_lock);
3110
3111         return 0;
3112 }
3113
3114 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3115                          struct btrfs_scrub_progress *progress)
3116 {
3117         struct btrfs_device *dev;
3118         struct scrub_ctx *sctx = NULL;
3119
3120         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3121         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3122         if (dev)
3123                 sctx = dev->scrub_device;
3124         if (sctx)
3125                 memcpy(progress, &sctx->stat, sizeof(*progress));
3126         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3127
3128         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3129 }
3130
3131 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3132                                u64 extent_logical, u64 extent_len,
3133                                u64 *extent_physical,
3134                                struct btrfs_device **extent_dev,
3135                                int *extent_mirror_num)
3136 {
3137         u64 mapped_length;
3138         struct btrfs_bio *bbio = NULL;
3139         int ret;
3140
3141         mapped_length = extent_len;
3142         ret = btrfs_map_block(fs_info, READ, extent_logical,
3143                               &mapped_length, &bbio, 0);
3144         if (ret || !bbio || mapped_length < extent_len ||
3145             !bbio->stripes[0].dev->bdev) {
3146                 kfree(bbio);
3147                 return;
3148         }
3149
3150         *extent_physical = bbio->stripes[0].physical;
3151         *extent_mirror_num = bbio->mirror_num;
3152         *extent_dev = bbio->stripes[0].dev;
3153         kfree(bbio);
3154 }
3155
3156 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3157                               struct scrub_wr_ctx *wr_ctx,
3158                               struct btrfs_fs_info *fs_info,
3159                               struct btrfs_device *dev,
3160                               int is_dev_replace)
3161 {
3162         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3163
3164         mutex_init(&wr_ctx->wr_lock);
3165         wr_ctx->wr_curr_bio = NULL;
3166         if (!is_dev_replace)
3167                 return 0;
3168
3169         WARN_ON(!dev->bdev);
3170         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3171                                          bio_get_nr_vecs(dev->bdev));
3172         wr_ctx->tgtdev = dev;
3173         atomic_set(&wr_ctx->flush_all_writes, 0);
3174         return 0;
3175 }
3176
3177 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3178 {
3179         mutex_lock(&wr_ctx->wr_lock);
3180         kfree(wr_ctx->wr_curr_bio);
3181         wr_ctx->wr_curr_bio = NULL;
3182         mutex_unlock(&wr_ctx->wr_lock);
3183 }
3184
3185 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3186                             int mirror_num, u64 physical_for_dev_replace)
3187 {
3188         struct scrub_copy_nocow_ctx *nocow_ctx;
3189         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3190
3191         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3192         if (!nocow_ctx) {
3193                 spin_lock(&sctx->stat_lock);
3194                 sctx->stat.malloc_errors++;
3195                 spin_unlock(&sctx->stat_lock);
3196                 return -ENOMEM;
3197         }
3198
3199         scrub_pending_trans_workers_inc(sctx);
3200
3201         nocow_ctx->sctx = sctx;
3202         nocow_ctx->logical = logical;
3203         nocow_ctx->len = len;
3204         nocow_ctx->mirror_num = mirror_num;
3205         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3206         btrfs_init_work(&nocow_ctx->work, copy_nocow_pages_worker, NULL, NULL);
3207         INIT_LIST_HEAD(&nocow_ctx->inodes);
3208         btrfs_queue_work(fs_info->scrub_nocow_workers,
3209                          &nocow_ctx->work);
3210
3211         return 0;
3212 }
3213
3214 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3215 {
3216         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3217         struct scrub_nocow_inode *nocow_inode;
3218
3219         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3220         if (!nocow_inode)
3221                 return -ENOMEM;
3222         nocow_inode->inum = inum;
3223         nocow_inode->offset = offset;
3224         nocow_inode->root = root;
3225         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3226         return 0;
3227 }
3228
3229 #define COPY_COMPLETE 1
3230
3231 static void copy_nocow_pages_worker(struct btrfs_work *work)
3232 {
3233         struct scrub_copy_nocow_ctx *nocow_ctx =
3234                 container_of(work, struct scrub_copy_nocow_ctx, work);
3235         struct scrub_ctx *sctx = nocow_ctx->sctx;
3236         u64 logical = nocow_ctx->logical;
3237         u64 len = nocow_ctx->len;
3238         int mirror_num = nocow_ctx->mirror_num;
3239         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3240         int ret;
3241         struct btrfs_trans_handle *trans = NULL;
3242         struct btrfs_fs_info *fs_info;
3243         struct btrfs_path *path;
3244         struct btrfs_root *root;
3245         int not_written = 0;
3246
3247         fs_info = sctx->dev_root->fs_info;
3248         root = fs_info->extent_root;
3249
3250         path = btrfs_alloc_path();
3251         if (!path) {
3252                 spin_lock(&sctx->stat_lock);
3253                 sctx->stat.malloc_errors++;
3254                 spin_unlock(&sctx->stat_lock);
3255                 not_written = 1;
3256                 goto out;
3257         }
3258
3259         trans = btrfs_join_transaction(root);
3260         if (IS_ERR(trans)) {
3261                 not_written = 1;
3262                 goto out;
3263         }
3264
3265         ret = iterate_inodes_from_logical(logical, fs_info, path,
3266                                           record_inode_for_nocow, nocow_ctx);
3267         if (ret != 0 && ret != -ENOENT) {
3268                 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3269                         "phys %llu, len %llu, mir %u, ret %d",
3270                         logical, physical_for_dev_replace, len, mirror_num,
3271                         ret);
3272                 not_written = 1;
3273                 goto out;
3274         }
3275
3276         btrfs_end_transaction(trans, root);
3277         trans = NULL;
3278         while (!list_empty(&nocow_ctx->inodes)) {
3279                 struct scrub_nocow_inode *entry;
3280                 entry = list_first_entry(&nocow_ctx->inodes,
3281                                          struct scrub_nocow_inode,
3282                                          list);
3283                 list_del_init(&entry->list);
3284                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3285                                                  entry->root, nocow_ctx);
3286                 kfree(entry);
3287                 if (ret == COPY_COMPLETE) {
3288                         ret = 0;
3289                         break;
3290                 } else if (ret) {
3291                         break;
3292                 }
3293         }
3294 out:
3295         while (!list_empty(&nocow_ctx->inodes)) {
3296                 struct scrub_nocow_inode *entry;
3297                 entry = list_first_entry(&nocow_ctx->inodes,
3298                                          struct scrub_nocow_inode,
3299                                          list);
3300                 list_del_init(&entry->list);
3301                 kfree(entry);
3302         }
3303         if (trans && !IS_ERR(trans))
3304                 btrfs_end_transaction(trans, root);
3305         if (not_written)
3306                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3307                                             num_uncorrectable_read_errors);
3308
3309         btrfs_free_path(path);
3310         kfree(nocow_ctx);
3311
3312         scrub_pending_trans_workers_dec(sctx);
3313 }
3314
3315 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3316                                       struct scrub_copy_nocow_ctx *nocow_ctx)
3317 {
3318         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3319         struct btrfs_key key;
3320         struct inode *inode;
3321         struct page *page;
3322         struct btrfs_root *local_root;
3323         struct btrfs_ordered_extent *ordered;
3324         struct extent_map *em;
3325         struct extent_state *cached_state = NULL;
3326         struct extent_io_tree *io_tree;
3327         u64 physical_for_dev_replace;
3328         u64 len = nocow_ctx->len;
3329         u64 lockstart = offset, lockend = offset + len - 1;
3330         unsigned long index;
3331         int srcu_index;
3332         int ret = 0;
3333         int err = 0;
3334
3335         key.objectid = root;
3336         key.type = BTRFS_ROOT_ITEM_KEY;
3337         key.offset = (u64)-1;
3338
3339         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3340
3341         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3342         if (IS_ERR(local_root)) {
3343                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3344                 return PTR_ERR(local_root);
3345         }
3346
3347         key.type = BTRFS_INODE_ITEM_KEY;
3348         key.objectid = inum;
3349         key.offset = 0;
3350         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3351         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3352         if (IS_ERR(inode))
3353                 return PTR_ERR(inode);
3354
3355         /* Avoid truncate/dio/punch hole.. */
3356         mutex_lock(&inode->i_mutex);
3357         inode_dio_wait(inode);
3358
3359         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3360         io_tree = &BTRFS_I(inode)->io_tree;
3361
3362         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3363         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3364         if (ordered) {
3365                 btrfs_put_ordered_extent(ordered);
3366                 goto out_unlock;
3367         }
3368
3369         em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3370         if (IS_ERR(em)) {
3371                 ret = PTR_ERR(em);
3372                 goto out_unlock;
3373         }
3374
3375         /*
3376          * This extent does not actually cover the logical extent anymore,
3377          * move on to the next inode.
3378          */
3379         if (em->block_start > nocow_ctx->logical ||
3380             em->block_start + em->block_len < nocow_ctx->logical + len) {
3381                 free_extent_map(em);
3382                 goto out_unlock;
3383         }
3384         free_extent_map(em);
3385
3386         while (len >= PAGE_CACHE_SIZE) {
3387                 index = offset >> PAGE_CACHE_SHIFT;
3388 again:
3389                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3390                 if (!page) {
3391                         btrfs_err(fs_info, "find_or_create_page() failed");
3392                         ret = -ENOMEM;
3393                         goto out;
3394                 }
3395
3396                 if (PageUptodate(page)) {
3397                         if (PageDirty(page))
3398                                 goto next_page;
3399                 } else {
3400                         ClearPageError(page);
3401                         err = extent_read_full_page_nolock(io_tree, page,
3402                                                            btrfs_get_extent,
3403                                                            nocow_ctx->mirror_num);
3404                         if (err) {
3405                                 ret = err;
3406                                 goto next_page;
3407                         }
3408
3409                         lock_page(page);
3410                         /*
3411                          * If the page has been remove from the page cache,
3412                          * the data on it is meaningless, because it may be
3413                          * old one, the new data may be written into the new
3414                          * page in the page cache.
3415                          */
3416                         if (page->mapping != inode->i_mapping) {
3417                                 unlock_page(page);
3418                                 page_cache_release(page);
3419                                 goto again;
3420                         }
3421                         if (!PageUptodate(page)) {
3422                                 ret = -EIO;
3423                                 goto next_page;
3424                         }
3425                 }
3426                 err = write_page_nocow(nocow_ctx->sctx,
3427                                        physical_for_dev_replace, page);
3428                 if (err)
3429                         ret = err;
3430 next_page:
3431                 unlock_page(page);
3432                 page_cache_release(page);
3433
3434                 if (ret)
3435                         break;
3436
3437                 offset += PAGE_CACHE_SIZE;
3438                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3439                 len -= PAGE_CACHE_SIZE;
3440         }
3441         ret = COPY_COMPLETE;
3442 out_unlock:
3443         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3444                              GFP_NOFS);
3445 out:
3446         mutex_unlock(&inode->i_mutex);
3447         iput(inode);
3448         return ret;
3449 }
3450
3451 static int write_page_nocow(struct scrub_ctx *sctx,
3452                             u64 physical_for_dev_replace, struct page *page)
3453 {
3454         struct bio *bio;
3455         struct btrfs_device *dev;
3456         int ret;
3457
3458         dev = sctx->wr_ctx.tgtdev;
3459         if (!dev)
3460                 return -EIO;
3461         if (!dev->bdev) {
3462                 printk_ratelimited(KERN_WARNING
3463                         "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3464                 return -EIO;
3465         }
3466         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3467         if (!bio) {
3468                 spin_lock(&sctx->stat_lock);
3469                 sctx->stat.malloc_errors++;
3470                 spin_unlock(&sctx->stat_lock);
3471                 return -ENOMEM;
3472         }
3473         bio->bi_iter.bi_size = 0;
3474         bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3475         bio->bi_bdev = dev->bdev;
3476         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3477         if (ret != PAGE_CACHE_SIZE) {
3478 leave_with_eio:
3479                 bio_put(bio);
3480                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3481                 return -EIO;
3482         }
3483
3484         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3485                 goto leave_with_eio;
3486
3487         bio_put(bio);
3488         return 0;
3489 }