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
 141         int                     is_dev_replace;
 142         struct scrub_wr_ctx     wr_ctx;
 143
 144         /*
 145          * statistics
 146          */
 147         struct btrfs_scrub_progress stat;
 148         spinlock_t              stat_lock;
 149 };
 150
 151 struct scrub_fixup_nodatasum {
 152         struct scrub_ctx        *sctx;
 153         struct btrfs_device     *dev;
 154         u64                     logical;
 155         struct btrfs_root       *root;
 156         struct btrfs_work       work;
 157         int                     mirror_num;
 158 };
 159
 160 struct scrub_nocow_inode {
 161         u64                     inum;
 162         u64                     offset;
 163         u64                     root;
 164         struct list_head        list;
 165 };
 166
 167 struct scrub_copy_nocow_ctx {
 168         struct scrub_ctx        *sctx;
 169         u64                     logical;
 170         u64                     len;
 171         int                     mirror_num;
 172         u64                     physical_for_dev_replace;
 173         struct list_head        inodes;
 174         struct btrfs_work       work;
 175 };
 176
 177 struct scrub_warning {
 178         struct btrfs_path       *path;
 179         u64                     extent_item_size;
 180         char                    *scratch_buf;
 181         char                    *msg_buf;
 182         const char              *errstr;
 183         sector_t                sector;
 184         u64                     logical;
 185         struct btrfs_device     *dev;
 186         int                     msg_bufsize;
 187         int                     scratch_bufsize;
 188 };
 189
 190
 191 static void scrub_pending_bio_inc(struct scrub_ctx *sctx);
 192 static void scrub_pending_bio_dec(struct scrub_ctx *sctx);
 193 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx);
 194 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx);
 195 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check);
 196 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
 197                                      struct btrfs_fs_info *fs_info,
 198                                      struct scrub_block *original_sblock,
 199                                      u64 length, u64 logical,
 200                                      struct scrub_block *sblocks_for_recheck);
 201 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
 202                                 struct scrub_block *sblock, int is_metadata,
 203                                 int have_csum, u8 *csum, u64 generation,
 204                                 u16 csum_size);
 205 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
 206                                          struct scrub_block *sblock,
 207                                          int is_metadata, int have_csum,
 208                                          const u8 *csum, u64 generation,
 209                                          u16 csum_size);
 210 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
 211                                              struct scrub_block *sblock_good,
 212                                              int force_write);
 213 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
 214                                             struct scrub_block *sblock_good,
 215                                             int page_num, int force_write);
 216 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock);
 217 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
 218                                            int page_num);
 219 static int scrub_checksum_data(struct scrub_block *sblock);
 220 static int scrub_checksum_tree_block(struct scrub_block *sblock);
 221 static int scrub_checksum_super(struct scrub_block *sblock);
 222 static void scrub_block_get(struct scrub_block *sblock);
 223 static void scrub_block_put(struct scrub_block *sblock);
 224 static void scrub_page_get(struct scrub_page *spage);
 225 static void scrub_page_put(struct scrub_page *spage);
 226 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
 227                                     struct scrub_page *spage);
 228 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 229                        u64 physical, struct btrfs_device *dev, u64 flags,
 230                        u64 gen, int mirror_num, u8 *csum, int force,
 231                        u64 physical_for_dev_replace);
 232 static void scrub_bio_end_io(struct bio *bio, int err);
 233 static void scrub_bio_end_io_worker(struct btrfs_work *work);
 234 static void scrub_block_complete(struct scrub_block *sblock);
 235 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
 236                                u64 extent_logical, u64 extent_len,
 237                                u64 *extent_physical,
 238                                struct btrfs_device **extent_dev,
 239                                int *extent_mirror_num);
 240 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
 241                               struct scrub_wr_ctx *wr_ctx,
 242                               struct btrfs_fs_info *fs_info,
 243                               struct btrfs_device *dev,
 244                               int is_dev_replace);
 245 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx);
 246 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
 247                                     struct scrub_page *spage);
 248 static void scrub_wr_submit(struct scrub_ctx *sctx);
 249 static void scrub_wr_bio_end_io(struct bio *bio, int err);
 250 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work);
 251 static int write_page_nocow(struct scrub_ctx *sctx,
 252                             u64 physical_for_dev_replace, struct page *page);
 253 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
 254                                       struct scrub_copy_nocow_ctx *ctx);
 255 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
 256                             int mirror_num, u64 physical_for_dev_replace);
 257 static void copy_nocow_pages_worker(struct btrfs_work *work);
 258 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 259 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info);
 260
 261
 262 static void scrub_pending_bio_inc(struct scrub_ctx *sctx)
 263 {
 264         atomic_inc(&sctx->bios_in_flight);
 265 }
 266
 267 static void scrub_pending_bio_dec(struct scrub_ctx *sctx)
 268 {
 269         atomic_dec(&sctx->bios_in_flight);
 270         wake_up(&sctx->list_wait);
 271 }
 272
 273 static void __scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 274 {
 275         while (atomic_read(&fs_info->scrub_pause_req)) {
 276                 mutex_unlock(&fs_info->scrub_lock);
 277                 wait_event(fs_info->scrub_pause_wait,
 278                    atomic_read(&fs_info->scrub_pause_req) == 0);
 279                 mutex_lock(&fs_info->scrub_lock);
 280         }
 281 }
 282
 283 static void scrub_blocked_if_needed(struct btrfs_fs_info *fs_info)
 284 {
 285         atomic_inc(&fs_info->scrubs_paused);
 286         wake_up(&fs_info->scrub_pause_wait);
 287
 288         mutex_lock(&fs_info->scrub_lock);
 289         __scrub_blocked_if_needed(fs_info);
 290         atomic_dec(&fs_info->scrubs_paused);
 291         mutex_unlock(&fs_info->scrub_lock);
 292
 293         wake_up(&fs_info->scrub_pause_wait);
 294 }
 295
 296 /*
 297  * used for workers that require transaction commits (i.e., for the
 298  * NOCOW case)
 299  */
 300 static void scrub_pending_trans_workers_inc(struct scrub_ctx *sctx)
 301 {
 302         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 303
 304         /*
 305          * increment scrubs_running to prevent cancel requests from
 306          * completing as long as a worker is running. we must also
 307          * increment scrubs_paused to prevent deadlocking on pause
 308          * requests used for transactions commits (as the worker uses a
 309          * transaction context). it is safe to regard the worker
 310          * as paused for all matters practical. effectively, we only
 311          * avoid cancellation requests from completing.
 312          */
 313         mutex_lock(&fs_info->scrub_lock);
 314         atomic_inc(&fs_info->scrubs_running);
 315         atomic_inc(&fs_info->scrubs_paused);
 316         mutex_unlock(&fs_info->scrub_lock);
 317
 318         /*
 319          * check if @scrubs_running=@scrubs_paused condition
 320          * inside wait_event() is not an atomic operation.
 321          * which means we may inc/dec @scrub_running/paused
 322          * at any time. Let's wake up @scrub_pause_wait as
 323          * much as we can to let commit transaction blocked less.
 324          */
 325         wake_up(&fs_info->scrub_pause_wait);
 326
 327         atomic_inc(&sctx->workers_pending);
 328 }
 329
 330 /* used for workers that require transaction commits */
 331 static void scrub_pending_trans_workers_dec(struct scrub_ctx *sctx)
 332 {
 333         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
 334
 335         /*
 336          * see scrub_pending_trans_workers_inc() why we're pretending
 337          * to be paused in the scrub counters
 338          */
 339         mutex_lock(&fs_info->scrub_lock);
 340         atomic_dec(&fs_info->scrubs_running);
 341         atomic_dec(&fs_info->scrubs_paused);
 342         mutex_unlock(&fs_info->scrub_lock);
 343         atomic_dec(&sctx->workers_pending);
 344         wake_up(&fs_info->scrub_pause_wait);
 345         wake_up(&sctx->list_wait);
 346 }
 347
 348 static void scrub_free_csums(struct scrub_ctx *sctx)
 349 {
 350         while (!list_empty(&sctx->csum_list)) {
 351                 struct btrfs_ordered_sum *sum;
 352                 sum = list_first_entry(&sctx->csum_list,
 353                                        struct btrfs_ordered_sum, list);
 354                 list_del(&sum->list);
 355                 kfree(sum);
 356         }
 357 }
 358
 359 static noinline_for_stack void scrub_free_ctx(struct scrub_ctx *sctx)
 360 {
 361         int i;
 362
 363         if (!sctx)
 364                 return;
 365
 366         scrub_free_wr_ctx(&sctx->wr_ctx);
 367
 368         /* this can happen when scrub is cancelled */
 369         if (sctx->curr != -1) {
 370                 struct scrub_bio *sbio = sctx->bios[sctx->curr];
 371
 372                 for (i = 0; i < sbio->page_count; i++) {
 373                         WARN_ON(!sbio->pagev[i]->page);
 374                         scrub_block_put(sbio->pagev[i]->sblock);
 375                 }
 376                 bio_put(sbio->bio);
 377         }
 378
 379         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 380                 struct scrub_bio *sbio = sctx->bios[i];
 381
 382                 if (!sbio)
 383                         break;
 384                 kfree(sbio);
 385         }
 386
 387         scrub_free_csums(sctx);
 388         kfree(sctx);
 389 }
 390
 391 static noinline_for_stack
 392 struct scrub_ctx *scrub_setup_ctx(struct btrfs_device *dev, int is_dev_replace)
 393 {
 394         struct scrub_ctx *sctx;
 395         int             i;
 396         struct btrfs_fs_info *fs_info = dev->dev_root->fs_info;
 397         int pages_per_rd_bio;
 398         int ret;
 399
 400         /*
 401          * the setting of pages_per_rd_bio is correct for scrub but might
 402          * be wrong for the dev_replace code where we might read from
 403          * different devices in the initial huge bios. However, that
 404          * code is able to correctly handle the case when adding a page
 405          * to a bio fails.
 406          */
 407         if (dev->bdev)
 408                 pages_per_rd_bio = min_t(int, SCRUB_PAGES_PER_RD_BIO,
 409                                          bio_get_nr_vecs(dev->bdev));
 410         else
 411                 pages_per_rd_bio = SCRUB_PAGES_PER_RD_BIO;
 412         sctx = kzalloc(sizeof(*sctx), GFP_NOFS);
 413         if (!sctx)
 414                 goto nomem;
 415         sctx->is_dev_replace = is_dev_replace;
 416         sctx->pages_per_rd_bio = pages_per_rd_bio;
 417         sctx->curr = -1;
 418         sctx->dev_root = dev->dev_root;
 419         for (i = 0; i < SCRUB_BIOS_PER_SCTX; ++i) {
 420                 struct scrub_bio *sbio;
 421
 422                 sbio = kzalloc(sizeof(*sbio), GFP_NOFS);
 423                 if (!sbio)
 424                         goto nomem;
 425                 sctx->bios[i] = sbio;
 426
 427                 sbio->index = i;
 428                 sbio->sctx = sctx;
 429                 sbio->page_count = 0;
 430                 btrfs_init_work(&sbio->work, btrfs_scrub_helper,
 431                                 scrub_bio_end_io_worker, NULL, NULL);
 432
 433                 if (i != SCRUB_BIOS_PER_SCTX - 1)
 434                         sctx->bios[i]->next_free = i + 1;
 435                 else
 436                         sctx->bios[i]->next_free = -1;
 437         }
 438         sctx->first_free = 0;
 439         sctx->nodesize = dev->dev_root->nodesize;
 440         sctx->sectorsize = dev->dev_root->sectorsize;
 441         atomic_set(&sctx->bios_in_flight, 0);
 442         atomic_set(&sctx->workers_pending, 0);
 443         atomic_set(&sctx->cancel_req, 0);
 444         sctx->csum_size = btrfs_super_csum_size(fs_info->super_copy);
 445         INIT_LIST_HEAD(&sctx->csum_list);
 446
 447         spin_lock_init(&sctx->list_lock);
 448         spin_lock_init(&sctx->stat_lock);
 449         init_waitqueue_head(&sctx->list_wait);
 450
 451         ret = scrub_setup_wr_ctx(sctx, &sctx->wr_ctx, fs_info,
 452                                  fs_info->dev_replace.tgtdev, is_dev_replace);
 453         if (ret) {
 454                 scrub_free_ctx(sctx);
 455                 return ERR_PTR(ret);
 456         }
 457         return sctx;
 458
 459 nomem:
 460         scrub_free_ctx(sctx);
 461         return ERR_PTR(-ENOMEM);
 462 }
 463
 464 static int scrub_print_warning_inode(u64 inum, u64 offset, u64 root,
 465                                      void *warn_ctx)
 466 {
 467         u64 isize;
 468         u32 nlink;
 469         int ret;
 470         int i;
 471         struct extent_buffer *eb;
 472         struct btrfs_inode_item *inode_item;
 473         struct scrub_warning *swarn = warn_ctx;
 474         struct btrfs_fs_info *fs_info = swarn->dev->dev_root->fs_info;
 475         struct inode_fs_paths *ipath = NULL;
 476         struct btrfs_root *local_root;
 477         struct btrfs_key root_key;
 478
 479         root_key.objectid = root;
 480         root_key.type = BTRFS_ROOT_ITEM_KEY;
 481         root_key.offset = (u64)-1;
 482         local_root = btrfs_read_fs_root_no_name(fs_info, &root_key);
 483         if (IS_ERR(local_root)) {
 484                 ret = PTR_ERR(local_root);
 485                 goto err;
 486         }
 487
 488         ret = inode_item_info(inum, 0, local_root, swarn->path);
 489         if (ret) {
 490                 btrfs_release_path(swarn->path);
 491                 goto err;
 492         }
 493
 494         eb = swarn->path->nodes[0];
 495         inode_item = btrfs_item_ptr(eb, swarn->path->slots[0],
 496                                         struct btrfs_inode_item);
 497         isize = btrfs_inode_size(eb, inode_item);
 498         nlink = btrfs_inode_nlink(eb, inode_item);
 499         btrfs_release_path(swarn->path);
 500
 501         ipath = init_ipath(4096, local_root, swarn->path);
 502         if (IS_ERR(ipath)) {
 503                 ret = PTR_ERR(ipath);
 504                 ipath = NULL;
 505                 goto err;
 506         }
 507         ret = paths_from_inode(inum, ipath);
 508
 509         if (ret < 0)
 510                 goto err;
 511
 512         /*
 513          * we deliberately ignore the bit ipath might have been too small to
 514          * hold all of the paths here
 515          */
 516         for (i = 0; i < ipath->fspath->elem_cnt; ++i)
 517                 printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
 518                         "%s, sector %llu, root %llu, inode %llu, offset %llu, "
 519                         "length %llu, links %u (path: %s)\n", swarn->errstr,
 520                         swarn->logical, rcu_str_deref(swarn->dev->name),
 521                         (unsigned long long)swarn->sector, root, inum, offset,
 522                         min(isize - offset, (u64)PAGE_SIZE), nlink,
 523                         (char *)(unsigned long)ipath->fspath->val[i]);
 524
 525         free_ipath(ipath);
 526         return 0;
 527
 528 err:
 529         printk_in_rcu(KERN_WARNING "BTRFS: %s at logical %llu on dev "
 530                 "%s, sector %llu, root %llu, inode %llu, offset %llu: path "
 531                 "resolving failed with ret=%d\n", swarn->errstr,
 532                 swarn->logical, rcu_str_deref(swarn->dev->name),
 533                 (unsigned long long)swarn->sector, root, inum, offset, ret);
 534
 535         free_ipath(ipath);
 536         return 0;
 537 }
 538
 539 static void scrub_print_warning(const char *errstr, struct scrub_block *sblock)
 540 {
 541         struct btrfs_device *dev;
 542         struct btrfs_fs_info *fs_info;
 543         struct btrfs_path *path;
 544         struct btrfs_key found_key;
 545         struct extent_buffer *eb;
 546         struct btrfs_extent_item *ei;
 547         struct scrub_warning swarn;
 548         unsigned long ptr = 0;
 549         u64 extent_item_pos;
 550         u64 flags = 0;
 551         u64 ref_root;
 552         u32 item_size;
 553         u8 ref_level;
 554         const int bufsize = 4096;
 555         int ret;
 556
 557         WARN_ON(sblock->page_count < 1);
 558         dev = sblock->pagev[0]->dev;
 559         fs_info = sblock->sctx->dev_root->fs_info;
 560
 561         path = btrfs_alloc_path();
 562
 563         swarn.scratch_buf = kmalloc(bufsize, GFP_NOFS);
 564         swarn.msg_buf = kmalloc(bufsize, GFP_NOFS);
 565         swarn.sector = (sblock->pagev[0]->physical) >> 9;
 566         swarn.logical = sblock->pagev[0]->logical;
 567         swarn.errstr = errstr;
 568         swarn.dev = NULL;
 569         swarn.msg_bufsize = bufsize;
 570         swarn.scratch_bufsize = bufsize;
 571
 572         if (!path || !swarn.scratch_buf || !swarn.msg_buf)
 573                 goto out;
 574
 575         ret = extent_from_logical(fs_info, swarn.logical, path, &found_key,
 576                                   &flags);
 577         if (ret < 0)
 578                 goto out;
 579
 580         extent_item_pos = swarn.logical - found_key.objectid;
 581         swarn.extent_item_size = found_key.offset;
 582
 583         eb = path->nodes[0];
 584         ei = btrfs_item_ptr(eb, path->slots[0], struct btrfs_extent_item);
 585         item_size = btrfs_item_size_nr(eb, path->slots[0]);
 586
 587         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
 588                 do {
 589                         ret = tree_backref_for_extent(&ptr, eb, &found_key, ei,
 590                                                       item_size, &ref_root,
 591                                                       &ref_level);
 592                         printk_in_rcu(KERN_WARNING
 593                                 "BTRFS: %s at logical %llu on dev %s, "
 594                                 "sector %llu: metadata %s (level %d) in tree "
 595                                 "%llu\n", errstr, swarn.logical,
 596                                 rcu_str_deref(dev->name),
 597                                 (unsigned long long)swarn.sector,
 598                                 ref_level ? "node" : "leaf",
 599                                 ret < 0 ? -1 : ref_level,
 600                                 ret < 0 ? -1 : ref_root);
 601                 } while (ret != 1);
 602                 btrfs_release_path(path);
 603         } else {
 604                 btrfs_release_path(path);
 605                 swarn.path = path;
 606                 swarn.dev = dev;
 607                 iterate_extent_inodes(fs_info, found_key.objectid,
 608                                         extent_item_pos, 1,
 609                                         scrub_print_warning_inode, &swarn);
 610         }
 611
 612 out:
 613         btrfs_free_path(path);
 614         kfree(swarn.scratch_buf);
 615         kfree(swarn.msg_buf);
 616 }
 617
 618 static int scrub_fixup_readpage(u64 inum, u64 offset, u64 root, void *fixup_ctx)
 619 {
 620         struct page *page = NULL;
 621         unsigned long index;
 622         struct scrub_fixup_nodatasum *fixup = fixup_ctx;
 623         int ret;
 624         int corrected = 0;
 625         struct btrfs_key key;
 626         struct inode *inode = NULL;
 627         struct btrfs_fs_info *fs_info;
 628         u64 end = offset + PAGE_SIZE - 1;
 629         struct btrfs_root *local_root;
 630         int srcu_index;
 631
 632         key.objectid = root;
 633         key.type = BTRFS_ROOT_ITEM_KEY;
 634         key.offset = (u64)-1;
 635
 636         fs_info = fixup->root->fs_info;
 637         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
 638
 639         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
 640         if (IS_ERR(local_root)) {
 641                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 642                 return PTR_ERR(local_root);
 643         }
 644
 645         key.type = BTRFS_INODE_ITEM_KEY;
 646         key.objectid = inum;
 647         key.offset = 0;
 648         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
 649         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
 650         if (IS_ERR(inode))
 651                 return PTR_ERR(inode);
 652
 653         index = offset >> PAGE_CACHE_SHIFT;
 654
 655         page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
 656         if (!page) {
 657                 ret = -ENOMEM;
 658                 goto out;
 659         }
 660
 661         if (PageUptodate(page)) {
 662                 if (PageDirty(page)) {
 663                         /*
 664                          * we need to write the data to the defect sector. the
 665                          * data that was in that sector is not in memory,
 666                          * because the page was modified. we must not write the
 667                          * modified page to that sector.
 668                          *
 669                          * TODO: what could be done here: wait for the delalloc
 670                          *       runner to write out that page (might involve
 671                          *       COW) and see whether the sector is still
 672                          *       referenced afterwards.
 673                          *
 674                          * For the meantime, we'll treat this error
 675                          * incorrectable, although there is a chance that a
 676                          * later scrub will find the bad sector again and that
 677                          * there's no dirty page in memory, then.
 678                          */
 679                         ret = -EIO;
 680                         goto out;
 681                 }
 682                 ret = repair_io_failure(inode, offset, PAGE_SIZE,
 683                                         fixup->logical, page,
 684                                         offset - page_offset(page),
 685                                         fixup->mirror_num);
 686                 unlock_page(page);
 687                 corrected = !ret;
 688         } else {
 689                 /*
 690                  * we need to get good data first. the general readpage path
 691                  * will call repair_io_failure for us, we just have to make
 692                  * sure we read the bad mirror.
 693                  */
 694                 ret = set_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 695                                         EXTENT_DAMAGED, GFP_NOFS);
 696                 if (ret) {
 697                         /* set_extent_bits should give proper error */
 698                         WARN_ON(ret > 0);
 699                         if (ret > 0)
 700                                 ret = -EFAULT;
 701                         goto out;
 702                 }
 703
 704                 ret = extent_read_full_page(&BTRFS_I(inode)->io_tree, page,
 705                                                 btrfs_get_extent,
 706                                                 fixup->mirror_num);
 707                 wait_on_page_locked(page);
 708
 709                 corrected = !test_range_bit(&BTRFS_I(inode)->io_tree, offset,
 710                                                 end, EXTENT_DAMAGED, 0, NULL);
 711                 if (!corrected)
 712                         clear_extent_bits(&BTRFS_I(inode)->io_tree, offset, end,
 713                                                 EXTENT_DAMAGED, GFP_NOFS);
 714         }
 715
 716 out:
 717         if (page)
 718                 put_page(page);
 719
 720         iput(inode);
 721
 722         if (ret < 0)
 723                 return ret;
 724
 725         if (ret == 0 && corrected) {
 726                 /*
 727                  * we only need to call readpage for one of the inodes belonging
 728                  * to this extent. so make iterate_extent_inodes stop
 729                  */
 730                 return 1;
 731         }
 732
 733         return -EIO;
 734 }
 735
 736 static void scrub_fixup_nodatasum(struct btrfs_work *work)
 737 {
 738         int ret;
 739         struct scrub_fixup_nodatasum *fixup;
 740         struct scrub_ctx *sctx;
 741         struct btrfs_trans_handle *trans = NULL;
 742         struct btrfs_path *path;
 743         int uncorrectable = 0;
 744
 745         fixup = container_of(work, struct scrub_fixup_nodatasum, work);
 746         sctx = fixup->sctx;
 747
 748         path = btrfs_alloc_path();
 749         if (!path) {
 750                 spin_lock(&sctx->stat_lock);
 751                 ++sctx->stat.malloc_errors;
 752                 spin_unlock(&sctx->stat_lock);
 753                 uncorrectable = 1;
 754                 goto out;
 755         }
 756
 757         trans = btrfs_join_transaction(fixup->root);
 758         if (IS_ERR(trans)) {
 759                 uncorrectable = 1;
 760                 goto out;
 761         }
 762
 763         /*
 764          * the idea is to trigger a regular read through the standard path. we
 765          * read a page from the (failed) logical address by specifying the
 766          * corresponding copynum of the failed sector. thus, that readpage is
 767          * expected to fail.
 768          * that is the point where on-the-fly error correction will kick in
 769          * (once it's finished) and rewrite the failed sector if a good copy
 770          * can be found.
 771          */
 772         ret = iterate_inodes_from_logical(fixup->logical, fixup->root->fs_info,
 773                                                 path, scrub_fixup_readpage,
 774                                                 fixup);
 775         if (ret < 0) {
 776                 uncorrectable = 1;
 777                 goto out;
 778         }
 779         WARN_ON(ret != 1);
 780
 781         spin_lock(&sctx->stat_lock);
 782         ++sctx->stat.corrected_errors;
 783         spin_unlock(&sctx->stat_lock);
 784
 785 out:
 786         if (trans && !IS_ERR(trans))
 787                 btrfs_end_transaction(trans, fixup->root);
 788         if (uncorrectable) {
 789                 spin_lock(&sctx->stat_lock);
 790                 ++sctx->stat.uncorrectable_errors;
 791                 spin_unlock(&sctx->stat_lock);
 792                 btrfs_dev_replace_stats_inc(
 793                         &sctx->dev_root->fs_info->dev_replace.
 794                         num_uncorrectable_read_errors);
 795                 printk_ratelimited_in_rcu(KERN_ERR "BTRFS: "
 796                     "unable to fixup (nodatasum) error at logical %llu on dev %s\n",
 797                         fixup->logical, rcu_str_deref(fixup->dev->name));
 798         }
 799
 800         btrfs_free_path(path);
 801         kfree(fixup);
 802
 803         scrub_pending_trans_workers_dec(sctx);
 804 }
 805
 806 /*
 807  * scrub_handle_errored_block gets called when either verification of the
 808  * pages failed or the bio failed to read, e.g. with EIO. In the latter
 809  * case, this function handles all pages in the bio, even though only one
 810  * may be bad.
 811  * The goal of this function is to repair the errored block by using the
 812  * contents of one of the mirrors.
 813  */
 814 static int scrub_handle_errored_block(struct scrub_block *sblock_to_check)
 815 {
 816         struct scrub_ctx *sctx = sblock_to_check->sctx;
 817         struct btrfs_device *dev;
 818         struct btrfs_fs_info *fs_info;
 819         u64 length;
 820         u64 logical;
 821         u64 generation;
 822         unsigned int failed_mirror_index;
 823         unsigned int is_metadata;
 824         unsigned int have_csum;
 825         u8 *csum;
 826         struct scrub_block *sblocks_for_recheck; /* holds one for each mirror */
 827         struct scrub_block *sblock_bad;
 828         int ret;
 829         int mirror_index;
 830         int page_num;
 831         int success;
 832         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
 833                                       DEFAULT_RATELIMIT_BURST);
 834
 835         BUG_ON(sblock_to_check->page_count < 1);
 836         fs_info = sctx->dev_root->fs_info;
 837         if (sblock_to_check->pagev[0]->flags & BTRFS_EXTENT_FLAG_SUPER) {
 838                 /*
 839                  * if we find an error in a super block, we just report it.
 840                  * They will get written with the next transaction commit
 841                  * anyway
 842                  */
 843                 spin_lock(&sctx->stat_lock);
 844                 ++sctx->stat.super_errors;
 845                 spin_unlock(&sctx->stat_lock);
 846                 return 0;
 847         }
 848         length = sblock_to_check->page_count * PAGE_SIZE;
 849         logical = sblock_to_check->pagev[0]->logical;
 850         generation = sblock_to_check->pagev[0]->generation;
 851         BUG_ON(sblock_to_check->pagev[0]->mirror_num < 1);
 852         failed_mirror_index = sblock_to_check->pagev[0]->mirror_num - 1;
 853         is_metadata = !(sblock_to_check->pagev[0]->flags &
 854                         BTRFS_EXTENT_FLAG_DATA);
 855         have_csum = sblock_to_check->pagev[0]->have_csum;
 856         csum = sblock_to_check->pagev[0]->csum;
 857         dev = sblock_to_check->pagev[0]->dev;
 858
 859         if (sctx->is_dev_replace && !is_metadata && !have_csum) {
 860                 sblocks_for_recheck = NULL;
 861                 goto nodatasum_case;
 862         }
 863
 864         /*
 865          * read all mirrors one after the other. This includes to
 866          * re-read the extent or metadata block that failed (that was
 867          * the cause that this fixup code is called) another time,
 868          * page by page this time in order to know which pages
 869          * caused I/O errors and which ones are good (for all mirrors).
 870          * It is the goal to handle the situation when more than one
 871          * mirror contains I/O errors, but the errors do not
 872          * overlap, i.e. the data can be repaired by selecting the
 873          * pages from those mirrors without I/O error on the
 874          * particular pages. One example (with blocks >= 2 * PAGE_SIZE)
 875          * would be that mirror #1 has an I/O error on the first page,
 876          * the second page is good, and mirror #2 has an I/O error on
 877          * the second page, but the first page is good.
 878          * Then the first page of the first mirror can be repaired by
 879          * taking the first page of the second mirror, and the
 880          * second page of the second mirror can be repaired by
 881          * copying the contents of the 2nd page of the 1st mirror.
 882          * One more note: if the pages of one mirror contain I/O
 883          * errors, the checksum cannot be verified. In order to get
 884          * the best data for repairing, the first attempt is to find
 885          * a mirror without I/O errors and with a validated checksum.
 886          * Only if this is not possible, the pages are picked from
 887          * mirrors with I/O errors without considering the checksum.
 888          * If the latter is the case, at the end, the checksum of the
 889          * repaired area is verified in order to correctly maintain
 890          * the statistics.
 891          */
 892
 893         sblocks_for_recheck = kzalloc(BTRFS_MAX_MIRRORS *
 894                                      sizeof(*sblocks_for_recheck),
 895                                      GFP_NOFS);
 896         if (!sblocks_for_recheck) {
 897                 spin_lock(&sctx->stat_lock);
 898                 sctx->stat.malloc_errors++;
 899                 sctx->stat.read_errors++;
 900                 sctx->stat.uncorrectable_errors++;
 901                 spin_unlock(&sctx->stat_lock);
 902                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 903                 goto out;
 904         }
 905
 906         /* setup the context, map the logical blocks and alloc the pages */
 907         ret = scrub_setup_recheck_block(sctx, fs_info, sblock_to_check, length,
 908                                         logical, sblocks_for_recheck);
 909         if (ret) {
 910                 spin_lock(&sctx->stat_lock);
 911                 sctx->stat.read_errors++;
 912                 sctx->stat.uncorrectable_errors++;
 913                 spin_unlock(&sctx->stat_lock);
 914                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 915                 goto out;
 916         }
 917         BUG_ON(failed_mirror_index >= BTRFS_MAX_MIRRORS);
 918         sblock_bad = sblocks_for_recheck + failed_mirror_index;
 919
 920         /* build and submit the bios for the failed mirror, check checksums */
 921         scrub_recheck_block(fs_info, sblock_bad, is_metadata, have_csum,
 922                             csum, generation, sctx->csum_size);
 923
 924         if (!sblock_bad->header_error && !sblock_bad->checksum_error &&
 925             sblock_bad->no_io_error_seen) {
 926                 /*
 927                  * the error disappeared after reading page by page, or
 928                  * the area was part of a huge bio and other parts of the
 929                  * bio caused I/O errors, or the block layer merged several
 930                  * read requests into one and the error is caused by a
 931                  * different bio (usually one of the two latter cases is
 932                  * the cause)
 933                  */
 934                 spin_lock(&sctx->stat_lock);
 935                 sctx->stat.unverified_errors++;
 936                 spin_unlock(&sctx->stat_lock);
 937
 938                 if (sctx->is_dev_replace)
 939                         scrub_write_block_to_dev_replace(sblock_bad);
 940                 goto out;
 941         }
 942
 943         if (!sblock_bad->no_io_error_seen) {
 944                 spin_lock(&sctx->stat_lock);
 945                 sctx->stat.read_errors++;
 946                 spin_unlock(&sctx->stat_lock);
 947                 if (__ratelimit(&_rs))
 948                         scrub_print_warning("i/o error", sblock_to_check);
 949                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_READ_ERRS);
 950         } else if (sblock_bad->checksum_error) {
 951                 spin_lock(&sctx->stat_lock);
 952                 sctx->stat.csum_errors++;
 953                 spin_unlock(&sctx->stat_lock);
 954                 if (__ratelimit(&_rs))
 955                         scrub_print_warning("checksum error", sblock_to_check);
 956                 btrfs_dev_stat_inc_and_print(dev,
 957                                              BTRFS_DEV_STAT_CORRUPTION_ERRS);
 958         } else if (sblock_bad->header_error) {
 959                 spin_lock(&sctx->stat_lock);
 960                 sctx->stat.verify_errors++;
 961                 spin_unlock(&sctx->stat_lock);
 962                 if (__ratelimit(&_rs))
 963                         scrub_print_warning("checksum/header error",
 964                                             sblock_to_check);
 965                 if (sblock_bad->generation_error)
 966                         btrfs_dev_stat_inc_and_print(dev,
 967                                 BTRFS_DEV_STAT_GENERATION_ERRS);
 968                 else
 969                         btrfs_dev_stat_inc_and_print(dev,
 970                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
 971         }
 972
 973         if (sctx->readonly) {
 974                 ASSERT(!sctx->is_dev_replace);
 975                 goto out;
 976         }
 977
 978         if (!is_metadata && !have_csum) {
 979                 struct scrub_fixup_nodatasum *fixup_nodatasum;
 980
 981 nodatasum_case:
 982                 WARN_ON(sctx->is_dev_replace);
 983
 984                 /*
 985                  * !is_metadata and !have_csum, this means that the data
 986                  * might not be COW'ed, that it might be modified
 987                  * concurrently. The general strategy to work on the
 988                  * commit root does not help in the case when COW is not
 989                  * used.
 990                  */
 991                 fixup_nodatasum = kzalloc(sizeof(*fixup_nodatasum), GFP_NOFS);
 992                 if (!fixup_nodatasum)
 993                         goto did_not_correct_error;
 994                 fixup_nodatasum->sctx = sctx;
 995                 fixup_nodatasum->dev = dev;
 996                 fixup_nodatasum->logical = logical;
 997                 fixup_nodatasum->root = fs_info->extent_root;
 998                 fixup_nodatasum->mirror_num = failed_mirror_index + 1;
 999                 scrub_pending_trans_workers_inc(sctx);
1000                 btrfs_init_work(&fixup_nodatasum->work, btrfs_scrub_helper,
1001                                 scrub_fixup_nodatasum, NULL, NULL);
1002                 btrfs_queue_work(fs_info->scrub_workers,
1003                                  &fixup_nodatasum->work);
1004                 goto out;
1005         }
1006
1007         /*
1008          * now build and submit the bios for the other mirrors, check
1009          * checksums.
1010          * First try to pick the mirror which is completely without I/O
1011          * errors and also does not have a checksum error.
1012          * If one is found, and if a checksum is present, the full block
1013          * that is known to contain an error is rewritten. Afterwards
1014          * the block is known to be corrected.
1015          * If a mirror is found which is completely correct, and no
1016          * checksum is present, only those pages are rewritten that had
1017          * an I/O error in the block to be repaired, since it cannot be
1018          * determined, which copy of the other pages is better (and it
1019          * could happen otherwise that a correct page would be
1020          * overwritten by a bad one).
1021          */
1022         for (mirror_index = 0;
1023              mirror_index < BTRFS_MAX_MIRRORS &&
1024              sblocks_for_recheck[mirror_index].page_count > 0;
1025              mirror_index++) {
1026                 struct scrub_block *sblock_other;
1027
1028                 if (mirror_index == failed_mirror_index)
1029                         continue;
1030                 sblock_other = sblocks_for_recheck + mirror_index;
1031
1032                 /* build and submit the bios, check checksums */
1033                 scrub_recheck_block(fs_info, sblock_other, is_metadata,
1034                                     have_csum, csum, generation,
1035                                     sctx->csum_size);
1036
1037                 if (!sblock_other->header_error &&
1038                     !sblock_other->checksum_error &&
1039                     sblock_other->no_io_error_seen) {
1040                         if (sctx->is_dev_replace) {
1041                                 scrub_write_block_to_dev_replace(sblock_other);
1042                         } else {
1043                                 int force_write = is_metadata || have_csum;
1044
1045                                 ret = scrub_repair_block_from_good_copy(
1046                                                 sblock_bad, sblock_other,
1047                                                 force_write);
1048                         }
1049                         if (0 == ret)
1050                                 goto corrected_error;
1051                 }
1052         }
1053
1054         /*
1055          * for dev_replace, pick good pages and write to the target device.
1056          */
1057         if (sctx->is_dev_replace) {
1058                 success = 1;
1059                 for (page_num = 0; page_num < sblock_bad->page_count;
1060                      page_num++) {
1061                         int sub_success;
1062
1063                         sub_success = 0;
1064                         for (mirror_index = 0;
1065                              mirror_index < BTRFS_MAX_MIRRORS &&
1066                              sblocks_for_recheck[mirror_index].page_count > 0;
1067                              mirror_index++) {
1068                                 struct scrub_block *sblock_other =
1069                                         sblocks_for_recheck + mirror_index;
1070                                 struct scrub_page *page_other =
1071                                         sblock_other->pagev[page_num];
1072
1073                                 if (!page_other->io_error) {
1074                                         ret = scrub_write_page_to_dev_replace(
1075                                                         sblock_other, page_num);
1076                                         if (ret == 0) {
1077                                                 /* succeeded for this page */
1078                                                 sub_success = 1;
1079                                                 break;
1080                                         } else {
1081                                                 btrfs_dev_replace_stats_inc(
1082                                                         &sctx->dev_root->
1083                                                         fs_info->dev_replace.
1084                                                         num_write_errors);
1085                                         }
1086                                 }
1087                         }
1088
1089                         if (!sub_success) {
1090                                 /*
1091                                  * did not find a mirror to fetch the page
1092                                  * from. scrub_write_page_to_dev_replace()
1093                                  * handles this case (page->io_error), by
1094                                  * filling the block with zeros before
1095                                  * submitting the write request
1096                                  */
1097                                 success = 0;
1098                                 ret = scrub_write_page_to_dev_replace(
1099                                                 sblock_bad, page_num);
1100                                 if (ret)
1101                                         btrfs_dev_replace_stats_inc(
1102                                                 &sctx->dev_root->fs_info->
1103                                                 dev_replace.num_write_errors);
1104                         }
1105                 }
1106
1107                 goto out;
1108         }
1109
1110         /*
1111          * for regular scrub, repair those pages that are errored.
1112          * In case of I/O errors in the area that is supposed to be
1113          * repaired, continue by picking good copies of those pages.
1114          * Select the good pages from mirrors to rewrite bad pages from
1115          * the area to fix. Afterwards verify the checksum of the block
1116          * that is supposed to be repaired. This verification step is
1117          * only done for the purpose of statistic counting and for the
1118          * final scrub report, whether errors remain.
1119          * A perfect algorithm could make use of the checksum and try
1120          * all possible combinations of pages from the different mirrors
1121          * until the checksum verification succeeds. For example, when
1122          * the 2nd page of mirror #1 faces I/O errors, and the 2nd page
1123          * of mirror #2 is readable but the final checksum test fails,
1124          * then the 2nd page of mirror #3 could be tried, whether now
1125          * the final checksum succeedes. But this would be a rare
1126          * exception and is therefore not implemented. At least it is
1127          * avoided that the good copy is overwritten.
1128          * A more useful improvement would be to pick the sectors
1129          * without I/O error based on sector sizes (512 bytes on legacy
1130          * disks) instead of on PAGE_SIZE. Then maybe 512 byte of one
1131          * mirror could be repaired by taking 512 byte of a different
1132          * mirror, even if other 512 byte sectors in the same PAGE_SIZE
1133          * area are unreadable.
1134          */
1135
1136         /* can only fix I/O errors from here on */
1137         if (sblock_bad->no_io_error_seen)
1138                 goto did_not_correct_error;
1139
1140         success = 1;
1141         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1142                 struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1143
1144                 if (!page_bad->io_error)
1145                         continue;
1146
1147                 for (mirror_index = 0;
1148                      mirror_index < BTRFS_MAX_MIRRORS &&
1149                      sblocks_for_recheck[mirror_index].page_count > 0;
1150                      mirror_index++) {
1151                         struct scrub_block *sblock_other = sblocks_for_recheck +
1152                                                            mirror_index;
1153                         struct scrub_page *page_other = sblock_other->pagev[
1154                                                         page_num];
1155
1156                         if (!page_other->io_error) {
1157                                 ret = scrub_repair_page_from_good_copy(
1158                                         sblock_bad, sblock_other, page_num, 0);
1159                                 if (0 == ret) {
1160                                         page_bad->io_error = 0;
1161                                         break; /* succeeded for this page */
1162                                 }
1163                         }
1164                 }
1165
1166                 if (page_bad->io_error) {
1167                         /* did not find a mirror to copy the page from */
1168                         success = 0;
1169                 }
1170         }
1171
1172         if (success) {
1173                 if (is_metadata || have_csum) {
1174                         /*
1175                          * need to verify the checksum now that all
1176                          * sectors on disk are repaired (the write
1177                          * request for data to be repaired is on its way).
1178                          * Just be lazy and use scrub_recheck_block()
1179                          * which re-reads the data before the checksum
1180                          * is verified, but most likely the data comes out
1181                          * of the page cache.
1182                          */
1183                         scrub_recheck_block(fs_info, sblock_bad,
1184                                             is_metadata, have_csum, csum,
1185                                             generation, sctx->csum_size);
1186                         if (!sblock_bad->header_error &&
1187                             !sblock_bad->checksum_error &&
1188                             sblock_bad->no_io_error_seen)
1189                                 goto corrected_error;
1190                         else
1191                                 goto did_not_correct_error;
1192                 } else {
1193 corrected_error:
1194                         spin_lock(&sctx->stat_lock);
1195                         sctx->stat.corrected_errors++;
1196                         spin_unlock(&sctx->stat_lock);
1197                         printk_ratelimited_in_rcu(KERN_ERR
1198                                 "BTRFS: fixed up error at logical %llu on dev %s\n",
1199                                 logical, rcu_str_deref(dev->name));
1200                 }
1201         } else {
1202 did_not_correct_error:
1203                 spin_lock(&sctx->stat_lock);
1204                 sctx->stat.uncorrectable_errors++;
1205                 spin_unlock(&sctx->stat_lock);
1206                 printk_ratelimited_in_rcu(KERN_ERR
1207                         "BTRFS: unable to fixup (regular) error at logical %llu on dev %s\n",
1208                         logical, rcu_str_deref(dev->name));
1209         }
1210
1211 out:
1212         if (sblocks_for_recheck) {
1213                 for (mirror_index = 0; mirror_index < BTRFS_MAX_MIRRORS;
1214                      mirror_index++) {
1215                         struct scrub_block *sblock = sblocks_for_recheck +
1216                                                      mirror_index;
1217                         int page_index;
1218
1219                         for (page_index = 0; page_index < sblock->page_count;
1220                              page_index++) {
1221                                 sblock->pagev[page_index]->sblock = NULL;
1222                                 scrub_page_put(sblock->pagev[page_index]);
1223                         }
1224                 }
1225                 kfree(sblocks_for_recheck);
1226         }
1227
1228         return 0;
1229 }
1230
1231 static int scrub_setup_recheck_block(struct scrub_ctx *sctx,
1232                                      struct btrfs_fs_info *fs_info,
1233                                      struct scrub_block *original_sblock,
1234                                      u64 length, u64 logical,
1235                                      struct scrub_block *sblocks_for_recheck)
1236 {
1237         int page_index;
1238         int mirror_index;
1239         int ret;
1240
1241         /*
1242          * note: the two members ref_count and outstanding_pages
1243          * are not used (and not set) in the blocks that are used for
1244          * the recheck procedure
1245          */
1246
1247         page_index = 0;
1248         while (length > 0) {
1249                 u64 sublen = min_t(u64, length, PAGE_SIZE);
1250                 u64 mapped_length = sublen;
1251                 struct btrfs_bio *bbio = NULL;
1252
1253                 /*
1254                  * with a length of PAGE_SIZE, each returned stripe
1255                  * represents one mirror
1256                  */
1257                 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical,
1258                                       &mapped_length, &bbio, 0);
1259                 if (ret || !bbio || mapped_length < sublen) {
1260                         kfree(bbio);
1261                         return -EIO;
1262                 }
1263
1264                 BUG_ON(page_index >= SCRUB_PAGES_PER_RD_BIO);
1265                 for (mirror_index = 0; mirror_index < (int)bbio->num_stripes;
1266                      mirror_index++) {
1267                         struct scrub_block *sblock;
1268                         struct scrub_page *page;
1269
1270                         if (mirror_index >= BTRFS_MAX_MIRRORS)
1271                                 continue;
1272
1273                         sblock = sblocks_for_recheck + mirror_index;
1274                         sblock->sctx = sctx;
1275                         page = kzalloc(sizeof(*page), GFP_NOFS);
1276                         if (!page) {
1277 leave_nomem:
1278                                 spin_lock(&sctx->stat_lock);
1279                                 sctx->stat.malloc_errors++;
1280                                 spin_unlock(&sctx->stat_lock);
1281                                 kfree(bbio);
1282                                 return -ENOMEM;
1283                         }
1284                         scrub_page_get(page);
1285                         sblock->pagev[page_index] = page;
1286                         page->logical = logical;
1287                         page->physical = bbio->stripes[mirror_index].physical;
1288                         BUG_ON(page_index >= original_sblock->page_count);
1289                         page->physical_for_dev_replace =
1290                                 original_sblock->pagev[page_index]->
1291                                 physical_for_dev_replace;
1292                         /* for missing devices, dev->bdev is NULL */
1293                         page->dev = bbio->stripes[mirror_index].dev;
1294                         page->mirror_num = mirror_index + 1;
1295                         sblock->page_count++;
1296                         page->page = alloc_page(GFP_NOFS);
1297                         if (!page->page)
1298                                 goto leave_nomem;
1299                 }
1300                 kfree(bbio);
1301                 length -= sublen;
1302                 logical += sublen;
1303                 page_index++;
1304         }
1305
1306         return 0;
1307 }
1308
1309 /*
1310  * this function will check the on disk data for checksum errors, header
1311  * errors and read I/O errors. If any I/O errors happen, the exact pages
1312  * which are errored are marked as being bad. The goal is to enable scrub
1313  * to take those pages that are not errored from all the mirrors so that
1314  * the pages that are errored in the just handled mirror can be repaired.
1315  */
1316 static void scrub_recheck_block(struct btrfs_fs_info *fs_info,
1317                                 struct scrub_block *sblock, int is_metadata,
1318                                 int have_csum, u8 *csum, u64 generation,
1319                                 u16 csum_size)
1320 {
1321         int page_num;
1322
1323         sblock->no_io_error_seen = 1;
1324         sblock->header_error = 0;
1325         sblock->checksum_error = 0;
1326
1327         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1328                 struct bio *bio;
1329                 struct scrub_page *page = sblock->pagev[page_num];
1330
1331                 if (page->dev->bdev == NULL) {
1332                         page->io_error = 1;
1333                         sblock->no_io_error_seen = 0;
1334                         continue;
1335                 }
1336
1337                 WARN_ON(!page->page);
1338                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1339                 if (!bio) {
1340                         page->io_error = 1;
1341                         sblock->no_io_error_seen = 0;
1342                         continue;
1343                 }
1344                 bio->bi_bdev = page->dev->bdev;
1345                 bio->bi_iter.bi_sector = page->physical >> 9;
1346
1347                 bio_add_page(bio, page->page, PAGE_SIZE, 0);
1348                 if (btrfsic_submit_bio_wait(READ, bio))
1349                         sblock->no_io_error_seen = 0;
1350
1351                 bio_put(bio);
1352         }
1353
1354         if (sblock->no_io_error_seen)
1355                 scrub_recheck_block_checksum(fs_info, sblock, is_metadata,
1356                                              have_csum, csum, generation,
1357                                              csum_size);
1358
1359         return;
1360 }
1361
1362 static inline int scrub_check_fsid(u8 fsid[],
1363                                    struct scrub_page *spage)
1364 {
1365         struct btrfs_fs_devices *fs_devices = spage->dev->fs_devices;
1366         int ret;
1367
1368         ret = memcmp(fsid, fs_devices->fsid, BTRFS_UUID_SIZE);
1369         return !ret;
1370 }
1371
1372 static void scrub_recheck_block_checksum(struct btrfs_fs_info *fs_info,
1373                                          struct scrub_block *sblock,
1374                                          int is_metadata, int have_csum,
1375                                          const u8 *csum, u64 generation,
1376                                          u16 csum_size)
1377 {
1378         int page_num;
1379         u8 calculated_csum[BTRFS_CSUM_SIZE];
1380         u32 crc = ~(u32)0;
1381         void *mapped_buffer;
1382
1383         WARN_ON(!sblock->pagev[0]->page);
1384         if (is_metadata) {
1385                 struct btrfs_header *h;
1386
1387                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1388                 h = (struct btrfs_header *)mapped_buffer;
1389
1390                 if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h) ||
1391                     !scrub_check_fsid(h->fsid, sblock->pagev[0]) ||
1392                     memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1393                            BTRFS_UUID_SIZE)) {
1394                         sblock->header_error = 1;
1395                 } else if (generation != btrfs_stack_header_generation(h)) {
1396                         sblock->header_error = 1;
1397                         sblock->generation_error = 1;
1398                 }
1399                 csum = h->csum;
1400         } else {
1401                 if (!have_csum)
1402                         return;
1403
1404                 mapped_buffer = kmap_atomic(sblock->pagev[0]->page);
1405         }
1406
1407         for (page_num = 0;;) {
1408                 if (page_num == 0 && is_metadata)
1409                         crc = btrfs_csum_data(
1410                                 ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE,
1411                                 crc, PAGE_SIZE - BTRFS_CSUM_SIZE);
1412                 else
1413                         crc = btrfs_csum_data(mapped_buffer, crc, PAGE_SIZE);
1414
1415                 kunmap_atomic(mapped_buffer);
1416                 page_num++;
1417                 if (page_num >= sblock->page_count)
1418                         break;
1419                 WARN_ON(!sblock->pagev[page_num]->page);
1420
1421                 mapped_buffer = kmap_atomic(sblock->pagev[page_num]->page);
1422         }
1423
1424         btrfs_csum_final(crc, calculated_csum);
1425         if (memcmp(calculated_csum, csum, csum_size))
1426                 sblock->checksum_error = 1;
1427 }
1428
1429 static int scrub_repair_block_from_good_copy(struct scrub_block *sblock_bad,
1430                                              struct scrub_block *sblock_good,
1431                                              int force_write)
1432 {
1433         int page_num;
1434         int ret = 0;
1435
1436         for (page_num = 0; page_num < sblock_bad->page_count; page_num++) {
1437                 int ret_sub;
1438
1439                 ret_sub = scrub_repair_page_from_good_copy(sblock_bad,
1440                                                            sblock_good,
1441                                                            page_num,
1442                                                            force_write);
1443                 if (ret_sub)
1444                         ret = ret_sub;
1445         }
1446
1447         return ret;
1448 }
1449
1450 static int scrub_repair_page_from_good_copy(struct scrub_block *sblock_bad,
1451                                             struct scrub_block *sblock_good,
1452                                             int page_num, int force_write)
1453 {
1454         struct scrub_page *page_bad = sblock_bad->pagev[page_num];
1455         struct scrub_page *page_good = sblock_good->pagev[page_num];
1456
1457         BUG_ON(page_bad->page == NULL);
1458         BUG_ON(page_good->page == NULL);
1459         if (force_write || sblock_bad->header_error ||
1460             sblock_bad->checksum_error || page_bad->io_error) {
1461                 struct bio *bio;
1462                 int ret;
1463
1464                 if (!page_bad->dev->bdev) {
1465                         printk_ratelimited(KERN_WARNING "BTRFS: "
1466                                 "scrub_repair_page_from_good_copy(bdev == NULL) "
1467                                 "is unexpected!\n");
1468                         return -EIO;
1469                 }
1470
1471                 bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
1472                 if (!bio)
1473                         return -EIO;
1474                 bio->bi_bdev = page_bad->dev->bdev;
1475                 bio->bi_iter.bi_sector = page_bad->physical >> 9;
1476
1477                 ret = bio_add_page(bio, page_good->page, PAGE_SIZE, 0);
1478                 if (PAGE_SIZE != ret) {
1479                         bio_put(bio);
1480                         return -EIO;
1481                 }
1482
1483                 if (btrfsic_submit_bio_wait(WRITE, bio)) {
1484                         btrfs_dev_stat_inc_and_print(page_bad->dev,
1485                                 BTRFS_DEV_STAT_WRITE_ERRS);
1486                         btrfs_dev_replace_stats_inc(
1487                                 &sblock_bad->sctx->dev_root->fs_info->
1488                                 dev_replace.num_write_errors);
1489                         bio_put(bio);
1490                         return -EIO;
1491                 }
1492                 bio_put(bio);
1493         }
1494
1495         return 0;
1496 }
1497
1498 static void scrub_write_block_to_dev_replace(struct scrub_block *sblock)
1499 {
1500         int page_num;
1501
1502         for (page_num = 0; page_num < sblock->page_count; page_num++) {
1503                 int ret;
1504
1505                 ret = scrub_write_page_to_dev_replace(sblock, page_num);
1506                 if (ret)
1507                         btrfs_dev_replace_stats_inc(
1508                                 &sblock->sctx->dev_root->fs_info->dev_replace.
1509                                 num_write_errors);
1510         }
1511 }
1512
1513 static int scrub_write_page_to_dev_replace(struct scrub_block *sblock,
1514                                            int page_num)
1515 {
1516         struct scrub_page *spage = sblock->pagev[page_num];
1517
1518         BUG_ON(spage->page == NULL);
1519         if (spage->io_error) {
1520                 void *mapped_buffer = kmap_atomic(spage->page);
1521
1522                 memset(mapped_buffer, 0, PAGE_CACHE_SIZE);
1523                 flush_dcache_page(spage->page);
1524                 kunmap_atomic(mapped_buffer);
1525         }
1526         return scrub_add_page_to_wr_bio(sblock->sctx, spage);
1527 }
1528
1529 static int scrub_add_page_to_wr_bio(struct scrub_ctx *sctx,
1530                                     struct scrub_page *spage)
1531 {
1532         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1533         struct scrub_bio *sbio;
1534         int ret;
1535
1536         mutex_lock(&wr_ctx->wr_lock);
1537 again:
1538         if (!wr_ctx->wr_curr_bio) {
1539                 wr_ctx->wr_curr_bio = kzalloc(sizeof(*wr_ctx->wr_curr_bio),
1540                                               GFP_NOFS);
1541                 if (!wr_ctx->wr_curr_bio) {
1542                         mutex_unlock(&wr_ctx->wr_lock);
1543                         return -ENOMEM;
1544                 }
1545                 wr_ctx->wr_curr_bio->sctx = sctx;
1546                 wr_ctx->wr_curr_bio->page_count = 0;
1547         }
1548         sbio = wr_ctx->wr_curr_bio;
1549         if (sbio->page_count == 0) {
1550                 struct bio *bio;
1551
1552                 sbio->physical = spage->physical_for_dev_replace;
1553                 sbio->logical = spage->logical;
1554                 sbio->dev = wr_ctx->tgtdev;
1555                 bio = sbio->bio;
1556                 if (!bio) {
1557                         bio = btrfs_io_bio_alloc(GFP_NOFS, wr_ctx->pages_per_wr_bio);
1558                         if (!bio) {
1559                                 mutex_unlock(&wr_ctx->wr_lock);
1560                                 return -ENOMEM;
1561                         }
1562                         sbio->bio = bio;
1563                 }
1564
1565                 bio->bi_private = sbio;
1566                 bio->bi_end_io = scrub_wr_bio_end_io;
1567                 bio->bi_bdev = sbio->dev->bdev;
1568                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1569                 sbio->err = 0;
1570         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1571                    spage->physical_for_dev_replace ||
1572                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1573                    spage->logical) {
1574                 scrub_wr_submit(sctx);
1575                 goto again;
1576         }
1577
1578         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1579         if (ret != PAGE_SIZE) {
1580                 if (sbio->page_count < 1) {
1581                         bio_put(sbio->bio);
1582                         sbio->bio = NULL;
1583                         mutex_unlock(&wr_ctx->wr_lock);
1584                         return -EIO;
1585                 }
1586                 scrub_wr_submit(sctx);
1587                 goto again;
1588         }
1589
1590         sbio->pagev[sbio->page_count] = spage;
1591         scrub_page_get(spage);
1592         sbio->page_count++;
1593         if (sbio->page_count == wr_ctx->pages_per_wr_bio)
1594                 scrub_wr_submit(sctx);
1595         mutex_unlock(&wr_ctx->wr_lock);
1596
1597         return 0;
1598 }
1599
1600 static void scrub_wr_submit(struct scrub_ctx *sctx)
1601 {
1602         struct scrub_wr_ctx *wr_ctx = &sctx->wr_ctx;
1603         struct scrub_bio *sbio;
1604
1605         if (!wr_ctx->wr_curr_bio)
1606                 return;
1607
1608         sbio = wr_ctx->wr_curr_bio;
1609         wr_ctx->wr_curr_bio = NULL;
1610         WARN_ON(!sbio->bio->bi_bdev);
1611         scrub_pending_bio_inc(sctx);
1612         /* process all writes in a single worker thread. Then the block layer
1613          * orders the requests before sending them to the driver which
1614          * doubled the write performance on spinning disks when measured
1615          * with Linux 3.5 */
1616         btrfsic_submit_bio(WRITE, sbio->bio);
1617 }
1618
1619 static void scrub_wr_bio_end_io(struct bio *bio, int err)
1620 {
1621         struct scrub_bio *sbio = bio->bi_private;
1622         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
1623
1624         sbio->err = err;
1625         sbio->bio = bio;
1626
1627         btrfs_init_work(&sbio->work, btrfs_scrubwrc_helper,
1628                          scrub_wr_bio_end_io_worker, NULL, NULL);
1629         btrfs_queue_work(fs_info->scrub_wr_completion_workers, &sbio->work);
1630 }
1631
1632 static void scrub_wr_bio_end_io_worker(struct btrfs_work *work)
1633 {
1634         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
1635         struct scrub_ctx *sctx = sbio->sctx;
1636         int i;
1637
1638         WARN_ON(sbio->page_count > SCRUB_PAGES_PER_WR_BIO);
1639         if (sbio->err) {
1640                 struct btrfs_dev_replace *dev_replace =
1641                         &sbio->sctx->dev_root->fs_info->dev_replace;
1642
1643                 for (i = 0; i < sbio->page_count; i++) {
1644                         struct scrub_page *spage = sbio->pagev[i];
1645
1646                         spage->io_error = 1;
1647                         btrfs_dev_replace_stats_inc(&dev_replace->
1648                                                     num_write_errors);
1649                 }
1650         }
1651
1652         for (i = 0; i < sbio->page_count; i++)
1653                 scrub_page_put(sbio->pagev[i]);
1654
1655         bio_put(sbio->bio);
1656         kfree(sbio);
1657         scrub_pending_bio_dec(sctx);
1658 }
1659
1660 static int scrub_checksum(struct scrub_block *sblock)
1661 {
1662         u64 flags;
1663         int ret;
1664
1665         WARN_ON(sblock->page_count < 1);
1666         flags = sblock->pagev[0]->flags;
1667         ret = 0;
1668         if (flags & BTRFS_EXTENT_FLAG_DATA)
1669                 ret = scrub_checksum_data(sblock);
1670         else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)
1671                 ret = scrub_checksum_tree_block(sblock);
1672         else if (flags & BTRFS_EXTENT_FLAG_SUPER)
1673                 (void)scrub_checksum_super(sblock);
1674         else
1675                 WARN_ON(1);
1676         if (ret)
1677                 scrub_handle_errored_block(sblock);
1678
1679         return ret;
1680 }
1681
1682 static int scrub_checksum_data(struct scrub_block *sblock)
1683 {
1684         struct scrub_ctx *sctx = sblock->sctx;
1685         u8 csum[BTRFS_CSUM_SIZE];
1686         u8 *on_disk_csum;
1687         struct page *page;
1688         void *buffer;
1689         u32 crc = ~(u32)0;
1690         int fail = 0;
1691         u64 len;
1692         int index;
1693
1694         BUG_ON(sblock->page_count < 1);
1695         if (!sblock->pagev[0]->have_csum)
1696                 return 0;
1697
1698         on_disk_csum = sblock->pagev[0]->csum;
1699         page = sblock->pagev[0]->page;
1700         buffer = kmap_atomic(page);
1701
1702         len = sctx->sectorsize;
1703         index = 0;
1704         for (;;) {
1705                 u64 l = min_t(u64, len, PAGE_SIZE);
1706
1707                 crc = btrfs_csum_data(buffer, crc, l);
1708                 kunmap_atomic(buffer);
1709                 len -= l;
1710                 if (len == 0)
1711                         break;
1712                 index++;
1713                 BUG_ON(index >= sblock->page_count);
1714                 BUG_ON(!sblock->pagev[index]->page);
1715                 page = sblock->pagev[index]->page;
1716                 buffer = kmap_atomic(page);
1717         }
1718
1719         btrfs_csum_final(crc, csum);
1720         if (memcmp(csum, on_disk_csum, sctx->csum_size))
1721                 fail = 1;
1722
1723         return fail;
1724 }
1725
1726 static int scrub_checksum_tree_block(struct scrub_block *sblock)
1727 {
1728         struct scrub_ctx *sctx = sblock->sctx;
1729         struct btrfs_header *h;
1730         struct btrfs_root *root = sctx->dev_root;
1731         struct btrfs_fs_info *fs_info = root->fs_info;
1732         u8 calculated_csum[BTRFS_CSUM_SIZE];
1733         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1734         struct page *page;
1735         void *mapped_buffer;
1736         u64 mapped_size;
1737         void *p;
1738         u32 crc = ~(u32)0;
1739         int fail = 0;
1740         int crc_fail = 0;
1741         u64 len;
1742         int index;
1743
1744         BUG_ON(sblock->page_count < 1);
1745         page = sblock->pagev[0]->page;
1746         mapped_buffer = kmap_atomic(page);
1747         h = (struct btrfs_header *)mapped_buffer;
1748         memcpy(on_disk_csum, h->csum, sctx->csum_size);
1749
1750         /*
1751          * we don't use the getter functions here, as we
1752          * a) don't have an extent buffer and
1753          * b) the page is already kmapped
1754          */
1755
1756         if (sblock->pagev[0]->logical != btrfs_stack_header_bytenr(h))
1757                 ++fail;
1758
1759         if (sblock->pagev[0]->generation != btrfs_stack_header_generation(h))
1760                 ++fail;
1761
1762         if (!scrub_check_fsid(h->fsid, sblock->pagev[0]))
1763                 ++fail;
1764
1765         if (memcmp(h->chunk_tree_uuid, fs_info->chunk_tree_uuid,
1766                    BTRFS_UUID_SIZE))
1767                 ++fail;
1768
1769         len = sctx->nodesize - BTRFS_CSUM_SIZE;
1770         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1771         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1772         index = 0;
1773         for (;;) {
1774                 u64 l = min_t(u64, len, mapped_size);
1775
1776                 crc = btrfs_csum_data(p, crc, l);
1777                 kunmap_atomic(mapped_buffer);
1778                 len -= l;
1779                 if (len == 0)
1780                         break;
1781                 index++;
1782                 BUG_ON(index >= sblock->page_count);
1783                 BUG_ON(!sblock->pagev[index]->page);
1784                 page = sblock->pagev[index]->page;
1785                 mapped_buffer = kmap_atomic(page);
1786                 mapped_size = PAGE_SIZE;
1787                 p = mapped_buffer;
1788         }
1789
1790         btrfs_csum_final(crc, calculated_csum);
1791         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1792                 ++crc_fail;
1793
1794         return fail || crc_fail;
1795 }
1796
1797 static int scrub_checksum_super(struct scrub_block *sblock)
1798 {
1799         struct btrfs_super_block *s;
1800         struct scrub_ctx *sctx = sblock->sctx;
1801         u8 calculated_csum[BTRFS_CSUM_SIZE];
1802         u8 on_disk_csum[BTRFS_CSUM_SIZE];
1803         struct page *page;
1804         void *mapped_buffer;
1805         u64 mapped_size;
1806         void *p;
1807         u32 crc = ~(u32)0;
1808         int fail_gen = 0;
1809         int fail_cor = 0;
1810         u64 len;
1811         int index;
1812
1813         BUG_ON(sblock->page_count < 1);
1814         page = sblock->pagev[0]->page;
1815         mapped_buffer = kmap_atomic(page);
1816         s = (struct btrfs_super_block *)mapped_buffer;
1817         memcpy(on_disk_csum, s->csum, sctx->csum_size);
1818
1819         if (sblock->pagev[0]->logical != btrfs_super_bytenr(s))
1820                 ++fail_cor;
1821
1822         if (sblock->pagev[0]->generation != btrfs_super_generation(s))
1823                 ++fail_gen;
1824
1825         if (!scrub_check_fsid(s->fsid, sblock->pagev[0]))
1826                 ++fail_cor;
1827
1828         len = BTRFS_SUPER_INFO_SIZE - BTRFS_CSUM_SIZE;
1829         mapped_size = PAGE_SIZE - BTRFS_CSUM_SIZE;
1830         p = ((u8 *)mapped_buffer) + BTRFS_CSUM_SIZE;
1831         index = 0;
1832         for (;;) {
1833                 u64 l = min_t(u64, len, mapped_size);
1834
1835                 crc = btrfs_csum_data(p, crc, l);
1836                 kunmap_atomic(mapped_buffer);
1837                 len -= l;
1838                 if (len == 0)
1839                         break;
1840                 index++;
1841                 BUG_ON(index >= sblock->page_count);
1842                 BUG_ON(!sblock->pagev[index]->page);
1843                 page = sblock->pagev[index]->page;
1844                 mapped_buffer = kmap_atomic(page);
1845                 mapped_size = PAGE_SIZE;
1846                 p = mapped_buffer;
1847         }
1848
1849         btrfs_csum_final(crc, calculated_csum);
1850         if (memcmp(calculated_csum, on_disk_csum, sctx->csum_size))
1851                 ++fail_cor;
1852
1853         if (fail_cor + fail_gen) {
1854                 /*
1855                  * if we find an error in a super block, we just report it.
1856                  * They will get written with the next transaction commit
1857                  * anyway
1858                  */
1859                 spin_lock(&sctx->stat_lock);
1860                 ++sctx->stat.super_errors;
1861                 spin_unlock(&sctx->stat_lock);
1862                 if (fail_cor)
1863                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1864                                 BTRFS_DEV_STAT_CORRUPTION_ERRS);
1865                 else
1866                         btrfs_dev_stat_inc_and_print(sblock->pagev[0]->dev,
1867                                 BTRFS_DEV_STAT_GENERATION_ERRS);
1868         }
1869
1870         return fail_cor + fail_gen;
1871 }
1872
1873 static void scrub_block_get(struct scrub_block *sblock)
1874 {
1875         atomic_inc(&sblock->ref_count);
1876 }
1877
1878 static void scrub_block_put(struct scrub_block *sblock)
1879 {
1880         if (atomic_dec_and_test(&sblock->ref_count)) {
1881                 int i;
1882
1883                 for (i = 0; i < sblock->page_count; i++)
1884                         scrub_page_put(sblock->pagev[i]);
1885                 kfree(sblock);
1886         }
1887 }
1888
1889 static void scrub_page_get(struct scrub_page *spage)
1890 {
1891         atomic_inc(&spage->ref_count);
1892 }
1893
1894 static void scrub_page_put(struct scrub_page *spage)
1895 {
1896         if (atomic_dec_and_test(&spage->ref_count)) {
1897                 if (spage->page)
1898                         __free_page(spage->page);
1899                 kfree(spage);
1900         }
1901 }
1902
1903 static void scrub_submit(struct scrub_ctx *sctx)
1904 {
1905         struct scrub_bio *sbio;
1906
1907         if (sctx->curr == -1)
1908                 return;
1909
1910         sbio = sctx->bios[sctx->curr];
1911         sctx->curr = -1;
1912         scrub_pending_bio_inc(sctx);
1913
1914         if (!sbio->bio->bi_bdev) {
1915                 /*
1916                  * this case should not happen. If btrfs_map_block() is
1917                  * wrong, it could happen for dev-replace operations on
1918                  * missing devices when no mirrors are available, but in
1919                  * this case it should already fail the mount.
1920                  * This case is handled correctly (but _very_ slowly).
1921                  */
1922                 printk_ratelimited(KERN_WARNING
1923                         "BTRFS: scrub_submit(bio bdev == NULL) is unexpected!\n");
1924                 bio_endio(sbio->bio, -EIO);
1925         } else {
1926                 btrfsic_submit_bio(READ, sbio->bio);
1927         }
1928 }
1929
1930 static int scrub_add_page_to_rd_bio(struct scrub_ctx *sctx,
1931                                     struct scrub_page *spage)
1932 {
1933         struct scrub_block *sblock = spage->sblock;
1934         struct scrub_bio *sbio;
1935         int ret;
1936
1937 again:
1938         /*
1939          * grab a fresh bio or wait for one to become available
1940          */
1941         while (sctx->curr == -1) {
1942                 spin_lock(&sctx->list_lock);
1943                 sctx->curr = sctx->first_free;
1944                 if (sctx->curr != -1) {
1945                         sctx->first_free = sctx->bios[sctx->curr]->next_free;
1946                         sctx->bios[sctx->curr]->next_free = -1;
1947                         sctx->bios[sctx->curr]->page_count = 0;
1948                         spin_unlock(&sctx->list_lock);
1949                 } else {
1950                         spin_unlock(&sctx->list_lock);
1951                         wait_event(sctx->list_wait, sctx->first_free != -1);
1952                 }
1953         }
1954         sbio = sctx->bios[sctx->curr];
1955         if (sbio->page_count == 0) {
1956                 struct bio *bio;
1957
1958                 sbio->physical = spage->physical;
1959                 sbio->logical = spage->logical;
1960                 sbio->dev = spage->dev;
1961                 bio = sbio->bio;
1962                 if (!bio) {
1963                         bio = btrfs_io_bio_alloc(GFP_NOFS, sctx->pages_per_rd_bio);
1964                         if (!bio)
1965                                 return -ENOMEM;
1966                         sbio->bio = bio;
1967                 }
1968
1969                 bio->bi_private = sbio;
1970                 bio->bi_end_io = scrub_bio_end_io;
1971                 bio->bi_bdev = sbio->dev->bdev;
1972                 bio->bi_iter.bi_sector = sbio->physical >> 9;
1973                 sbio->err = 0;
1974         } else if (sbio->physical + sbio->page_count * PAGE_SIZE !=
1975                    spage->physical ||
1976                    sbio->logical + sbio->page_count * PAGE_SIZE !=
1977                    spage->logical ||
1978                    sbio->dev != spage->dev) {
1979                 scrub_submit(sctx);
1980                 goto again;
1981         }
1982
1983         sbio->pagev[sbio->page_count] = spage;
1984         ret = bio_add_page(sbio->bio, spage->page, PAGE_SIZE, 0);
1985         if (ret != PAGE_SIZE) {
1986                 if (sbio->page_count < 1) {
1987                         bio_put(sbio->bio);
1988                         sbio->bio = NULL;
1989                         return -EIO;
1990                 }
1991                 scrub_submit(sctx);
1992                 goto again;
1993         }
1994
1995         scrub_block_get(sblock); /* one for the page added to the bio */
1996         atomic_inc(&sblock->outstanding_pages);
1997         sbio->page_count++;
1998         if (sbio->page_count == sctx->pages_per_rd_bio)
1999                 scrub_submit(sctx);
2000
2001         return 0;
2002 }
2003
2004 static int scrub_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
2005                        u64 physical, struct btrfs_device *dev, u64 flags,
2006                        u64 gen, int mirror_num, u8 *csum, int force,
2007                        u64 physical_for_dev_replace)
2008 {
2009         struct scrub_block *sblock;
2010         int index;
2011
2012         sblock = kzalloc(sizeof(*sblock), GFP_NOFS);
2013         if (!sblock) {
2014                 spin_lock(&sctx->stat_lock);
2015                 sctx->stat.malloc_errors++;
2016                 spin_unlock(&sctx->stat_lock);
2017                 return -ENOMEM;
2018         }
2019
2020         /* one ref inside this function, plus one for each page added to
2021          * a bio later on */
2022         atomic_set(&sblock->ref_count, 1);
2023         sblock->sctx = sctx;
2024         sblock->no_io_error_seen = 1;
2025
2026         for (index = 0; len > 0; index++) {
2027                 struct scrub_page *spage;
2028                 u64 l = min_t(u64, len, PAGE_SIZE);
2029
2030                 spage = kzalloc(sizeof(*spage), GFP_NOFS);
2031                 if (!spage) {
2032 leave_nomem:
2033                         spin_lock(&sctx->stat_lock);
2034                         sctx->stat.malloc_errors++;
2035                         spin_unlock(&sctx->stat_lock);
2036                         scrub_block_put(sblock);
2037                         return -ENOMEM;
2038                 }
2039                 BUG_ON(index >= SCRUB_MAX_PAGES_PER_BLOCK);
2040                 scrub_page_get(spage);
2041                 sblock->pagev[index] = spage;
2042                 spage->sblock = sblock;
2043                 spage->dev = dev;
2044                 spage->flags = flags;
2045                 spage->generation = gen;
2046                 spage->logical = logical;
2047                 spage->physical = physical;
2048                 spage->physical_for_dev_replace = physical_for_dev_replace;
2049                 spage->mirror_num = mirror_num;
2050                 if (csum) {
2051                         spage->have_csum = 1;
2052                         memcpy(spage->csum, csum, sctx->csum_size);
2053                 } else {
2054                         spage->have_csum = 0;
2055                 }
2056                 sblock->page_count++;
2057                 spage->page = alloc_page(GFP_NOFS);
2058                 if (!spage->page)
2059                         goto leave_nomem;
2060                 len -= l;
2061                 logical += l;
2062                 physical += l;
2063                 physical_for_dev_replace += l;
2064         }
2065
2066         WARN_ON(sblock->page_count == 0);
2067         for (index = 0; index < sblock->page_count; index++) {
2068                 struct scrub_page *spage = sblock->pagev[index];
2069                 int ret;
2070
2071                 ret = scrub_add_page_to_rd_bio(sctx, spage);
2072                 if (ret) {
2073                         scrub_block_put(sblock);
2074                         return ret;
2075                 }
2076         }
2077
2078         if (force)
2079                 scrub_submit(sctx);
2080
2081         /* last one frees, either here or in bio completion for last page */
2082         scrub_block_put(sblock);
2083         return 0;
2084 }
2085
2086 static void scrub_bio_end_io(struct bio *bio, int err)
2087 {
2088         struct scrub_bio *sbio = bio->bi_private;
2089         struct btrfs_fs_info *fs_info = sbio->dev->dev_root->fs_info;
2090
2091         sbio->err = err;
2092         sbio->bio = bio;
2093
2094         btrfs_queue_work(fs_info->scrub_workers, &sbio->work);
2095 }
2096
2097 static void scrub_bio_end_io_worker(struct btrfs_work *work)
2098 {
2099         struct scrub_bio *sbio = container_of(work, struct scrub_bio, work);
2100         struct scrub_ctx *sctx = sbio->sctx;
2101         int i;
2102
2103         BUG_ON(sbio->page_count > SCRUB_PAGES_PER_RD_BIO);
2104         if (sbio->err) {
2105                 for (i = 0; i < sbio->page_count; i++) {
2106                         struct scrub_page *spage = sbio->pagev[i];
2107
2108                         spage->io_error = 1;
2109                         spage->sblock->no_io_error_seen = 0;
2110                 }
2111         }
2112
2113         /* now complete the scrub_block items that have all pages completed */
2114         for (i = 0; i < sbio->page_count; i++) {
2115                 struct scrub_page *spage = sbio->pagev[i];
2116                 struct scrub_block *sblock = spage->sblock;
2117
2118                 if (atomic_dec_and_test(&sblock->outstanding_pages))
2119                         scrub_block_complete(sblock);
2120                 scrub_block_put(sblock);
2121         }
2122
2123         bio_put(sbio->bio);
2124         sbio->bio = NULL;
2125         spin_lock(&sctx->list_lock);
2126         sbio->next_free = sctx->first_free;
2127         sctx->first_free = sbio->index;
2128         spin_unlock(&sctx->list_lock);
2129
2130         if (sctx->is_dev_replace &&
2131             atomic_read(&sctx->wr_ctx.flush_all_writes)) {
2132                 mutex_lock(&sctx->wr_ctx.wr_lock);
2133                 scrub_wr_submit(sctx);
2134                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2135         }
2136
2137         scrub_pending_bio_dec(sctx);
2138 }
2139
2140 static void scrub_block_complete(struct scrub_block *sblock)
2141 {
2142         if (!sblock->no_io_error_seen) {
2143                 scrub_handle_errored_block(sblock);
2144         } else {
2145                 /*
2146                  * if has checksum error, write via repair mechanism in
2147                  * dev replace case, otherwise write here in dev replace
2148                  * case.
2149                  */
2150                 if (!scrub_checksum(sblock) && sblock->sctx->is_dev_replace)
2151                         scrub_write_block_to_dev_replace(sblock);
2152         }
2153 }
2154
2155 static int scrub_find_csum(struct scrub_ctx *sctx, u64 logical, u64 len,
2156                            u8 *csum)
2157 {
2158         struct btrfs_ordered_sum *sum = NULL;
2159         unsigned long index;
2160         unsigned long num_sectors;
2161
2162         while (!list_empty(&sctx->csum_list)) {
2163                 sum = list_first_entry(&sctx->csum_list,
2164                                        struct btrfs_ordered_sum, list);
2165                 if (sum->bytenr > logical)
2166                         return 0;
2167                 if (sum->bytenr + sum->len > logical)
2168                         break;
2169
2170                 ++sctx->stat.csum_discards;
2171                 list_del(&sum->list);
2172                 kfree(sum);
2173                 sum = NULL;
2174         }
2175         if (!sum)
2176                 return 0;
2177
2178         index = ((u32)(logical - sum->bytenr)) / sctx->sectorsize;
2179         num_sectors = sum->len / sctx->sectorsize;
2180         memcpy(csum, sum->sums + index, sctx->csum_size);
2181         if (index == num_sectors - 1) {
2182                 list_del(&sum->list);
2183                 kfree(sum);
2184         }
2185         return 1;
2186 }
2187
2188 /* scrub extent tries to collect up to 64 kB for each bio */
2189 static int scrub_extent(struct scrub_ctx *sctx, u64 logical, u64 len,
2190                         u64 physical, struct btrfs_device *dev, u64 flags,
2191                         u64 gen, int mirror_num, u64 physical_for_dev_replace)
2192 {
2193         int ret;
2194         u8 csum[BTRFS_CSUM_SIZE];
2195         u32 blocksize;
2196
2197         if (flags & BTRFS_EXTENT_FLAG_DATA) {
2198                 blocksize = sctx->sectorsize;
2199                 spin_lock(&sctx->stat_lock);
2200                 sctx->stat.data_extents_scrubbed++;
2201                 sctx->stat.data_bytes_scrubbed += len;
2202                 spin_unlock(&sctx->stat_lock);
2203         } else if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
2204                 blocksize = sctx->nodesize;
2205                 spin_lock(&sctx->stat_lock);
2206                 sctx->stat.tree_extents_scrubbed++;
2207                 sctx->stat.tree_bytes_scrubbed += len;
2208                 spin_unlock(&sctx->stat_lock);
2209         } else {
2210                 blocksize = sctx->sectorsize;
2211                 WARN_ON(1);
2212         }
2213
2214         while (len) {
2215                 u64 l = min_t(u64, len, blocksize);
2216                 int have_csum = 0;
2217
2218                 if (flags & BTRFS_EXTENT_FLAG_DATA) {
2219                         /* push csums to sbio */
2220                         have_csum = scrub_find_csum(sctx, logical, l, csum);
2221                         if (have_csum == 0)
2222                                 ++sctx->stat.no_csum;
2223                         if (sctx->is_dev_replace && !have_csum) {
2224                                 ret = copy_nocow_pages(sctx, logical, l,
2225                                                        mirror_num,
2226                                                       physical_for_dev_replace);
2227                                 goto behind_scrub_pages;
2228                         }
2229                 }
2230                 ret = scrub_pages(sctx, logical, l, physical, dev, flags, gen,
2231                                   mirror_num, have_csum ? csum : NULL, 0,
2232                                   physical_for_dev_replace);
2233 behind_scrub_pages:
2234                 if (ret)
2235                         return ret;
2236                 len -= l;
2237                 logical += l;
2238                 physical += l;
2239                 physical_for_dev_replace += l;
2240         }
2241         return 0;
2242 }
2243
2244 /*
2245  * Given a physical address, this will calculate it's
2246  * logical offset. if this is a parity stripe, it will return
2247  * the most left data stripe's logical offset.
2248  *
2249  * return 0 if it is a data stripe, 1 means parity stripe.
2250  */
2251 static int get_raid56_logic_offset(u64 physical, int num,
2252                                    struct map_lookup *map, u64 *offset)
2253 {
2254         int i;
2255         int j = 0;
2256         u64 stripe_nr;
2257         u64 last_offset;
2258         int stripe_index;
2259         int rot;
2260
2261         last_offset = (physical - map->stripes[num].physical) *
2262                       nr_data_stripes(map);
2263         *offset = last_offset;
2264         for (i = 0; i < nr_data_stripes(map); i++) {
2265                 *offset = last_offset + i * map->stripe_len;
2266
2267                 stripe_nr = *offset;
2268                 do_div(stripe_nr, map->stripe_len);
2269                 do_div(stripe_nr, nr_data_stripes(map));
2270
2271                 /* Work out the disk rotation on this stripe-set */
2272                 rot = do_div(stripe_nr, map->num_stripes);
2273                 /* calculate which stripe this data locates */
2274                 rot += i;
2275                 stripe_index = rot % map->num_stripes;
2276                 if (stripe_index == num)
2277                         return 0;
2278                 if (stripe_index < num)
2279                         j++;
2280         }
2281         *offset = last_offset + j * map->stripe_len;
2282         return 1;
2283 }
2284
2285 static noinline_for_stack int scrub_stripe(struct scrub_ctx *sctx,
2286                                            struct map_lookup *map,
2287                                            struct btrfs_device *scrub_dev,
2288                                            int num, u64 base, u64 length,
2289                                            int is_dev_replace)
2290 {
2291         struct btrfs_path *path;
2292         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
2293         struct btrfs_root *root = fs_info->extent_root;
2294         struct btrfs_root *csum_root = fs_info->csum_root;
2295         struct btrfs_extent_item *extent;
2296         struct blk_plug plug;
2297         u64 flags;
2298         int ret;
2299         int slot;
2300         u64 nstripes;
2301         struct extent_buffer *l;
2302         struct btrfs_key key;
2303         u64 physical;
2304         u64 logical;
2305         u64 logic_end;
2306         u64 physical_end;
2307         u64 generation;
2308         int mirror_num;
2309         struct reada_control *reada1;
2310         struct reada_control *reada2;
2311         struct btrfs_key key_start;
2312         struct btrfs_key key_end;
2313         u64 increment = map->stripe_len;
2314         u64 offset;
2315         u64 extent_logical;
2316         u64 extent_physical;
2317         u64 extent_len;
2318         struct btrfs_device *extent_dev;
2319         int extent_mirror_num;
2320         int stop_loop = 0;
2321
2322         nstripes = length;
2323         physical = map->stripes[num].physical;
2324         offset = 0;
2325         do_div(nstripes, map->stripe_len);
2326         if (map->type & BTRFS_BLOCK_GROUP_RAID0) {
2327                 offset = map->stripe_len * num;
2328                 increment = map->stripe_len * map->num_stripes;
2329                 mirror_num = 1;
2330         } else if (map->type & BTRFS_BLOCK_GROUP_RAID10) {
2331                 int factor = map->num_stripes / map->sub_stripes;
2332                 offset = map->stripe_len * (num / map->sub_stripes);
2333                 increment = map->stripe_len * factor;
2334                 mirror_num = num % map->sub_stripes + 1;
2335         } else if (map->type & BTRFS_BLOCK_GROUP_RAID1) {
2336                 increment = map->stripe_len;
2337                 mirror_num = num % map->num_stripes + 1;
2338         } else if (map->type & BTRFS_BLOCK_GROUP_DUP) {
2339                 increment = map->stripe_len;
2340                 mirror_num = num % map->num_stripes + 1;
2341         } else if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2342                                 BTRFS_BLOCK_GROUP_RAID6)) {
2343                 get_raid56_logic_offset(physical, num, map, &offset);
2344                 increment = map->stripe_len * nr_data_stripes(map);
2345                 mirror_num = 1;
2346         } else {
2347                 increment = map->stripe_len;
2348                 mirror_num = 1;
2349         }
2350
2351         path = btrfs_alloc_path();
2352         if (!path)
2353                 return -ENOMEM;
2354
2355         /*
2356          * work on commit root. The related disk blocks are static as
2357          * long as COW is applied. This means, it is save to rewrite
2358          * them to repair disk errors without any race conditions
2359          */
2360         path->search_commit_root = 1;
2361         path->skip_locking = 1;
2362
2363         /*
2364          * trigger the readahead for extent tree csum tree and wait for
2365          * completion. During readahead, the scrub is officially paused
2366          * to not hold off transaction commits
2367          */
2368         logical = base + offset;
2369         physical_end = physical + nstripes * map->stripe_len;
2370         if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2371                          BTRFS_BLOCK_GROUP_RAID6)) {
2372                 get_raid56_logic_offset(physical_end, num,
2373                                         map, &logic_end);
2374                 logic_end += base;
2375         } else {
2376                 logic_end = logical + increment * nstripes;
2377         }
2378         wait_event(sctx->list_wait,
2379                    atomic_read(&sctx->bios_in_flight) == 0);
2380         scrub_blocked_if_needed(fs_info);
2381
2382         /* FIXME it might be better to start readahead at commit root */
2383         key_start.objectid = logical;
2384         key_start.type = BTRFS_EXTENT_ITEM_KEY;
2385         key_start.offset = (u64)0;
2386         key_end.objectid = logic_end;
2387         key_end.type = BTRFS_METADATA_ITEM_KEY;
2388         key_end.offset = (u64)-1;
2389         reada1 = btrfs_reada_add(root, &key_start, &key_end);
2390
2391         key_start.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2392         key_start.type = BTRFS_EXTENT_CSUM_KEY;
2393         key_start.offset = logical;
2394         key_end.objectid = BTRFS_EXTENT_CSUM_OBJECTID;
2395         key_end.type = BTRFS_EXTENT_CSUM_KEY;
2396         key_end.offset = logic_end;
2397         reada2 = btrfs_reada_add(csum_root, &key_start, &key_end);
2398
2399         if (!IS_ERR(reada1))
2400                 btrfs_reada_wait(reada1);
2401         if (!IS_ERR(reada2))
2402                 btrfs_reada_wait(reada2);
2403
2404
2405         /*
2406          * collect all data csums for the stripe to avoid seeking during
2407          * the scrub. This might currently (crc32) end up to be about 1MB
2408          */
2409         blk_start_plug(&plug);
2410
2411         /*
2412          * now find all extents for each stripe and scrub them
2413          */
2414         ret = 0;
2415         while (physical < physical_end) {
2416                 /* for raid56, we skip parity stripe */
2417                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2418                                 BTRFS_BLOCK_GROUP_RAID6)) {
2419                         ret = get_raid56_logic_offset(physical, num,
2420                                         map, &logical);
2421                         logical += base;
2422                         if (ret)
2423                                 goto skip;
2424                 }
2425                 /*
2426                  * canceled?
2427                  */
2428                 if (atomic_read(&fs_info->scrub_cancel_req) ||
2429                     atomic_read(&sctx->cancel_req)) {
2430                         ret = -ECANCELED;
2431                         goto out;
2432                 }
2433                 /*
2434                  * check to see if we have to pause
2435                  */
2436                 if (atomic_read(&fs_info->scrub_pause_req)) {
2437                         /* push queued extents */
2438                         atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2439                         scrub_submit(sctx);
2440                         mutex_lock(&sctx->wr_ctx.wr_lock);
2441                         scrub_wr_submit(sctx);
2442                         mutex_unlock(&sctx->wr_ctx.wr_lock);
2443                         wait_event(sctx->list_wait,
2444                                    atomic_read(&sctx->bios_in_flight) == 0);
2445                         atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2446                         scrub_blocked_if_needed(fs_info);
2447                 }
2448
2449                 if (btrfs_fs_incompat(fs_info, SKINNY_METADATA))
2450                         key.type = BTRFS_METADATA_ITEM_KEY;
2451                 else
2452                         key.type = BTRFS_EXTENT_ITEM_KEY;
2453                 key.objectid = logical;
2454                 key.offset = (u64)-1;
2455
2456                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2457                 if (ret < 0)
2458                         goto out;
2459
2460                 if (ret > 0) {
2461                         ret = btrfs_previous_extent_item(root, path, 0);
2462                         if (ret < 0)
2463                                 goto out;
2464                         if (ret > 0) {
2465                                 /* there's no smaller item, so stick with the
2466                                  * larger one */
2467                                 btrfs_release_path(path);
2468                                 ret = btrfs_search_slot(NULL, root, &key,
2469                                                         path, 0, 0);
2470                                 if (ret < 0)
2471                                         goto out;
2472                         }
2473                 }
2474
2475                 stop_loop = 0;
2476                 while (1) {
2477                         u64 bytes;
2478
2479                         l = path->nodes[0];
2480                         slot = path->slots[0];
2481                         if (slot >= btrfs_header_nritems(l)) {
2482                                 ret = btrfs_next_leaf(root, path);
2483                                 if (ret == 0)
2484                                         continue;
2485                                 if (ret < 0)
2486                                         goto out;
2487
2488                                 stop_loop = 1;
2489                                 break;
2490                         }
2491                         btrfs_item_key_to_cpu(l, &key, slot);
2492
2493                         if (key.type == BTRFS_METADATA_ITEM_KEY)
2494                                 bytes = root->nodesize;
2495                         else
2496                                 bytes = key.offset;
2497
2498                         if (key.objectid + bytes <= logical)
2499                                 goto next;
2500
2501                         if (key.type != BTRFS_EXTENT_ITEM_KEY &&
2502                             key.type != BTRFS_METADATA_ITEM_KEY)
2503                                 goto next;
2504
2505                         if (key.objectid >= logical + map->stripe_len) {
2506                                 /* out of this device extent */
2507                                 if (key.objectid >= logic_end)
2508                                         stop_loop = 1;
2509                                 break;
2510                         }
2511
2512                         extent = btrfs_item_ptr(l, slot,
2513                                                 struct btrfs_extent_item);
2514                         flags = btrfs_extent_flags(l, extent);
2515                         generation = btrfs_extent_generation(l, extent);
2516
2517                         if (key.objectid < logical &&
2518                             (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK)) {
2519                                 btrfs_err(fs_info,
2520                                            "scrub: tree block %llu spanning "
2521                                            "stripes, ignored. logical=%llu",
2522                                        key.objectid, logical);
2523                                 goto next;
2524                         }
2525
2526 again:
2527                         extent_logical = key.objectid;
2528                         extent_len = bytes;
2529
2530                         /*
2531                          * trim extent to this stripe
2532                          */
2533                         if (extent_logical < logical) {
2534                                 extent_len -= logical - extent_logical;
2535                                 extent_logical = logical;
2536                         }
2537                         if (extent_logical + extent_len >
2538                             logical + map->stripe_len) {
2539                                 extent_len = logical + map->stripe_len -
2540                                              extent_logical;
2541                         }
2542
2543                         extent_physical = extent_logical - logical + physical;
2544                         extent_dev = scrub_dev;
2545                         extent_mirror_num = mirror_num;
2546                         if (is_dev_replace)
2547                                 scrub_remap_extent(fs_info, extent_logical,
2548                                                    extent_len, &extent_physical,
2549                                                    &extent_dev,
2550                                                    &extent_mirror_num);
2551
2552                         ret = btrfs_lookup_csums_range(csum_root, logical,
2553                                                 logical + map->stripe_len - 1,
2554                                                 &sctx->csum_list, 1);
2555                         if (ret)
2556                                 goto out;
2557
2558                         ret = scrub_extent(sctx, extent_logical, extent_len,
2559                                            extent_physical, extent_dev, flags,
2560                                            generation, extent_mirror_num,
2561                                            extent_logical - logical + physical);
2562                         if (ret)
2563                                 goto out;
2564
2565                         scrub_free_csums(sctx);
2566                         if (extent_logical + extent_len <
2567                             key.objectid + bytes) {
2568                                 if (map->type & (BTRFS_BLOCK_GROUP_RAID5 |
2569                                         BTRFS_BLOCK_GROUP_RAID6)) {
2570                                         /*
2571                                          * loop until we find next data stripe
2572                                          * or we have finished all stripes.
2573                                          */
2574                                         do {
2575                                                 physical += map->stripe_len;
2576                                                 ret = get_raid56_logic_offset(
2577                                                                 physical, num,
2578                                                                 map, &logical);
2579                                                 logical += base;
2580                                         } while (physical < physical_end && ret);
2581                                 } else {
2582                                         physical += map->stripe_len;
2583                                         logical += increment;
2584                                 }
2585                                 if (logical < key.objectid + bytes) {
2586                                         cond_resched();
2587                                         goto again;
2588                                 }
2589
2590                                 if (physical >= physical_end) {
2591                                         stop_loop = 1;
2592                                         break;
2593                                 }
2594                         }
2595 next:
2596                         path->slots[0]++;
2597                 }
2598                 btrfs_release_path(path);
2599 skip:
2600                 logical += increment;
2601                 physical += map->stripe_len;
2602                 spin_lock(&sctx->stat_lock);
2603                 if (stop_loop)
2604                         sctx->stat.last_physical = map->stripes[num].physical +
2605                                                    length;
2606                 else
2607                         sctx->stat.last_physical = physical;
2608                 spin_unlock(&sctx->stat_lock);
2609                 if (stop_loop)
2610                         break;
2611         }
2612 out:
2613         /* push queued extents */
2614         scrub_submit(sctx);
2615         mutex_lock(&sctx->wr_ctx.wr_lock);
2616         scrub_wr_submit(sctx);
2617         mutex_unlock(&sctx->wr_ctx.wr_lock);
2618
2619         blk_finish_plug(&plug);
2620         btrfs_free_path(path);
2621         return ret < 0 ? ret : 0;
2622 }
2623
2624 static noinline_for_stack int scrub_chunk(struct scrub_ctx *sctx,
2625                                           struct btrfs_device *scrub_dev,
2626                                           u64 chunk_tree, u64 chunk_objectid,
2627                                           u64 chunk_offset, u64 length,
2628                                           u64 dev_offset, int is_dev_replace)
2629 {
2630         struct btrfs_mapping_tree *map_tree =
2631                 &sctx->dev_root->fs_info->mapping_tree;
2632         struct map_lookup *map;
2633         struct extent_map *em;
2634         int i;
2635         int ret = 0;
2636
2637         read_lock(&map_tree->map_tree.lock);
2638         em = lookup_extent_mapping(&map_tree->map_tree, chunk_offset, 1);
2639         read_unlock(&map_tree->map_tree.lock);
2640
2641         if (!em)
2642                 return -EINVAL;
2643
2644         map = (struct map_lookup *)em->bdev;
2645         if (em->start != chunk_offset)
2646                 goto out;
2647
2648         if (em->len < length)
2649                 goto out;
2650
2651         for (i = 0; i < map->num_stripes; ++i) {
2652                 if (map->stripes[i].dev->bdev == scrub_dev->bdev &&
2653                     map->stripes[i].physical == dev_offset) {
2654                         ret = scrub_stripe(sctx, map, scrub_dev, i,
2655                                            chunk_offset, length,
2656                                            is_dev_replace);
2657                         if (ret)
2658                                 goto out;
2659                 }
2660         }
2661 out:
2662         free_extent_map(em);
2663
2664         return ret;
2665 }
2666
2667 static noinline_for_stack
2668 int scrub_enumerate_chunks(struct scrub_ctx *sctx,
2669                            struct btrfs_device *scrub_dev, u64 start, u64 end,
2670                            int is_dev_replace)
2671 {
2672         struct btrfs_dev_extent *dev_extent = NULL;
2673         struct btrfs_path *path;
2674         struct btrfs_root *root = sctx->dev_root;
2675         struct btrfs_fs_info *fs_info = root->fs_info;
2676         u64 length;
2677         u64 chunk_tree;
2678         u64 chunk_objectid;
2679         u64 chunk_offset;
2680         int ret;
2681         int slot;
2682         struct extent_buffer *l;
2683         struct btrfs_key key;
2684         struct btrfs_key found_key;
2685         struct btrfs_block_group_cache *cache;
2686         struct btrfs_dev_replace *dev_replace = &fs_info->dev_replace;
2687
2688         path = btrfs_alloc_path();
2689         if (!path)
2690                 return -ENOMEM;
2691
2692         path->reada = 2;
2693         path->search_commit_root = 1;
2694         path->skip_locking = 1;
2695
2696         key.objectid = scrub_dev->devid;
2697         key.offset = 0ull;
2698         key.type = BTRFS_DEV_EXTENT_KEY;
2699
2700         while (1) {
2701                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2702                 if (ret < 0)
2703                         break;
2704                 if (ret > 0) {
2705                         if (path->slots[0] >=
2706                             btrfs_header_nritems(path->nodes[0])) {
2707                                 ret = btrfs_next_leaf(root, path);
2708                                 if (ret)
2709                                         break;
2710                         }
2711                 }
2712
2713                 l = path->nodes[0];
2714                 slot = path->slots[0];
2715
2716                 btrfs_item_key_to_cpu(l, &found_key, slot);
2717
2718                 if (found_key.objectid != scrub_dev->devid)
2719                         break;
2720
2721                 if (found_key.type != BTRFS_DEV_EXTENT_KEY)
2722                         break;
2723
2724                 if (found_key.offset >= end)
2725                         break;
2726
2727                 if (found_key.offset < key.offset)
2728                         break;
2729
2730                 dev_extent = btrfs_item_ptr(l, slot, struct btrfs_dev_extent);
2731                 length = btrfs_dev_extent_length(l, dev_extent);
2732
2733                 if (found_key.offset + length <= start)
2734                         goto skip;
2735
2736                 chunk_tree = btrfs_dev_extent_chunk_tree(l, dev_extent);
2737                 chunk_objectid = btrfs_dev_extent_chunk_objectid(l, dev_extent);
2738                 chunk_offset = btrfs_dev_extent_chunk_offset(l, dev_extent);
2739
2740                 /*
2741                  * get a reference on the corresponding block group to prevent
2742                  * the chunk from going away while we scrub it
2743                  */
2744                 cache = btrfs_lookup_block_group(fs_info, chunk_offset);
2745
2746                 /* some chunks are removed but not committed to disk yet,
2747                  * continue scrubbing */
2748                 if (!cache)
2749                         goto skip;
2750
2751                 dev_replace->cursor_right = found_key.offset + length;
2752                 dev_replace->cursor_left = found_key.offset;
2753                 dev_replace->item_needs_writeback = 1;
2754                 ret = scrub_chunk(sctx, scrub_dev, chunk_tree, chunk_objectid,
2755                                   chunk_offset, length, found_key.offset,
2756                                   is_dev_replace);
2757
2758                 /*
2759                  * flush, submit all pending read and write bios, afterwards
2760                  * wait for them.
2761                  * Note that in the dev replace case, a read request causes
2762                  * write requests that are submitted in the read completion
2763                  * worker. Therefore in the current situation, it is required
2764                  * that all write requests are flushed, so that all read and
2765                  * write requests are really completed when bios_in_flight
2766                  * changes to 0.
2767                  */
2768                 atomic_set(&sctx->wr_ctx.flush_all_writes, 1);
2769                 scrub_submit(sctx);
2770                 mutex_lock(&sctx->wr_ctx.wr_lock);
2771                 scrub_wr_submit(sctx);
2772                 mutex_unlock(&sctx->wr_ctx.wr_lock);
2773
2774                 wait_event(sctx->list_wait,
2775                            atomic_read(&sctx->bios_in_flight) == 0);
2776                 atomic_inc(&fs_info->scrubs_paused);
2777                 wake_up(&fs_info->scrub_pause_wait);
2778
2779                 /*
2780                  * must be called before we decrease @scrub_paused.
2781                  * make sure we don't block transaction commit while
2782                  * we are waiting pending workers finished.
2783                  */
2784                 wait_event(sctx->list_wait,
2785                            atomic_read(&sctx->workers_pending) == 0);
2786                 atomic_set(&sctx->wr_ctx.flush_all_writes, 0);
2787
2788                 mutex_lock(&fs_info->scrub_lock);
2789                 __scrub_blocked_if_needed(fs_info);
2790                 atomic_dec(&fs_info->scrubs_paused);
2791                 mutex_unlock(&fs_info->scrub_lock);
2792                 wake_up(&fs_info->scrub_pause_wait);
2793
2794                 btrfs_put_block_group(cache);
2795                 if (ret)
2796                         break;
2797                 if (is_dev_replace &&
2798                     atomic64_read(&dev_replace->num_write_errors) > 0) {
2799                         ret = -EIO;
2800                         break;
2801                 }
2802                 if (sctx->stat.malloc_errors > 0) {
2803                         ret = -ENOMEM;
2804                         break;
2805                 }
2806
2807                 dev_replace->cursor_left = dev_replace->cursor_right;
2808                 dev_replace->item_needs_writeback = 1;
2809 skip:
2810                 key.offset = found_key.offset + length;
2811                 btrfs_release_path(path);
2812         }
2813
2814         btrfs_free_path(path);
2815
2816         /*
2817          * ret can still be 1 from search_slot or next_leaf,
2818          * that's not an error
2819          */
2820         return ret < 0 ? ret : 0;
2821 }
2822
2823 static noinline_for_stack int scrub_supers(struct scrub_ctx *sctx,
2824                                            struct btrfs_device *scrub_dev)
2825 {
2826         int     i;
2827         u64     bytenr;
2828         u64     gen;
2829         int     ret;
2830         struct btrfs_root *root = sctx->dev_root;
2831
2832         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
2833                 return -EIO;
2834
2835         /* Seed devices of a new filesystem has their own generation. */
2836         if (scrub_dev->fs_devices != root->fs_info->fs_devices)
2837                 gen = scrub_dev->generation;
2838         else
2839                 gen = root->fs_info->last_trans_committed;
2840
2841         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2842                 bytenr = btrfs_sb_offset(i);
2843                 if (bytenr + BTRFS_SUPER_INFO_SIZE >
2844                     scrub_dev->commit_total_bytes)
2845                         break;
2846
2847                 ret = scrub_pages(sctx, bytenr, BTRFS_SUPER_INFO_SIZE, bytenr,
2848                                   scrub_dev, BTRFS_EXTENT_FLAG_SUPER, gen, i,
2849                                   NULL, 1, bytenr);
2850                 if (ret)
2851                         return ret;
2852         }
2853         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
2854
2855         return 0;
2856 }
2857
2858 /*
2859  * get a reference count on fs_info->scrub_workers. start worker if necessary
2860  */
2861 static noinline_for_stack int scrub_workers_get(struct btrfs_fs_info *fs_info,
2862                                                 int is_dev_replace)
2863 {
2864         int ret = 0;
2865         int flags = WQ_FREEZABLE | WQ_UNBOUND;
2866         int max_active = fs_info->thread_pool_size;
2867
2868         if (fs_info->scrub_workers_refcnt == 0) {
2869                 if (is_dev_replace)
2870                         fs_info->scrub_workers =
2871                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2872                                                       1, 4);
2873                 else
2874                         fs_info->scrub_workers =
2875                                 btrfs_alloc_workqueue("btrfs-scrub", flags,
2876                                                       max_active, 4);
2877                 if (!fs_info->scrub_workers) {
2878                         ret = -ENOMEM;
2879                         goto out;
2880                 }
2881                 fs_info->scrub_wr_completion_workers =
2882                         btrfs_alloc_workqueue("btrfs-scrubwrc", flags,
2883                                               max_active, 2);
2884                 if (!fs_info->scrub_wr_completion_workers) {
2885                         ret = -ENOMEM;
2886                         goto out;
2887                 }
2888                 fs_info->scrub_nocow_workers =
2889                         btrfs_alloc_workqueue("btrfs-scrubnc", flags, 1, 0);
2890                 if (!fs_info->scrub_nocow_workers) {
2891                         ret = -ENOMEM;
2892                         goto out;
2893                 }
2894         }
2895         ++fs_info->scrub_workers_refcnt;
2896 out:
2897         return ret;
2898 }
2899
2900 static noinline_for_stack void scrub_workers_put(struct btrfs_fs_info *fs_info)
2901 {
2902         if (--fs_info->scrub_workers_refcnt == 0) {
2903                 btrfs_destroy_workqueue(fs_info->scrub_workers);
2904                 btrfs_destroy_workqueue(fs_info->scrub_wr_completion_workers);
2905                 btrfs_destroy_workqueue(fs_info->scrub_nocow_workers);
2906         }
2907         WARN_ON(fs_info->scrub_workers_refcnt < 0);
2908 }
2909
2910 int btrfs_scrub_dev(struct btrfs_fs_info *fs_info, u64 devid, u64 start,
2911                     u64 end, struct btrfs_scrub_progress *progress,
2912                     int readonly, int is_dev_replace)
2913 {
2914         struct scrub_ctx *sctx;
2915         int ret;
2916         struct btrfs_device *dev;
2917         struct rcu_string *name;
2918
2919         if (btrfs_fs_closing(fs_info))
2920                 return -EINVAL;
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         if (!is_dev_replace && !readonly && !dev->writeable) {
2969                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2970                 rcu_read_lock();
2971                 name = rcu_dereference(dev->name);
2972                 btrfs_err(fs_info, "scrub: device %s is not writable",
2973                           name->str);
2974                 rcu_read_unlock();
2975                 return -EROFS;
2976         }
2977
2978         mutex_lock(&fs_info->scrub_lock);
2979         if (!dev->in_fs_metadata || dev->is_tgtdev_for_dev_replace) {
2980                 mutex_unlock(&fs_info->scrub_lock);
2981                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2982                 return -EIO;
2983         }
2984
2985         btrfs_dev_replace_lock(&fs_info->dev_replace);
2986         if (dev->scrub_device ||
2987             (!is_dev_replace &&
2988              btrfs_dev_replace_is_ongoing(&fs_info->dev_replace))) {
2989                 btrfs_dev_replace_unlock(&fs_info->dev_replace);
2990                 mutex_unlock(&fs_info->scrub_lock);
2991                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2992                 return -EINPROGRESS;
2993         }
2994         btrfs_dev_replace_unlock(&fs_info->dev_replace);
2995
2996         ret = scrub_workers_get(fs_info, is_dev_replace);
2997         if (ret) {
2998                 mutex_unlock(&fs_info->scrub_lock);
2999                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3000                 return ret;
3001         }
3002
3003         sctx = scrub_setup_ctx(dev, is_dev_replace);
3004         if (IS_ERR(sctx)) {
3005                 mutex_unlock(&fs_info->scrub_lock);
3006                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3007                 scrub_workers_put(fs_info);
3008                 return PTR_ERR(sctx);
3009         }
3010         sctx->readonly = readonly;
3011         dev->scrub_device = sctx;
3012         mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3013
3014         /*
3015          * checking @scrub_pause_req here, we can avoid
3016          * race between committing transaction and scrubbing.
3017          */
3018         __scrub_blocked_if_needed(fs_info);
3019         atomic_inc(&fs_info->scrubs_running);
3020         mutex_unlock(&fs_info->scrub_lock);
3021
3022         if (!is_dev_replace) {
3023                 /*
3024                  * by holding device list mutex, we can
3025                  * kick off writing super in log tree sync.
3026                  */
3027                 mutex_lock(&fs_info->fs_devices->device_list_mutex);
3028                 ret = scrub_supers(sctx, dev);
3029                 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
3030         }
3031
3032         if (!ret)
3033                 ret = scrub_enumerate_chunks(sctx, dev, start, end,
3034                                              is_dev_replace);
3035
3036         wait_event(sctx->list_wait, atomic_read(&sctx->bios_in_flight) == 0);
3037         atomic_dec(&fs_info->scrubs_running);
3038         wake_up(&fs_info->scrub_pause_wait);
3039
3040         wait_event(sctx->list_wait, atomic_read(&sctx->workers_pending) == 0);
3041
3042         if (progress)
3043                 memcpy(progress, &sctx->stat, sizeof(*progress));
3044
3045         mutex_lock(&fs_info->scrub_lock);
3046         dev->scrub_device = NULL;
3047         scrub_workers_put(fs_info);
3048         mutex_unlock(&fs_info->scrub_lock);
3049
3050         scrub_free_ctx(sctx);
3051
3052         return ret;
3053 }
3054
3055 void btrfs_scrub_pause(struct btrfs_root *root)
3056 {
3057         struct btrfs_fs_info *fs_info = root->fs_info;
3058
3059         mutex_lock(&fs_info->scrub_lock);
3060         atomic_inc(&fs_info->scrub_pause_req);
3061         while (atomic_read(&fs_info->scrubs_paused) !=
3062                atomic_read(&fs_info->scrubs_running)) {
3063                 mutex_unlock(&fs_info->scrub_lock);
3064                 wait_event(fs_info->scrub_pause_wait,
3065                            atomic_read(&fs_info->scrubs_paused) ==
3066                            atomic_read(&fs_info->scrubs_running));
3067                 mutex_lock(&fs_info->scrub_lock);
3068         }
3069         mutex_unlock(&fs_info->scrub_lock);
3070 }
3071
3072 void btrfs_scrub_continue(struct btrfs_root *root)
3073 {
3074         struct btrfs_fs_info *fs_info = root->fs_info;
3075
3076         atomic_dec(&fs_info->scrub_pause_req);
3077         wake_up(&fs_info->scrub_pause_wait);
3078 }
3079
3080 int btrfs_scrub_cancel(struct btrfs_fs_info *fs_info)
3081 {
3082         mutex_lock(&fs_info->scrub_lock);
3083         if (!atomic_read(&fs_info->scrubs_running)) {
3084                 mutex_unlock(&fs_info->scrub_lock);
3085                 return -ENOTCONN;
3086         }
3087
3088         atomic_inc(&fs_info->scrub_cancel_req);
3089         while (atomic_read(&fs_info->scrubs_running)) {
3090                 mutex_unlock(&fs_info->scrub_lock);
3091                 wait_event(fs_info->scrub_pause_wait,
3092                            atomic_read(&fs_info->scrubs_running) == 0);
3093                 mutex_lock(&fs_info->scrub_lock);
3094         }
3095         atomic_dec(&fs_info->scrub_cancel_req);
3096         mutex_unlock(&fs_info->scrub_lock);
3097
3098         return 0;
3099 }
3100
3101 int btrfs_scrub_cancel_dev(struct btrfs_fs_info *fs_info,
3102                            struct btrfs_device *dev)
3103 {
3104         struct scrub_ctx *sctx;
3105
3106         mutex_lock(&fs_info->scrub_lock);
3107         sctx = dev->scrub_device;
3108         if (!sctx) {
3109                 mutex_unlock(&fs_info->scrub_lock);
3110                 return -ENOTCONN;
3111         }
3112         atomic_inc(&sctx->cancel_req);
3113         while (dev->scrub_device) {
3114                 mutex_unlock(&fs_info->scrub_lock);
3115                 wait_event(fs_info->scrub_pause_wait,
3116                            dev->scrub_device == NULL);
3117                 mutex_lock(&fs_info->scrub_lock);
3118         }
3119         mutex_unlock(&fs_info->scrub_lock);
3120
3121         return 0;
3122 }
3123
3124 int btrfs_scrub_progress(struct btrfs_root *root, u64 devid,
3125                          struct btrfs_scrub_progress *progress)
3126 {
3127         struct btrfs_device *dev;
3128         struct scrub_ctx *sctx = NULL;
3129
3130         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
3131         dev = btrfs_find_device(root->fs_info, devid, NULL, NULL);
3132         if (dev)
3133                 sctx = dev->scrub_device;
3134         if (sctx)
3135                 memcpy(progress, &sctx->stat, sizeof(*progress));
3136         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
3137
3138         return dev ? (sctx ? 0 : -ENOTCONN) : -ENODEV;
3139 }
3140
3141 static void scrub_remap_extent(struct btrfs_fs_info *fs_info,
3142                                u64 extent_logical, u64 extent_len,
3143                                u64 *extent_physical,
3144                                struct btrfs_device **extent_dev,
3145                                int *extent_mirror_num)
3146 {
3147         u64 mapped_length;
3148         struct btrfs_bio *bbio = NULL;
3149         int ret;
3150
3151         mapped_length = extent_len;
3152         ret = btrfs_map_block(fs_info, READ, extent_logical,
3153                               &mapped_length, &bbio, 0);
3154         if (ret || !bbio || mapped_length < extent_len ||
3155             !bbio->stripes[0].dev->bdev) {
3156                 kfree(bbio);
3157                 return;
3158         }
3159
3160         *extent_physical = bbio->stripes[0].physical;
3161         *extent_mirror_num = bbio->mirror_num;
3162         *extent_dev = bbio->stripes[0].dev;
3163         kfree(bbio);
3164 }
3165
3166 static int scrub_setup_wr_ctx(struct scrub_ctx *sctx,
3167                               struct scrub_wr_ctx *wr_ctx,
3168                               struct btrfs_fs_info *fs_info,
3169                               struct btrfs_device *dev,
3170                               int is_dev_replace)
3171 {
3172         WARN_ON(wr_ctx->wr_curr_bio != NULL);
3173
3174         mutex_init(&wr_ctx->wr_lock);
3175         wr_ctx->wr_curr_bio = NULL;
3176         if (!is_dev_replace)
3177                 return 0;
3178
3179         WARN_ON(!dev->bdev);
3180         wr_ctx->pages_per_wr_bio = min_t(int, SCRUB_PAGES_PER_WR_BIO,
3181                                          bio_get_nr_vecs(dev->bdev));
3182         wr_ctx->tgtdev = dev;
3183         atomic_set(&wr_ctx->flush_all_writes, 0);
3184         return 0;
3185 }
3186
3187 static void scrub_free_wr_ctx(struct scrub_wr_ctx *wr_ctx)
3188 {
3189         mutex_lock(&wr_ctx->wr_lock);
3190         kfree(wr_ctx->wr_curr_bio);
3191         wr_ctx->wr_curr_bio = NULL;
3192         mutex_unlock(&wr_ctx->wr_lock);
3193 }
3194
3195 static int copy_nocow_pages(struct scrub_ctx *sctx, u64 logical, u64 len,
3196                             int mirror_num, u64 physical_for_dev_replace)
3197 {
3198         struct scrub_copy_nocow_ctx *nocow_ctx;
3199         struct btrfs_fs_info *fs_info = sctx->dev_root->fs_info;
3200
3201         nocow_ctx = kzalloc(sizeof(*nocow_ctx), GFP_NOFS);
3202         if (!nocow_ctx) {
3203                 spin_lock(&sctx->stat_lock);
3204                 sctx->stat.malloc_errors++;
3205                 spin_unlock(&sctx->stat_lock);
3206                 return -ENOMEM;
3207         }
3208
3209         scrub_pending_trans_workers_inc(sctx);
3210
3211         nocow_ctx->sctx = sctx;
3212         nocow_ctx->logical = logical;
3213         nocow_ctx->len = len;
3214         nocow_ctx->mirror_num = mirror_num;
3215         nocow_ctx->physical_for_dev_replace = physical_for_dev_replace;
3216         btrfs_init_work(&nocow_ctx->work, btrfs_scrubnc_helper,
3217                         copy_nocow_pages_worker, NULL, NULL);
3218         INIT_LIST_HEAD(&nocow_ctx->inodes);
3219         btrfs_queue_work(fs_info->scrub_nocow_workers,
3220                          &nocow_ctx->work);
3221
3222         return 0;
3223 }
3224
3225 static int record_inode_for_nocow(u64 inum, u64 offset, u64 root, void *ctx)
3226 {
3227         struct scrub_copy_nocow_ctx *nocow_ctx = ctx;
3228         struct scrub_nocow_inode *nocow_inode;
3229
3230         nocow_inode = kzalloc(sizeof(*nocow_inode), GFP_NOFS);
3231         if (!nocow_inode)
3232                 return -ENOMEM;
3233         nocow_inode->inum = inum;
3234         nocow_inode->offset = offset;
3235         nocow_inode->root = root;
3236         list_add_tail(&nocow_inode->list, &nocow_ctx->inodes);
3237         return 0;
3238 }
3239
3240 #define COPY_COMPLETE 1
3241
3242 static void copy_nocow_pages_worker(struct btrfs_work *work)
3243 {
3244         struct scrub_copy_nocow_ctx *nocow_ctx =
3245                 container_of(work, struct scrub_copy_nocow_ctx, work);
3246         struct scrub_ctx *sctx = nocow_ctx->sctx;
3247         u64 logical = nocow_ctx->logical;
3248         u64 len = nocow_ctx->len;
3249         int mirror_num = nocow_ctx->mirror_num;
3250         u64 physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3251         int ret;
3252         struct btrfs_trans_handle *trans = NULL;
3253         struct btrfs_fs_info *fs_info;
3254         struct btrfs_path *path;
3255         struct btrfs_root *root;
3256         int not_written = 0;
3257
3258         fs_info = sctx->dev_root->fs_info;
3259         root = fs_info->extent_root;
3260
3261         path = btrfs_alloc_path();
3262         if (!path) {
3263                 spin_lock(&sctx->stat_lock);
3264                 sctx->stat.malloc_errors++;
3265                 spin_unlock(&sctx->stat_lock);
3266                 not_written = 1;
3267                 goto out;
3268         }
3269
3270         trans = btrfs_join_transaction(root);
3271         if (IS_ERR(trans)) {
3272                 not_written = 1;
3273                 goto out;
3274         }
3275
3276         ret = iterate_inodes_from_logical(logical, fs_info, path,
3277                                           record_inode_for_nocow, nocow_ctx);
3278         if (ret != 0 && ret != -ENOENT) {
3279                 btrfs_warn(fs_info, "iterate_inodes_from_logical() failed: log %llu, "
3280                         "phys %llu, len %llu, mir %u, ret %d",
3281                         logical, physical_for_dev_replace, len, mirror_num,
3282                         ret);
3283                 not_written = 1;
3284                 goto out;
3285         }
3286
3287         btrfs_end_transaction(trans, root);
3288         trans = NULL;
3289         while (!list_empty(&nocow_ctx->inodes)) {
3290                 struct scrub_nocow_inode *entry;
3291                 entry = list_first_entry(&nocow_ctx->inodes,
3292                                          struct scrub_nocow_inode,
3293                                          list);
3294                 list_del_init(&entry->list);
3295                 ret = copy_nocow_pages_for_inode(entry->inum, entry->offset,
3296                                                  entry->root, nocow_ctx);
3297                 kfree(entry);
3298                 if (ret == COPY_COMPLETE) {
3299                         ret = 0;
3300                         break;
3301                 } else if (ret) {
3302                         break;
3303                 }
3304         }
3305 out:
3306         while (!list_empty(&nocow_ctx->inodes)) {
3307                 struct scrub_nocow_inode *entry;
3308                 entry = list_first_entry(&nocow_ctx->inodes,
3309                                          struct scrub_nocow_inode,
3310                                          list);
3311                 list_del_init(&entry->list);
3312                 kfree(entry);
3313         }
3314         if (trans && !IS_ERR(trans))
3315                 btrfs_end_transaction(trans, root);
3316         if (not_written)
3317                 btrfs_dev_replace_stats_inc(&fs_info->dev_replace.
3318                                             num_uncorrectable_read_errors);
3319
3320         btrfs_free_path(path);
3321         kfree(nocow_ctx);
3322
3323         scrub_pending_trans_workers_dec(sctx);
3324 }
3325
3326 static int copy_nocow_pages_for_inode(u64 inum, u64 offset, u64 root,
3327                                       struct scrub_copy_nocow_ctx *nocow_ctx)
3328 {
3329         struct btrfs_fs_info *fs_info = nocow_ctx->sctx->dev_root->fs_info;
3330         struct btrfs_key key;
3331         struct inode *inode;
3332         struct page *page;
3333         struct btrfs_root *local_root;
3334         struct btrfs_ordered_extent *ordered;
3335         struct extent_map *em;
3336         struct extent_state *cached_state = NULL;
3337         struct extent_io_tree *io_tree;
3338         u64 physical_for_dev_replace;
3339         u64 len = nocow_ctx->len;
3340         u64 lockstart = offset, lockend = offset + len - 1;
3341         unsigned long index;
3342         int srcu_index;
3343         int ret = 0;
3344         int err = 0;
3345
3346         key.objectid = root;
3347         key.type = BTRFS_ROOT_ITEM_KEY;
3348         key.offset = (u64)-1;
3349
3350         srcu_index = srcu_read_lock(&fs_info->subvol_srcu);
3351
3352         local_root = btrfs_read_fs_root_no_name(fs_info, &key);
3353         if (IS_ERR(local_root)) {
3354                 srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3355                 return PTR_ERR(local_root);
3356         }
3357
3358         key.type = BTRFS_INODE_ITEM_KEY;
3359         key.objectid = inum;
3360         key.offset = 0;
3361         inode = btrfs_iget(fs_info->sb, &key, local_root, NULL);
3362         srcu_read_unlock(&fs_info->subvol_srcu, srcu_index);
3363         if (IS_ERR(inode))
3364                 return PTR_ERR(inode);
3365
3366         /* Avoid truncate/dio/punch hole.. */
3367         mutex_lock(&inode->i_mutex);
3368         inode_dio_wait(inode);
3369
3370         physical_for_dev_replace = nocow_ctx->physical_for_dev_replace;
3371         io_tree = &BTRFS_I(inode)->io_tree;
3372
3373         lock_extent_bits(io_tree, lockstart, lockend, 0, &cached_state);
3374         ordered = btrfs_lookup_ordered_range(inode, lockstart, len);
3375         if (ordered) {
3376                 btrfs_put_ordered_extent(ordered);
3377                 goto out_unlock;
3378         }
3379
3380         em = btrfs_get_extent(inode, NULL, 0, lockstart, len, 0);
3381         if (IS_ERR(em)) {
3382                 ret = PTR_ERR(em);
3383                 goto out_unlock;
3384         }
3385
3386         /*
3387          * This extent does not actually cover the logical extent anymore,
3388          * move on to the next inode.
3389          */
3390         if (em->block_start > nocow_ctx->logical ||
3391             em->block_start + em->block_len < nocow_ctx->logical + len) {
3392                 free_extent_map(em);
3393                 goto out_unlock;
3394         }
3395         free_extent_map(em);
3396
3397         while (len >= PAGE_CACHE_SIZE) {
3398                 index = offset >> PAGE_CACHE_SHIFT;
3399 again:
3400                 page = find_or_create_page(inode->i_mapping, index, GFP_NOFS);
3401                 if (!page) {
3402                         btrfs_err(fs_info, "find_or_create_page() failed");
3403                         ret = -ENOMEM;
3404                         goto out;
3405                 }
3406
3407                 if (PageUptodate(page)) {
3408                         if (PageDirty(page))
3409                                 goto next_page;
3410                 } else {
3411                         ClearPageError(page);
3412                         err = extent_read_full_page_nolock(io_tree, page,
3413                                                            btrfs_get_extent,
3414                                                            nocow_ctx->mirror_num);
3415                         if (err) {
3416                                 ret = err;
3417                                 goto next_page;
3418                         }
3419
3420                         lock_page(page);
3421                         /*
3422                          * If the page has been remove from the page cache,
3423                          * the data on it is meaningless, because it may be
3424                          * old one, the new data may be written into the new
3425                          * page in the page cache.
3426                          */
3427                         if (page->mapping != inode->i_mapping) {
3428                                 unlock_page(page);
3429                                 page_cache_release(page);
3430                                 goto again;
3431                         }
3432                         if (!PageUptodate(page)) {
3433                                 ret = -EIO;
3434                                 goto next_page;
3435                         }
3436                 }
3437                 err = write_page_nocow(nocow_ctx->sctx,
3438                                        physical_for_dev_replace, page);
3439                 if (err)
3440                         ret = err;
3441 next_page:
3442                 unlock_page(page);
3443                 page_cache_release(page);
3444
3445                 if (ret)
3446                         break;
3447
3448                 offset += PAGE_CACHE_SIZE;
3449                 physical_for_dev_replace += PAGE_CACHE_SIZE;
3450                 len -= PAGE_CACHE_SIZE;
3451         }
3452         ret = COPY_COMPLETE;
3453 out_unlock:
3454         unlock_extent_cached(io_tree, lockstart, lockend, &cached_state,
3455                              GFP_NOFS);
3456 out:
3457         mutex_unlock(&inode->i_mutex);
3458         iput(inode);
3459         return ret;
3460 }
3461
3462 static int write_page_nocow(struct scrub_ctx *sctx,
3463                             u64 physical_for_dev_replace, struct page *page)
3464 {
3465         struct bio *bio;
3466         struct btrfs_device *dev;
3467         int ret;
3468
3469         dev = sctx->wr_ctx.tgtdev;
3470         if (!dev)
3471                 return -EIO;
3472         if (!dev->bdev) {
3473                 printk_ratelimited(KERN_WARNING
3474                         "BTRFS: scrub write_page_nocow(bdev == NULL) is unexpected!\n");
3475                 return -EIO;
3476         }
3477         bio = btrfs_io_bio_alloc(GFP_NOFS, 1);
3478         if (!bio) {
3479                 spin_lock(&sctx->stat_lock);
3480                 sctx->stat.malloc_errors++;
3481                 spin_unlock(&sctx->stat_lock);
3482                 return -ENOMEM;
3483         }
3484         bio->bi_iter.bi_size = 0;
3485         bio->bi_iter.bi_sector = physical_for_dev_replace >> 9;
3486         bio->bi_bdev = dev->bdev;
3487         ret = bio_add_page(bio, page, PAGE_CACHE_SIZE, 0);
3488         if (ret != PAGE_CACHE_SIZE) {
3489 leave_with_eio:
3490                 bio_put(bio);
3491                 btrfs_dev_stat_inc_and_print(dev, BTRFS_DEV_STAT_WRITE_ERRS);
3492                 return -EIO;
3493         }
3494
3495         if (btrfsic_submit_bio_wait(WRITE_SYNC, bio))
3496                 goto leave_with_eio;
3497
3498         bio_put(bio);
3499         return 0;
3500 }